Research On The Concept Drift Oriented Recommender System

With the rapid development of the Internet,a personalized algorithm for recommending items to users based on their own characteristics has been widely used in various fields.Meanwhile,it has become a hot research direction in academia.Although the relevant researches and applications of the recommender system have obtained fruitful achievements,it still faces many problems such as sparsity of data and cold start problem.To deal with these challenges,research institutions and commercial companies throughout the world have put forward a plenty of solutions.In this paper,we mainly study the concept drift in the recommender system.The traditional concept drift problem usually can be defined as the change of implicit content will more or less fundamentally lead to the change of target concept.The problem of concept drift in the recommender system is that when the data in the recommender system accumulates over time,the system's recommender model cannot capture the changes of users and items and the recommended results deviate from the actual needs of users.In order to improve the performance of the recommender system,this article mainly aim to reduce the impact of the concept drift problem in the recommender system and carry out the study on the following subjects:(1)A vertical web crawler is proposed for expanding the content data of items.According to the comparison method,the climb rate of our method is much higher than the other results.(2)An item clustering method based on user-item relationship and item content features is proposed.Compared with the traditional item clustering algorithm,our method can better distinguish the unpopular items from the popular items in the implicit feedback data.Also our method can make the clustering results are as balanced as possible,and make the clustering category retain the long-tail distribution of items.(3)The unbalanced relationship between items is verified.This paper verifies the relationship between the items is not as balanced as the traditional method thinks in time series system through experiments.And this relationship has a greater correlation with the features of the item itself;(4)The concept drift of user behavior is analyzed combined with real data.In this paper,the user's concept drift can be expressed as the process that the user's preferencefor certain categories of items changes slowly over time while having a long-term continuous behavior on the same category of items.(5)A pre-state based recommendation algorithm is designed in this paper.This algorithm designs the transition probability based on the pre-state of user behavior.And based on this transition probability we complete the design of our recommender algorithm.The results of the experiment show that our algorithm can greatly improve the accuracy of the recommender system in the timing state.In this paper,the concept drift in the recommendation system is systematically studied and summarized.Based on the research and analysis results of the concept drift,we design a recommendation algorithm that can effectively improve the accuracy of the recommendation system under the timing state.