hildar / recsys-retail

Recommender system in retail

Two-layer Hybrid Recommender System for retail

About

Two-layer hybrid recommender system for retail. Layer 1 uses an Implicit library for sparse data (KNN and ALS approaches). Level 2 uses a ranking model using the CatBoost (gradient boosting). This gave double the growth compared to the baseline. Evaluated by custom precision metric.

Stack:

• 1-st layer: Implicit, ItemItemRecommender, ALS, sklearn, pandas, numpy, matplotlib
• 2-nd layer: CatBoost, LightGBM

Data: from Retail X5 Hero Competition

Steps:

1. Prepare data: prefiltering
2. Matching model (initialize MainRecommender 1-st layer model as baseline)
3. Evaluate Top@k Recall
4. Ranking model (choose 2-nd layer model)
5. Feature engineering for ranking

Usage

Please, open train.ipynb Jupiter notebook file and explore how to create Recommender system step-by-step.

Project has next few steps:

1. Prepare data

First is looking at datasets and prefiltering data

2. Matching model

Learn first-layer model as baseline. In MainRecommender class we have two base models from implicit lib - ItemItemRecommender and AlternatingLeastSquares:

ALS used to find similar users, items and als recommendations. ItemItemRecommender used to find own item recommendations among user's purchases.

3. Evaluate Top@k Recall

For first-layer model we have taken Recall@k metric because it is show the proportion of correct answers from real purchases. With this approach we going to significantly cut dataset size for second-layer model.

Here we are evaluating different types of recommendations:

And are selecting optimal value of Recall:

4. Ranking model

In that step we are making new X_train dataset with target based on purchases:

Here we are choosing classifier from LightGBM and CatBoost, evaluate it by Precision@k at test data. In this step we have not impressive result.

5. Feature engineering for ranking

Adding new features for ranking model based on user, item and paired user-item data.

Controling overfitting for CatBoost and cutting extra estimators:

Ranking model gave us double the growth compared to the baseline..

As we see the best feature importance is paired user-item features: