| The traditional recommendation systems mainly use offline user data to train offline models,and then recommend items for online users.There are three intrinsic problems,which are the unreliable estimation of user preferences based on sparse and noisy historical data,the ignorance of online contextual factors that affect user behavior,and the unreliable assumption that users are aware of their preferences by default.While the online recommendation system can update in real time based on the current behavior data of users,it is still the traditional recommendation mode in essence,that is,the system passively accepts the data generated by the user.With the development of dialogue system,dialogue technology provides a new idea to solve the intrinsic problems of traditional recommendation,which is Conversational Recommendation.Conversational recommendation allows the recommendation system to dynamically capture user preferences in the interactive session with users.In this way,online and active conversational recommendation can guide users to explore their preferences and assist users in decision-making.A major research challenge of conversational recommendation is to balance the benefits of exploring users’ preferences behavior and recommending items behavior.Long session of interaction can indeed obtain more user preferences,so as to improve the accuracy of recommendation,but it also brings a burden to users,which causes the loss of users.However,due to the lack of context information modeling,the existing conversational recommendation methods lack initiative and cannot deal with the problem of exploration and exploration,resulting in excessive burden on users.Therefore,from the perspective of improving the sample quality,this research work introduces Knowledge Graph as the interactive environment,designs the interactive task to actively obtain the user’s online feedback under the condition of minimizing the user’s burden,and realizes an efficient conversational recommendation algorithm based on high-quality samples.The research contents and main contributions of this paper are as follows:1.This paper proposes a knowledge-enhanced single-session active interactive recommendation algorithm.The data of KG is enhanced to solve the problem of exploration and income imbalance caused by insufficient environmental information modeling.The user’s historical behavior data is constructed into a heterogeneous graph,which is integrated with the external item attribute knowledge graph to serve as the context environment.The idea of active learning and hard negative sampling are introduced to enhance the uncertainty signal and negative sample signal in the environmental graph respectively.Based on the enhanced environmental knowledge,an efficient single-session active interactive recommendation algorithm is realized.2.This paper proposes a knowledge-guided multi-session active interactive recommendation algorithm.Multi-session scenario is introduced to expand the algorithm framework to solve the problem of sparse interactive data in a single session.Each single session is constructed as a subgraph,then the representations of session strategies are learned from the subgraphs of sessions.By taking the historical sessions as the important interactive experience,the strategy candidate set is generated to provide strategic guidance for the current session.Based on the guidance of strategy knowledge,an efficient multi-session active interactive recommendation algorithm is realized.In this paper,two research scenarios,single-session scenario and multi-session scenario,are discussed.Based on the enhanced environmental knowledge and the guidance of strategy knowledge,the proposed active interactive recommendation algorithm interacts with users efficiently and recommends items accurately.Experimental results demonstrate the superiority of proposed algorithm with significant improvements over state-of-the-art methods. |
