Research And Implementation Of Recommender System On Heterogeneous Implicit Feedbacks

Recommender system as another method of information filtering,take the initiative to push information for the user to provide personalized recommendations in the field of e-commerce,the media and other fields which gained more and more attentions.The recommender system helps users finding items which they prefered quickly or help company to enhance the user experience.At the same time,the service provider are easier to locate the target clients and achieve the effect of precision marketing.In a word,recommender system achieves a win-win situation for both client and business company.At present,most recommendation algorithms were designed for explicit feedbacks.While in real world,the explicit feedbacks require the user to replicate actively.Of course.the numbers and types of explicit feedback are alse less than implicit feedback.Implicit feedback data captures user behaviors which are more common and more representative.On the other hand,implicit feedback data represents user behavior record,which is more common and more representative.Therefore,the research of recommendation algorithm based on implicit feedback has drawn more and more attention.Secondly,the implicit feedback data has a great variety and influence each other.How to integrate multidimensional implicit feedback reasonably and efficiently is an urgent problem to be solved in the field of recommended systems.Therefore,the recommendation algorithm based on multi-dimensional implicit feedback data has important research significance.In this paper,we mainly research about the recommendation algorithms under multidimensional implicit feedbacs.I also proposed a improved itembased similarity model named HCoM to genernazed a heterogeneous constraint for certain implicit feedback and uncertain implicit feedback.At the same time,I also researched about time drift and the novelty.The contents of this research are displayed as follows:Firstly,we discussed about the differences between explicit feedbacks and implicit feedbacks.Aiming at the multidimensional implicit feedbacks,we proposed an algorithm named HCoM.This model considers about the similarity of items learns it with data-driving method which is learned directly from the data.Secondly,we analyze how the different aspects of structured modules affect the recommender results.We compared the loss function based on scoring and ranking respectly.On the other hand,the influence of the regularization constraint on the recommended quality are also beening discussed.What's more,we propose a regularization named HC to constrain the relationship between certain implicit feedbacks and uncertain implicit feedbacks.Thirdly,we concern about the time drift which is generated by the user's preference,and using the time decay function to adjust the implicit feedback data.At the same time,we study the problem of novelty of recommended results,and adjust the original model according to the frequency of user behavior.Fourthly,we generate the experiment of our algorithm and implement the simple recommendation system based on e-commerce.The evaluation methods are HR and ARHR.From the experiment result,we provecd that the algorithm proposed can achieve better recommender results than other start-of-art methods on the two real world data set.In a whole,this paper introduced the research background firstly.Then,the related literatures of the system are reviewed,and the literature of the recommendation system based on the implicit feedback data is discussed.This paper also analyzes the problems and the shortcomings of the of reality existing literature.To solve those problems,we proposed an algorithm named HCoM which consider about both certain implicit feedbacks and uncertain implicit feedbacks.At the same time,the time and numerical characteristics of the data are used for the time drift and the novelty problems.Finally,this paper introduces the implementation of the recommendation system based on multi-dimensional implicit feedback data.The algorithm is experimented on the data of Tmall and the IJCAI contest.Compared with other commonly used algorithms,our algorithm can achieve better recommender quality.