Research On Sequential Recommendation Of Dynamic And Static Interest Modeling

The rapid development of cloud computing and the Internet of Things has resulted in a massive amount of online data and a proliferation of network applications and services,which lead us into the era of big data.However,this situation has also resulted in a serious "information overload" problem,where users find it difficult to find content or services that they are interested in.To solve the problem,recommender systems,which use user historical behavior data to provide personalized services and improve customer experience,are becoming increasingly popular in e-commerce,multimedia,and many other fields.Usually,users interact with different items in a chronological order,which brings huge potential to sequential recommendation research in both academic and industrial fields.The core of sequential recommendation is how to build an effective and accurate representation of user interests.Existing sequential recommendation works typically model user interests from the perspective of long-term and short-term interests.Particularly,longterm interests are also referred to as static interests,and short-term interests are also referred to as dynamic interests.However,the existing distinction between long-term and short-term interests often only lies in the different sequence lengths used for interest modeling,thereby preventing relevant methods from fully exploring the representation of interests from both dynamic and static perspectives.To solve the problem proposed above,this dissertation conducts research on mining more efficient users’ dynamic and static interests to achieve more accurate personalized sequential recommendation.The main contributions of this dissertation include two parts:(1)Sequential recommendation model based on single-level dynamic-static interest mining:The dissertation presents a sequential recommendation model called DSIM(Dynamic and Static Interest Mining)for pure dynamic and static interest mining,the single-level refers to DSIM directly modeling the user’s dynamic and static interests for sequential recommendation.It models user’s dynamic interests through a neural Hawkes process that perceives duration time and user’s static interests through non-invasive self-attention mechanism.DSIM combines dynamic and static interests with a gating mechanism and generates the hybrid interest for achieving better personalized recommendations.(2)Sequential recommendation based on multi-level dynamic-static interest mining:to effectively and deeply mine the dynamic-static interests of users,this dissertation proposes a Hierarchical Dynamic and Static interest Mining(HDSM)based sequential recommendation model,the multi-level refers to HDSM decoupling the dynamic and static interests within long-term and short-term interests respectively,and recombining them to make recommendations..Note that the user’s interests are not instantly changing and there is a certain process of transition between interests.Therefore,this dissertation also considers the evolution of interests from static to dynamic interests.Specifically,HDSM first models the respective dynamic,static and evolve interests in long-term interest and short-term interest through two independent self-attention modules.Then,the long-term dynamic interest and short-term dynamic interest are combined to obtain the user’s dynamic interest,and the user’s static interest and evolving interest are modeled similarly.Finally,the complete dynamic interest,static interest,and evolving interest are fused to obtain the final interest representation and realize personalized recommendations.