SEPARATION IN DATA MINING BASED ON FRACTAL NATURE OF DATA

The separation of the searched data from the rest is an important task in data mining. Three separation/classification methods are presented. We use a singularity exponent in classifiers that are based on distances of patterns to a given (classified) pattern. The approximation of so called probability distribution mapping function of the distribution of points from the viewpoint of distances from a given point in the form of a scaling exponent power of a distance is presented together with a way how to state it. Considering data as points in a metric space, three methods are based on transformed distances of neighbors of a given point in a multidimensional space via functions that use different estimates of scaling exponent. Classifiers a€“ data separators utilizing knowledge about explored data distribution in a space and suggested expressions of the scaling exponent are presented. Experimental results on both synthetic and real-life data show interesting behavior (classification accuracy) of classifiers in comparison with other well-known approaches.