考虑负面评价的个性化推荐算法研究

Personalized Recommendation Algorithm by Considering the Negative Ratings

利用用户的负面评价信息,本文提出了一种新的推荐算法结构。算法首先将用户选择过的产品分为喜欢和不喜欢两类。其次利用用户的喜好信息构建推荐列表,同时利用负面评价信息构建厌恶列表。最后将推荐列表中的厌恶产品进行过滤,精炼推荐列表。Movielens数据上的实验结果显示,当采用90%数据作为训练集时,推荐列表的排序打分可以达到已知算法的最大值0.077,推荐列表的长度为10时,推荐列表的多样性和推荐新信息的能力相对不考虑负面信息的算法分别提高了16.08%和28.83%。同时,算法可以识别出根据喜好信息构建的推荐列表中19.15%的产品是用户不喜欢的。新算法结构不仅是目前已知的准确度和多样性都最高的算法,而且可以极大地降低系统的计算复杂度,节约存储空间。该工作开辟了利用用户负面评价提高推荐效果的新思路。

By considering users＇ negative ratings,this paper introduces a new recommendation algorithm architec- ture. According to the rating scores, the algorithm divides all the objects rated by users into two categories, liked one and disliked one. Both two sets are used to obtain users＇ recommended lists and disliked list by the mass- diffusion-based algorithm. Then the lists disliked are used to filter out the disliked objects in the recommended list. The numerical results on one benchmark dataset show when the training set is set as 90% percent of data, the average ranking score of the recommended list could be improved to O. 077. Compared with the standard mass-diffusion-based algorithm, when the length of recommended list is 10, the diversity and predicting new in- formation capability could be respectively improved by 16.08% and 28.83%. In addition, the algorithm could i- dentify 19.15% disliked objects in the recommended list. The numerical results indicate that the negative ratings are crucial for improving the effectiveness of the recommendation algorithm.