Research On POI Recommendation Method Based On Hawkes Process

In recent years,in the era of data explosion and information overload,POI(point-of-interest)recommendation tasks have become one of the most important tasks in location-based social networks(LBSNs).POI recommendation tasks usually help users select places of interest by modeling their historical check-in data and contextual information.This will not only help users explore POIs and promote outdoor activities for users,but also help businesses accurately place commercial advertisements so that they can obtain higher profits.Since POI check-in activities are closely related to people's daily life behaviors,more and more research work has begun to simulate users' behavior habits from two perspectives of time and space.But the existing work usually only models the time interval between POI check-ins to consider the time dependence between check-in points,but does not consider that the influence between check-in points may change with time.To improve this problem,we will use the Hawkes process to capture the time-dependent dependencies between POI check-in points,and integrate them in expression learning models,deep learning models,and optimization matrix solution models.Three POI recommendation methods based on Hawkes process give users recommendations.At the same time,we also consider factors such as user preferences,spatial impact and other factors that can affect the performance of POI recommendation to improve the final recommendation performance.First,we propose a new hawkes process based self-exciting embedding model(HPSEE),which embeds users and places into the Hawkes space to capture individual preferences and check-ins of users Periodic time pattern between points.On the basis of HPSEE,we also proposed a hawkes process based self-modulating embedding model(HPSME)to jointly capture the excitatory and inhibitory effects between check-in points.Second,we propose a new continuous-time LSTM(long short-term memory)model Nhaw ST(neural hawkes with spatio and temporal influence)that incorporates a space gate,using neural networks to capture more complex dependencies between check-in points.Due to the consideration that the influence between points of interest will change with time,we changed the traditional LSTM to a continuous-time LSTM.At the same time,we also added a space gate to the proposed LSTM variant to model the distance between adjacent check-in points to capture the user's short-term and long-term dependence between geographic intervals.Finally,we propose a graph regularized multi-dimensional hawkes process POI recommend model GRMHP(graph regularized multi-dimensional hawkes process),which effectively integrates the a priori spatial structure into the multi-dimensional hawks process.We represent the prior spatial structure as a connection graph,and then use the multi-view method to compare the prior connection graph and the impact matrix.Our colleagues maintain the sparsity and low rank of the kernel matrix.In addition,we also use the alternating direction multiplier method to solve the resulting optimization problem.