Prediction

Apply the model trained by the machine learning operator to prediction data. This is typically linked after the machine learning operator. By specifying feature column mapping and customizing the prediction field name, this operator outputs the model's prediction results for new data.

Model files

In AI algorithm prediction scenarios, this is used to select a trained model file. By configuring input feature fields, viewing model target features, and customizing the prediction field name, the model performs predictions on the entire prediction dataset and outputs the results.

One-hot model application

After connecting to the One-hot encoding operator, you can apply the transformation model saved by the One-hot encoding operator to perform encoding conversion on fields with the same meaning in new data. Specify the correspondence between the fields to be encoded in the new data and the fields used during training.

Classification

Supports multiple classification models (such as logistic regression, Xgboost, Catboost, and more). Models are trained using feature columns and label columns, and are used to predict the category of the target variable. Each model has its own characteristics and is suitable for different scenarios, such as binary classification and multi-class classification. The choice can be made based on data characteristics and business requirements.

Clustering

Includes unsupervised learning models such as K-means clustering, Gaussian mixture model, and bisecting K-means, which are used to automatically divide the dataset into several clusters. By analyzing the similarity between samples, samples within the same cluster have high similarity, while samples in different clusters have low similarity, which helps to discover natural groupings in the data.

Regression

Provides models such as Xgboost regression, linear regression, and decision tree regression, which are used to establish the relationship between independent variables and continuous dependent variables, and to predict continuous target values. The models can capture linear or nonlinear relationships and support mechanisms such as regularization to control overfitting. They are suitable for scenarios such as price prediction and quantity estimation.

Evaluation

Includes operators for binary classification evaluation, multi-class classification evaluation, clustering evaluation, regression evaluation, and time series evaluation, which are used to evaluate the performance of the corresponding types of models. By comparing the true labels with the prediction results, various evaluation metrics (such as accuracy, error values, curves, and more) are calculated and output to help determine the quality of the model.

Feature column mapping

Specifies the feature columns in the input data used for prediction, which must correspond to the feature fields used during model training.

Prediction field name

Specifies the field name for the prediction result in the output data, which is used to distinguish between the original data and the prediction result.

Model information

Specifies the trained model file used for prediction, supporting selection via keyword search.

Select feature columns

Specify the feature columns in the input data used for prediction. These must correspond to the feature columns used during model training, and the field types must remain consistent.

Model target feature

Displays the target field used during model training (that is, the core metric predicted by the model). This is a read-only parameter and cannot be modified.

Prediction field name

Customize the field name for prediction results in the output data to distinguish between original data and prediction results.

Feature column mapping

Specify the fields in new data that require encoding, corresponding to the fields used during training of the One-hot encoding operator (ensure that the same encoding rules are applied to fields with the same meaning).

Logistic regression

Used to build a binary classification prediction model. By inputting feature columns and label columns, the model is trained to predict the target variable. Applies the logistic function to a linear combination of features to map the result to the [0,1] range and outputs classification probabilities. Suitable for binary classification scenarios, regularization can be used to control overfitting, and training is highly efficient.

Xgboost

An ensemble learning algorithm based on gradient boosting decision trees (GBDT), suitable for classification tasks. Builds multiple decision trees and combines them through serial iteration, with each tree focusing on correcting the errors of the previous model. The final output is a weighted combination of results. Features efficient data processing and strong resistance to overfitting, supports various regularization parameters and sampling strategies.

Catboost

A gradient boosting framework based on symmetric decision trees, specifically designed for classification problems. The core advantages are efficient handling of categorical features (no manual preprocessing required), solving gradient bias and prediction shift issues, and providing high accuracy and generalization ability. Default parameters perform excellently, support GPU acceleration, and offer fast prediction speed, making it suitable for latency-sensitive scenarios.

Naive Bayes

A classification algorithm based on Bayes' theorem and the conditional independence assumption of features, suitable for binary and multiclass tasks. Predicts categories by calculating posterior probabilities. The model is simple and fast to train, making it suitable for large-scale data. The advantage is high computational efficiency; the disadvantage is reliance on the feature independence assumption, which may affect performance in practice if the assumption does not hold. Supports both multinomial and Bernoulli model types.

Support vector machine

A classification model based on statistical learning theory, with the core objective of finding the separating hyperplane with the maximum geometric margin. Handles nonlinear problems through kernel functions and is suitable for small sample sizes and high-dimensional data classification. During training, model complexity is controlled through regularization. It is sensitive to feature scaling and requires standardization by default. Delivers excellent performance in scenarios such as text classification.

Decision tree

A supervised learning model based on a tree structure, which constructs the decision process by recursively partitioning the feature space. Internal nodes represent feature-based decision criteria, and leaf nodes represent classification results. The principle is intuitive and highly interpretable. It can handle various types of features and does not require prior knowledge of feature relationships. However, a single tree is prone to overfitting and requires optimization by controlling parameters such as tree depth and node splitting conditions.

Random forest

A bagging ensemble method based on multiple decision trees, which builds multiple independent decision trees through sampling with replacement and aggregates the results by majority voting. Each tree is trained on a randomly selected subset of samples and features, reducing the risk of overfitting while retaining the interpretability of decision trees and improving performance. Suitable for high-dimensional data; training is efficient and can be parallelized, and it has strong noise resistance.

Gradient Boosted Trees

An ensemble learning algorithm that builds decision trees in a sequential manner, with each tree correcting the errors of the previous model and minimizing the loss function based on gradient descent. As the number of trees increases, the model's expressive power gradually improves, and the final result is obtained by weighted summation. Unlike the parallel training of random forests, its serial mechanism places greater emphasis on error correction, making it suitable for handling complex patterns, but the training speed is relatively slower.

Feature columns

Input feature fields used for model training, serving as the basis for model prediction. These features should be correlated with the prediction target; selection should be based on business knowledge and feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass classification scenarios, categorical features need to be processed with one-hot encoding first.

Maximum number of iterations

The maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training will stop regardless of whether convergence is achieved.

• If the model finds the optimal hyperplane in fewer iterations, this value can be reduced to improve efficiency;
• If convergence is slow (for example, due to complex data), this value should be increased to ensure sufficient training.

Convergence threshold

The threshold for determining convergence during model training; when the change in loss value between two consecutive iterations is less than this threshold, iteration stops early. A smaller value indicates a higher requirement for convergence accuracy, but may increase training time. For scenarios requiring high accuracy, set a smaller value; to pursue training efficiency, a larger value can be set as appropriate.

Regularization type

The regularization method used to prevent model overfitting; this is an optional parameter.

• 0 means no regularization, suitable for scenarios with large data volumes and low correlation between features.
• 1 usually refers to L1 regularization (Lasso), which sets some feature coefficients to 0 and can be used for feature selection.
• 2 usually refers to L2 regularization (Ridge), which shrinks feature coefficients towards smaller values and is suitable for handling multicollinearity.

Regularization coefficient

The coefficient for regularization strength, used together with the regularization type; this is an optional parameter. The larger the value, the stronger the regularization effect. Setting it to 0 disables regularization. After selecting the regularization type, adjust this value according to the degree of overfitting. If overfitting is severe, increase this value.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

Sorting method for label category indices.

• alphabetAsc: Sort in ascending order according to the alphabetical order of labels (A-Z).
• alphabetDesc: Sort in descending order according to the alphabetical order of labels (Z-A).
• frequencyAsc: Sort labels in ascending order by frequency of occurrence, that is, labels with fewer occurrences are placed first.
• frequencyDesc: Sort labels in descending order by frequency of occurrence, that is, labels with more occurrences are placed first.

