H&M Personalized Fashion Recommendation System

Implementation of different models and analyzing their accuracy as per the kaggle competition standard. After research and evaluating different models ,the conclusion reached was that a hybrid approch seems the way to go and my proposal for the recommendation model is hybrid pipeline of SASRec+GNN+Faiss

1. Performance Considerations:

• Implemented: Added dask support for loading large datasets (Step 6).
• Implemented: Reduced batch sizes for Sentence-BERT (64), GNN (64), SASRec (32), and TabNet (512) to handle memory constraints.
• The script uses Parquet files for faster read/write and smaller file sizes.
• GNN and SASRec training use efficient data loaders (NeighborLoader, batching).

2. Model Improvements:

• Implemented: SASRec now uses 4 blocks and 8 heads for better sequential modeling (Step 10).
• Implemented: Faiss now uses an IVF index with 100 clusters for faster search (Step 11).
• Implemented: TabNet hyperparameters adjusted (n_d=128, n_a=128, n_steps=5) for better performance (Step 12).
• GNN already includes dropout, L2 regularization, and early stopping to prevent overfitting.

3. Embedding Quality:

• Implemented: Upgraded Sentence-BERT to all-mpnet-base-v2 (768D) for better text embeddings (Step 9a).
• Implemented: Adjusted UMAP n_components to retain at least 98% variance based on PCA (Step 9c).
• User embeddings combine GNN and SASRec for relational and sequential patterns.

4. Evaluation Insights:

• Implemented: Tested different ensemble weights (TabNet vs. SASRec) and selected the best based on MAP@12 (Step 13).
• Implemented: Increased the number of candidate items retrieved by Faiss to 200 (Step 13).
• SASRec and TabNet have been fine-tuned (more blocks/heads for SASRec, adjusted hyperparameters for TabNet).
• Metrics (RMSE, MSE, MAE, R-squared, MAP@12) are computed and saved for analysis.

5. Next Steps:

• Implemented: Tested different ensemble weights and saved the best submission (Step 13).
• Use Faiss indices for real-time retrieval of similar users/items.
• Example: Retrieve top-k similar items for a given item item_index = faiss.read_index(f"{Config.processed_dir}/item_faiss_index.bin") query_embedding = item_combined_embeddings[0:1] # Example query distances, indices = item_index.search(query_embedding, k=5)
• Use TabNet feature importance for business insights (e.g., which features drive purchases).
• Use the evaluation metrics to compare different model configurations or ensemble weights.