Research On Algorithms Of Recommendation Based On User Behavior Diversity

As an effective and efficient tool for solving the information overload problem,Recommender Systems(RS)plays an important role in this information explosion era.RS can help users filter information and find potential interest promptly.The core of RS is the ecosystem of users,platforms and recommendation algorithms.This recommendation ecosystem consists of three parts:(1)users visit the platforms and interact with the items provided by the platforms;(2)platforms collect the interaction data between users and provided items;(3)recommendation algorithms perform recommendation models to analyze user preferences and recommend items that users possibly be interested in to users.Among them,as the basis of this recommendation ecosystems,interaction behaviors between users and items play important role in analyzing user preferences.Therefore,making full and reasonable use of historical interaction behavior data play critical role in improving user satisfaction,platform profits and social financial status.However,traditional recommendation models only consider whether there are interactions between users and items,ignoring the diversity of user behavior.That is,apart from the singular interaction behavior between users and items,there are also social behaviors between users and users,diversity of user intents when choosing items and diversity of interaction types between users and items.Considering these problems,modeling the diversity of user behavior has important theoretical value and practical significance for improving recommendation ecosystem and promote the economic development of the society.However,modeling the diversity of user behavior has several challenges:(1)As important supplementary information for modeling user preferences,the social behavior between users and users is of great significance for improving recommendation performance.However,the degree to which different users are influenced by social factors varies.For example,some users pretend to believe in what their friends recommend to them while some users insist on what they choose.Modeling social influence for all users uniformly cannot discriminate the diversity of social influence,further limiting recommendation performance.(2)User intents varies when choosing items.For instance,why users interact with items can be attributed to interest,conformity or randomly selecting.Learning user properties and item attributes in a unified way may results in the disentanglement of user intents.As a result,model cannot capture the factors that influence the interactions between users and items,thereby affecting the recommendation performance.To handle this problem,currently,there are a handful research attempt to develop disentangled models.However,they disentangle user intents in an implicit way,which lead to suboptimal performance.(3)There are interaction overlaps between multiple type of behavior data.For example,a user first adds an item to cart and purchases it subsequently.Then,the interaction record between this user and item co-exists in both “add-to-cart” and“purchase” behavior data.The recommendation models based on the overlapping interaction data can lead to the confusion of user interest and semantic information of multi-type behaviors.Furthermore,it can increase the time and memory consumption of models.This thesis studies these three problems and the contributions and innovations of this thesis are as follows:Firstly,in terms of the diversity of social influence,this thesis proposes an Adaptive Social Information Fused Graph Convolutional Network Recommendation Model(shorted as ASI-GCN).Specifically,proposed model consists of three parts:embedding(“embedding”,“representation” and “feature vector” represent the same meaning,“embedding” is used in the rest of the thesis)propagation process in user-item interaction graph,embedding propagation process in user-user social graph and adaptive social information aggregation process.ASI-GCN resorts to Gumbel-Softmax trick to generated categorical distributions between embeddings from user-item graph and user-user graph.Then,these embeddings are aggregated proportionally according to the generated categorical distributions when update user and item embeddings.It can model the diversity of social influence.ASI-GCN outperforms competitive methods by3.21%-13.21% in terms of several evaluation metrics(Recall,Precision,HR and NDCG).Secondly,in view of the diversity of user intents,this thesis proposes an Explicit Rating Based Disentangled Graph Convolutional Network Recommendation Model(shorted as ERDGCN).Concretely,ERDGCN has an interest degree calculation module,which estimate users’ interest degrees to different items according to the explicit feedback.Based on the calculated interest degrees,holistic user-item interaction graph can be divided into several intent-aware subgraphs.In each graph,the graph convolutional network is performed independently,so as to prevent the interference of multiple user intents.In the end,multi-task learning is used to optimize model parameters in different graphs jointly,so as to capture user intents comprehensively.ERDGCN outperforms competitive methods by 5.17%-14.35% in terms of several evaluation metrics(Recall,Precision,HR and NDCG).Finally,in terms of the diversity of interaction type between users and items,this thesis proposes a Non-overlapping Heterogeneous Graph Convolutional Network Multi-behavior Recommendation Model(shorted as No-HGMR).To be specific,a multi-type behavior data reduction module is designed to process multiple types of interaction data to make them independent and non-overlapping.Based on the nonoverlapping behavior data,heterogeneous graph convolutional network is performed learn user properties and item attributes.Apart from the discrimination of multi-type behaviors,there are correlations between different type of behaviors.Hence,NoHGMR fuse the behavior inter-correlation embedding into the graph convolution process,so as to capture the commonality between multiple types of behaviors.Specifically,No-HGMR outperforms competitive methods by 5.12%-45.98% in terms of several evaluation metrics(HR and NDCG).