Item Collaborative Recommendation Algorithm Based On Implicit Associations And Timing Relation

With the explosive development of electronic commerce businesses in the last decade,the items(movies,music,news,etc.)information flood has almost inundated users so that users hardly get access to the content what they really want.Thus,the recommendation system came into being.The recommendation system predicts the user's preference for the product through the user's past behavior,characteristics,hobbies and social relationships,to guide the user to browse or purchase the products which they need or have interested in.The recommendation system not only saves time for users to search for information or products,greatly improving the user experience,but also brings considerable revenue increase for information or product providers.So more and more researchers spare no effort to make the recommendation system smart and humanized.Collaborative filtering algorithm was firstly proposed and the most widely used recommendation method so far.Collaboration refers to taking a large number of users' choice or rating of products into consideration to recommend the items.Compared with the traditional collaborative filtering recommendation algorithm,the current collaborative filtering recommendation algorithm generally assumes that the items or the users do not exist independently.The recommendation system fully considers structural and feature connection among the items and the users to improve the recommendation accuracy.However,the relation among users is usually not easy to collect and obtain in some websites,and the calculation methods of the relationship among items are relatively simple.Thus,both aspects result in unsatisfactory recommendation results.In this paper,we adopt association rule algorithm to mine the implicit association relation among multiple items.We fuse implicit Item Correlation into the recommendation framework of probabilistic matrix factorization(PMF)model.Item Correlation is calculated by the multiple items correlation formula proposed in our research.Our proposed algorithm obtains more accurate recommendation result than those which only calculate the similarities between any two items,and meanwhile the presented method avoids the consideration of extra dimensional information,which reduces the difficulty of data acquisition and makes our proposed algorithm has better scalability.In addition,we found that the phenomenon of user's interest drift is ubiquitous.As time goes on,the user's purchase tendency and hobbies will change imperceptibly.In order to solve the problem of user's interest drift,we extend the thought of reading leader to the timing influence among items.In the research field of news recommendation,reading leaders refer to the users who always browse some news earlier than the other users,called followers.There is an obvious timing relationship,where the reading followers tend to read or browse the news that the reading leaders have read before.Therefore,we propose a timing relation model of items purchase,to calculate the probabilities of other items being followed to purchase when the current item is purchased.So the whole users' purchasing preferences tendency over time can be obtained,which is efficient to improve the final recommendation.Based on the above recommendation algorithms which utilize association rules mining and timing relations,our research integrates implicit multiple items correlation and the item timing-related network into PMF,which is excavated by association rule mining and generated by timing relation model,respectively.And then,the timing impact factor is applied to combine the implicit Item Correlation and the directional item purchase influence.Our proposed method makes the choice of item neighbors more accurate and diverse.Relationship among items is fully utilized to improve the accuracy of recommendation and to alleviate the rating data sparseness,which causes similarities calculation of item neighbors inaccurate and rough.Experimental results on the public movie dataset Movie Lens show that the algorithm proposed in our research can effectively make full use of the items relationship to reduce the prediction error in the case that user relationship data cannot be obtained.Thus,the research in our paper not only presents an open framework and novel thought to the following researchers,but also is of great guiding significance to the item recommendation in the real e-commerce website.