Research On Multi-task Personalized Recommendation Method Based On Attention Mechanism

As web technology and mobile communication continue to advance,people are increasingly inundated with information,leading to the problem of information overload.This makes it difficult for individuals to quickly and accurately locate information of interest.To address this issue,personalized recommendation systems leverage historical user interaction data to mine users’ personal preferences and provide relevant content.Currently,research on personalized recommendation methods is divided into point-of-interest recommendation and item recommendation for different recommendation tasks.Among these,mainstream point-ofinterest recommendation methods overlook the immediate preferences of users influenced by the geographical environment when traveling.These preferences are manifested as users may make choices different from their long-term preference patterns when they are in an area of interest.Furthermore,these point-of-interest recommendation studies only consider the provision of desired points of interest,ignoring the various items that users may be interested in included in these points of interest.Therefore,to address the problem that the personalized recommendation system cannot provide more real-time and complete recommendation services for users when they travel due to these limitations,this thesis models users’ long-and shortterm preferences and immediate preference expressions in the point-of-interest recommendation task and provides personalized point-of-interest recommendation services for users by weighted fusion of multiple preferences.In the item recommendation task,this thesis uses a conversational approach to mine users’ long-and short-term preferences about items and implements an item recommendation service after users select interest points by embedding interest point information.The main research work of this thesis is as follows:(1)Contextual information extraction methods for different user history interaction sequences are studied.In order to analyze users’ personalized preferences from multiple perspectives and thus improve the accuracy of recommendation methods,this thesis adopts different extraction methods for the multi-factor contextual information contained in the sequences to meet the personalized needs of users in different recommendation tasks,based on the differences between interest point history check-in sequences and item history interaction sequences.(2)A point-of-interest recommendation method that incorporates users’ compound preferences is proposed.The method extracts long-and short-term preferences using a twolayer attention network,and uses long-term preferences to filter users’ historical check-in records to get instant sequences,and then combines the user’s current geography and the twolayer attention network to mine the instant preferences in the instant sequences.Finally,the user’s composite preference representation is obtained by weighted fusion of the three preferences,and the score with the target interest point is calculated.This method solves the problem that the current interest point recommendation methods cannot instantly express the user’s current preference changes and thus the interest drift.(3)A method for item recommendation based on a session approach is proposed.The method extracts contextual information of different sessions based on the characteristics of the user’s historical item interaction sequences and uses trained self-encoders to obtain the corresponding embedding representations.Then uses a multi-headed attention mechanism and a Bi GRU recurrent neural network to extract the user’s item long-and short-term preference expressions.In the recommendation session,the preference expressions related to the current scene in the user’s item preferences are highlighted by embedding interest point information,which makes the method applicable to item recommendation in travel scenarios,thus realizing a multi-task recommendation service from interest point recommendation to item recommendation in user travel scenarios and improving the user’s leisure and entertainment experience.This thesis uses publicly available interest point dataset and item dataset to experimentally evaluate and analyze the two methods to verify the feasibility and effectiveness of the proposed method.The experimental results show that the proposed method has better performance in recommendation compared with traditional recommendation methods,and can provide realtime and complete recommendation services for user travel.