CMI-PIU

A repository containing the submission code for the kaggle competition Child Mind Institute — Problematic Internet Use (Relating Physical Activity to Problematic Internet Use)

Step-by-Step Plan for Building the Combined Tabular-Actigraphy Prediction Model

Objective: Develop a model that integrates tabular and actigraphy data, manages missing values, and optimizes for the Quadratic Weighted Kappa (QWK) score.

Phase 1: Data Preparation and Imputation (Done)

1. Impute Missing Values:
   • Apply iterative imputation methods (e.g., MICE) for complex relationships.
   • Use unsupervised learning (e.g., K-means, PCA, or autoencoders) to impute values based on data structure.
   • For missing target values, consider filtering them out or using models that can handle partial supervision.
2. Outlier Detection and Noise Reduction:
   • Identify and handle outliers in both tabular and actigraphy data, potentially using isolation forests or statistical thresholds.
   • Denoise actigraphy data (e.g., with smoothing or Fourier transformations) to retain essential patterns while reducing noise.

Phase 2: Feature Selection and Engineering for Tabular Data (Done)

1. Correlation-Based Feature Selection:
   • Calculate correlations between features and the target variable to identify and keep only the most relevant features.
2. Feature Interaction Creation:
   • Generate new features by interacting key features with each other, such as polynomial features or multiplicative combinations.
3. Advanced Feature Selection: -Use methods like Lasso regression, Recursive Feature Elimination (RFE), or tree-based feature importance to refine the feature set further.
4. Feature Importance Evaluation:
   • Use models like gradient boosting (e.g., CatBoost, XGBoost) or SHAP values to validate and refine features based on their contribution to the model.

Phase 3: Neural Network Encoding for Actigraphy Data

1. Actigraphy Data Encoding with a Neural Network:
   • Build a neural network encoder (e.g., with RNNs or LSTMs) to transform actigraphy sequences into fixed-size representations.
   • Optionally, integrate attention mechanisms to help the model focus on important time points within the sequences.
2. Combining Encoded Actigraphy with Tabular Data:
   • Concatenate the actigraphy encoding with selected tabular features, creating a unified feature set for the final model.

Phase 4: Model Training and Optimization

1. Model Selection and Initial Training:
   • Use gradient boosting models like CatBoost, XGBoost, or LightGBM to leverage their robustness with tabular data.
   • Optionally, employ a voting regressor that combines multiple models (e.g., CatBoost, XGBoost, NN encoder) to improve generalization.
2. Custom QWK Loss Function (if feasible):
   • Adjust training to optimize directly for QWK by implementing it as a custom loss function if the model framework allows.
3. Hyperparameter Optimization:
   • Perform hyperparameter tuning for gradient boosting models and neural networks using grid search or Bayesian optimization to maximize the QWK score.
4. Cross-Validation:
   • Use K-Fold cross-validation to train the model on multiple splits of the dataset, improving robustness and reliability in performance evaluation.

Phase 5: Model Evaluation and Submission

1. Final Model Evaluation:
   • After training, apply the model to the test dataset (if target labels are available) and calculate the QWK score to validate the model’s effectiveness.
2. Generate Submission File:
   • Using the final model configuration, generate predictions for the test set.
   • Save the predictions in the required submission format, ensuring that all necessary columns are included.