YSDA RecSys 2025 Lavka Solution

This repository contains a solution for the Kaggle competition YSDA RecSys 2025 Lavka. The goal is to predict user-item interactions for a grocery recommendation system.

Getting started

# Download competition data
kaggle competitions download -c ysda-recsys-2025-lavka
unzip ysda-recsys-2025-lavka.zip

# Install dependencies
pip install -r requirements.txt

# Prepare daily features
python precalc_features.py

# Run training
python trainer.py

# Optimizes model params automatically
python optimize.py

Key Components

1. calculate_stats.py

• Generates collaborative and statistical features:
  • Collaborative filtering: NPMI, Jaccard similarity, SVD embeddings
  • Log1p-counts for actions (views, purchases) grouped by user/product/store/city
  • Click-through rate (CTR) features
  • Temporal features (hour, weekday, time_of_day, etc.)

2. Featurizer

• Computes features in causal mode (no future data leakage)
• Stores daily feature snapshots for efficient joins
• Uses sliding windows (3/7/30/all days) for time-decay modeling

3. Preprocessor

• Handles competition-specific rules:
  • Makes train/val split
  • Filters small request_ids

4. Trainer

• Supports three training modes:
  • Random lag: Random historical offset (1-30 days) for the each sample
  • Constant lag: Fixed historical offset (e.g., 7 days)
  • Ensemble: Combines random + all constant lags
• Uses CatBoost:
  • CatBoostClassifier for binary classification (view/purchase)
  • CatBoostRanker with pairwise NDCG optimization
• Implements smart caching for fast iteration

5. Hyperparameter Optimization

• optimize.py: Uses Optuna for GPU-accelerated parameter tuning

6. Metrics & Submission

• Calculates NDCG@10 for ranking evaluation
• Handles ensemble result merging
• Produces Kaggle-ready CSV submissions

Key Implementation Details

1. Causal Feature Engineering

• All features are computed per day with no future data leakage
• Windows handle time series changes

2. Ensemble Strategy

• Combines 31 models (random lag + 30 constant lags)
• Uses day-specific weighting for constant lags

3. Efficiency Features

• Parquet caching for fast feature reuse
• GPU acceleration for CatBoost and collaborative features