| Static point of interest recommendation has always been the focus of research and commercial application.Nowadays,the research based on dynamic point of interest is also gradually rising,and has become the focus of research and practical needs.For example,the automatic selling truck that can serve many users is a typical dynamic point of interest recommendation problem.It can be seen that there are essential differences between mobility points of interest recommendation and traditional static interest point recommendation,such as merchant recommendation.In order to solve the problem of mobile interest points recommendation,this paper proposes a recommendation algorithm for mobile interest points.It mainly includes the following contents:(1)We propose a recommended method for implementing mobile interest points with composite neural network structures including 3 network modules.The overall algorithm framework includes the gated recurrent neural network GRU expert module and the portal decision system,and the recurrent neural network expert module consists of two sub-modules,mode GRU and point GRU.Mode GRU module integrates transfer learning strategy to learn dynamic spatial and temporal mode in other data sets,responsible for mode prediction;the point GRU module integrates contrast learning strategy to expand the number of samples of the target training set and improve the network generalization ability;the portal decision module is responsible for selecting the output of the corresponding GRU module according to the form of input samples to realize the neural network expert decision system.(2)The process of training two GRU expert networks adopts the method of training and freezing parameters respectively.The training data set is expanded by using the method of comparative learning,so that the data samples can better drive the above point GRU experts.At the same time,the method of transfer learning is also used to drive the training process of GRU experts in the above mode after eliminating the geographic coordinate information and point information from the external data set.By using comparative learning and transfer learning,the training samples of the two GRU expert networks are fully expanded.(3)A two-step filtering method for extended samples is also proposed.In the first step,the MMD method is used for the first step filtering,and in the second step,the combination method of Tr Ada Boost and EM is used for scoring and screening.This two-step sample filtering method,combined with comparative learning and transfer learning strategy,makes effective data sample selection on the effect of generating large training data set based on small samples,which can well drive the network training process.Finally,this paper carries out experiments and result analysis on the social data set Gowalla,and uses other external data sets for migration learning,including Facebook V:predicting check ins data set and yelp data set check ins data set.These two data sets are used to provide spatio-temporal trajectory information of different users,including latitude and time and longitude,.Finally,the experimental results of the method proposed in this paper show that the prediction and recommendation effect of this method is better than that of the traditional method in the short and medium term,and the comparative test is carried out with the Markov method.In the comparative experiment of short-term prediction,the statistical results show that the method proposed in this paper has a low median error;In the comparison experiment of medium-term prediction,the method proposed in this paper has less error. |
