Back to research

RESEARCH EXPLORATION / EXPLAINABLE AI

Explainable Machine Learning for Predicting Problematic Short-Form Video Use Among University Students

A synthetic-data, explainable-ML workflow examining how model comparison and feature-level explanations can make an applied model easier to inspect and challenge.

Open reproducibility notebook

01 / Research question

How can an explainable modelling workflow estimate elevated short-form-video-use risk while making its predictions, thresholds, and feature contributions inspectable?

02 / Dataset

The notebook explicitly labels the final validation data as synthetic. It uses 100,000 rows, with 80,000 development rows and 20,000 holdout rows, targeting risk_high_flag. The final validation summary records 81 approved predictors.

IMPORTANT BOUNDARY

These are synthetic modelling results, not clinical, causal, or population-level conclusions about university students.

03 / Methodology

Synthetic student data→
Cleaning and feature preparation→
Development / holdout split→
Model comparison→
Calibration and threshold selection→
SHAP / LIME→
Robustness and fairness checks

04 / Models and explainability

The notebook records logistic-regression, decision-tree, Random Forest, and HistGradientBoosting experiments. A calibrated logistic-regression model is selected for final validation, with SHAP feature-attribution outputs and a LIME local-explanation artifact retained in the reproducibility materials.

05 / Evaluation

The following figures are copied from the notebook's final validation summary and are intentionally labeled as synthetic-data results.

100,000

synthetic dataset rows

81

approved predictors

0.811

holdout ROC AUC

0.613

holdout PR AUC

0.145

holdout Brier score

0.729

selected-threshold balanced accuracy

06 / Interpretability and checks

  • - SHAP summary and feature-ranking outputs.
  • - LIME local explanation artifact.
  • - Threshold sensitivity, negative-control, robustness, and fairness outputs.

07 / Limitations and reproducibility

The project is a methodological exploration over synthetic data. The notebook is the source of code, generated-data methodology, tracked outputs, and the exact implementation needed to reproduce or challenge these results.