Location Prediction Based On Deep Factorization Machine And Recurrent Neural Network

With the rapid development of Location-based Social Networks(LBSN),Location prediction is an important research problem for both academia and industry in recent years.Location prediction plays an important role in many areas.People's future mobility patterns can be reflected through their personal preferences and check-in history.Nowadays,the popularity of smart media allows people to share experiences and strategies related to tourist attractions,restaurants,hotels,etc.on social platforms.Location prediction is to analyze and learns these data to predict the user's future location information.The prediction results can be applied to areas such as attraction recommendation and path planning.This thesis analyzes and researches from two aspects: predicting the number of potential users from the perspective of location.And predicting the user's next location from the perspective of user.The main contributions of this thesis are as follows:1.We propose a method called DFMLP(Deep Factorization Machine for Location Prediction)for predicting the number of potential users from the perspective of location.DFMLP considers check-in features from multiple dimensions(i.e user's preference,location attraction,temporal features,spatial features,weather features)of social network data.More specifically,we use a Factorization Machine component to remove the sparsity of the user-location check-in matrix and learn low-order feature interactions between users and locations.And we design a feature extraction modeling component to process the temporal-spatial-weather features.Then,we use a deep neural network component to bridge all check-in features to capture high-order latent attributes between users and locations.In addition,we propose to non-uniformly weight the location based on the number of times the user has checked-in on each location.Further,considering the sparsity of user data,we propose a new negative instances sampling algorithm.Experiments on two typical real-world datasets show that DFMLP is superior to the baselines.2.We propose a method called RLPN(Recurrent Location Prediction Network)for predicting the user's next location from the perspective of user.We use matrix factorization to infer the missing values of the user-location check-in matrix to obtain the user's global preference,and then use Long Short-Term Memory(LSTM)to learn theuser's long-term and short-term preference.In order to solve the problem of data sparseness and cold start,we consider two aspects of social relationship,namely,the friend relationship with the user who has a high degree of similarity with the user check-in location and the neighbor relationship that is close to the user in the geographical location.In order to examine the robustness and accuracy of the method.Experiments on two typical real-world datasets show that RLPN has the best prediction performance compared to the baselines.