PHKT model beats 9 baselines in multi-behavior recommendation
A new hybrid model uses hypergraphs and KAN to predict purchase behavior with 14 pages of evidence.
In multi-behavior recommendation systems, auxiliary user actions (clicks, add-to-cart, purchases) provide rich signals to predict the ultimate target behavior (e.g., a purchase). While graph and hypergraph methods capture high-order relationships among users, items, and behaviors, they often struggle with heterogeneous semantics, user-specific weighting, and temporal dynamics. Standard Transformers excel at sequence modeling but use a shared feedforward network that cannot differentiate between diverse latent patterns across behaviors.
To address this, the authors introduce PHKT (Personalized Hypergraph-enhanced Kolmogorov-Arnold Network Transformer). The model features a personalized dynamic hypergraph module that weights item similarities based on each user's historical behavior, capturing heterogeneous high-order relationships. The temporal backbone is a Transformer enhanced with KAN layers in its feedforward network, replacing traditional MLPs to better model nonlinear responses to different behavior patterns. Experiments on three real-world datasets – Tmall, RetailRocket, and IJCAI – show PHKT consistently outperforming nine strong baselines across multiple evaluation metrics, demonstrating its effectiveness for target behavior prediction in multi-behavior scenarios. The paper is 14 pages with 6 figures and 6 tables.
- PHKT outperforms 9 baseline models on 3 real-world datasets (Tmall, RetailRocket, IJCAI).
- Replaces standard Transformer MLPs with Kolmogorov-Arnold Networks (KAN) for finer nonlinear modeling.
- Personalized dynamic hypergraph module captures user-specific heterogeneous relationships across behavior types.
Why It Matters
Better multi-behavior recommendation means e-commerce platforms can predict purchases from clicks and cart additions more accurately.