Binary classification threshold

The threshold for determining the predicted class in a binary classification model. When the probability of the positive class is greater than this value, the prediction is 1 (positive class); otherwise, it is 0 (negative class). 0.5 is the default and commonly used threshold.

• In scenarios where precision is emphasized (such as spam detection, to reduce misclassification of normal emails), the threshold can be increased.
• In scenarios where recall is emphasized (such as disease screening, to reduce missed diagnoses), the threshold can be decreased.

Feature normalization

A preprocessing switch used to convert input features to a unified scale. Enabled by default.

• Enable scenario: When there are large differences in the value ranges among features, normalization can accelerate model convergence and prevent features with large value ranges from dominating model training.
• Disable scenario: When features have clear physical meaning and the scale differences have actual business significance, consider disabling normalization.

treeAggregation depth

The depth of the tree aggregation method, used for model optimization. It affects the model's ability to capture feature interactions and should be adjusted according to data complexity.

• The smaller the value, the simpler the model and the faster the training speed, but complex feature relationships may not be captured.
• The larger the value, the more complex the model, which may improve performance but can also lead to overfitting.

Add intercept term

Whether to include an intercept term (bias term) in the model. The intercept term is used to adjust the model's baseline prediction value.

• Enabled by default, which allows the model to fit the data more flexibly.
• In certain special scenarios (such as when the data has already been centered and no baseline adjustment is needed), this can be disabled.

Feature columns

Input feature fields used for model training, which serve as the basis for model prediction. These features should be correlated with the prediction target; selection should be based on business knowledge and feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass classification scenarios, categorical features need to be processed with one-hot encoding first.

Number of boosting rounds

The number of decision trees constructed in ensemble learning, which should be adjusted according to data complexity and model performance. If set to 100, this means 100 decision trees will be constructed for the ensemble.

• Too few rounds may cause the model to underfit and fail to fully capture data patterns;
• too many rounds may lead to overfitting and increase training time.

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to choose a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns;
• The deeper the tree, the more complex it becomes, which may improve performance but also increases the risk of overfitting.

Automatic validation set split

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Learning rate

Controls the contribution of each tree to the final prediction result, also known as the step size.

• A smaller learning rate can make the model more robust and reduce the risk of overfitting, but requires more boosting rounds to achieve better performance;
• A larger learning rate results in faster training, but may cause the model to oscillate and make convergence more difficult.

Minimum number of samples per leaf

The minimum number of samples required for each leaf node. For data with significant noise, it is recommended to set a larger value.

• The larger this value, the harder it is to generate leaf nodes, resulting in lower tree complexity and preventing overfitting;
• The smaller the value, the more complex the tree structure, which may fit the training data better but could reduce generalization ability.

Minimum loss reduction per leaf

The minimum loss reduction required to split a leaf node. A split will only occur if the loss reduction from the split is greater than or equal to this value.

• Setting a larger value can simplify the tree structure and prevent overfitting;
• Setting a smaller value allows for more detailed splits, which may improve training accuracy.

Sample sampling ratio

The proportion of samples used to train a single tree in each iteration. If set to 1, all samples are used. The value range is (0,1]. Sampling increases model randomness, prevents overfitting, and speeds up training. When the dataset is large, this ratio can be reduced appropriately.

Feature sampling ratio

The proportion of features used to train a single tree in each iteration. If set to 1, all features are used. The value range is (0,1]. Feature sampling reduces correlation between features and improves model generalization, which is especially effective for high-dimensional data.

Use the best model

The best model is selected based on the evaluation metric used for the optimal model.

Automatic missing value imputation

Whether to automatically impute missing values in the data. The XGBoost algorithm has a robust mechanism for handling missing values. Enabling this option simplifies data preprocessing and is very useful for datasets with missing values. If missing values have already been handled manually, this feature can be disabled.

Label index sorting method

Sorting methods for tag category indexes.

• alphabetAsc: Sort tags in ascending order based on alphabetical sequence (A-Z).
• alphabetDesc: Sort tags in descending order based on alphabetical sequence (Z-A).
• frequencyAsc: Sort tags from lowest to highest frequency of occurrence, with tags appearing less frequently listed first.
• frequencyDesc: Sort tags from highest to lowest frequency of occurrence, with tags appearing more frequently listed first.

L2 regularization term for weights

L2 regularization parameter used to control model complexity, applied to leaf node weights. The larger the value, the stronger the regularization effect. This can prevent overfitting, but may make the model too simple and lead to underfitting. Adjust according to the degree of overfitting.

L1 regularization coefficient

L1 regularization parameter used to control model complexity, applied to leaf node weights. Setting the value to 0 means L1 regularization is not used. L1 regularization will cause some leaf node weights to become zero, serving as feature selection. This is suitable for scenarios where the goal is to simplify the model and select important features.

Number of Workers

Number of worker processes used for parallel computation. 0 means parallel computation is not performed. When the dataset is large or there are many features, increasing the number of Workers appropriately can improve training speed. However, hardware resources are limited, and too many Workers may cause resource contention and reduce efficiency.

Worker connection timeout

Timeout duration for worker process connections (in seconds). 0 means no timeout is set. Used to control the connection wait time for worker processes during parallel computation. In unstable network environments or when computing resources are tight, setting a timeout can help avoid prolonged waiting.

Feature columns

Input feature fields used for model training, which serve as the basis for model prediction. These features should be correlated with the prediction target. Selection should be based on business knowledge and feature importance analysis.

Label columns

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• In multiclass scenarios, categorical features must first be processed using one-hot encoding.

Maximum number of trees

Total number of decision trees in the ensemble model. The greater the number of trees, the more complex patterns the model may capture; however, this also increases training time. Thanks to CatBoost's optimization mechanism, even with a large number of trees, overfitting is less likely.

Maximum tree depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to select a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training speed, but it may not capture complex data patterns;
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Learning rate

Controls the contribution weight (step size) of each tree to the final prediction. In this case, the learning rate is set to 0.03. A smaller learning rate combined with a larger number of trees yields a more robust model, while a larger learning rate results in faster training but may affect convergence stability. CatBoost's default learning rate setting has already been optimized and usually does not require significant adjustment.

Use the best model

The best model will be selected based on the evaluation metric used to determine the optimal model.

Automatic missing value filling

Whether to automatically handle missing values in the data. In this case, this option is enabled. CatBoost has a built-in efficient mechanism for handling missing values, which can automatically fill missing values appropriately, simplifying data preprocessing steps. This is very useful for incomplete datasets commonly encountered in real-world business scenarios.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

Sorting method for label category indices.

• alphabetAsc: Sort in ascending order according to the alphabetical order of labels (A-Z).
• alphabetDesc: Sort in descending order according to the alphabetical order of labels (Z-A).
• frequencyAsc: Sort by label frequency from low to high, that is, labels with fewer occurrences are placed first.
• frequencyDesc: Sorts tags in descending order of frequency, meaning tags that appear more frequently are listed first.

Random seed

Controls the initial value of the random process during training. Fixing the seed ensures that model training results are reproducible, which facilitates experimental comparison and troubleshooting; to average results over multiple experiments, different seed values can be used.

Feature column

Input feature fields used for model training, serving as the basis for model prediction. These features should be correlated with the prediction target; selection should be based on business knowledge and feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass scenarios, categorical features must first be processed with one-hot encoding.

