A Multiple-Phase Stratification-Based Hierarchical Clustering Over a Deep Web Data Source

摘要

Compared with surface web, deep web stores more high-quality data, and data mining over deep web is more valuable. Nevertheless, in deep web, the entire data sets are stored in back-end databases and cannot be accessed directly, and data can only be retrieved over the Internet through query forms. The only particular method for mining a deep web data source is to sample the data set, which caused several unique challenges. In this paper, according to active learning, instead of traditional one-time sample allocation, we use multiple phases of sample allocation, which improves the representativeness of our gained samples. At the step of stratified sampling in each phase, we sample parts of representative samples for initial clustering. Using gained clusters, we can explore boundary points in them. A boundary point owns much uncertainty than others; for example, it contains more information. Sampling on a boundary point is useful to gain more representative samples. According to our experiments, our method performs better than random sampling and two-phase sampling in Liu and Agrawal (Int Conf Data Mining 70-81, 2012) at the same sampling costs.