Study On Personalized Recommendation Algorithms Via Recommending Users To Items

The rapid development of Internet technology brings people into the era of information explosion, how to get the valuable information becomes more and more difficult. A typical information filtering technology is recommender systems, by which users can find the relevant information. The traditional recommender systems are based on user or item similarity. Such approaches are subject to a high risk of exposing user to a narrowing band of popular items. Users can not get more information from the more accurate recommendations. Diversity and novelty become important criteria of algorithm performance. A easy way to increase the recommendation diversity is sacrificing accuracy. How to balance the recommendation accuracy and diversity is still a challenge in the field of recommender systems.Motivated by improving the recommendation diversity and novelty, we aim to procure a fair opportunity for most items to be recommended. We may consider selecting which users each item should be recommended to. This view entails a symmetric swap of the recommendation task, whereby users are recommended to items rather than the opposite. The personalized recommendation algorithms based on reverse recommendation will be studied in this paper. The main work of this paper showed as follows:(1)We propose the item-based inverted collaborative filtering method. In our method, all user's opinions count to the same extent overall in the produced recommendations, thereby indirectly enhancing a more even distribution of items. However, popular items still account for a large proportion of the recommended list. Furthermore, we introduce a free parameter to control and reduce the effect of item degree on the recommended results. Our method can provide more personal recommendation.(2)We propose the user-based inverted collaborative filtering method. Our method directly consider selecting which users each item should be recommended to, therefore, the long tail items will get more chances of being recommended. The similarity between items is based on the number of common users, apparently popular items are more similar to all the items in the user's preferences. In order to eliminate the impact of items on the recommended results, our method quantifies the extent to which the target user is fond of the item relative to the other user. Long tail items, by getting a more equal opportunity to be recommended with respect to popular items, make for novelty and diversity enhancement of recommendations.(3)We propose a recommendation algorithm based on the inverted diffusion process on a user-item bipartite network. Our algorithm follows the law of conservation of energy, the degree to which the user's affection for a particular item is proportional to the degree of the item, so the effect of item degree on the recommended results is quantified. The effect can be controlled and reduced through smoothing technique. Our method can provide more accurate, diversity and novelty recommendations.Numerical analyses on two benchmark data sets, MovieLens and Netflix, indicate that our recommendation algorithms in the inverted recommendation task not only provide more accurate recommendations, but also generate more diverse and novel recommendations. This paper provides a new perspective to balance the accuracy and diversity of recommendations.