Recommendation Algorithms Based On User Behavior Sequence Modeling

With the ever-growing volume of online information,recommender systems have been a fundamental tool to overcome information overload and play an essential role in almost every daily aspect of working,study,entertaining and business.The con-ventional recommendation algorithms,including the collaborative filtering method and content-based filtering method,utilize all historical interactions(e.g.,a user's clicks on items)to learn user preferences.These methods generally assume that user preferences are static.However,the static assumption is insufficient to reflect the user's dynamic interests over time.In real-world scenarios,a user's choice on items not only depends on his/her static interest but also depends on his/her short-term dynamic interest.To address this problem,the recommendation algorithms based on user behavior sequence modeling arrange the user's historical interactions into a sequence in chronological or-der and characterizes the user's long-and short-term preferences by considering the temporal dependency of items in the sequence,thereby improving the recommendation accuracy.This thesis is based on user behavior sequence modeling to study how to character-ize user interests to improve recommendation accuracy.Since user behavior sequences in different scenarios have different characteristics,this thesis conducts sufficient in-vestigation on user behavior sequences in single-domain scenarios and cross-domain scenarios.The main contribution of this thesis can be summarized into three folds:First,this thesis proposes a temporal hierarchical attention model for the long-term user behavior sequence modeling in single-domain recommendation scenarios.The user interactions are divided into multiple temporal windows,which can not only cap-ture the short-term dynamics of user interests but also reduce the computational cost for long-term behavior sequence modeling.Within each window,a category-and item-level attention mechanism is proposed to characterize user interests.The category-level attention is to describe user's diverse interests,and the item-level attention is to profile fine-grained user interests.Furthermore,a forward multi-head self-attention mecha-nism is proposed to identify and integrate the long-term correlation between the pre-viously split windows.The proposed method is tested on a new dataset of 1.7 million micro-videos,coming from real data of a micro-video sharing service in China.Ex-perimental results demonstrate the effectiveness of the proposed method in comparison with the state-of-the-art solutions.Second,a pre-training recommendation algorithm based on a bidirectional encoder is proposed to model short-term user behavior sequences in single-domain recommen-dation scenarios.A bidirectional encoder based on the self-attention model is adequate to capture item co-occurrence patterns in short-term sequences.Furthermore,the pro-posed method is built upon a pre-training and fine-tuning framework,which trains the bidirectional encoder in a self-supervised manner through the mask language model to avoid the impact of the imbalance of the click-through data.The proposed method is tested on the dataset of content-based video relevance prediction challenge held at the ACM international conference on Multimedia 2019.The experimental results verify the superior performance of our proposed method compared with state-of-the-art methods.Third,this thesis proposes a domain-transfer network based on adversarial pre-training for the user cross-domain behavior sequence in cross-domain recommendation scenarios.First,since user behaviors in the content domain are sparser than in the prod-uct domain,this thesis proposes an attentional domain-transfer network,which effec-tively selects the relevant items in the two domains to characterize user preferences.Furthermore,to bridge the domain gap,a dual-domain contrastive adversarial learning method is proposed to pre-train the feature extractors for content and product simulta-neously.We conduct offline and online experiments on the e-commerce platform,and experimental results demonstrate the advantage of our proposed method that consis-tently outperforms the state-of-the-art methods.