Research On User Behavior Sequence:Modeling And Applications

User behavior sequence is the user behavior organized and recorded in the form of sequence,which contains user's habits and preferences and has great application value for many network services.In real scenarios,there are various types of user behavior data,but not all of the data in the sequence is related to the user's preference and in-tent.From the perspective of contents,user behavior can be divided as macro objects and micro-operations.The macro object is the content concerned by users.That is,the product provided to users by network services,such as items in online shopping,songs in music websites,movies in video websites and more.The micro-operation refers to the specific behavior that the user locates the macro objects,such as clicking,purchase,and commenting for items,and auditioning,skipping,and liking for songs.Macro ob-jects and micro-operations show user preferences and intentions at different levels and granularities,but existing works often focus on single granularity,unable to capture the multi-level user preferences.In addition,from the perspective of users,user behavior can be divided into single-person sequences and community behavior sequences.Dif-ferent from the personalized single user sequence,the behavior of the community is a mixed result.While the users in the community have their own unique habits,they also contain some common characteristics and trends.In summary,modeling techniques for user behavior sequences still face the following challenges :(1)In macro object se-quences,it is difficult to model the users' current preferences due to the interference of unrelated data.(2)There are complex nonlinear relations and different behavior patterns in macro object and micro-operation sequences,which are difficult to directly model.(3)In the community behavior sequence,it is difficult to capture the trend of behavior.In view of the challenges above,from the perspective of macro objects and micro-operations,single user and community sequence,this paper takes three typical applica-tions as examples for the modeling technique of user behavior sequence.The specific research contents include:(1)For macro object sequence modeling,we propose a dual sparse attention net-work.This model introduces a special index for the target object,and uses a self-attention mechanism to capture the collaborative information between sequential ob-jects to model the real intent and interest of users.Meanwhile,an adaptive sparse trans-formation function is proposed to automatically ignore unrelated data in the sequence with the attention network.In the scenario of the session-based recommendation,ex-perimental results demonstrate the ability of the proposed method to model real interests and the ability to distinguish unrelated data better.(2)For the complex relations between macro objects and micro-operations,we propose a neural tensor factorization technique combining the attention mechanism.In the framework of tensor factorization,neural networks are introduced to capture the nonlinear relations between user-operation and object-operation.Meanwhile,an atten-tion network is introduced for users' historical micro-operation sequence to capture the sequential pattern.In the scenario of tag recommendation,the effectiveness of the pro-posed method for user behavior modeling is verified by experiments.(3)For the different behavior patterns in macro and micro behavior sequences,we propose a unified framework based on graph neural network for encoding behaviors.Firstly,the sequence of macro objects is transformed into a multiple graph,and the graph neural network is used to aggregate the object information.For micro-operations,the recurrent neural network is used to incorporate the micro-operation information into the graph neural network to capture the sequence pattern.In addition,we propose an improved self-attention mechanism to capture the pattern of the dyadic relation between micro behaviors for the pairwise relationships.In the scenario of the session-based recommendation,the model effectively captures different behavior patterns.(4)For the behavior trend in the community,we propose a dynamic evolution framework that combines two graph structures to predict the popular elements of the behavior object.Specifically,we construct a dynamic community-attribute bipartite graph to learn the collaboration of different communities.Next,we transform the bi-partite graph into a hypergraph to exploit the associations of different attribute tags in one community.Lastly,a dynamic evolution component based on the RNN is intro-duced to capture the behavior trend.In the community trend prediction scenario,the model can effectively discover the community trend of attribute words in advance.