Research On Ranking Algorithms Of Recommender Systems Based On Deep Learning And Historical Interaction Sequence Modeling

In the context of the rapid development of Internet technology,people are facing the dilemma of information overload while enjoying the convenience of acquiring abundant resources.The emergence of recommender systems provides new possibilities for people to select information of interest from massive amounts of data.Therefore,in recent years,recommender systems have played an increasingly important role in practical applications in e-commerce,online advertising systems,and social media.In practice,a complete recommender system usually contains a candidate generation stage and a ranking stage.This article mainly focuses on the algorithm research work in the ranking stage.The key to the ranking models of recommender systems is how to effectively utilize rich features and their complex relationships to achieve the goal of accurately predicting the user preference for each item.To realize the personalized recommendation,in recent years,researchers design a lot of refined ranking models based on advanced technologies such as the attention mechanism to strengthen the use of historical interaction information.However,the long-tail problem of items widely exists in actual recommendation scenarios.Due to the insufficiency of interaction information,the recommendation for long-tail items is more difficult than recommendations for highly exposed items.Besides,new users and new items keep appearing in the system.How to make recommendations in the absence of historical interaction information has become a more difficult challenge than the long-tail item recommendation problem,and gradually formed the research direction of cold-start recommendation.This article takes historical interaction sequence information as a starting point to study how to make better use of historical interaction sequence information to improve the prediction performance of the ranking models of recommender systems,and how to design algorithms to make full use of only information to quickly generate high-quality alternative features when the historical interactive sequence information is insufficient.On this basis,we propose two recommendation algorithms based on deep learning to specifically solve the above problems,and apply them in the actual book recommendation system.The main research contents and contributions of this article are as follows:Firstly,from the perspective of effective utilization of historical interactive information,we first model the interaction between users and items as a dynamic bipartite graph,and then propose a recommendation model called multi-directional interactive graph attention network(MIGAT).The MIGAT model respectively considers the importance of each interaction in the historical interaction sequence to its interactive objects,recommended objects,and recommended context through the attention mechanism.At the same time,the impact of the occurring time of interactions on the interaction confidence is taken into the consideration.The confidence embedding vectors are used to distinguish the confidence of different interactions.Integrating multiple importance measurements,the MIGAT model generates more expressive user feature representations,item feature representations,and adaptive interaction sequence feature representations based on graph neural network and feed-forward neural network structures.Experimental shows that this model performs better than the state-of-the-art recommendation ranking models on multiple datasets and the performance improvement is more obvious when the long-tail problem is more serious.Secondly,to handle the problem that new users or new items lack historical interaction,we propose a meta-learned pseudo interaction generator model(Me PIGen).The Me PIGen model learns the parameters based on the meta-learning algorithm so that the model can generate high-quality pseudo interaction sequence features after a few explorations of new users and new items.The pseudo interaction sequence features will be used as alternative features of the real interaction sequence in ranking models,to improve the prediction performance of the ranking models.The structure of the Me PIGen is designed based on the Transformer.It takes the existing historical interaction sequence and profile features as input and obtains the output pseudo interaction sequence.Such structure ensures the quality and efficiency of the generated pseudo interaction sequence.Experimental shows that this model improves the cold-start recommendation performance of the state-of-the-art ranking models on multiple datasets.Finally,we apply the two recommendation models proposed in this article to the actual book recommender system we built.Results show that recommendation performance is satisfied.In our book recommendation system,a complete recommendation environment is constituted from the user cold-start phase to the regular phase that the new user has accumulated a certain amount of interaction data,which demonstrates the practical application value of the proposed models.