Study On Diversity-oriented Recommendation Algorithms

Nowadays,artificial intelligence technology has been making tremendous contribution to the development of science and technology.The recommender system,as an important branch of the artificial intelligence and machine learning research fields,can effectively help users to find potential valuable information from the massive data,thus has received considerable attention from academic community and industry enterprises.However,the traditional recommendation algorithms,putting more emphasis on the accuracy of recommendation results,usually push a small amount of popular items to the majority of users,thus bring about poor personalization experience.In the past decade,recommender system designers have put more effort to the diversity of recommendation results.High diversity means that more recommendation opportunities will be given to a large number of non-popular items.According to the literature,simply increasing the diversity of recommendation results will result in a remarkable accuracy decrease.Therefore,how to improve the diversity of recommendation results while ensuring high accuracy is still an important issue in the research.In order to alleviate the trade-off of the diversity and accuracy,we conduct a series of research by distinguishing the user group by their degrees(how many items a user has rated).Empirical analysis on real datasets shows that,the large-degree users like to accept the recommendation results with high diversity,while the small-degree users enjoy accurate recommendation results.Based on this observation,we have completed the following major work and contributions.(1)We proposed some improvement strategies on typical recommendation algorithms.For original user-based collaborative filtering algorithm,we combine the original static parameter with the dynamic user degree together to decrease some neighbor similarity weights.Experimental results show that the diversity of recommendation results for all users has been significantly improved.What's more,for low-degree users,their accuracy has been even improved slightly;For original item-based collaborative filtering algorithm,we adaptively deemphasize each user's historical interactive items based on their own item degree(item population).The experimental results show that the diversity of recommendation results for all users has been slightly improved;For original mass diffusion algorithm,we introduce a random perturbation in the propagation probability of the first and third steps of energy propagation process,which brings an energy propagation bias to non-popular item nodes.The experimental results show that the diversity of recommendation results for all users,especially high-degree users,has been obviously improved.(2)We proposed two re-ranking methods for the existing recommendation results.The first method re-ranks the recommendation list by distinguishing the item degrees.The second method takes into account the reversed recommendation idea of recommending users to the target item.The experimental results on real datasets show that two re-ranking methods can improve the diversity of recommendation results at different levels of confidence.The first one can improve the diversity a lot,while the second one can provide a more moderate recommendation result in regard to both diversity and accuracy.(3)We proposed two new diversity evaluation metrics based on existing evaluation measures.The first new metric is an improved version of the original hamming distance metric,and the second one is a linear hybrid of the ranking precision metric and the hamming distance metric.We pick some state of the art recommendation algorithms adapted for high diversity,and compare the experimental results of new metrics with the original ones in such algorithms.Finally,the results show that the first new metric can give a more strict evaluation of the algorithmic ability of recommending diverse,and the hybrid metric can better evaluate algorithmic comprehensive recommending ability.