Research On POI Recommendation Based On Self-paced Learning

With the rapid development of the fourth generation mobile network and the improvement of the built-in positioning system,some social media platforms can easily obtain the user's location information.And users can also share their location tour records on social platforms according to their preferences.Therefore,recommending a point of interest(POI)to users has become a popular direction in the field of recommendation.More and more researchers have proposed advanced POI recommendation algorithms that have practical application significance.However,some classic recommendation algorithms often face non-convex optimization problems,such as matrix factorization and neural networks.Especially under the influence of noise data or outliers,it will lead to poor performance of the recommendation algorithm.In order to solve this problem,this article attempts to introduce the concept of self-paced learning into the recommendation algorithm,researches and observes its feasibility under the real data set.And self-paced learning is a new model learning,which adopts the idea of learning new things in our daily life,turning the process of model training samples into a process from simple to complex,thereby improving the effect of the model.This article mainly studies the characteristics of self-paced learning,and then combines some representative algorithms in the field of POI recommendation to study the application of self-paced learning in recommendation algorithms in order to alleviate the impact of noise data on the model and improve the recommendation effect.The main content and innovations of this article are as follows:1)In order to analyze the characteristics of self-paced learning,we combined the traditional matrix factorization algorithm and confirmed that under the influence of noise data,self-paced learning can indeed help the matrix factorization to achieve a good effect.At the same time,under the real data set MovieLens,combined with generalized matrix factorization,we analyze the four disadvantages of traditional self-paced learning soft weights.And from the perspective of noise,the causes and solutions of these problems are analyzed.Finally,a new type of self-paced learning was proposed,which we named Bounded Self-Paced Learning.2)The research focuses on two classic POI recommendation algorithms,the USG model and the MGMPFM model,and analyzes several main factors affecting user sign-in in a location-based social network.At the same time,under the real data set Gowalla and Foursquare,the MGMPFM model integrated with self-paced learning has indeed performed well.3)In order to further promote self-paced learning and verify the generality of selfpaced learning,we first focus on the classic recommendation algorithm neural matrix factorization(NeuMF)under deep learning,and then combine the traditional self-paced learning soft weights and bounded self-paced learning is integrated into GMF,MLP,and NeuMF,respectively.Experiments on the real data sets Brightkite and Brightkite with noise have confirmed that self-paced learning with boundaries can indeed improve the recommendation effect,especially under the influence of noise.Compared with traditional models and traditional self-paced learning,the improvement of self-paced learning with boundaries is more obvious.