基于多重特征的双层Web用户聚类方法

摘要

Web log clustering is a general task of exploratory data mining to find user group patterns on the Website and even predict user access behavior.It supplies personalized services for different user groups further.This paper introduced a twolayer Web user clustering method based on multiple feature selection to solve common problems that current algorithms had,including little user interests bias revealed by single feature selection,low convergence efficiency and the diversity of user access impact on traditional hierarchical clusters.This method took use of multiple characteristics for user similarity measurement,and conducted the two-layer clustering on the basis of this metric.First,the method eliminated outliers and detected irregular clusters by DBSCAN clustering.The second layer used the previous clusters to produce the final clusters by a bottomup hierarchical algorithm.Experiments confirm that this method is stable and effective.%通过对Web日志的聚类分析,可以发现用户的群体特征,甚至可以预测用户将来的访问模式,进而为不同的用户群提供个性化服务.针对现有方法的一般缺陷,包括特征选择单一无法充分体现用户兴趣偏好和传统Hierarchical算法在用户聚类时存在的收敛效率低、易受用户访问多样性影响的问题,提出了基于多重特征的双层用户聚类方法.该方法采用多重特征对用户相似性进行度量,并在此基础上进行双层聚类.采用基于密度的DBSCAN算法来排除用户会话中的离群对象和发现不规则簇,然后再采用自底向上的Hierarchical方法对第一层的聚类结果进行聚类.实验结果表明,该方法具有良好的稳定性和聚类效果.