Smoothing option

The Laplace smoothing parameter used to address the zero probability problem, also known as a pseudo-count. When the occurrence count of a feature under a certain class is 0, the smoothing parameter prevents the posterior probability from being calculated as 0. The larger the value, the stronger the smoothing effect and the more pronounced the regularization effect on the model; it is typically set as a positive integer.

Model type

Specific implementation types of Naive Bayes, with two options:

• multinomial (multinomial model): Suitable for discrete features, assumes features follow a multinomial distribution, and is commonly used in scenarios such as text classification;
• bernoulli (Bernoulli model): Suitable for binary features (0 or 1), assumes features follow a Bernoulli distribution, and is more appropriate for handling binarized features.

Automatic validation set split

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

The sorting method for label category indices.

• alphabetAsc: Sorts labels in ascending alphabetical order (A-Z).
• alphabetDesc: Sorts labels in descending alphabetical order (Z-A).
• frequencyAsc: Sorts tags in ascending order of frequency, meaning tags that appear less frequently are listed first.
• frequencyDesc: Sorts tags in descending order of frequency, meaning tags that appear more frequently are listed first.

Feature columns

Input feature fields used for model training, which serve as the basis for model prediction. These features should be correlated with the prediction target; selection should be based on business knowledge and feature importance analysis.

Label columns

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass classification scenarios, categorical features must be processed with one-hot encoding first.

Maximum number of iterations

The maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training will stop regardless of whether convergence has been achieved.

• If the model finds the optimal hyperplane within fewer iterations, this value can be reduced to improve efficiency;
• If convergence is slow (for example, due to complex data), this value should be increased to ensure sufficient training.

Regularization coefficient

The coefficient for regularization strength. The larger the value, the stronger the regularization effect. If set to 0, regularization is disabled. This value should be adjusted based on the degree of overfitting; if overfitting is severe, increase this value.

Feature standardization

A preprocessing option used to convert input features to a unified scale. Enabled by default.

• When to enable: When there are large differences in value ranges among features, standardization can accelerate model convergence and prevent features with larger value ranges from dominating the model training process.
• When to disable: If features have clear physical meanings and the scale differences have actual business significance, consider disabling standardization.

Iteration termination threshold

The threshold for determining convergence during model training; if the change in loss value between two consecutive iterations is less than this threshold, iteration stops early. The smaller the value, the higher the requirement for convergence accuracy, but it may increase training time. For scenarios requiring high accuracy, set a smaller value; to prioritize training efficiency, increase the value appropriately.

Depth of the tree aggregation method

The depth of the tree aggregation method, used for model optimization. It affects the model's ability to capture feature interactions and should be adjusted according to data complexity.

• A smaller value results in a simpler model and faster training, but may fail to capture complex feature relationships.
• A larger value results in a more complex model, which may improve performance but can also lead to overfitting.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

The sorting method for label category indices.

• alphabetAsc: Sorts labels in ascending alphabetical order (A-Z).
• alphabetDesc: Sorts labels in descending alphabetical order (Z-A).
• frequencyAsc: Sorts labels by frequency from low to high, meaning labels with fewer occurrences are listed first.
• frequencyDesc: Sorts labels by frequency from high to low, meaning labels with more occurrences are listed first.

Binary classification threshold

The threshold for determining predicted classes in a binary classification model. When the probability of the positive class exceeds this value, the prediction is 1 (positive class); otherwise, it is 0 (negative class). 0.5 is the default and commonly used threshold.

• In scenarios emphasizing precision (such as spam detection, to reduce misclassification of normal emails), the threshold can be increased.
• In scenarios emphasizing recall (such as disease screening, to reduce missed diagnoses), the threshold can be lowered.

Add intercept term

Whether to include an intercept term (bias term) in the model. The intercept term is used to adjust the model's baseline prediction value.

• It is enabled by default, which allows the model to fit data more flexibly.
• In certain special scenarios (such as when the data has already been centralized and baseline adjustment is unnecessary), it can be disabled.

Feature columns

The input feature fields used for model training, which serve as the basis for model prediction. These features should have some correlation with the prediction target. When selecting features, combine business knowledge with feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass classification scenarios, categorical features need to be processed with the one-hot operator first.

VectorIndexer

Feature auto-indexing switch, used for handling categorical features:

• Off: Categorical features must be handled manually
• On: Categorical features are automatically converted to indices, with additional parameter settings available

Maximum number of categories (when VectorIndexer is enabled)

Threshold for automatically identifying categorical features. If set to 32, when the number of distinct values for a feature is less than or equal to 32, it is treated as a categorical feature.

Handling of unseen category values (when VectorIndexer is enabled)

How to handle unseen category values. Choose based on data quality and business requirements: select 'error' for clean data, and 'skip' when there is noise in the data.

• skip: Skip samples containing unseen category values
• error: Throw an error when encountering unseen category values

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to choose a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns;
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Maximum number of bins

The number of bins for discretizing continuous features. Used to divide continuous features into a limited number of intervals (bins) to facilitate tree splitting. The larger the value, the finer the partitioning of continuous features.

Minimum number of samples for splitting

The minimum number of samples required to split a node. If set to 1, splitting is allowed when the number of samples at a node is greater than or equal to 1. The larger the value, the simpler the tree, which helps prevent overfitting.

Minimum gain

The minimum information gain required to split a node. If set to 0, splitting is allowed as long as there is any gain. Setting a larger value can prevent meaningless splits and simplify the tree structure.

Maximum memory

The maximum memory allowed during training (in MB). In this example, it is set to 256 to control memory usage and prevent memory overflow.

Cache node ID

Whether to cache node IDs to accelerate training. In this example, it is set to off. Enabling this option can increase training speed but will increase memory consumption.

Checkpoint interval

The iteration interval for saving checkpoints during model training. If set to 10, it means intermediate results are saved every 10 splits, making it easier to resume training after interruption.

Impurity

A metric used to measure the quality of node splits, with two options available:

• entropy: Based on information theory, measures the reduction in information uncertainty
• gini: Measures the degree of class disorder in the dataset

Random seed

Controls the initial value of random processes during training. Setting a fixed seed ensures that model training results are reproducible, facilitating experiment comparison and troubleshooting. To average results across multiple experiments, different seed values can be used.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

The sorting method for label category indices.

• alphabetAsc: Sorts labels in ascending alphabetical order (A-Z).
• alphabetDesc: Sorts labels in descending alphabetical order (Z-A).
• frequencyAsc: Sorts labels by frequency from low to high, meaning labels with fewer occurrences are listed first.
• frequencyDesc: Sorts labels by frequency from high to low, meaning labels with more occurrences are listed first.

Feature columns

The input feature fields used for model training, which serve as the basis for model prediction. These features should be correlated with the prediction target. Selection should be based on business knowledge and feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, the labels are usually 0 and 1 or similar identifiers representing two classes.
• For multi-class classification scenarios, categorical features need to be processed with one-hot encoding first.

VectorIndexer

Automatic feature indexing switch, used for handling categorical features:

• Off: Categorical features must be processed manually
• On: Categorical features are automatically converted to indices, with additional parameter settings provided

Maximum number of categories (when VectorIndexer is enabled)

Threshold for automatically identifying categorical features. If set to 32, features with 32 or fewer distinct values are treated as categorical features.

Handling of unseen category values (when VectorIndexer is enabled)

The processing method for unseen category values can be chosen based on data quality and business requirements: select "error" for clean data, or "skip" if there is noise in the data.

