Course-Drop Prediction

This repository contains the project submission for the Introduction to Machine Learning course within the Digital Science for High-Tech program (Department of Engineering) at Tel Aviv University (Spring 2026). The project applies the CRISP-DM methodology to predict the probability that a B2B course registration will be cancelled (Dropped_Course) prior to the course start date. The official guidelines and grading criteria are detailed in the project instructions.

Quick Navigation

• Interactive Presentations
  • Interactive Notebook (HTML) – Full exploratory analysis, validation diagnostics, and model tuning.
  • HTML Overview (No Code) – High-level summary of findings and output figures.
  • Markdown Notebook Export – Static markdown version.
• Code & Scripts
  • Jupyter Notebook (ipynb) – Submitted notebook file (and the Marimo script version).
  • Inference Pipeline (pipeline.py) – Production script replicating the final predictions.

Project Overview

The prediction task is structured as a future-window forecasting problem. The dataset represents historical course bookings, where the validation and test sets are chronologically separated from the training data.

Technical Summary

• Categorical Normalization: Consolidated 2,670 raw categorical levels (containing typographic variations, casing noise, and special characters) into 201 clean levels.
• Validation Strategy: Used adversarial validation to identify temporal and distribution drift between the training and test sets (adversarial AUC of 0.935). Consequently, a chronological split (training on 2015-2016, validating on 2017) was used instead of random cross-validation.
• Feature Engineering: Formulated domain-specific indicators including cancellation histories (prev_drop_rate), lab assignment matches (got_requested_lab), group composition features, and linear time indices (days_since_epoch). High-cardinality identifiers (Agent, Country, Company) were represented using frequency encoding.
• Model Blend: Evaluated Logistic Regression (AUC 0.881), Multi-Layer Perceptron (AUC 0.877), and gradient boosted tree architectures. The final pipeline implements a rank-average ensemble of XGBoost, LightGBM, and CatBoost (validation AUC of 0.9159, Average Precision of 0.897).
• Submission Details: The current scored submission achieves a pre-submission test AUC score of 0.889314.

Repository Structure

• submission - Files submitted for grading:
  • Group_27_Notebook.ipynb - Jupyter notebook containing EDA, training, and evaluation.
  • Group_27_Submission.csv - Pre-submission prediction CSV.
  • submission-report.md - Project report (written in Hebrew).
  • Train_Data.csv and Test_Data_No_Target.csv - Source datasets.
• docs - Documentation and interactive views:
  • index.html - Landing page for GitHub Pages.
  • notebook.html - Interactive notebook export.
  • overview.html - HTML overview containing prose and figures without code blocks.
  • notebook.md - Markdown copy of the notebook.
• versions - Archived iterations and previous pipeline/notebook versions (v1, v2, v3, ron_version).
• pipeline.py - Script containing the inference pipeline.
• pyproject.toml and uv.lock - Python dependencies managed via uv.
• References - Original assignments, instructions, and course materials.

Setup and Execution

The project requires Python 3.13 and uses uv for dependency management.

Environment Setup

Sync the dependencies and build the virtual environment:

uv sync

Running the Notebook

Marimo version:

uv run marimo edit notebook.py

Jupyter version:

uv run --with jupyter jupyter notebook notebook.ipynb

Running the Pipeline

To run a dry-run check of the pipeline without writing predictions:

uv run python pipeline.py

To run the pipeline and output a prediction file:

uv run python pipeline.py --write --out data/tmp_submission.csv