Ing. Marcel Jiřina, DrSc.

Vědecký pracovník

Oddělení: Oddělení strojového učení
Telefon: +420 266 05 3350
Kancelář:
Email: (email)

Project IINC - Classifiers based on multifractal nature of data

There is a set of new classification methods that use distances in a multidimensional data space and are derived on the basis of suitable polynomial transformation with noninteger exponent.

News

It has been proved that any change of weights of neighbors in the Inverted Indexes of Neighbors Classifier leads to larger classification error. Then the IINC is the best NN rule among all k-NN rules. This proof is independent on the distance function or metric used.

Motivation

We found that near-neighbor based methods can for some tasks outperform much sophisticated methods. Also, there is a reliance on locall uniformity around the point to be classified. As the local uniformity is impossible for multivariate data of higher dimensions and with limited number of samples in practice, we searched for a suitable transformation. We finally found that a polynomial transformation with noninteger exponent exist, and leads, in the end, to very good results.

The exponent mentioned is called the distribution mapping exponent and is very close to correlation dimension in the theory of fractals and to scaling exponent known from the theory of multifractals.

Software

IINC package

IINC is a package of classification methods based on notion of the distribution mapping exponent. The exponent is very similar to singularity exponent or fractal or multifractal dimension used for study of fractals and multifractals.

Added are standard nearest neighbors methods 1-NN and k-NN. User can choose one of fifteen metrics including Mahalanobis distance and Hassanat metric

In papers and report attached to the package theoretical basis and details of methods are described.

The software is freeware with no warranty and for non-commercial use only. If used successfuly please cite an appropriate paper of this package.

Download: IINC Package

Please report bugs to my mail. For any help don't hesitate write to my mail. Let us know about citation to my mail.

IINC in Excel Document

The IINC-Excel software package (for MS Excel only) contains classification method (IINC) that can outperform a range of standard classification methods. In Excel workbook the IINC is implemented as a macro. The program is designed for easy exprimentation. Corresponding papers and technical reports are included as well.

This software is freeware with no warranty and for non-commercial use only.

Download: IINC Excel Document

Please report bugs to my mail. For any help don't hesitate write to my mail. If used successfuly please cite an appropriate paper in this package. Let us know about citation to my mail.