• skip: Skip samples containing unseen category values
• error: Throw an error when encountering unseen category values

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small number of samples, it is recommended to select a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns;
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Maximum number of bins

The number of bins for discretizing continuous features. Used to divide continuous features into a limited number of intervals (bins) to facilitate tree splitting. The larger the value, the finer the discretization of continuous features.

Minimum number of samples for a split

The minimum number of samples required to split a node. If set to 1, a node can be split when it contains at least 1 sample. The larger the value, the simpler the tree, which helps prevent overfitting.

Minimum gain

The minimum information gain required to split a node. If set to 0, as long as there is any gain, a split can occur. Setting a larger value can avoid meaningless splits and simplify the tree structure.

Maximum memory

The maximum memory allowed to be used during training (in MB). This case is set to 256 to control memory usage and prevent memory overflow.

Cache node ID

Whether to cache node IDs to accelerate training. In this case, this option is disabled. Enabling it can improve training speed but increases memory consumption.

Checkpoint interval

The iteration interval for saving checkpoints during model training. If set to 10, an intermediate result is saved every 10 splits, making it easier to resume training after interruption.

Impurity

A metric for evaluating the quality of node splits. There are two options:

• entropy: Based on information theory, measures the reduction in information uncertainty.
• gini: Measures the degree of class disorder in the dataset.

Subset sampling ratio

The sampling ratio for each tree during training. If set to 1, all samples are sampled with replacement. Typically set between 0.6 and 1.0.

Random seed

Controls the initial value of the random process during training. Fixing the seed ensures reproducible model training results, which facilitates experiment comparison and troubleshooting. To average results over multiple experiments, use different seed values.

Number of trees

The total number of decision trees in the ensemble model. A larger number of trees generally leads to more stable model performance, but also increases training time.

Feature selection strategy

The strategy for determining the size of the feature subset selected at each tree split. Multiple options are available:

• auto: Automatically selects an appropriate strategy.
• all: Uses all features.
• onethird: Uses one third of the total number of features.
• sqrt: Uses the square root of the total number of features.
• log2: Uses the base-2 logarithm of the total number of features.

This parameter controls feature randomness and helps reduce correlation between trees.

Automatic validation set split

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

The sorting method for label category indices.

• alphabetAsc: Sorts in ascending order according to the label's alphabetical order (A-Z).
• alphabetDesc: Sorts in descending order according to the label's alphabetical order (Z-A).
• frequencyAsc: Sorts tags in ascending order of frequency, meaning tags that appear less frequently are listed first.
• frequencyDesc: Sorts tags in descending order of frequency, meaning tags that appear more frequently are listed first.

Feature columns

Input feature fields used for model training, which serve as the basis for model prediction. These features should be correlated with the prediction target; selection should be based on business knowledge and feature importance analysis.

Label column

The target field that the model needs to predict, that is, the dependent variable.

• In binary classification scenarios, this is usually 0 and 1 or similar identifiers representing two classes.
• For multiclass scenarios, categorical features must be processed with one-hot encoding first.

VectorIndexer

Automatic feature indexing switch, used for handling categorical features:

• Off: Categorical features must be processed manually
• On: Categorical features are automatically converted to indices, with additional parameter settings available

Maximum number of categories (when VectorIndexer is enabled)

Threshold for automatically identifying categorical features. If set to 32, features with 32 or fewer distinct values are treated as categorical features.

Handling of invalid category values (when VectorIndexer is enabled)

How to handle unseen category values; choose based on data quality and business requirements. Select 'error' for clean data, or 'skip' if noise is present.

• skip: Skip samples containing invalid values
• error: Throw an error when encountering invalid values

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to choose a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns;
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Maximum number of bins

The number of bins for discretizing continuous features. Used to divide continuous features into a limited number of intervals (bins) to facilitate tree splitting; the larger the value, the finer the partitioning of continuous features.

Minimum number of samples per split

Minimum number of samples required for node splitting. If set to 1, node splitting is allowed when the number of samples at the node is greater than or equal to 1. A larger value results in a simpler tree and helps prevent overfitting.

Minimum gain

Minimum information gain required for node splitting. If set to 0, splitting is allowed as long as there is any gain. Setting a larger value can prevent meaningless splits and simplify the tree structure.

Maximum memory

Maximum memory allowed during training (in MB). In this case, it is set to 256 to control memory usage and prevent memory overflow.

Cache node ID

Whether to cache node IDs to accelerate training. In this case, it is disabled. Enabling it can improve training speed but increases memory consumption.

Checkpoint interval

Iteration interval for saving checkpoints during model training. If set to 10, intermediate results are saved every 10 splits, making it easier to resume training after interruption.

Impurity

Metric for evaluating the quality of node splits, with two options:

• entropy: Based on information theory, measures the reduction in information uncertainty
• gini: Measures the degree of class impurity in the dataset

Subset sampling ratio

Sampling ratio for each tree during training. If set to 1, all samples are used for sampling with replacement. Typically set between 0.6 and 1.0.

Random seed

Controls the initial value of the random process during training. Fixing the seed ensures reproducible model training results, which facilitates experiment comparison and troubleshooting. To average results over multiple experiments, use different seed values.

Maximum number of iterations

Maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training stops regardless of convergence.

• If the model finds the optimal hyperplane in fewer iterations, reduce this value to improve efficiency.
• If convergence is slow (for example, due to complex data), increase this value to ensure sufficient training.

Step size

The contribution weight of each tree (learning rate). Controls the influence of each tree. A smaller step size requires more iterations, but results in a more robust model.

Loss function type

A function used to measure prediction error. Only logistic can be selected. It is suitable for binary classification problems and predicts probabilities using cross-entropy.

Feature selection strategy

The strategy for selecting the size of the feature subset when each tree splits. Multiple options are available:

• auto: Automatically selects an appropriate strategy
• all: Uses all features
• onethird: Uses one third of the total number of features
• sqrt: Uses the square root of the total number of features
• log2: Uses the base-2 logarithm of the total number of features

This parameter controls feature randomness and helps reduce correlation between trees.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Label index sorting method

The sorting method for label category indices.

• alphabetAsc: Sorts labels in alphabetical order (A-Z).
• alphabetDesc: Sorts labels in reverse alphabetical order (Z-A).
• frequencyAsc: Sorts labels in ascending order of frequency, with less frequent labels appearing first.
• frequencyDesc: Sorts labels in descending order of frequency, with more frequent labels appearing first.

K-means clustering

The K-means algorithm is a typical distance-based clustering algorithm that uses distance as a similarity metric. After specifying K initial cluster centers, calculate the distance between each sample and the cluster centers, assign each sample to the nearest cluster, update the cluster centers after each assignment, and repeat until all samples are clustered and the centers no longer change.

Gaussian mixture model

Gaussian mixture clustering uses the Gaussian probability density function (normal distribution curve) to precisely quantify entities, and can cluster sample points into several clusters formed based on the Gaussian probability density function (normal distribution curve). In theory, it can fit any continuous function; the drawback is that it requires significant computational resources.

Bisecting K-means

The bisecting K-means clustering algorithm is a variant of the K-means clustering algorithm, primarily designed to address the uncertainty in clustering results caused by the randomness of initial centroid selection in the K-means algorithm. The bisecting K-means algorithm is less affected by the randomness of initial centroid selection.

Feature columns

Input feature fields used for model training; here, this refers to the fields used for cluster analysis. The scale of features affects the distance calculation results. Standardization is generally recommended.

