Conference publications

Abstracts

XVII conference

Algorothm of signs set optimization for solving classification problems

Akinin P.V., Karp V.P.

MIREA, Russia, 119454, Moscow, prospekt Vernadskogo 78, +7(909)923-22-33, E-mail: pavak@mail.ru

One of the most important problems solving by intellectual information systems is the classification problem. Generally, a researcher consider obviously superfluous features set because of the absence of a sufficient a priori information. This set of features bears noisy information about classes and generates unreliable dependences. In such cases, there is a need to optimize the feature space [1].

The idea of the algorithm is based on two iterative procedures: features recruitment and features exception. Each procedure consists of several stages. Features recruitment procedure begins with the sorting of all features by estimation of their diagnostics quality in decreasing order (Q). The feature, which has maximum value of Q, becomes the first feature in the set. Then for a given feature another one, which gives together with the first one the best estimate Q, is added. Далее к этим двум признакам аналогичным путем подбирается тот, который дает в сочетании с ними наилучшее Q (превышающее качество уже имеющегося набора признаков). The process continues until among excluded signs will not remain a single one, whose inclusion in the set would improve the evaluation of Q. After this the reverse procedure: algorithm excludes from the resulting set those features that do not impair the evaluation of Q formed by the set. Upon completion, another attempt to find out features that would improve Q is taken. After completion begins features elimination process. Thus, formed an iterative system, which is an odd iteration strive to include features that improve the assessment of the quality of diagnosis, and even - to exclude evidence, not worsen it. This results in a set of features, which gives the best diagnostic results for the chosen method of classification. It is important to say that described optimization engine is universal as it does not imply the use of any specific classification algorithm. Testing of the algorithm is planned, primarily on the formation of a diagnostic feature space by sorting conjunctions.

References

1. Karp V.P. Metody i sredstva kontrolya i diagnostiki slognoorganizovannih system. – M.: MIREA, 2008.

• view as PDF (in Russian) (PDF)