Dual conditional diffusion transforms how we model user preferences in recommendation systems
Bridging discrete-continuous gap in recommendation systems using dual diffusion
📚 https://arxiv.org/abs/2410.21967
🎯 Original Problem:
Current diffusion-based sequential recommendation systems face two limitations: They model diffusion for item embeddings instead of discrete items, creating inconsistency. And they use either implicit or explicit conditional models alone, limiting their ability to capture user behavior context.
-----
🔧 Solution in this Paper:
→ Introduces DCRec (Dual Conditional Diffusion Models for Sequential Recommendation) with:
- A discrete-to-continuous framework bridging item spaces using complete Markov chain
- A Dual Conditional Diffusion Transformer combining both implicit and explicit conditioning
→ Key mechanisms:
- Forward Process: Adds noise to both history sequence and target items
- Reverse Process: Uses dual conditional mechanism for denoising
- Efficient few-step inference process reducing computational overhead
-----
💡 Key Insights:
→ Dual conditioning (implicit + explicit) outperforms single conditioning approaches
- Implicit captures global preferences
- Explicit preserves sequential dynamics
- Combined approach prevents overfitting to noise
→ Complete Markov chain bridges discrete-continuous gap effectively
→ Few-step inference achieves optimal results with reduced computation
-----
📊 Results:
→ Outperforms SOTA across multiple datasets:
- 3.45% improvement in HR@5 on Beauty dataset
- 3.16% improvement in NDCG@10 on Beauty dataset
- 17.35% improvement in HR@5 on Toys dataset
→ Achieves better results with fewer sampling steps
Share this post