Research On Session-Based Recommendation With Multi-Task Learning

As an effective technology to alleviate the information explosion problem,recommenda-tion systems are becoming increasingly important in many real-world scenarios,such as e-com-merce platforms,media streaming websites,and search engines.Session-based recommenda-tion,which aims to predict the next interested item based anonymous behavior sequences(e.g.,item click sequences)of users,is critical in modern recommender systems.With the increasing study of the session-based recommendation problem,many kinds of proposals have been de-veloped.While prior works have made efforts to addressing the session-based recommendation problem,two significant limitations exist:(1)The existing methods ignore the fact that items may be correlated with others across different sessions;(2)The existing methods are also lim-ited in their assumption of rigidly ordered pattern over intra-session item transition,which may not be true in practiceIn order to address these above limitations,this paper proposes a multi-task learning frame-work for joint modeling the local and global transitional information of items for session-based recommendation.It includes the following two well designed modules:(1)Cross-session item dependency encoder module.This module combines all historical sessions to form an item graph,and uses a graph embedding algorithm to capture global transitional information of items across different sessions;(2)Session-based recommendation module.This module includes a dual-stage attentive encoder and a bilinear decoder to learn local item transitional information and capture both current interest and main purpose of users in current session.With the explo-ration of both complex intra-and inter-session interest transitional regularities,the designed model enables the representation learning of user behavior dynamics via jointly mapping local and global signals into the same latent space.In order to explore the impact of different graph embedding algorithms on the model,we design two different cross-session item-dependent en-coders based on random walks and deep graph informax,respectivelyIn the experimental part,we carry out extensive contrast experiments with state-of-art al-gorithms on two real world datasets.The experimental results demonstrate the feasibility and superiority of the designed model.