Visual Data Mining and Discovery with Binarized Vectors
Department or Administrative Unit
The emerging field of Visual Analytics combines several fields where Data Mining and Visualization play leading roles. The fundamental departure of visual analytics from other approaches is in extensive use of visual analytical tools to discover patterns not only to visualize pattern that have been discovered by traditional data mining methods. High complexity data mining tasks often require employing a multi-level top-down approach, where first at the top levels a qualitative analysis of the complex situation is conducted and top-level patterns are discovered. This paper presents the concept of Monotone Boolean Function Visual Analytics (MBFVA) for such top level pattern discovery. This approach employs binarization and monotonization of quantitative attributes to get a top level data representation. The top level discoveries form a foundation for next more detailed data mining levels where patterns are refined. The approach is illustrated with application to the medical, law enforcement and security domains. The medical application is concerned with discovering breast cancer diagnostic rules (i) interactively with a radiologist, (ii) analytically with data mining algorithms, and (iii) visually. The coordinated visualization of these rules opens an opportunity to coordinate the multi-source rules, and to come up with rules that are meaningful for the expert in the field, and are confirmed with the database. Often experts and data mining algorithms operate at the very different and incomparable levels of detail and produce incomparable patterns. The proposed MBFVA approach allows solving this problem. This paper shows how to represent and visualize binary multivariate data in 2-D and 3-D. This representation preserves the structural relations that exist in multivariate data. It creates a new opportunity to guide the visual discovery of unknown patterns in the data. In particular, the structural representation allows us to convert a complex border between the patterns in multidimensional space into visual 2-D and 3-D forms. This decreases the information overload on the user. The visualization shows not only the border between classes, but also shows a location of the case of interest relative to the border between the patterns. A user does not need to see the thousands of previous cases that have been used to build a border between the patterns. If the abnormal case is deeply inside in the abnormal area, far away from the border between “normal” and “abnormal” classes, then this shows that this case is very abnormal and needs immediate attention. The paper concludes with the outline of the scaling of the algorithm for the large data sets and expanding the approach for non-monotone data.
Kovalerchuk B., Delizy F., Riggs L., Vityaev E. (2012) Visual Data Mining and Discovery with Binarized Vectors. In: Holmes D.E., Jain L.C. (eds) Data Mining: Foundations and Intelligent Paradigms. Intelligent Systems Reference Library, vol 24. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-23241-1_7
© Springer-Verlag Berlin Heidelberg 2012