Number of clusters

Preset number of clusters (K value). If set to 3, this means the data is intended to be divided into 3 clusters. The selection of the K value should be determined based on business knowledge or methods such as the elbow method. A K value that is too small will result in excessive intra-cluster differences, while a K value that is too large will lead to over-segmentation.

Centroid initialization method

There are two options for generating the initial cluster centers (centroids):

• random: Randomly selects K points from the samples as the initial centroids. This method is simple but may be greatly affected by randomness.
• k-means: An improved initialization method that generates better initial centroids through multiple samplings, offering greater stability.

You can choose based on the required stability of the results; k-means is recommended for large datasets.

Maximum number of iterations

The maximum number of iterations allowed by the algorithm. When the number of iterations reaches this value, the calculation will stop regardless of whether the centroids have stabilized. For complex datasets, this value can be appropriately increased to ensure convergence.

Iteration termination threshold

The convergence threshold for model training. If the change in loss value between two consecutive iterations is less than this threshold, the iteration will stop early. A smaller value indicates a higher requirement for convergence accuracy, but may increase training time. For scenarios requiring high accuracy, set a smaller value; to improve training efficiency, increase the value appropriately.

Random seed

Controls the initial value of the random process during training. Fixing the seed ensures that model training results are reproducible, which facilitates experiment comparison and troubleshooting. To average results over multiple experiments, use different seed values.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Feature columns

Input feature fields used for clustering, which must be numerical and continuous features. Because the Gaussian model is sensitive to data scale, it is recommended to perform standardization first.

Number of clusters

The preset number of clusters, that is, the number of components in the Gaussian mixture distribution. If set to 3, it means the data is assumed to be a mixture of three Gaussian distributions. The selection of the number of clusters should be based on domain knowledge and can be evaluated using information criteria such as BIC.

Maximum number of iterations

The maximum number of optimization iterations for the algorithm, used to train the parameters of the Gaussian mixture model. When the number of iterations reaches this value, computation will stop regardless of whether convergence is achieved. For complex data distributions, increase this value appropriately to ensure the model is sufficiently trained.

Iteration stopping threshold

The convergence threshold for model training. If the change in loss value between two consecutive iterations is less than this threshold, the iteration will stop early. A smaller value indicates a higher requirement for convergence accuracy, but may increase training time. For scenarios requiring high accuracy, set a smaller value; to improve training efficiency, increase the value appropriately.

Random seed

Control the initial value of the random process during training. Setting a fixed seed ensures that model training results are reproducible, facilitating experimental comparison and troubleshooting; to average results over multiple experiments, different seed values can be used.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Feature columns

Input feature fields used for model training; here, these refer to the fields used for cluster analysis. The scale of features affects the results of distance calculations; standardization is generally recommended.

Number of clusters

The number of clusters to be obtained in the end. If set to 3, the data will ultimately be split into 3 clusters through recursive bisection. The choice of the number of clusters should be based on business requirements and can typically be adjusted by observing clustering performance, such as intra-cluster compactness.

Maximum number of iterations

The maximum number of iterations when performing K-means with K=2 for each cluster to be split. In this case, it is set to 10, which controls the maximum number of centroid updates during a single bisection to ensure stable sample allocation within clusters.

Minimum cluster split size

The minimum number of samples allowed in a cluster to be split. If set to 1, a cluster can be split when the number of samples in the cluster is greater than or equal to 1. Setting a larger value (such as 5) can prevent excessively small clusters from being further split, avoiding over-segmentation.

Distance metric

The distance formula for calculating similarity between samples; two options are available:

• euclidean (Euclidean distance): The most commonly used distance metric, suitable for continuous features, reflects the straight-line distance between samples in space;
• cosine (cosine distance): Measures the similarity of vector directions, suitable for high-dimensional features, and is insensitive to magnitude.

Random seed

Control the initial value of the random process during training. Setting a fixed seed ensures that model training results are reproducible, facilitating experimental comparison and troubleshooting; to average results over multiple experiments, different seed values can be used.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Xgboost regression

An ensemble learning regression algorithm based on gradient boosting decision trees (GBDT). It constructs multiple decision trees in series and continuously corrects the errors of previous models, ultimately outputting continuous prediction results. It has strong nonlinear fitting capability, supports regularization, sample and feature sampling, and is suitable for regression problems where features and target variables have complex relationships. It offers excellent prediction accuracy and generalization performance.

Linear regression

A classic supervised learning regression algorithm used to establish a linear relationship between independent variables and continuous dependent variables. By minimizing a loss function (such as squared error or Huber loss), the regression coefficients are estimated. The model is simple and highly interpretable, and can intuitively reflect the impact of features on the target (coefficient magnitude and sign). Supports regularization (L1, L2) to control overfitting, is sensitive to feature scales, and requires standardization by default.

Decision tree regression

A supervised learning regression algorithm based on a tree structure. It recursively partitions the feature space using a greedy algorithm to construct the decision process: internal nodes represent feature-based decision criteria, and leaf nodes are continuous prediction values. The principle is intuitive and highly interpretable, does not require assuming a linear relationship, can capture nonlinearity and feature interactions, and is well adapted to high-dimensional data. Splitting points are determined using criteria such as mean squared error. Parameters such as tree depth must be controlled to avoid overfitting.

Random forest regression

An ensemble learning regression algorithm based on multiple decision trees that improves performance through the bagging strategy: multiple trees are trained independently, each using a sample subset with replacement and a random feature subset. The final prediction result is the average of all trees. Effectively reduces the risk of overfitting, preserves the interpretability of decision trees, enables efficient parallel training, provides strong noise resistance, and is suitable for complex data scenarios. The number of trees and the feature selection strategy influence model performance.

Gradient boosting decision tree regression

An ensemble learning regression algorithm that constructs decision trees through serial iteration: each new tree focuses on correcting the errors of the previous model, minimizes the loss function (squared loss, absolute loss) based on the gradient descent method, and finally outputs the result by weighted summation. Compared with the parallel training of random forests, its serial mechanism emphasizes error correction, can accurately capture nonlinear patterns, but has a slower training speed, and parameters such as step size and number of iterations require fine-tuning.

Monotonic regression

A specialized regression algorithm that enforces the predicted results to satisfy a monotonic increase (or a preset order relationship) constraint when fitting data. While minimizing prediction error, ensures that the output follows a specific order, addressing the issue in traditional regression where the relationship between features and the target may contradict common sense, and is suitable for scenarios with strict requirements on the order relationship between variables (such as the positive correlation between age and income).

Feature columns

Input feature fields used for model training, which serve as the basis for regression prediction. Features can be continuous or discrete. XGBoost is insensitive to feature scale, but it is recommended to remove redundant features to improve efficiency.

Label column

The target field that the model needs to predict, that is, the dependent variable.

Number of boosting rounds

The number of decision trees constructed in ensemble learning, which should be adjusted according to data complexity and model performance. Setting this to 100 means that 100 decision trees will be constructed for the ensemble.

• Too few rounds may lead to underfitting and insufficient capture of data patterns;
• Too many rounds may lead to overfitting and increased training time.

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to select a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training speed, but it may not capture complex data patterns.
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Learning rate

Controls the contribution of each tree to the final prediction result, also known as the step size.

• A smaller learning rate can make the model more robust and reduce the risk of overfitting, but requires more boosting rounds to achieve better performance.
• A larger learning rate results in faster training, but may cause the model to oscillate and make convergence difficult.

Minimum samples per leaf

The minimum number of samples required for each leaf node. For data with a high level of noise, it is recommended to set a larger value.

