User Preference Estimation In Conversational Recommendation

In recent years,recommender systems have gradually become one of the most important applications of artificial intelligence.Traditional recommender systems generally use historical behaviors of users(such as user’s click records,historical purchase records or scoring records,etc.)to predict user preference and recommend items to users.However,this recommendation scenario is static,and it is difficult to make accurate estimation of users’ current short-term preference from users’ historical behaviors solely.The emergence of conversational recommender systems can fundamentally alleviate the problem above.Conversational recommender systems interact with users in multiple turns of dialogue,guiding users to describe short-term preferences and address users’ current needs during the conversation.Different from general recommendation scenarios,the conversational recommender system needs to model the feedback given by the user in the conversation history to make predictions about the user’s current preferences.In this thesis,the research on the multiround conversational recommendation is listed as the following three points:First,this thesis proposes a user preference prediction model based on attribute information,which uses the self-attention mechanism to model the high-order interaction information between attributes.In multi-round conversational recommendation,the system usually uses attributes as a medium to guide users to give feedback signals,such as directly asking the user about his preference on a certain attribute,and then utilizes the answers given by the user to make better predictions of user’s current preference.After obtaining the user’s preference on attributes,it is necessary to utilize the attributes associated with each item to make more accurate estimation of preference on different items.Therefore,this thesis uses the self-attention mechanism to model the different combinations of attribute information.The high-order interaction relationship between the attributes can help system learn more accurate representations of both users and items,thereby improving preference prediction.Experimental results on two public datasets showed that the proposed model,which models higher-order interactions between attributes,achieves better results than other baseline models.Second,this thesis proposes a user preference prediction model based on user feedback,which utilizes the gating mechanism to model the relationship between different kinds of user feedback.In multi-round conversational recommendation,the system can choose to ask the user whether he likes the given attribute or recommend an item list to the user in each dialogue turn,and the user will give corresponding reply based on his preference.The attribute-level feedback comes from the user’s reply to the system’s inquiries about attributes,and the item-level feedback comes from his reply to the system’s recommended items.In order to model the relationship between item-level feedback and attribute-level feedback,the model designs a gating module that modifies the embedding of rejected items according to the attributes that the user likes in the dialogue,so as to obtain representations with items that the user currently dislikes.The model also uses another gating module to modify the user’s long-term preference representation according to the disliked attributes in the dialogue,and then integrates the different types of feedback information to predict the user’s preference.Experimental results showed that the proposed model outperforms the baseline model in both preference prediction task and user interaction task.Finally,this thesis proposes an explainable method for conversational recommendation,which can generate corresponding explanations for the prediction results given by the recommendation model.During the conversation,the system can provide the user with the explanations of current recommendation,which can not only help the user understand how the recommendation model works,but also guide the user to provide feedback signals and correct the system’s mistakes.Given the current recommendation of the model,the proposed method can inform the user how the recommendation system models his preference for tags,and calculate the contributions of different tag preference to the current recommendation.The proposed method can also modify the tag preferences and generate new recommendation results.We conducted experiments on both simulated datasets and real datasets,and the results verified the effectiveness of the proposed method.