L.I. Dvoyris - Dr. Sc. (Eng.), Professor M.V. Kobzar K.D. Galev

A number of known methods is reduced to ranking features for their information according to different criteria (Student, Fisher et al.). However, this ignores the possibility of a relationship between them, which significantly affects the final results. RELIEF method is reduced to finding the weighting factors for each of the signs in accordance with their ability to determine the difference of neighboring objects. Analysis of the results of the method according to the algorithm shown RELIEF weights signs (or artificial source) depending on the number of iterations. From the initial signs are most useful 3rd and 4th, artificially introduced features are not informative. To solve the problem of classifying all of the original features have weight 1, all artificial - 0. It also became obvious that the calculation results are practically unchanged after ≈100 iterations.

 

1. Robnik-Sikonja M., Kononenko I. Theoretical and Empirical Analysis of ReliefF and R ReliefF // Machine Learning. 2003. 53. 23−69.
2. Sun Y., Lou X., Bao B. A Novel Relief Feature Selection Algorithm Based onMean-Variance Model // Journal of Information & Computational Science. 2011. 8: 16. 3921−3929.
3. Sun Y., Li J. Iterative RELIEF for Feature Weighting // Proceedings of the 23rd International Conference on Machine Learning. Pittsburgh, PA. USA. 2006.