Study On Coverage-Oriented Recommendation Algorithms

In the information era of Web2.0,the development of network technology makes the information production popular,the amount of network information is growing explosively,and users need to pay more and more price to find valuable information from massive information.In the face of such information overload,recommendation system,as an information filtering tool,has been proposed by researchers and gradually evolved into a discipline,which has attracted the attention and research of academia and industry.Nowadays,researchers are more focused on the accuracy metric of recommendation,but only to improve the accuracy of the algorithm,will not cause the user to improve the satisfaction of the recommendation results,but at the same time,it will cause the homogenization of users' interests,popular items appear in a large number of users' recommendation list,a large number of long tail items can not be mined by the recommendation algorithm.Many evaluation metrics have appeared in the research of recommendation algorithm to measure the quality of a recommendation algorithm.Among them,the coverage metric describes the ability of the recommendation system to mine long tail items,which is of great significance to businesses interests.The improvement of the metric can avoid the limitation of the recommended range,and meanwhile,users can get more more chance to see long-tailed items.Therefore,the research on the recommendation algorithm of coverage metric is of great significance to both academia and industry.In order to mine the long tail items in the system,this paper studies a series of traditional recommendation algorithms,including adding punishment items of neighbor users,inhibiting the spread of popular items and other improved ways to improve the exposure of long tail items.The re-ranking method is also used to calibrate the proportion of users' interests,so as to avoid the narrow interests of users.This paper mainly Carries the following work and innovation.(1)A direct improvement method for the traditional recommendation algorithm is proposed.First of all,the collaborative filtering algorithm based on users is improved.When calculating the recommended score of items,the contribution value of neighbor users is punished to reduce the recommended score of popular items.The offline experimental results metrics show that after the algorithm is improved,long tail items get more exposure.Secondly,the collaborative filtering algorithm based on items is improved,and considering the influence of item popularity,the similarity of items is adjusted adaptively when calculating the item recommendation score.The experimental results show that the improved algorithm can greatly improve the recommendation coverage metric.In the end,the material diffusion algorithm is improved to increase the disturbance in the process of energy diffusion,so that more long-tailed objects can obtain energy.The off-line experimental results show that under certain loss,the recommended coverage metric of the algorithm can be significantly improved.(2)A new re-ranking method for recommendation results is proposed.In order to suit the different needs of different users,researchers proposed a way to re-rank the recommendation results to provide users with a new recommendation list.However,different re-ranking strategies have different effects on user satisfaction.In this paper,we propose a re-ranking strategy aiming at calibration,and select the items that meet the historical interest ratio of users from the candidate items as a new recommendation list.The offline experiments on real datasets show that this new re-ranking strategy can greatly improve the recommendation coverage and maintain the recommendation accuracy.Experiments on two real datasets show that the improved algorithm proposed in this paper can effectively improve the ability of mining long tail objects.A new reordering method based on the calibration idea significantly improves the recommendation coverage,satisfies the proportion of user interest,and avoids narrowing user interests,and meanwhile,expand user interests.