• The larger this value, the harder it is to generate leaf nodes, resulting in lower tree complexity and preventing overfitting.
• The smaller the value, the more complex the tree structure, which may fit the training data better but reduce generalization ability.

Minimum loss reduction per leaf

The minimum loss reduction required when splitting a leaf node. Splitting is performed only when the loss reduction from the split is greater than or equal to this value.

• Setting a larger value can simplify the tree structure and prevent overfitting.
• Setting a smaller value allows for more fine-grained splits, which may improve training accuracy.

Sample subsampling ratio

The proportion of samples used to train each tree in every iteration. If set to 1, all samples are used. The value range is (0,1]. Subsampling increases model randomness, prevents overfitting, and speeds up training. When the dataset is large, this ratio can be appropriately reduced.

Feature subsampling ratio

The proportion of features used to train each tree in every iteration. If set to 1, all features are used. The value range is (0,1]. Feature sampling can reduce the correlation between features and improve the model's generalization capability, with a particularly significant effect on high-dimensional data.

Use the best model

The best model is selected based on the evaluation metric specified for the optimal model.

Automatic missing value imputation

Whether to automatically fill missing values in the data. The XGBoost algorithm has a robust mechanism for handling missing values. Enabling this option can simplify data preprocessing steps and is very useful for datasets containing missing values. If missing values have already been handled manually, this feature can be disabled.

L2 regularization term for weights

The L2 regularization parameter used to control model complexity, applied to leaf node weights. The larger the value, the stronger the regularization effect, which can prevent overfitting but may make the model too simple and lead to underfitting. Adjust this parameter based on the degree of overfitting.

L1 regularization coefficient

The L1 regularization parameter used to control model complexity, applied to leaf node weights. Setting this to 0 means L1 regularization is not used. L1 regularization will set some leaf node weights to 0, serving as feature selection. This is suitable for scenarios where simplifying the model and selecting important features is desired.

Number of workers

The number of worker processes used for parallel computation. 0 means parallel computation is not performed. When the dataset is large or there are many features, appropriately increasing the number of workers can improve training speed. However, due to hardware resource limitations, too many workers may cause resource contention and reduce efficiency.

Worker connection timeout

The timeout period for worker process connections (in seconds). 0 means no timeout is set. This parameter controls the connection wait time for worker processes during parallel computation. In unstable network environments or when computing resources are limited, setting a timeout can help avoid long waits.

Feature columns

Input independent variables used to build the regression model. Features must be numeric, and the magnitude of the coefficients can be used to determine their impact on the target.

Label column

The continuous dependent variable that the model needs to predict. This field is the output variable of the linear regression equation and must be a continuous numeric value (such as price, quantity, and more).

Loss function

A function used to measure the difference between the predicted value and the actual value. There are two options:

• squaredError (squared error): The most commonly used loss function, which calculates the sum of squared differences between the predicted and actual values. It is sensitive to outliers.
• huber (Huber loss): Combines the advantages of squared error and absolute error. When the error is small, it behaves as squared error; when the error is large, it behaves as linear error, making it more robust to outliers.

Maximum number of iterations

The maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training will stop regardless of whether convergence is achieved.

• To improve efficiency, reduce this value if the model finds the optimal hyperplane within fewer iterations.
• To ensure sufficient training when convergence is slow (for example, when the data is complex), increase this value.

Convergence threshold for iterations

The threshold for determining convergence during model training. If the change in loss value between two consecutive iterations is less than this threshold, iteration stops early. A smaller value indicates a higher requirement for convergence accuracy but may increase training time. Set a smaller value for scenarios requiring high accuracy, or increase it appropriately to improve training efficiency.

Regularization type

The regularization method used to prevent model overfitting. This is an optional parameter.

• 0 means no regularization, suitable for scenarios with large data volumes and low correlation between features.
• 1 usually refers to L1 regularization (Lasso), which sets some feature coefficients to 0 and can be used for feature selection.
• 2 usually refers to L2 regularization (Ridge), which shrinks feature coefficients toward smaller values and is suitable for handling multicollinearity.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Regularization coefficient

The coefficient for regularization strength. The larger the value, the stronger the regularization effect. Setting this to 0 disables regularization. Adjust this value based on the degree of overfitting; increase it if overfitting is severe.

Feature normalization

A preprocessing switch for converting input features to a unified scale. Enabled by default.

• When to enable: If the value ranges of features differ significantly, normalization can accelerate model convergence and prevent features with large value ranges from dominating model training.
• When to disable: If the features have clear physical meaning and the scale differences are meaningful for the business, consider disabling normalization.

Add intercept term

Whether to include an intercept term (bias term) in the model. The intercept term is used to adjust the model's baseline prediction value.

• Enabled by default. This allows the model to fit the data more flexibly.
• In certain special scenarios (such as when the data has already been centered and no baseline adjustment is needed), this can be disabled.

Feature column

The input independent variables used to build the regression model. Features should be numeric; the magnitude of the coefficient indicates the degree of influence on the target variable.

Label column

The continuous dependent variable that the model needs to predict.

Loss function

The function used to measure the difference between predicted and actual values. There are two options:

• squaredError (squared error): The most commonly used loss function, calculates the sum of the squared differences between predicted and actual values, and is sensitive to outliers.
• huber (Huber loss): Combines the advantages of squared error and absolute error; uses squared error for small errors and absolute error for large errors, making it more robust to outliers.

Maximum number of iterations

The maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training will stop regardless of whether convergence is achieved.

• If the model finds the optimal hyperplane within fewer iterations, this value can be reduced to improve efficiency;
• If convergence is slow (for example, due to complex data), this value should be increased to ensure sufficient training.

Iteration stopping threshold

The threshold for determining convergence during model training. When the change in loss value between two consecutive iterations is less than this threshold, iteration stops early. A smaller value indicates higher convergence accuracy, but may increase training time. For scenarios requiring high accuracy, set a smaller value. To prioritize training efficiency, increase this value appropriately.

Regularization type

The regularization method used to prevent model overfitting. This is an optional parameter.

• 0 indicates no regularization, suitable for scenarios with large data volumes and low correlation between features.
• 1 usually refers to L1 regularization (Lasso), which sets some feature coefficients to 0 and can be used for feature selection.
• 2 usually refers to L2 regularization (Ridge), which drives feature coefficients toward smaller values and is suitable for handling multicollinearity.

Automatic validation split

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Regularization coefficient

The coefficient for regularization strength. The larger the value, the stronger the regularization effect. Setting this to 0 disables regularization. This value should be adjusted according to the degree of overfitting. If overfitting is severe, increase this value.

Feature normalization

A preprocessing switch used to convert input features to a unified scale. Enabled by default.

• Enable when there are large differences in the value ranges of features. Normalization can accelerate model convergence and prevent features with large value ranges from dominating model training.
• Disable when features have clear physical meaning and the scale differences are significant for business logic.

Add an intercept term

Whether to include an intercept term (bias term) in the model. The intercept term is used to adjust the model's baseline prediction value.

• Enabled by default, which allows the model to fit data more flexibly.
• Can be disabled in certain special scenarios, such as when the data has already been centralized and baseline adjustment is not required.

Feature columns

Input feature fields used for model training.

Label columns

The target variable that the model needs to predict.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Use VectorIndexer

Automatic feature indexing switch for handling categorical features:

• Disabled: Categorical features must be handled manually
• Enabled: Categorical features are automatically converted to indices, with additional parameter settings provided

Maximum number of categories

Threshold for automatically identifying categorical features. If set to 32, it means that when the number of distinct values for a feature is less than or equal to 32, it is treated as a categorical feature. (Effective only when VectorIndexer is enabled)

Handling of unseen category values

The method for handling unseen category values. Choose based on data quality and business requirements: select 'error' when the data is clean, or 'skip' when the data contains noise.

