Historical User Interest Model

This project implements a hierarchical user interest modeling system for news recommendation, focusing on combining both historical user behaviors and multimodal content representations.

Core Components

1. Data Processing (data_processing.py)

• save_imgs_reduced(): Reduces image embeddings using PCA
• save_articles_reduced(): Reduces text embeddings using PCA
• prepare_reduced_embeddings(): Main function to prepare both image and text reduced embeddings
• compute_engagement_bins(): Creates bins for user engagement metrics (read times and scroll percentages)

2. History Embeddings (history_embeddings.py)

ArticleEmbedder

• Creates base article embeddings (e_i) using categorical features
• Handles article metadata including category, subcategory, article type, and sentiment
• Produces fixed-dimension embeddings for each article

ArticleEmbeddingManager

• Manages precomputation and caching of article embeddings
• Provides efficient lookup of embeddings by article ID
• Saves embeddings in HDF5 format for efficient storage and retrieval

UserHistoryEmbedder

• Combines article embeddings with user engagement data
• Processes historical user interactions including read times and scroll percentages
• Transforms engagement metrics into learned embeddings

3. Historical Interest Model (historical_interest.py)

Base Components

• MLP: Multi-layer perceptron for feature transformation
• PointwiseAttention: Implements attention mechanism without softmax normalization

Main Models

• UserSideInterest: Combines article embeddings (e_i) with user history (e_j) to create e_u
• MultiModalInterest: Processes reduced image and text embeddings to create e_u'
• HistoricUserInterest: Final model that combines both e_u and e_u' into a unified user interest representation

Usage Flow

1. First, prepare the reduced embeddings:

prepare_reduced_embeddings(img_path='path/to/images.parquet', 
                         word2vec_path='path/to/word2vec.parquet')

2. Compute engagement bins:

compute_engagement_bins(history_path='path/to/history.parquet')

3. Initialize the models:

# Initialize base components
article_embedder = ArticleEmbedder(embedding_dim=64)
article_embedder.fit_from_parquet('path/to/articles.parquet')

embedding_manager = ArticleEmbeddingManager(article_embedder)
user_history_embedder = UserHistoryEmbedder(embedding_manager, 'path/to/bins.pkl')

# Initialize interest models
user_side = UserSideInterest(user_history_embedder, article_embedder, embedding_dim=64)
multimodal = MultiModalInterest('path/to/imgs_reduced.pkl', 
                              'path/to/word2vec_reduced.pkl', 
                              device)

# Create final combined model
historic_model = HistoricUserInterest(user_side, multimodal)

Architecture Details

The system follows a hierarchical architecture for modeling user interests:

1. Base Level: Article representations using categorical features and sentiment
2. Engagement Level: User interaction patterns through read times and scroll behavior
3. Historical Level: Attention-based combination of user history and target articles
4. Multimodal Level: Integration of reduced image and text representations
5. Final Level: Unified representation combining behavioral and content-based interests

The model aims to capture both long-term user preferences through historical behaviors and short-term interests through multimodal content understanding.