All Notebooks

A data-driven approach to predicting the Estimated Time of Arrival (ETA) for Universitas Indonesia's campus shuttle (Bis Kuning). This notebook analyzes travel patterns and route efficiency to improve campus mobility.

Segments human presence from millimeter-wave radar heatmaps using a custom ERASENet-style model. The notebook covers radar tensor visualization, weighted pixel scoring, and a 30-epoch PyTorch training run that reached a best validation score of 0.9733 and a test score of 0.9394.

Builds a slice-based crowd analysis pipeline that predicts both head counts and head-center coordinates in dense scenes. It combines density-map regression, SAHI-based patching, and localization evaluation, with a best validation loss of 0.0002 across the training run.

Classifies four activity classes from accelerometer and gyroscope time-series signals. The notebook combines exploratory analysis, feature engineering, and gradient-boosted modeling for competition-style sensor classification.

Detects fraudulent users in a fintech lending platform using anonymized user attributes and loan activity records. The notebook combines user profiles and transaction history to classify users as fraud or non-fraud.

Classifies multiple attributes from fashion product images, including clothing type and product color. The notebook builds an image-based multilabel classification pipeline to predict all relevant attributes from a single image.

Analyzes questionnaire-based mental health indicators to predict diagnosed mental health conditions. The notebook focuses on structured-data cleaning, survey response exploration, and supervised classification over multi-class mental health labels.

Investigates which health and demographic factors most strongly relate to total cholesterol in an employee wellness dataset. The notebook emphasizes anomaly correction, feature engineering, and Random Forest feature-importance analysis over 1,336 employee records.

Predicts whether one scientific paper should cite another by combining SPECTER-style document embeddings, sliding-window local similarity, metadata priors, and XGBoost. The final fused feature table reaches a mean 5-fold MCC of 0.6097 and produces a 336,021-row competition submission.