ML Capabilities

Classification: Random Forest, LDA, QDA, KNN, SVM | Validation: Temporal CV, Walk-Forward, K-Fold | Explainability: SHAP values, feature importance, decision boundaries

Classification (7 tools)

train_swing_classifier

Train RF/DT/SVM/logistic model for FH/BH/Serve classification using Zepp sensor data

train_stroke_classifier

Train model using linked Apple Watch sessions with Zepp as ground truth

train_discriminant_classifier

Linear or Quadratic Discriminant Analysis with boundary visualization

train_knn_classifier

K-Nearest Neighbors with optimal k selection via cross-validation

predict_stroke_type

Predict stroke types for Apple Watch session using trained model

compare_classifiers

Benchmark KNN, LDA, QDA, RF, DT and rank by performance

suggest_classifier

Recommend best classifier based on problem characteristics

Validation (5 tools)

train_swing_classifier_cv

Proper temporal cross-validation preventing future data leakage

run_cross_validation

Systematic CV with stratified/kfold/loo and multiple metrics

run_walk_forward_validation

Sliding window validation mimicking production deployment

create_temporal_cv_splits

Time-aware folds with purge gap to prevent temporal leakage

evaluate_temporal_stability

Test if classifier degrades over time (monthly accuracy trend)

Explainability (4 tools)

compute_swing_shap_values

SHAP values using TreeExplainer for feature contributions

explain_swing_prediction

SHAP waterfall for single swing showing push toward each class

get_swing_feature_importance

Compare SHAP importance vs gain importance

analyze_feature_interaction

SHAP interaction values between two sensor features

Tuning & Diagnostics (3 tools)

tune_hyperparameters

Grid search for optimal k, regularization, tree depth, etc.

generate_learning_curves

Learning curves for bias/variance diagnosis

get_ml_capabilities

List all available ML tools and techniques by category

Example Workflow: Train and Explain Classifier

Step Query Tool
1 "train stroke classifier with random forest" train_swing_classifier
2 "compare classifiers performance" compare_classifiers
3 "train with temporal cross validation" train_swing_classifier_cv
4 "compute SHAP values" compute_swing_shap_values
5 "explain prediction for swing 0" explain_swing_prediction
6 "flag ambiguous swings for review" flag_ambiguous_swings

Audio-Based Training

Use voice labels as ground truth for higher accuracy than sensor-inferred labels.

train_with_audio_labels - Train classifier using spoken shot labels aligned to motion peaks.