Model-Based Clustering of Multivariate EMA Time-Series Data

Bachelor's thesis project exploring clustering approaches for multivariate Ecological Momentary Assessment (EMA) time-series data.

Overview

This project implements a pipeline for analyzing EMA time-series data through feature extraction, dimensionality reduction, and clustering. The goal is to identify distinct behavioral patterns in longitudinal self-report data.

Pipeline

1. Preprocessing — Data cleaning, transformation, and structured/unstructured representations of EMA time series
2. Feature Extraction — Coefficient extraction using multiple regression models (Linear, Polynomial, Lasso, Ridge, XGBoost, Random Forest, SVR, MLP, VAR)
3. Dimensionality Reduction — PCA and PLSR with cross-validation
4. Clustering — KMeans, KMedoids, Agglomerative Clustering, Gaussian Mixture Models, and DTW-based time-series clustering
5. Evaluation — Silhouette Score, Calinski-Harabasz Index, Davies-Bouldin Index, BIC/AIC

Repository Structure

├── Clustering.ipynb       # Main analysis notebook (feature extraction, reduction, clustering)
├── Preprocessing.ipynb    # Data preprocessing pipeline
├── clustering.py          # Reusable clustering utility functions
└── Data/
    ├── Data/              # Raw data (codebook + original datasets)
    ├── preprocessed/      # Cleaned and transformed datasets
    ├── coefficients/      # Extracted regression coefficients
    └── cluster/           # Clustering results

Tech Stack

• Python 3
• scikit-learn / scikit-learn-extra
• tslearn (DTW clustering)
• pandas / NumPy
• matplotlib

Usage

Open the Jupyter notebooks in order:

1. Preprocessing.ipynb — Run preprocessing steps
2. Clustering.ipynb — Run the full analysis pipeline (feature extraction → clustering)

The clustering.py module provides helper functions used by the notebooks.

License

This project is part of an academic thesis. Please cite appropriately if you use this work.