Similarity-driven sampling for data mining

Industrial databases often contain millions of tuples but most data mining algorithms suffer from limited applicability to only small sets of examples. In this paper, we propose to utilize data reduction before data mining to overcome this deficit. We specifically present a novel similarity-driven sampling approach which applies two preparation steps, sorting and stratification, and reuses an improved variant of leader clustering. We experimentally evaluate similarity-driven sampling in comparison to statistical sampling techniques in different classification domains using C4.5 and instance-based learning as data mining algorithms. Experimental results show that similarity-driven sampling often outperforms statistical sampling techniques in terms of error rates using smaller samples.