A Sparse Sinkhorn Transformer for Multi-Behaviour Recommendation

Tendai Mukande, Esraa Ali, Annalina Caputo, Ruihai Dong, Noel E. O'Connor

Abstract: Modelling dynamic user-item interactions using GNN in Recommender Systems has attracted significant research interest recently. Existing research has focused on single type user-item interactions such as “purchases or “user ratings in ecommerce domain, whereas real-world scenarios involve multi-type interactions such as “view”, “add-to-favourites”, “add-to-cart” and “product purchase” that are not considered by these approaches. To address this problem, Vanilla transformer-based self-attention mechanism have been adopted in literature for multi-behaviour recommendation. These methods suffer from high computational cost which results in scalability issues especially on large datasets. To overcome this limitation, we propose the Sinkhorn Transformer for Multi-Behaviour Recommendation (STMBR) model. Our proposed method uses a differentiable algorithm to sort and re-arrange elements of the input sequence based on Sparse Sinkhorn Attention. Experimental results on three real-world-datasets show our model’s competitive recommendation performance against the state-of-the-art baselines at lower computational cost.