TARDIS — Predicting the Unpredictable

SNCF TGV delay analysis and prediction platform built with scikit-learn, XGBoost, LightGBM, PyTorch, TensorFlow, and Streamlit.

Project structure

Tardis/
├── cleaned_dataset.csv       Cleaned dataset (output of EDA notebook)
├── model.joblib              Trained model pipeline (output of model notebook)
├── route_stats.csv           Route-level median stats (output of model notebook)
├── requirements.txt
├── tardis_eda.ipynb          Data cleaning, exploration & feature engineering
├── tardis_model.ipynb        Model training orchestrator (calls scripts/model/)
├── tardis_dashboard.py       Streamlit dashboard
└── scripts/
    ├── cleaning/             Data cleaning modules
    ├── visualization/        EDA plot modules
    ├── merging/              Dataset merging utilities
    ├── audit/                Data quality tracking
    └── model/
        ├── config.py         Constants (target, features, season map)
        ├── preprocessing.py  Preprocessor pipeline + route stats builder
        ├── sklearn_models.py sklearn / XGBoost / LightGBM training & tuning
        ├── dl_models.py      PyTorch & TensorFlow MLP models
        ├── evaluation.py     Metrics and comparison utilities
        ├── predict.py        predict_delay(dep, arr, date, model, stats)
        └── train.py          Full training orchestrator (also CLI-runnable)

Setup

git clone <repo-url>
cd Tardis

python -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

Usage

1. Data cleaning & EDA

Open and run tardis_eda.ipynb from top to bottom. Produces cleaned_dataset.csv.

2. Model training

Open and run tardis_model.ipynb from top to bottom. Produces model.joblib and route_stats.csv.

Or train from the command line:

python -m scripts.model.train cleaned_dataset.csv

3. Dashboard

streamlit run tardis_dashboard.py

The dashboard opens at http://localhost:8501 with three tabs:

Tab	Content
🎯 Predict a delay	Select two stations + date → predicted delay + historical distribution + seasonal comparison
📊 Explore data	KPIs, delay distribution, monthly trend, station ranking, year×season heatmap, correlation matrix, CSV export
🔍 Model insights	Active model info, feature list, feature importance chart, prediction explanation

Models trained

Framework	Model
scikit-learn	Linear Regression, Ridge, Random Forest, Gradient Boosting
XGBoost	XGBRegressor
LightGBM	LGBMRegressor
PyTorch	MLP — 256 → 128 → 64 → 1 (BatchNorm + Dropout)
TensorFlow	MLP — 256 → 128 → 64 → 1 (BatchNorm + Dropout + EarlyStopping)

The dashboard always uses the best sklearn/boosting model after GridSearchCV tuning (serialised with joblib). DL models are compared in the notebook but not deployed.

Dataset format

The pipeline expects a CSV named dataset.csv (or cleaned_dataset.csv post-EDA) with at minimum:

• Date — ISO date string
• Departure station, Arrival station — station names
• Average delay of all trains at arrival — prediction target (minutes)
• See tardis_eda.ipynb for the full column list and cleaning steps