The goal of this project is to develop a predictive model that analyzes children's physical activity and fitness data to identify early signs of problematic internet use. Identifying these patterns can help trigger interventions to encourage healthier digital habits.
this project analyzes and predicts problematic internet use (piu) using the child mind institute dataset we use sleep, activity, demographic, and survey data to build machine-learning models that classify piu risk levels.
dataset: https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/data
- project-objectives
- dataset
- project-structure
- data-preparation
- feature-engineering
- modeling-approach
- results
- limitations
- future-work
- license
- references
- develop predictive models to identify youth at risk for problematic internet use (piu).
- investigate relationships between physical activity / sleep (actigraphy) and sii
- produce reproducible pipelines (eda → fe → models → evaluation).
- publish a clear kaggle notebook + submission ready pipeline.
-
competition: child mind institute — problematic internet use
https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/data -
key files:
train.csv,test.csv,sample_submission.csvseries_train.parquet/*(actigraphy time-series)series_test.parquet/*(actigraphy time-series)
read csv & parquet (polars/pandas).
per-id aggregation of time-series (mean/std/skew/kurtosis/enmo).
non-wear detection & masking.
imputation: knn / simple mean/median for numeric; mode for categoricals.
save processed sets to data/processed/.
time-series aggregates (per-day / per-week): mean, std, max, min, pctiles.
derived features: bmiage, internet-hoursage, enmo-norm, sleep-efficiency.
autoencoder embedding for time-series (optional).
feature selection via correlation thresholding & tree-based importance.
target: sii (ordinal 0..3). approaches:
regression + threshold tuning (optimize QWK)
ordinal classification (optional)
models:
catboost, xgboost, lightgbm
voting regressor ensemble
optional: MLP / CNN / LSTM on raw actigraphy
validation:
stratified k-fold (n=5) based on sii
track oof preds and use scipy.optimize for threshold tuning
| model | accuracy | precision | recall | f1-score | cohen-kappa |
|---|---|---|---|---|---|
| catboost | 0.3289 | 0.34 | 0.33 | 0.325 | 0.10 |
| lightgbm | 0.4794 | 0.48 | 0.48 | 0.47 | 0.21 |
| xgboost | 0.4066 | 0.41 | 0.40 | 0.40 | 0.16 |
| voting-regressor | 0.5474 | 0.5946 | 0.5474 | 0.5387 | 0.3465 |
missing data and imputation bias.
cross-sectional data: cannot infer causality.
limited demographic diversity may limit generalizability.
sensor noise and variations in wear-time may affect actigraphy features.
ordinal regression specialized models (proportional odds)
end-to-end deep learning on raw actigraphy (1D-CNN / transformer).
longitudinal study to probe causality.
explainability (shap for per-subject explanations).
this project is released under the mit license. see LICENSE for details.
child mind institute — problematic internet use competition (kaggle): https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/data
literature on actigraphy and adolescent behavior (see REFERENCES.md in repo)..