用余弦相似度修正评分的协同过滤推荐算法

在用户对项目进行评分的时候,有时会出现不合理因素导致用户对项目做出不合理评分,使得推荐过程出现偏差。为修正这一偏差,采用评分矩阵的多种维度进行相似度比较以修正不合理评分,再用修正后的评分进行协同过滤推荐。而在采用变维度评分矩阵进行相似度对比时,主要利用同一用户对相似项目评分的相近性,对比2个用户对多个相似项目评分数组在不同维度下的余弦相似度。首先将多个评分构建成等维度的几个数组,对比2个用户的各个评分数组相似度,当某个相似度与其它相似度差别较大时,认为该相似度对应的2个用户的数组中至少有1个包含不合理评分;然后将2个数组按同样的方式均分为维度更低的数组,以此类推,最终确定不合理评分;最后以所有合理评分数组对应的相似度均值作为不合理评分数组对应的相似度,从而修正不合理评分。利用MovieLens和Bookcrossing数据库进行实验,结果表明:带修正评分的协同过滤算法相比未修正前的评分具有更高的推荐精度,其推荐评分MAE明显下降,本文算法相比对照算法获得了更优的MAE,Precision和Coverage。

A collaborative filtering recommendation algorithm with revised rating by cosine similarity

When the user scores items, sometimes there are unreasonable factors that cause the user to make an unreasonable score on items, which may cause bias in the recommendation process. In order to correct this deviation, multiple dimensions of the scoring matrix are used to compare the similarity to correct the unreasonable score, and then the revised score is used for collaborative filtering recommendation. When using the variable dimension scoring matrix for similarity comparison, the similarity of the same user's scoring similar items is used, and the cosine similarity of two users' similarity of multiple similar items in different dimensions is compared. Firstly, multiple scores are constructed into several arrays of equal dimensions, and the similarity of each score array of the two users is compared. When a similarity differs greatly from other similarities, it is considered that the similarity corresponds to at least one of the two user arrays containing an unreasonable score. Secondly, the two arrays are divided into smaller arrays in the same way, and the reset can be done in the same manner, finally determining the unreasonable score. Finally, the similarity mean of all reasonable score arrays is used as the similarity of the corresponding array of unreasonable scores, thereby correcting the unreasonable score. Experiments using the MovieLens and Bookcrossing datasets show that the collaborative filtering algorithm with revised scoring has higher recommendation accuracy than the unmodified scoring, and its recommendation score MAE is reduced significantly. Compared with the comparison algorithm, the proposed algorithm can obtain better recommendation performance on MAE, Precision and Coverage.