Research On Neural Network Modeling For Cross-Domain Sequential Recommendation Scenarios

In e-commerce systems and social networks,users’ historical interactions are usually recorded as sequences,which makes Sequential Recommendation(SR)a popular research topic.Moreover,as users tend to sign up for different platforms to access domain-specific services,e.g.,news subscription and video watching,Cross-domain Sequential Recommendation(CSR)that aims at making the next item recommendation via leveraging users’ historical sequential behaviors from multiple domains is gaining immense attention.Although many State-Of-The-Art(SOTA)CSR solutions have been proposed,there are still several challenges in the current CSR scenario: 1)the problem of "Shared-account",i.e.,users are sometimes willing to share their personal accounts with friends or family members,resulting in a mixed historical sequence.2)Since user’s behaviors from two domains are mixed in a same sequence,the noisy items from both domains are also combined,which affects the performance of cross-domain sequential recommenders.3)Existing studies on CSR mainly focus on using composite or in-depth structures that achieve significant improvement in accuracy but bring a huge burden to the model training.4)Moreover,to learn the user-specific sequence-level representations,existing works usually adopt the global relevance weighting strategy(e.g.,self-attention mechanism),which has quadratic computational complexity.To address the above challenges,this paper has proposed two novel CSR solutions,namely RL-ISN(Reinforcement Learning-enhanced Information Sharing Network)and LEA-GCN(Lightweight External Attention-enhanced Graph Convolution Network).Specifically,in RL-ISN,an MLP-based smoothing attention mechanism is devised to achieve the clustering of potential user from the shared-accounts.Then,to remove the impact of the noisy items during the crossdomain information sharing process,a hierarchical reinforcement learning network is exploited,which trains an agent to filter the "inter-domain" noisy information through a hierarchical Markov decision process.In LEA-GCN,a lightweight Single Layer Aggregation Protocols(SLAP)is proposed to optimize the GCN-based cross-domain sequential recommenders.Meanwhile,LEAGCN also integrates a dual-channels external attention-based model to reduce the computational complexity to the linear level.To verify the effectiveness of above two solutions,extensive experiments have been conducted on several real-world large-scale datasets.According to the results,both proposed CSR solutions achieved better performance compared with several SOTA baselines.Especially,the RL-ISN model successfully filters of inter-domain noise,while the LEAGCN model smoothly reduces the memory overhead and model size of CSR and increases the migrability of such models.Finally,this paper summarizes the problems and challenges that still exist in CSR scenario and provide an outlook for the future work.