Abstract:

Traditional recommendation algorithms mainly focus on the recommendation accuracy, while user different preferences for items and diversity needs also affect user experience and satisfaction. Aiming at the problem, a new algorithm is proposed. Firstly, to improve the accuracy of the similarity between items, the algorithm combines the rating differences between different items by users when computing the similarity. Secondly, according to user historical rating data and item category data, a user-category weight matrix is concluded. On the one hand, a definition of diversity based on entropy depends on the matrix. In addition, according to the calculation formula of the user’s interest in the item, an initial recommendation sequence in descending order is generated,a recommendation based on user diversity preference can be implemented by combining a setting of error threshold for user preference and the user-category weight matrix, N recommended items are finally generated, which aims at improving user satisfaction on the premise of ensuring accuracy and diversity. Experiments on the movielens datasets with different versions show that the proposed algorithm can effectively improve the recommendation accuracy and user satisfaction compared with several classic algorithms.

Key words: diversity, user preference, user satisfaction, similarity

摘要：

传统推荐算法主要关注推荐准确性，而用户对项目的不同偏好和多样性需求也影响着用户体验和满意度。针对该问题，提出了一种新的算法，在计算项目相似度时结合了用户对不同项目的评分差异，以此可以提高项目相似度计算的准确性，根据用户历史评分数据和项目类别数据得到用户-类别权重矩阵，一方面以此计算基于熵的多样性，另外根据用户对项目的兴趣计算公式，生成一个降序排列的初始推荐序列，根据用户偏好误差门限，并结合用户-类别权重矩阵实现基于用户偏好的推荐，最终生成[N]个推荐的项目，同时保证准确率和多样性的前提下，提高用户满意度。在数据集movielens的多个版本上，与多个经典算法比较，实验结果表明，提出的算法可以有效提高推荐精度和用户满意度。

关键词: 多样性, 用户偏好, 用户满意度, 相似度