Skip to content

Train Model

This class is responsible for training a machine learning model to classify emotions or behaviors based on movement data. It automates the entire training process, including preparing the data, finding the best model settings, evaluating performance, and saving the results.

Step-by-Step Explanation:

  1. Input Data:
    • The input is a dataset with:
      • Features: Data that describes movements (e.g., accelerometer readings).
      • Target Label: The emotion or state you want the model to predict (e.g., arousal or valence).
  2. Target Label Encoding:
    • The target label (e.g., "high," "medium," "low") is converted into numbers (e.g., 0, 1, 2) so the model can understand it.
    • A mapping is created to show how the original labels correspond to these numbers.
  3. Model Selection:
    • You can choose from three machine learning models:
      • XGBoost: Great for handling complex datasets.
      • Support Vector Machine (SVM): Works well for smaller datasets.
      • Random Forest: Easy to use and interpretable.
    • This choice is specified in the configuration file.
  4. Hyperparameter Tuning:
    • Models have settings (called hyperparameters) that affect how they learn.
    • The class automatically tries different combinations of these settings to find the best one using Bayesian optimization. This saves time compared to manual tuning.
  5. Cross-Validation:
    • The class uses Stratified Group K-Fold Cross-Validation to test the model’s performance during training.
    • This ensures the model is evaluated fairly and works well across different groups of data (e.g., participants).
  6. Model Training:
    • The best settings from hyperparameter tuning are used to train the model.
    • The class outputs the model’s accuracy and a detailed classification report (e.g., how well it predicts each label).
  7. Saving Results:
    • The class saves:
      • The trained model (for future use).
      • Metadata, such as:
        • The best settings.
        • The model's accuracy.
        • Label mappings and feature importances (if available).
      • A detailed classification report as a JSON file.
  8. Output Files:
    • The saved files include:
      • Model File: Contains the trained model.
      • Metadata File: Includes information about the model’s performance and settings.
      • Classification Report: Shows how well the model predicts each label.