DeepFM-ANN Recommendation System

This repository contains an optimized recommendation system for the MovieLens dataset, with a specific focus on cold-start user scenarios using a hybrid approach combining DeepFM and Approximate Nearest Neighbors (ANN).

Overview

The recommendation system combines several state-of-the-art techniques:

• DeepFM Model: Combines factorization machines for recommendation with deep neural networks
• HNSW Approximate Nearest Neighbors: For fast candidate retrieval during recommendation
• Cold-start Optimization: Specialized handling for users with limited rating history

Key Features

• Efficient Recommendations: ANN-based retrieval enables fast recommendations with minimal accuracy loss
• Cold-start Handling: Optimized for new users with limited rating history
• Comprehensive Evaluation: Leave-one-out methodology with relevant metrics (Hit Rate, MRR, NDCG)
• Memory-efficient Design: Batch processing for handling large datasets
• Model Persistence: Save/load capabilities for model weights and mappings

Requirements

numpy
pandas
torch
hnswlib
scikit-learn
matplotlib
tqdm

Usage

Basic Usage

# Load and preprocess data
ratings_df, movies_df = load_movielens_data("path/to/movielens")
data = preprocess_for_recommendation(ratings_df)

# Create recommendation system
rec_system = MovieLensRecommendationSystem(data["num_users"], data["num_items"])
rec_system.set_mapping(data["reverse_user_map"], data["reverse_movie_map"], movies_df, data["ratings_df"])

# Train model
rec_system.train(data["user_ids"], data["movie_ids"], data["labels"], epochs=10)

# Build ANN index for fast recommendations
rec_system.build_ann_index()

# Get recommendations for a user
recommendations = rec_system.recommend_items_ann(user_id, top_k=10)

Cold-start Evaluation

# Run cold-start evaluation
results = run_improved_coldstart_evaluation(
    "path/to/movielens", 
    sample_size=None,
    num_test_users=100,
    test_ratio=0.2
)

Structure

The code is organized into several key components:

1. Data Loading and Processing

   • load_movielens_data(): Loads and samples MovieLens dataset
   • prepare_coldstart_evaluation(): Creates training/testing splits for cold-start evaluation
   • preprocess_for_recommendation(): Prepares data for the recommendation model

2. Core Models

   • DeepFM: Neural recommendation model combining factorization machines and deep learning
   • HNSWIndex: Wrapper for HNSW approximate nearest neighbors index
   • CollisionlessEmbeddingTable: Embedding table with expiration mechanism

3. Recommendation System

   • MovieLensRecommendationSystem: Main class that integrates models for recommendations
   • Methods for both standard and ANN-based recommendations

4. Evaluation

   • create_leave_one_out_testset(): Creates test data for leave-one-out evaluation
   • evaluate_leave_one_out(): Evaluates recommendation methods using leave-one-out methodology
   • compare_coldstart_methods_leave_one_out(): Compares different recommendation approaches

Cold-start Optimization

The system handles cold-start users with several strategies:

1. Strategic initial rating selection based on user preferences
2. Larger candidate pool during ANN retrieval for better recall
3. Hybrid re-ranking approach combining popularity and similarity
4. Optimized thresholds for seen/unseen item determination

Evaluation Metrics

The system is evaluated using:

• Hit Rate (HR@k): Percentage of users for whom the held-out item is in the top-k recommendations
• Mean Reciprocal Rank (MRR): Average of reciprocal ranks of the held-out items
• Normalized Discounted Cumulative Gain (NDCG): Measures the ranking quality of recommendations

Results

Comparative evaluation between standard and ANN-based recommendation methods:

• ANN method is significantly faster (typically 5-10x speedup)
• Loss in recommendation quality (comparable Hit Rate, MRR, and NDCG)

License

MIT License