Research On Real-time POI Recommendation Based On Users' Longterm And Short-term Preferences

With the development of network technology and the wide application of positioning technology in people's life,Location-Based Social Network(LBSN)has gained rapid development and is increasingly integrated into people's daily life.The geographical location information in LBSN provides rich data support for researching people's behavior patterns and mining people's rules of movement.In recent years,Point-of-Interest(POI)Recommendation,as one of applications in LBSN,has gained much attention.It mines users' preferences from users' historical check-in location sequences to recommend POIs that users have interests in,which reduces users' trouble in searching those POIs from numerous sites and brings convenience for users.In real life,POI recommendation usually recommends the next POI which the user may visit next according to his check-in sequence called next POI recommendation.At present,most next-POI recommendation methods usually focus on the influence of users' check-in sequences on the following visit preferences,but ignore the users' preferences changing over time.So,the recommendation results cannot match the realtime interests of users at the current time,which not only affects the accuracy of the recommendation results,but also influences the effectiveness of POI recommendation.Therefore,how to mine the real-time preferences of users based on the influence of check-in sequences has become a key issue in the real-time POI recommendation.To address the key issue above,this paper proposes a real-time POI recommendation algorithm based on Long Short-Term Memory(LSTM),which mines users' preferences from their long-term and short-term preferences.The specific contents are as follows:Users' long-term preferences are often reflected in the historical activity patterns,and current states are likely to have similar preferences with the similar states in the past.Based on this idea,this paper uses a non-local network framework to model users' long-term preferences by constructing temporal and spatial similarities between current and historical states.At the same time,the everyday activity patterns of users often show periodicity and take one week as a cycle.Therefore,this paper adopts a periodic function as the weighting function to better express the activity patterns of users.Users' short-term preferences is highly real-time and will change over time.In addition,people's preferences will be affected by public preferences to some extent.Hence,when modeling users' short-term preferences,a transfer vector is introduced for each time slot to represent the influence of public preferences on users in the corresponding time slot.Besides,before the final POI recommendation,the places that do not conform to people's living rules in the current time slot will be eliminated according to the visiting rules of the POI categories,so as to reduce the search space and generate the recommendation lists which match users' living habits more.Moreover,we design relevant experiments to verify the effectiveness of each component of the model,and compare the model with the baseline methods on realworld datasets to verify the effectiveness of the model.