• skip: Skip samples containing unseen category values
• error: Throw an error when encountering unseen category values

(Effective only when VectorIndexer is enabled)

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to choose a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns.
• The greater the depth, the more complex the tree, which may improve performance but also increases the risk of overfitting.

Maximum number of bins

The number of bins for discretizing continuous features. Used to divide continuous features into a limited number of intervals (bins) to facilitate tree splitting. The larger the value, the finer the division of continuous features.

Minimum number of samples for splitting

The minimum number of samples required for node splitting. If set to 1, it means that node splitting can occur when the number of samples at the node is greater than or equal to 1. The larger the value, the simpler the tree, which can help prevent overfitting.

Minimum gain

Minimum information gain required for node splitting. When set to 0, any gain allows a split. Setting a larger value can prevent meaningless splits and simplify the tree structure.

Maximum memory

Maximum memory allowed during training (in MB). In this case, it is set to 256 to control memory usage and prevent memory overflow.

Checkpoint interval

Iteration interval for saving checkpoints during model training. When set to 10, an intermediate result is saved every 10 splits, making it easier to resume training after interruption.

Random seed

Controls the initial value of the random process during training. Fixing the seed ensures reproducible model training results, which facilitates experiment comparison and troubleshooting; to average results over multiple experiments, different seed values can be used.

Cache node ID

Whether to cache node IDs to accelerate training. In this case, it is set to off. Enabling it can improve training speed but increases memory consumption.

Subset sampling ratio

Sample ratio for each tree during training. When set to 1, all samples are sampled with replacement. Typically set between 0.6 and 1.0.

Number of trees

Total number of decision trees in the ensemble model. The more trees, the more stable the model performance usually is, but training time also increases.

Feature selection strategy

Strategy for selecting the feature subset size when splitting each tree, with several options:

• auto: Automatically select an appropriate strategy
• all: Use all features
• onethird: Use one third of the total number of features
• sqrt: Use the square root of the total number of features
• log2: Use the base-2 logarithm of the total number of features

This parameter controls feature randomness and helps reduce correlation between trees.

Feature column

Input feature field used for model training.

Label column

Target variable (continuous value) to be predicted by the model.

Automatic validation set splitting

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Use VectorIndexer

Feature automatic indexing switch, used for handling categorical features:

• Off: categorical features must be processed manually
• On: categorical features are automatically converted to indices, with additional parameter settings available

Maximum number of categories

Threshold for automatic identification of categorical features. If set to 32, it means that when the number of distinct values for a feature is less than or equal to 32, it is treated as a categorical feature. Only effective when VectorIndexer is enabled.

Handling of unseen category values

Choose the method for handling unseen category values based on data quality and business requirements. If the data is clean, choose error; if the data contains noise, choose skip.

• skip: skip samples containing unseen category values
• error: throw an error when encountering unseen category values

Only effective when VectorIndexer is enabled.

Maximum depth

The maximum depth of each decision tree, used to control the complexity of the tree. For datasets with a small sample size, it is recommended to select a smaller depth.

• The smaller the depth, the simpler the tree and the faster the training, but it may not capture complex data patterns.
• The greater the depth, the more complex the tree, which may improve performance but also makes overfitting more likely.

Maximum number of buckets

The number of buckets for discretizing continuous features. Used to divide continuous features into a limited number of intervals (buckets) to facilitate tree splitting. The larger the value, the finer the division of continuous features.

Minimum number of samples for splitting

The minimum number of samples required for node splitting. If set to 1, it means that node splitting can occur when the number of samples at a node is greater than or equal to 1. The larger the value, the simpler the tree, which can prevent overfitting.

Minimum gain

The minimum information gain required for node splitting. If set to 0, it means that splitting can occur as long as there is any gain. Setting a larger value can avoid meaningless splits and simplify the tree structure.

Maximum memory

The maximum memory allowed during training (in MB). In this example, it is set to 256 to control memory usage and prevent memory overflow.

Checkpoint interval

The iteration interval for saving checkpoints during model training. If set to 10, an intermediate result is saved every 10 splits, making it easier to resume training after an interruption.

Random seed

Controls the initial value for random processes during training. Fixing the seed ensures reproducible model training results, facilitating experimental comparison and troubleshooting. To average results over multiple experiments, use different seed values.

Cache node ID

Determines whether to cache node IDs to accelerate training. In this case, this option is disabled. Enabling it can improve training speed but increases memory consumption.

Subset sampling ratio

The sampling ratio for each tree during training. If set to 1, all samples are used with replacement. Typically, this value is set between 0.6 and 1.0.

Maximum number of iterations

The maximum number of optimization iterations allowed during model training. When the number of iterations reaches this value, training stops regardless of convergence.

• If the model finds the optimal hyperplane in fewer iterations, reduce this value to improve efficiency.
• If convergence is slow (for example, due to complex data), increase this value to ensure sufficient training.

Step size

The contribution weight (learning rate) of each tree. Controls the influence of each tree. A smaller step size requires more iterations but results in a more robust model.

Loss function type

The function used to measure prediction error. There are two options:

• squared (squared loss): A commonly used loss function, more sensitive to outliers.
• absolute (absolute loss): More robust to outliers, suitable for scenarios with outliers in the data.

Feature selection strategy

The strategy for selecting the feature subset size when splitting each tree. Multiple options are available:

• auto: Automatically selects an appropriate strategy
• all: Uses all features
• onethird: Uses one third of the total number of features
• sqrt: Uses the square root of the total number of features
• log2: Uses the base-2 logarithm of the total number of features

This parameter controls feature randomness and helps reduce correlation between trees.

Feature column

The input feature field used for model training.

Label column

The target variable that the model needs to predict (continuous value).

Enable order-preserving

Enable the order-preserving constraint feature. When enabled, the model enforces the predicted results to satisfy the preset order relationship (such as monotonic increase) during training, ensuring that the output conforms to the sequential logic in domain knowledge; when disabled, it becomes a standard regression model without order constraints.

Automatic validation set split

Whether to automatically split the input dataset into a training set (for model training) and a validation set (for evaluating model performance).

Time column

The field used to indicate the chronological order of the data. This field must contain time information (such as date or timestamp), and the data must be arranged in chronological order. This field is the foundation for time series modeling.

Label column

The target time series to be predicted.

Autoregressive order (p)

The order of the AR part, indicating that the model uses the observed values from the previous p time steps to predict the current value. If set to 1, it means the model will use the revenue data from the previous 1 time point to predict the current revenue.

Differencing order (d)

The number of differencing operations, used to convert a non-stationary time series into a stationary series. If set to 0, it means the original series is already stationary and no differencing is needed; if the series has a trend or seasonality, d should be set to 1 or greater.

Moving average order (q)

The order of the MA part, indicating that the model uses the prediction errors from the previous q time steps to correct the current prediction. If set to 1, it uses the error from the previous 1 time point to optimize the current prediction.

Maximum number of objective function evaluations

The maximum number of evaluations allowed during model parameter estimation. Controls the search range for parameter optimization. A higher number may yield better parameters, but requires more computation time.

Maximum number of iterations

The maximum number of training iterations for the model. When the number of iterations reaches this value, training will stop regardless of convergence, ensuring that the model completes training within a reasonable time.

