Proposal of Recommender System Removed Popularity Bias by Using Information Gain
Proposal of Recommender System Removed Popularity Bias by Using Information Gain
复制标题
推荐系统的提议通过使用信息增益消除了流行度偏差
DOI：
10.1527/tjsai.30_647
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
T. Takagi
中科院分区：
文献类型：
--
作者：
Satoshi Yoshida;T. Takagi
最近，推荐系统作为在网络上收集大量信息并向用户推荐信息的系统引起了人们的关注。推荐系统帮助用户找到他们想要的产品。推荐系统和长尾之间存在密切的关系，因为它们的性能不仅由准确性指标评估，还由长尾指标评估。协同过滤（CF）是一种典型的推荐系统。它被描述为用来支持长尾的技术。然而，CF倾向于推荐热门产品。在本文中，我们提出了一个系统，如果商品与用户的偏好相似，则推荐利基产品。我们将通过使用对关键字的兴趣来减少top-N推荐中的偏差。从信息增益中计算兴趣值，用于决策树学习中选择属性和机器学习中选择特征。实验结果表明，该系统在推荐小众产品方面优于基于项目的CF。在大多数关注长尾的现有研究中，利基产品的推荐是以准确性为代价的。然而，在我们的研究中，不仅推荐了小众产品，而且准确性也得到了提高。
Recently, recommender systems have attracted attention as systems that collect the enormous amount of information on the Web and suggests information to users. Recommender systems help users find the products that they want. There is a close relationship between a recommender system and the long tail because the performance of them is evaluated by not only accuracy metrics but also long tail metrics. Collaborative filtering (CF) is a typical recommender system. It is described as technology used to support the long tail. However, CF is prone to be biased towards recommending hit products. In this paper, we propose a system that recommends niche products if an item is similar to the user’s preference. We will reduce the bias in top-N recommendation by using the interest in a keyword. The interest is computed from information gain, which is used to choose attributes in decision tree learning and to select features in machine learning. The results from the experiments show that the proposed system outperformed item-based CF in recommending niche products. In most existing studies focused on the long tail, niche products are recommended at the cost of accuracy. However, in our study, not only are niche products recommended but accuracy is also improved.