Ted Mihalisin, John Timlin
TempleMVV a system for visually mining very large high dimensional datasets is presented. The system first developed at Temple University’s Department of Physics is based on U.S. Patent No. 5,228,119 and utilizes nested dimensions and hierarchical graphics. The system achieves very high performance which is independent of the size of the dataset by utilizing discrete recursive computing to the maximum degree possible. Data involving any mix of continuous or discrete numeric variables and nominal or ordinal categorical string variables can be mined. In this paper we will try to convey some of the types of knowledge that can be mined utilizing human pattern recognition skills when suitable graphic data representations are chosen. These include but are not limited to 2, 3, ..., 10 way interactions; complex correlations which may be linear or non-linear, marginal or highly constrained, over an entire range or any sub-range of one or more variables; anomalously large and statistically significant frequencies for multi-dimensionally non-contiguous cells (nuggets); clustering and discriminate function analysis in up to ten dimensions. Recent enhancements to the system allow one to deal with datasets involving thousands of variables (e.g. marketbasket data). The system is superior to neural nets, CART, CHAID and clustering algorithms in several respects.