Triple-Attention

Dynamic Heterogeneous Hypergraphs with Triple Attention for Multi-View Recommendation

Dynamic Hypergraph Neural Networks with Triple Attention for Recommendation Systems

This repository contains the implementation of a novel recommendation system that combines dynamic hypergraph neural networks, triple attention mechanisms, and contrastive learning for enhanced recommendation accuracy.

Overview

Our framework introduces several innovative components:

• Dynamic hypergraph construction with triple attention mechanism
• Hierarchical LightGCN backbone for multi-view learning
• Adaptive edge-dropping view augmentation
• Multi-view contrastive learning representations

Key Components:

1. Dynamic Hypergraph with Triple Attention

   • Multi-relational attention mechanism
   • Sparse attention for computational efficiency
   • Higher-order attention for complex relationships
   • Dynamic weight matrix updates

2. Hierarchical LightGCN Backbone

   • Multi-view framework integration
   • Contrastive learning strategy
   • Efficient message passing mechanism
   • Streamlined architecture without feature transformation

3. Adaptive Edge-Dropping

   • Attention-based edge dropping
   • Learnable parameters for optimization
   • Preservation of essential connections
   • Dynamic adaptation during training

4. Multi-View Contrastive Learning

   • Local contrastive learning for within-view similarities
   • Hierarchical global contrastive learning across views
   • Combined loss function with regularization

Embedding Generation

• First view embedding
• Second view embedding
• Combined embedding