Training the Classification Model

The SAP HANA PAL Unified Classification function [1] provides an efficient way to train classification models with enhanced features such as:

• Seamless switching between classification algorithms
• Automatic dataset partitioning
• Built-in model evaluation procedures
• Support for additional evaluation metrics

For this task, the Hybrid Gradient Boosting Tree (HGBT) algorithm is selected by setting the 'func' parameter to 'HybridGradientBoostingTree'. Finally, the training time, that is, the time taken to fit the model to the training dataset, is displayed.

123456789101112131415161718192021
# Train the classifer model using PAL HybridGradientBoostingTree
# Initialize the model object 

hgbc = UnifiedClassification(func='HybridGradientBoostingTree',
                            n_estimators = 101, split_threshold=0.1,
                            learning_rate=0.1, max_depth=6,
                            split_method='histogram', max_bin_num=256, feature_grouping=True,
                            tolerant_iter_num=5,
                            resampling_method='cv', fold_num=5, ref_metric=['auc'],
                            evaluation_metric = 'error_rate')

# Execute the training of the model
# key= 'EMPLOYEE_ID',

hgbc.fit(data=df_train.drop('EMPLOYEE_ID'), 
         label='FLIGHT_RISK', 
         partition_method='stratified', stratified_column='FLIGHT_RISK', training_percent=0.8,
         ntiles=20,  
         build_report=True)

display(hgbc.runtime)

1.82820272445678

The PAL/APL report generator is used to evaluate model performance. Currently, it supports only UnifiedClassification and UnifiedRegression models [1].

A variety of classification performance metrics exist, with AUC-ROC (Area Under the Receiver Operating Characteristic Curve) [2] being one of the most widely used. This metric, ranging from 0 to 1, assesses a binary classification model’s ability to distinguish between positive and negative classes. The closer the AUC value is to 1, the better the model performs in separating the classes.

In the report shown below, under the Statistics tab in the Stats Table, an AUC value of 0.95 is observed. Additionally, the AUC-ROC plot is available for visualization under the 'Scoring Metrics' tab.

The Variable Importance tab presents a feature importance pie chart, which assigns scores to input features based on their contribution to predicting the target variable [3]. Higher scores indicate a greater influence on the model’s predictions, with importance values ranging between 0 and 1.

From the results, 'PREVIOUS_COUNTRY' has the highest importance score of 0.23, followed by 'FUNCTIONALAREACHANGETYPE' at 0.21.

123
# Build Model Report

UnifiedReport(hgbc).build().display()