Session-Based Prefix-Target Modeling

Colab-first pipeline for turning the RecSys 2015 / YOOCHOOSE click log into prefix→target training pairs. Everything happens inside notebooks/session_based_prefix_target_datasets.ipynb, which you can rerun end-to-end to recreate data/processed/.

Notebook in a nutshell

1. Setup & download – mounts Drive, installs Kaggle CLI, pulls the 33M-row yoochoose-clicks.dat, and prints file sizes for quick checks.
2. EDA & sorting – loads the clicks table, reports key stats (sessions, items, session-length histogram), and globally sorts events by session_id + timestamp.
3. Temporal split – derives session end-times and slices train/valid/test by chronology (80/10/10) to avoid leakage.
4. Filtering & remapping – keeps train items with ≥5 interactions, drops sessions <2 events, optional session-length cap, and remaps item_id to contiguous ids.
5. Pair generation – streams prefix lists and next-item targets into Parquet (train|valid|test_pairs.parquet) plus an accompanying item_map.json.
6. Validation & viz – reloads the Parquet files, checks prefix length distributions/ID bounds, and includes optional Plotly timelines for sampled sessions.

Outputs

Artifact	Purpose
data/processed/train_pairs.parquet	Training prefixes/targets post-filtering
data/processed/valid_pairs.parquet	Chronological validation split
data/processed/test_pairs.parquet	Held-out evaluation split
data/processed/item_map.json	Original item → contiguous id lookup

Quick start

1. pip install -r requirements.txt
2. Open the notebook (locally or in Colab), upload your kaggle.json, run all cells.
3. Consume the Parquet outputs:

   import pandas as pd
   train_pairs = pd.read_parquet("data/processed/train_pairs.parquet")

Repo overview

├── data/        # external/raw/interim/processed
├── notebooks/   # notebook named session_based_prefix_target_datasets (1).ipynb
├── src/         # placeholder package for downstream code
├── requirements.txt
└── README.md

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience