| With the rise of social networks and online communities,group activities have become increasingly common,which has stimulated a growing interest in improving the effectiveness and efficiency of group recommendation systems.Unlike personalized recommendation,the goal of group recommendation is to provide recommendations that meet the common preferences of a group of users.The large number of explicit and implicit interactions among groups,users,and items in group decision-making makes it more challenging to recommend for groups.There are several key issues that need to be addressed in group recommendation research.Firstly,group discovery methods often fail to simultaneously consider user behavioral preferences and social relationships,which may make it difficult to accurately reflect users’ needs and preferences.Moreover,these methods may not be trained in a group-discovery-oriented manner,which may result in user features being unsuitable for group division.Secondly,preference aggregation methods often overlook user differences inside and outside groups,lack effective utilization of the complex relationships among groups,members,and items,and do not fully model user interactions,resulting in poor recommendation effectiveness.With the purpose of addressing the above problems,this thesis conducts research from the perspectives of group discovery and preference aggregation:From the perspective of group discovery,this thesis proposes a group discovery method based on user behavior and social relationship,aiming to address two problems in this field: the underutilization of user features and the difficulty of performing end-toend training for models.The method combines user behaviors and social relationships to mine users’ latent features and detects groups through self-optimized clustering analysis.Specifically,this thesis proposes three tasks: user preference feature learning task,user social feature learning task,and clustering discovery task.The user preference feature learning task analyzes users’ historical behavior records and learns the cross-feature of user and item to obtain user preference features.The user social feature learning task analyzes user social networks and learns the interactions among users to mine user social features.The clustering discovery task combines the two types of features and performs group division.Through multi-task joint training,the method realizes end-to-end training guided by group discovery and improves the quality of user feature embedding and the accuracy of group division.Simulation experiments show that the proposed method has better performance both in group discovery and group recommendation gain.From the perspective of preference aggregation,this thesis proposes a group recommendation method based on decision process omni-modeling,aiming to overcome the difficulty of comprehensively considering multiple factors in group decision-making modeling.This method combines user preferences,group preferences,member contributions,and complex relationships among users,groups,and items to build group preferences model and generate recommendations.Specifically,the method introduces the concept of independent preferences for groups and users and the concepts of groupmember relevance and item-member relevance,proposes a relevance-attention strategy,and considers the mutual influence among members.Firstly,the method analyzes the user-item and group-item interactions separately in a pre-training manner to learn the independent preference features of users and groups.Then,it constructs and analyzes the member interaction graphs within groups to learn the mutual influence among members.Finally,it explores the relationship between groups and members,as well as between items and members,to discover potential connections.At the same time,the method models the importance of members in the same group to identify member contributions and perform aggregate accordingly.Simulation experiments show that the proposed method outperforms baseline methods and is reasonable and robust. |
