Research Of Sequential Recommendation Under Multi-dimensional Criteria

Sequential Recommendation,which predicts users’ future behaviors based on their long-term behavior sequences or short-term sessions,has become an important area of recommender systems.With the development of deep learning,researchers have made progress in accurately predicting users’ future behaviors by modeling behavior sequences using deep neural networks.However,existing studies have pointed out that only considering accuracy of recommendations is not enough to measure the impact of the recommendation algorithm on user experience.On the one hand,pursuing accuracy alone may cause problems such as homogenization of contents and excessive tendency on popular items,which narrow the range of content for users,reduce users’ willingness to use the system,and also limit the long-term development of the platform.On the other hand,considering criteria such as diversity and explainability can enhance the richness of the content received by users,increase their trust,and have a positive effect on the platform.This thesis investigates sequential recommendations under multi-dimensional criteria.At the level of the recommendation results,this thesis focus on diversity and fairness.This thesis investigates calibrated sequential recommendation and long-tail sequential recommendation from the perspectives of item category and popularity,respectively.At the level of the related information of recommended items,this thesis focuses on the explainable sequential recommendation,which provides reasons for recommended items.Evaluating sequential recommendation algorithms by multi-dimensional criteria at the two levels can mitigate the problem of homogeneous content and make results more convincing.It is positive for users’ satisfaction and trust and the development of platforms.The novel contributions of this work are summarized as follows:(1)Calibrated Sequential Recommendation based on Decouple-Aggregated Framework: This thesis focuses on calibrated sequence recommendation which improves the diversity and fairness of recommendations in terms of item category.Due to the limitations of existing re-ranking-based calibrated recommendation algorithms,this thesis proposes an end-to-end calibrated sequential recommendation algorithm.This thesis designs a loss function for calibration,which aligns the preference distribution of the recommendation list with that of the behavior sequence.Meanwhile,due to the concern of diversity and imbalanced interests,this thesis provides modifications to the preference distribution of the behavior sequence.In addition,this thesis designs a decoupled-aggregated framework to enhance theability to handle two optimization objectives.Experiments on benchmark datasets validate the effectiveness of the proposed algorithm.(2)Calibrated Sequential Recommendation based on Cheap Causal Convolutional Network: This thesis focuses on calibrated sequential recommendation by enhancing the ability of the sequence encoder to better handle the constrained optimization objective of accuracy and calibration.To address the problem of the inability to perceive the item context in the point-by-point calculation of self attention and sequence representation generation,this thesis utilizes a causal convolution mechanism to fuse the context information of the items to enhance the ability of the self-attentive sequence encoder.Meanwhile,to handle the redundant information in the convolutional layers,this thesis utilizes a cheap convolutional mechanism.By reducing channels and augmenting representations,it can obtain better item and sequence representations with a lightweight structure.Experiments on benchmark datasets validate the effectiveness of proposed cheap causal convolutions.(3)Long-tail Session-based Recommendation from Calibration: This thesis focuses on long-tail sequential recommendation,which aims to mitigate the problem that sequential recommendation algorithms tend to recommend popular items due to popularity bias,to improve the aggregate diversity and fairness of recommendations in terms of item popularity.This thesis proposes a long-tail session-based recommendation algorithm from calibration in the session-based recommendation scenario.This thesis proposes a calibration module containing a feed-forward neural network that learns the connection between the proportion of long-tail items in the recommendation list and the session representation.It further aligns the predicted distribution with the proportion of actual long-tail items in the session.In addition,this thesis proposes a pre-training and fine-tuning mechanism to train the calibration module,and a distillation mechanism to keep the recommendation accuracy.Experiments on benchmark datasets show that the proposed framework can keep the accuracy and give more exposure to long-tail items.(4)Explainable Session-based Recommendation: Finally,this thesis focuses on the explainable sequential recommendation at the level of the related information of recommendations.To address the lack of explainability of deep-learningbased models,this thesis proposes an explainable session-based recommendation algorithm.This thesis predicts the next item for a session from three perspectives:sequential patterns,repeated clicks,and item similarity,and proposes a recommendation framework based on candidate selection and re-ranking.Experimental results on two benchmark datasets show that the proposed model not only achievescompetitive recommendation accuracy compared to the advanced deep learningbased models,but also provides reasonable explanations for recommended items according to their scores.