Research On Atypical Sequential Recommendation Problems

In recent years,with the development of information technology,contents and information on the Internet rapidly increase.In the information explosion era,users commonly suffer from the information overload issue.As one of the main ways to alleviate the information overload issue,personalized recommendation systems have been paid more and more attention by most contents and information service platforms.Different from traditional static recommendation methods that can only model users' generalized interests,sequential recommendation methods are able to capture dynamic interests of users and sequential correlations of user behaviors by analyzing and modeling their historical behavior sequences.Although many works have studied sequential recommendation problems in various scenarios,devoted many efforts to model user behavior sequences with powerful sequential models.However,these typical sequential recommendation approaches are limited to model single behavior sequence of users,which ignore the dynamic interests of groups and the characteristics of different behavior sequences.In many complex recommendation scenarios,the interaction behavior records cannot be simply generalized to a single behavior sequence,but need to construct different behavior sequences for different interactions.For example,in sequential group recommendation problem,for the purpose of capturing the common dynamic interests of the group,it is necessary to model the behavior sequences of each member in the group.In multi-channel sequential recommendation problem,in order to handle interactions under different channels and capture the dynamic interests of users in the current channel,it is essential to model users' multiple behavior sequences under different channels.In the sparse target behavior(e.g.,buying or sharing)sequence recommendation problem,the sparsity of target behaviors leads to insufficient information in the target behavior sequences.to more accurately capture users' target behavior preferences,it is required to employ another dense behavior(e.g.,clicking)as auxiliary information,and fuse target and auxiliary behavior sequence.In this thesis,we study several atypical sequential recommendation problems that require to fuse multiple behavior sequences.In summaries,the main contents and major contributions of this thesis can be summarized as follows:First,we study sequential group recommendation problem.The major challenge of this problem is how to effectively learn dynamic group representations based on the group members' behavior sequences.To address this,we devise a GLSGRL approach for sequential group recommendation.Specifically,for a target group,we construct a group-aware long-term and short-term graph to capture user-item interactions and item-item co-occurrence in the whole history and only the current time frame,respectively.Based on the graphs,GLS-GRL performs graph representation learning to obtain longterm and short-term user representations,and further adaptively fuse them to gain integrated user representations.Finally,group representations are obtained by a constrained user-interacted attention mechanism which encodes the correlations between group members.Second,we study multi-channel sequential recommendation problem.this problem requires considering the user's interaction records in different channels,fusing the user's behavior sequences in multiple channels,and capturing the user's interest preferences in the current channel.To this end,we propose a novel Sequential MultiFusion Network(SMFN)by classifying all the channels into two categories:(1)target channel which current candidate videos belong to,and(2)context channel which includes all the left channels.For each category,SMFN leverages a recurrent neural network to model the corresponding behavior sequence.The hidden interactions between the two categories are characterized by correlating each video of a sequence with the overall representation of another sequence through a simple but effective fusion unit.Finally,the two fused sequence representations are obtained as the user's interest representation in the current channel by sum-pooling operation,respectively.Third,we study sparse target behavior sequential recommendation problem.It is difficult to predict the next item based on only target behavior sequence when the target behavior is sparse(e.g.,buying or sharing an item).The data sparsity issue can effectively alleviate by taking other dense behavior sequence as auxiliary information.Therefore,the main challenge of this problem is how to integrate multiple types of behavior sequences to accurately capture the user's target behavior preferences.To this end,we first build a Multi-Relational Item Graph(MRIG)based on all behavior sequences from all users,involving target and auxiliary behavior types.Then,two models are progressively proposed to solve the problem.(1)We first propose a novel MultiRelational Graph Neural Network(MRGNN)to learn global item-to-item relations and further obtains user preferences w.r.t.current target and auxiliary behavior sequences,respectively.Afterwards,MRGNN leverages a gating mechanism to adaptively fuse user representations for predicting next item interacted with target behavior.(2)In order to facilitate the training and further boost the performance of the recommendation,we then propose LP-MRGNN.In this model,we incorporate link prediction into multi-relational item graph modeling,acting as a simple but relevant task to sequential recommendation.The extensive experiments on real-world datasets demonstrate the superiority of the model over diverse and competitive baselines,validating its main components' significant contributions.