一种使用全部评分提高推荐精度的方法

A Method for Improving Recommendation Accuracy via All Rating History

传统的基于用户的协同过滤推荐算法只能使用用户在共同评价过的项目上的评分记录进行推荐,由于推荐系统中数据稀疏和冷启动问题的存在,用户共同评价的项目较少,导致了用户的大量评分记录中只有少部分的数据得到了利用,限制了推荐系统预测用户偏好的精度。为了利用用户的全部评分提高推荐系统的精度,定义了用于描述和区分不同项目的内部子信息,提出了将用户对项目的评分分解为对内部子信息评分的方法,该方法能够使用用户的全部评分记录分析用户的相似度,同时设计了考虑用户间共同评价项目比例的动态调节权重用于将基于全部评分的用户相似度与传统的基于共同评分的用户相似度进行混合,并将混合相似度用于预测用户对项目评分。实验结果表明:使用用户的全部评分记录能够提高推荐系统预测精度,动态调节权重比静态的混合权重更能改善推荐效果。

In traditional user-based collaborative filtering algorithms, only users＇ ratings to shared items are utilized in recommendation. Due to the existence of data sparsity and cold-start problems, the amount of users＇ common ra- ted items is not sufficient. As a result, only a small quantity of data in users＇ massive rating records can be consid- ered, which limits the accuracy of recommender systems in predicting user＇s preferences. In order to use all the rat- ing records to improve recommendation, this paper introduces items＇ internal description information （IDI） for de- scribing and discriminating different items. Based on IDI, a method is p ternal description information from the ratings to items, so that users＇ similarity on users＇ all existed ratings can be calculated. For mixing users＇ similarity based on all the ratings with traditional similarity based on those ratings to users＇ shared items, we design a dynamic adjusting weight considering the proportion of users＇ common items in their all rated items. Then the mixed similarity is used to predict users＇ ratings to unobserved items. The experiment results show that, all the ratings can be used to improve the accuracy of recommender systems, and the proposed dynamic adjusting weight overpowers the static hybrid weight.