Study On Recommendation Algorithm Based On Temporal Behavior

Nowadays,users cannot find their needed information quickly and accurately with the abundant Internet resources.In order to raise user's research efficiency and meet user's individual requirements,personalized recommendation system comes into being.The system recommends suitable items to user through mining user's interests.User's interests calculation is an important basis for recommendation algorithm.The algorithm can make the efficient recommendation if user's interests can be grasped accurately.However,there are still a lot of challenges for user's interests calculation.Firstly,user's interests are of great diversity and it is hard to express.Secondly,user's interests change over time which is difficult to capture.In order to solve these problems,on one hand,the researchers express user's interests with item genres,latent topics,and latent semantemes.On the other hand,the researchers deal the dynamic change of interests with modeling the temporal information,and introducing the time-window and time delay function.Based on the problems and related research,this paper has done the following research.To solve the problem that user's interests is difficult to express,this paper proposes that we use the genre combination to express the user's interests.Firstly,this paper establishes the genre combination space according to the item genre combination.Secondly,this paper defines the distance between genre combinations according to the spatial structure and the relationship between elements.The user interests represented by the genre combination can not only express the diversity of user interests,but also can be used to express the correlation degree of different interests by the distance of the genre combination.Aiming at the problem that the user's interestsare difficult to capture,this paper abstracts the temporal behavior in a graph model,and uses random walk algorithm to calculate the user's preference for the genre combination.Firstly,the user interest segment is defined based on the genre combination distance.Therefore,the user's temporal information is divided.Secondly,this paper calculates the genre combination transition matrix with consideration of ingterests changing.Finally,this paper uses personal random walk model to calculate the genre combination preference.The proposed method takes into account both the overall interests preference and the period interest change.It can capture the user's interests better.Thirdly,for each user,he may not browse an item repeatedly.But,he can browse a genre combination repeatedly.It mans that the user come back to some interests again.In view of the fact that users may come back to some genre combinations for several times,this paper introduces the commute time kernel function to measure the transition difficulty of genrecombinations,which can improve the genre combinations transition matrix.Finally,we use the matrix decomposition model to predict the rating,and generate a list of recommendations.In this paper,the MovieLens data sets are selected to carry out the experiment.The results show that the proposed algorithm has a good improvement in the accuracy of recommendation.