Research On Multi-Objective Recommender Model Based On User-Item Matrix

With the rapid development of the Internet and mobile communication technology,more and more businesses and individuals have opportunities to carry on various commercial activities on online platforms,which speeds up people’s access to information.The Internet produces numerous data every hour of the day,and today’s society has entered the era of information explosion.However,the speed of people’s access to information has far exceeded the limit of their processing speed due to the amount of information and the uneven quality of information.It is difficult for individuals to find valuable information from massive data.Therefore,as an important mean to alleviate the problem of information overload,the research of recommender system has attracted much attention from both academia and industry.Early recommendation algorithms and researchers focused on single-target recommendation,whose core was to improve the accuracy of recommendations.However,excessive pursuit of the improvement on the recommendation accuracy makes the recommended items bear some commonalities,and the recommendation results appear too deliberate.20% of the products receive 80% of the recommended exposures,and the recommendation results are excessively concentrated on a small number of products with high popularity.These recommendations are accurate,but may not live up to users’ expectation because his/her horizon is limited to a few popular items and it is hard to obtain serendipity in the recommended items.Many researchers have claimed that accuracy related metrics are not enough for measuring the quality of a recommender system,and other evaluation criteria,such as diversity and calibration,should be identified as an important ingredient to improve recommendation quality.Because of this,the thesis carries out a series of exploratory researches on recommendation accuracy,diversity and calibration.The main contents and contributions of this thesis can be summarized as follows:(1)Based on the previous work and the definition of user-coverage,a user-coverage model based on rating differences is proposed.During generating user’s interest domain(usercoverage),on the one hand,the model combines rating differences between users across an item with use-coverage model effectively,thus obtaining a more precise interest domain of the user.On the other hand,objective function is constructed in the form of vector by mapping users’ and itemsets’ interest domain to m-dimensional vectors(called user vector and itemset vector respectively),which can reduce the number of iterations in the calculation process.In addition,a new items selection strategy is provided by similarity relationship between user vectors and itemset vectors.The experimental results of the proposed model are superior to the user-coverage model in all aspects.(2)Compared with the user-coverage model,the item-coverage model is proposed.From the perspective of the model,item-coverage model transforms user’s interests and itemset’s interests to two n-dimensional row vectors and provides the trade-off between recommendation relevance and diversity.The model is parameter-free and suitable for either implicit data or explicit data,which has a good performance in adaptability.However,identifying the optimal recommendation list from a large pool of candidate items is a NP-hard problem,which is difficult to work out in limited time and space resources.Therefore,on the one hand,the greedy algorithm can be applied to handle these intractable problems with solid theoretical supports.On the other hand,by limiting the original objective function to a special case where top-K items can be directly picked out instead of using greedy algorithm,a more efficient approach has been proposed.(3)For the purpose of improving the recommendation accuracy and calibration,the triple calibration distance design for neighbor-based recommender system is proposed.Most of the existing calibrated-oriented recommendations take an extra post-processing step to rerank the initial outputs.However,applying this post-processing strategy may decrease the recommendation relevance.In remedy of this weakness,this thesis is dedicated to modifying the criterion of neighbor selection.The thesis provides the first-order,second-order,and the third-order calibration distance based on the motivation that if a user has a similar genre distribution or genre rating schema toward the target user,then his/her suggestions will be more useful for rating prediction.Besides,an equivalent transformation for the original method to speed up the algorithm with solid theoretical proof also be proposed.