Consider the intercept term

Whether to include a constant term (intercept) in the model. The intercept term captures the baseline level of the time series. It is recommended to enable this in most scenarios; consider disabling it only when the series mean is 0.

Method

Parameter estimation optimization algorithm. There are two options:

• css-bobyqa: An algorithm based on constrained optimization, suitable for complex objective functions, with good numerical stability.
• css-cgd: An algorithm based on conjugate gradient descent, with faster convergence, suitable for simple scenarios.

Number of predictions

The number of future time units to predict. If set to 1, this predicts the revenue value for the next 1 time unit (such as 1 day or 1 month, depending on the granularity of the time column).

Binary classification evaluation

Used to evaluate the predictive performance of binary classification models. By comparing the true labels (label) in the test set with the model predictions, various evaluation metrics are calculated and output (such as accuracy, ROC curve, K-S value, and more) to help assess the quality of the model.

Multiclass classification evaluation

The multiclass classification evaluation operator is used to assess the predictive performance of multiclass classification models. By comparing the true labels (label) in the test set with the model predictions, various evaluation metrics are calculated and output (such as accuracy, weighted recall, and more) to help assess the model's ability to distinguish among multiple classes.

Clustering evaluation

The clustering evaluation operator is used to quantitatively assess the performance of clustering models. By inputting data containing clustering results and original features, the rationality and effectiveness of the clustering results (silhouette coefficient) are analyzed to help assess the clustering quality of the model.

Regression evaluation

This operator is used to evaluate the performance of regression models and is applicable to scenarios where the label attribute is numeric. By inputting a test sample set containing true labels (label role) and prediction results (prediction role), performance metrics reflecting the model's predictive effect are calculated and output.

Time series evaluation

The time series evaluation operator is used to assess the performance of time series forecasting models and is suitable for time series forecasting scenarios where the label attribute is numeric (such as sales forecasting, stock price forecasting, temperature forecasting, and more). By comparing the true values in the test set (label role) with the model prediction values (prediction role), various error metrics are calculated and output to quantify the prediction accuracy of the model.

Label column

Specifies the field in the test set that stores the true class (that is, the "label role" field).

Positive label value

Defines the value of the "positive class" in binary classification (used to distinguish between positive and negative classes).

Confusion matrix

Controls whether to generate a confusion matrix (shows the matching between true and predicted classes, including TP/TN/FP/FN).

Number of plotting bins

Controls the smoothness of visualization charts such as ROC curves and PR curves (the more bins, the smoother the curve).

ROC curve

Controls whether to generate an ROC curve (shows the relationship between the false positive rate (FPR) and true positive rate (TPR) at different thresholds).

K-S

Controls whether to calculate the K-S value (measures the degree of separation between the distributions of positive and negative classes, with a value range of 0-1).

Lift

Controls whether to calculate Lift (measures the improvement of the model compared to random prediction).

PR curve

Controls whether to generate a PR curve (shows the relationship between Precision and Recall at different thresholds).

Main criteria

Specifies the core metrics to be prioritized in the evaluation results:

• ROC: Suitable for overall discrimination
• K-S: Suitable for distinguishing extreme samples
• Lift: Suitable for resource optimization
• PR: Suitable for evaluating the positive class in imbalanced data

Label column

Specifies the field in the test set that stores the true class (that is, the "label role" field).

Confusion matrix

Controls whether to generate a confusion matrix (shows the matching between true and predicted classes, including TP/TN/FP/FN).

Weighted recall

Controls whether to calculate weighted recall (Recall, the weighted average of recall considering class imbalance).

Weighted precision

Controls whether to calculate weighted precision (Precision, the weighted average of precision considering class imbalance).

Weighted F-measure

Controls whether to calculate the weighted F1 score (the harmonic mean of precision and recall, providing a combined assessment of both).

Accuracy

Controls whether to calculate accuracy (the proportion of correctly predicted samples to the total number of samples).

Main criteria

Specifies the core metrics to be prioritized in the evaluation results:

• weightedRecall (weighted recall): Aggregates recall across all categories, with emphasis on majority classes. Applicable to scenarios focusing on "coverage completeness" (such as ensuring that most samples in the majority class are correctly identified).
• weightedPrecision (weighted precision): Aggregates precision across all categories, with emphasis on majority classes. Applicable to scenarios focusing on "prediction accuracy" (such as reducing misclassification of majority class samples).
• weightedFMeasure (weighted F1 score): Aggregates precision and recall, with emphasis on majority classes. Applicable to scenarios requiring a balance between accuracy and completeness (such as accurately identifying majority classes while ensuring comprehensive coverage).
• accuracy (accuracy): Overall correctness rate. Applicable to tasks with balanced classes (such as when the proportion of samples in each class is similar and overall prediction correctness needs to be evaluated).

Label column

Specifies the field in the test set that stores the true category (that is, the "label role" field).

Mean absolute error

Controls whether to calculate mean absolute error (MAE), which is the average of the absolute differences between predicted values and true values.

Mean squared error

Controls whether to calculate mean squared error (MSE), which is the average of the squared differences between predicted values and true values.

r2

Controls whether to calculate the coefficient of determination (R²), which measures the model's ability to explain data variance, with values ranging from 0 to 1.

Root mean squared error

Controls whether to calculate root mean squared error (RMSE), which is the square root of MSE and is consistent with the unit of the original data.

Main criteria

Specifies the core metrics to be prioritized in the evaluation results:

• meanAbsoluteError (mean absolute error): Applicable to scenarios focusing on the absolute deviation between predicted and true values, with lower sensitivity to outliers, such as evaluating the average scale of errors in sales forecasts.
• meanSquaredError (mean squared error): Suitable for scenarios where large errors need to be penalized (sensitive to outliers), such as when strict control of significant deviations is required in house price prediction;
• r2 (coefficient of determination): Suitable for evaluating the model's ability to explain overall data variation, such as exploring the overall impact of features on the target variable;
• rootMeanSquaredError (root mean squared error): Suitable for scenarios where the error unit needs to be consistent with the original data, making it easier to intuitively understand the magnitude of the error, such as measuring the actual degree of error in temperature prediction.

Label column

Specifies the field in the test set that stores the true category (that is, the "label role" field).

Mean absolute error

Controls whether to calculate the mean absolute error (MAE), which is the average of the absolute differences between predicted values and true values.

Mean squared error

Controls whether to calculate the mean squared error (MSE), which is the average of the squared differences between predicted values and true values.

r2

Controls whether to calculate the coefficient of determination (R²), which measures the model's ability to explain data variation, with a value range of 0–1.

Root mean squared error

Controls whether to calculate the root mean squared error (RMSE), which is the square root of MSE and is consistent with the unit of the original data.

Main metrics

Specifies the core metrics to be prioritized in the evaluation results:

• meanAbsoluteError (mean absolute error): Suitable for scenarios focusing on the absolute deviation between predicted and true values, with lower sensitivity to outliers, such as evaluating the average magnitude of error in sales prediction;
• meanSquaredError (mean squared error): Suitable for scenarios where large errors need to be penalized (sensitive to outliers), such as when strict control of significant deviations is required in house price prediction;
• r2 (coefficient of determination): Suitable for evaluating the model's ability to explain overall data variation, such as exploring the overall impact of features on the target variable;
• rootMeanSquaredError (root mean squared error): Suitable for scenarios where the error unit needs to be consistent with the original data, making it easier to intuitively understand the magnitude of the error, such as measuring the actual degree of error in temperature prediction.