project

Intent-Aware Graph Contrastive Learning for Recommendation

This project implements a self-supervised learning approach for recommendation systems that combines Graph Neural Networks (GNNs), disentangled intent modeling, and contrastive learning.

Project Structure

project/
├── data/
│   └── gowalla/            # Gowalla dataset
├── data_processing/
│   └── dataset.py          # Dataset loading and processing
├── model/
│   ├── intentgcl.py        # Main model implementation
│   └── utils.py            # Utility functions
└── run_training_testing.py # Main training script

Core Components

1. Intent-Aware GNN

   • Multiple GNN layers for message passing
   • Learnable intent prototypes
   • Attention-based intent aggregation

2. Disentangled Representations

   • Intent prototype learning
   • Multi-intent attention mechanism
   • SVD-enhanced message passing

3. Contrastive Learning

   • InfoNCE loss with temperature scaling
   • Local and global view contrasting
   • Intent-aware negative sampling

Model Architecture

The model combines three key components:

1. Graph Neural Network Layers

# Message passing aggregation
msg_user = torch.spmm(adj_norm, item_embeddings)
msg_item = torch.spmm(adj_norm.t(), user_embeddings)

2. Intent Modeling

# Intent attention computation
attention = softmax(embeddings @ intent_prototypes.t() / temp)
intent_aware_emb = attention @ intent_prototypes

3. Contrastive Learning

# Contrastive loss with temperature scaling
pos_score = (intent_emb * user_emb).sum(1) / temp
neg_score = log(exp(intent_emb @ all_emb.t() / temp).sum(1))

Usage

1. Data Preparation

   • The Gowalla dataset is used
   • Processed into training and testing matrices
   • Normalized adjacency matrix computation

2. Training

python run_training_testing.py

3. Configuration
   • Embedding dimension: 64
   • Number of GNN layers: 2
   • Number of intents: 128
   • Temperature: 0.2
   • Learning rate: 0.001

Results

The model is evaluated using standard recommendation metrics:

• Recall@20, NDCG@20
• Recall@40, NDCG@40

Results are saved after training with timestamp.

References

The implementation is based on:

1. LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation
2. Learning intents behind interactions with knowledge graph for recommendation
3. Neural Graph Collaborative Filtering