Research On POI Conversational Recommendation Method With Reinforcement Learning

With the rapid development of the mobile Internet and the advent of the big data age,the amount of information that users can obtain through terminals has exploded,resulting in"information overload".In order to mine the information that users are interested in from the massive information,the recommendation system emerges as the times require.POI(Point of Interest)recommendation is an important scenario in the recommendation system.POI recommendation aims to learn more accurate user interest embeddings based on the users’ historical interaction sequences,so as to recommend POIs that users are interested in.However,in real life,historical interaction sequences cannot reflect the current dynamic preferences of users,and conversational recommendation can collect users’ dynamic preferences through dialogue,which can alleviate this problem to a certain extent.Reinforcement learning simulates the reward and punishment mechanism in the learning process,it can make decisions with more long-term benefits,and can better solve the problems of sequential decision-making.Now,many works apply reinforcement learning to conversational recommendation,and train efficient dialogue strategies through reinforcement learning,so as to obtain the users’ current preference information faster and help make better recommendations.This article is devoted to studying how to integrate the users’ historical interaction sequences into the conversational recommendation model in a suitable way,exploring around spatio-temporal information and users’ multi-interest information,and training an efficient dialogue strategy through reinforcement learning to achieve better POI conversational recommendation.The main work of this article is as follows:(1)This article analyzes the current research background and significance of POI recommendation and conversational recommendation,as well as the research status at home and abroad,and then introduces related concepts and technologies such as the deep reinforcement learning.(2)In order to solve the problem that the traditional POI recommendation methods cannot obtain the dynamic user preferences and the existing conversational recommendation methods do not consider the spatio-temporal information,this article proposes a POI conversational recommendation method based on PG strategy fused with the spatio-temporal information.This method uses the self-attention network to learn embeddings that incorporate the spatio-temporal information.Based on the embeddings and users’ interaction sequences,a dialogue state fused with the spatio-temporal information is designed,and the dialogue strategy considering the spatio-temporal information is learned through the policy gradient algorithm(PG)in reinforcement learning.The dialogue strategy based on reinforcement learning selects dialogue actions according to the dialogue states fused with spatio-temporal information,and obtains users’ dynamic preferences by asking questions or recommendations to help make accurate POI recommendations.Finally,experiments are conducted on two real datasets to verify the effectiveness of the model.(3)For the diversification of users’ interests in POI conversational recommendation,this article proposes a POI conversational recommendation method based on users’ multiinterest and node-level AC strategy.This method uses graph path reasoning for conversational recommendation,and learns interest embeddings and their weights represented by the set of user preference attributes in the current dialogue based on the self-attention network and the gated recurrent units.The weighted average of interest embeddings is fed into the Actor-Critic network(AC)in reinforcement learning to learn a fine-grained dialogue strategy that considers users’ multi-interests information.The fine-grained dialogue strategy based on reinforcement learning calculates the probability of graph nodes representing each candidate action according to the weighted average of interest embeddings,and performs path inference according to the probability.This results in more efficient conversations and more accurate POI recommendations.Finally,experiments are conducted on two real datasets to verify the effectiveness of the model.(4)In order to learn the complex relationship between users,POIs and attributes and alleviate the noise interference in interest embeddings,this article proposes a POI conversational recommendation method based on the spatio-temporal graph convolution interest and weight denoising AC strategy.This method obtains the graph embeddings fused with the spatio-temporal information based on the powerful representation ability of the graph convolutional network,and obtains interest embeddings fused with the spatio-temporal information and their weights through the self-attention network and the gated recurrent units.Only the spatio-temporal interest embeddings with weights Top-k are fed into the ActorCritic network(AC)in reinforcement learning to remove noise from the low-weight parts.Different spatio-temporal interest embeddings are fed into different Actor-Critic networks to learn the probabilities of candidate actions under different spatio-temporal interest embeddings.Perform reinforcement learning training on multiple Actor-Critic networks,and select the action with the highest probability under the condition of Top-k spatio-temporal interest embeddings for dialogue each time.Better POI conversational recommendation results are obtained by using more informative graph embeddings and denoising spatio-temporal interest embeddings.Finally,experiments are conducted on two real datasets to verify the effectiveness of the model.