Research And Implementation Of Conversational Recommendation Method On User Intention Understanding

The recommender system is the result of the rapid development of Internet,it can help users to find items that meet the preferences from the item pool when the information is overloaded.Because the traditional recommender system uses interaction history to estimate user intention,this static method has some defects,such as unable to capture interest drift,passive access to user preferences,etc.These limitations inevitably affect the performance of the recommender system.The above problems can be alleviated by introducing conversation into the recommender system.This thesis is devoted to the research of conversational recommender system,that is,actively engaging in conversation with users to dynamically capture user preferences,and finally recommending suitable items to users in a limited number of rounds.Therefore,how to find a good conversation strategy and fully understand the user intention are important issues for conversational recommendation.The main research work of this thesis is as follows:(1)This thesis analyzes the development background and significance of the current conversational recommendation method,as well as the research status at home and abroad,and then introduces the related concepts and technologies in detail.(2)To make full use of the interactive information of user history in the conversation,this thesis designs a conversational recommendation method considering the interactive sequence of users’ history items,that is,firstly,the user preferences is learned by modeling the interactive sequence of users through the self-attention network and forward feedback network,then the user intention is updated through multiple rounds of conversations with users through reinforcement learning,and appropriate conversation actions are learned to realize the items recommendation,in which the conversation history and the number of candidate items are taken as the state input of reinforcement learning.Finally,the effectiveness of this method is verified by the success rate and average turn on the datasets.(3)To make full use of conversation information and realize fine-grained understanding of user intention.In this thesis,entity embedding representation is obtained by graph neural network,and then use gating to process attribute-level feedback and item-level feedback to adapt user embedding.In order to improve the performance of reinforcement learning,we limit the size of action space,and take the attributes accepted by users as the state of reinforcement learning after passing through graph convolution network,Transformer and average pool.Finally,experiments on datasets show that this method has a higher success rate and a lower average turn,and achieves a better ranking of target items.(4)Based on the two proposed methods,this thesis implements a movie conversational recommender system,which updates and understands user intentions through conversation and provides personalized recommendation services to users.