State Modeling And Representation Learning For Temporal Behavior

The temporal behavioral data is widely available in various applications.How to mine the meaningful behavioral patterns in the temporal behavioral data,capture the potential needs of users,and then propose strategies for better prediction,recommendation or operation has received much attention from researchers.In this thesis,the latent state learning method for temporal behavior was proposed to realize the feature representation of discrete and sparse behavioral data,based on which the latent states of users and their complex temporal dependencies can be modeled and learned to achieve better temporal recommendations.However,state modeling and representation learning of temporal behaviors face three main challenges:(1)the real behavioral states of users are often invisible,and how to capture and infer the hidden states of behaviors based on explicit behavioral data is one of the challenges;(2)the behavioral event space is high-dimensional,which means that the samples of events that can be observed in a fixed time-segment are often sparse,and how to learn the state representation with low time complexity is another challenge;(3)the patterns of behaviors that change along with time often have individualized differences,and how to learn the changing patterns of user states is a challenge.To address the above challenges,the temporal dynamics of user behaviors were modeled into intra-and inter-period dependency patterns,respectively,from the unique perspective of the latent states of user behavior,and then a template-based graph neural network structure was used to capture the state changes of temporal behaviors.Specifically,a series of methods for deep structured state learning were proposed,including: Structure State Learning based on Hierarchical Representation(HR-SSL)and Structure State Learning based on Variational Inference(VI-SSL).The main works of the thesis include:1)To solve the problems of sparse behaviors during fixed time segments and complex behavioral dependencies in sequential recommendation,the HR-SSL model was proposed.The temporal behaviors of users on items were divided according to time segments,and then the low-dimensional representations of items were learned using the maximum pooling hierarchy to characterize the latent states of behavioral sequences.2)In order to effectively improve the state modeling of temporal behavior,the VI-SSL model was considered directly from the perspective of state representation of temporal data.Unlike the HR-SSL model based on learning item representations,the VI-SSL assumed that the prior distribution of states obeys the standard Gaussian distribution and directly learned the states of temporal sequential behaviors at different time segments,and then jointly learned collaborative filtering-based VAE and GNN-based state dependency networks.And personalized signals were introduced in state dependency networks.3)The time complexity of the proposed methods was analyzed theoretically and the proposed methods can be applied to high-dimensional time-series data.And the advantages of the proposed algorithms in various evaluation metrics are analyzed from the experimental results,verifying that the proposed algorithms have significant advantages over existing benchmark algorithms for temporal recommendation task.