Research And Implementation Of Ranking Based Personalized Recommender Algorithms

With the development of Internet science and technology, Internet has been a vital way to access different kinds of information and digital resources. While the increasing amount of information gives people more choices, it also makes people spend much more time and energy on choosing their needed information and resources. To address this kind of information overload problem and help people find their interested information efficiently is becoming a heated research topic recently. Personalized recommendation system, providing a helpful way to address information overload problem, can generates recommendation from a large collection in favor of user preferences. In recent years, personalized recommender systems have become a de facto standard and must-own tool for e-commerce to promote business and help customers find products. Prominent examples include Amazon, Netflix, and Douban.Collaborative filtering approaches generally fall into two categories:rating-based and ranking-based. The former bases on users’historical rating score to compute similarity between user pairs and then make prediction according to the target user’s neighbors’rating score. However, similar users considered by rating based similarity measure may have great differences in their preferences, which means that the rating based similarity measures lose the capability to capture users preference. And existing rating based collaborative filtering algorithms only predict users’rating score individually without considering users’preference on item pairs. In contrast, ranking based collaborative filtering algorithms can compute users’preference similarity according to users rank list. However, existing raking based collaborative filtering algorithms only consider whether users have the same preference, without considering users preference degree and preference popularity. To address above problems, our proposed ranking based collaborative filtering algorithms can combine degree and popularity of users preference and generate final recommendation list according to users’preferences. As collaborative filtering and content based algorithms have their own limitations, recommender systems for business usually combine different kinds of recommender systems and proposed hybrid recommender system to take advantage of all kinds of recommender systems’strengthens. We view this hybrid recommendation problem as machine learning problem, which can take user and items content information and historical rating information into consideration. Unlike Information Retrieval, the relevance between queries and documents can be measured by the similarity between queries and documents, users and items are not content comparable in recommender systems, which means that we cannot compute the relevance between users and items. To address above problems, we extract meta features and rating-based features of users and items, and then represent user and items in a same dimension so that they can be content comparable. On our proposed feature framework, any learning to rank algorithms can be adopted to recommendation systems.In this paper, we1) analyze the limitations of existing ranking based collaborative filtering algorithms and then define users’preference with degree and popularity, based on which, we use two kinds of similarity measures to compute the similarity between user pairs, at last use two preference aggregation algorithms to do rank prediction and generate the final recommendation list.2) In hybrid recommendation systems, represent users and items in a same dictionary and view recommendation problem as a learning to rank problem. Based on this feature framework, any learning to rank algorithms can be adopted to do recommendation. To evaluate the effectiveness of our proposed algorithms, we conducted a extensive experiments on EachMovie and MovieLens in comparison with the state-of-the-art approaches to demonstrate the promise of our approach. The experimental results show that our methods are effective in improving the recommendation accuracy.