数据稀疏环境下基于用户主题偏好的协同过滤算法

在B2C电子商务中,user-based协同过滤算法是一种重要的推荐方法,但用户共同评价项目数据稀疏影响了user-based协同过滤算法的应用。鉴于此,在考虑用户消费水平的基础上,利用关联规则挖掘形式化描述商品间的替代相似性;利用基于时间的贝叶斯概率描述商品间的关联关系构建商品网络,通过社会网络分析中的成份分析方法对商品网分析,得到面向用户主题偏好的商品间互补性关系,进而利用这两种商品间关系构建用户主题偏好项目集,最后在数据极度稀疏的环境下通过F₁方法和多样性测量方法与传统推荐算法进行对比实验分析,实验结果显示提高了推荐结果的准确性与新颖性。研究用的所有数据均采集于京东商城网站。本文为缓解数据稀疏问题提出了一种新的方法,扩展了整体网分析方法在商品关系分析中的应用,含有理论与实践双重意义。

Topic Preference Based Method for Collaborative Filtering Algorithm in Sparse Datasets

User-based collaborative filtering algorithm is an important method for B2C electronic commerce to recommend commodity, but it has been limited to some extent because of the sparsity of common rations between users. In order to resolve these problems,the paper first adopts association rule mining to formalize similarity among competitive goods based on considering consumption level, and then constructs time-based Bayesian goods relation network, based on the network the paper takes advantage of components analysis of whole network to find complementary similarity of the goods and the topic preference of consumer for expanding common rating sets. At last, through comparative experiments based on F₁ method and diversity method, the result shows that the accuracy and diversity have been improved significantly in sparsity environment. The data are collected from the site of JingDong Mall. In conclusion, the model provides a new way of dealing with sparsity problem, enriches examples using whole network approach in goods relationship analysis, and has the significance in theory and practice.