Abstract: Collaborative filtering recommendation, as a relatively mature information filtering technology in the recommendation algorithm, is widely used in the field of commodity recommendation, but it faces problems such as data sparsity and cold start, which will reduce the recommendation quality of the algorithm . In view of the low prediction accuracy and low recommendation quality of traditional collaborative filtering recommendation algorithms, many scholars have proposed some improved algorithms to improve the accuracy of a single algorithm. Some scholars have improved the similarity measure to improve the effect of a single collaborative filtering recommendation algorithm, but they do not comprehensively use the recommendation information of multiple single recommendation algorithms to further improve the quality of recommendation, and hybrid recommendation algorithm is an effective strategy to solve this problem. Although the idea of mixed strategy alleviates the sparsity of data and improves the recommendation accuracy, the weight determination lacks theoretical basis and is too subjective. So this paper proposes a collaborative filtering hybrid recommendation algorithm (BEST-CF) based on the idea of optimal combination prediction.
Collaborative filtering recommendation mainly uses the similarity between users or the similarity between items to predict the user's rating of the item, its essence is the prediction problem. Different recommendation algorithms have different score prediction. In order to make effective use of different score prediction results, overcome the subjectivity existing in determining algorithm weight, and improve the quality of mixed recommendation more effectively, this paper applies the idea of optimal combination prediction to collaborative filtering mixed recommendation, so as to improve the accuracy of score prediction. The BEST-CF algorithm obtains the optimal weight by constructing the optimal combination prediction model, and uses the optimal combination of the user-based collaborative filtering recommendation algorithm (User-CF) and the item-based collaborative filtering recommendation algorithm (Item-CF) in the Movielens 100K data set.
The experimental results show that the MAE value of the BEST-CF algorithm is 7% lower than that of Item-CF, 11% lower than that of User-CF, and the RMSE value is 11% lower than that of Item-CF, and 15% lower than that of User-CF.Therefore, the BEST-CF algorithm significantly improves the scoring prediction accuracy and can improve the recommendation quality. Finally, BEST-CF is used for the recommendation of insurance products. The experimental results show that the recommendation accuracy of BEST-CF is significantly higher than that of Item-CF and User-CF. BEST-CF can effectively improve the accuracy of product recommendation results, be more accurate to recommend products to users who need more accurate recommendations for customer preference of insurance products, and alleviate the problem of whether insurance products are fit for customers who cannot be determined because of lack of insurance knowledge.
BEST-CF algorithm advantage is determined by its weight. The algorithm overcomes the subjectivity, and can make the combination forecast score to minimize the sum of squares fitting error.Its performance is superior to the single recommendation algorithm, and any of the multiple single recommendation algorithm weighted combination, but this does not take into account the user's interest in the dynamic change or fitting problems, which is the next issue we will study.

Key words: collaborative filtering recommendation, optimal combination prediction, algorithm

摘要： 针对传统协同过滤推荐算法预测精度不高、推荐质量低的问题,提出了一种基于最优组合预测思想的协同过滤混合推荐算法(BEST-CF),并利用基于用户的协同过滤推荐算法(User-CF)和基于项目的协同过滤推荐算法(Item-CF)的最优组合在Movielens 100K数据集上验证了BEST-CF的有效性,实验结果表明:BEST-CF算法明显提高了评分预测精度,能够提升推荐质量。最后,将BEST-CF用于保险产品的推荐,实验结果表明,BEST-CF的推荐准确度明显高于Item-CF和User-CF的,能为客户更加精准地推荐所偏好的保险产品。

关键词: 协同过滤推荐, 最优组合预测, 算法