A Geometric Approach to Feature Ranking Based Upon Results of Effective Decision Boundary Feature Matrix - Feature Selection for Data and Pattern Recognition - page 65

Table 4.6 Overall comparison of the five algorithms
Sign test—all datasets

ˆ

BVQ-FR Relief G.Ratio Onerule

4/9 7/6 6/7 3/10 w/l

OLDA-FR 0.733 0.0 0.0 0.908 P

10/3 10/3 7/6 w/l

BVQ-FR 0.908 0.908 0.0 P

8/5 6/7 w/l

Relief 0.419 0.0 P

5/8 w/l

GainRatio 0.419 P

4.6 Conclusions

This chapter focuses on a novel ranking procedure. We considered that the premise
for integrating the feature ranking models into domain knowledge is their repre-
sentation in terms of real world features. This principle is the fundamental premise
of the study conducted which leads to a computational model that is accurate and
humanly understandable. A new approach to Feature Ranking (FR) based on fea-
tures extraction (FE) and properties of the decision border has been discussed. This
method uses Effective Decision Boundary Feature Matrix (EDBFM) to measure the
relevance of the real world features thus maintaining the readability of the knowl-
edge model extracted. The method has been tested on classification problems and
cost-benefit analysis of features. While maintaining the geometric procedure which
yields the ranking of features, this method allows to choose between alternative core
FE algorithms, such as BVQ, when extracting the EDBFM, that allows to optimize
the method application on datasets with different complexity. In particular BVQ-FR
has proven to be more effective in applications to dataset of non-linearly separable
points. Benchmarking tests, supported by the calculation of index of performance,
show that BVQ-FR and OLDA-FR are generally more effective than other solutions.
Furthermore, the comparison with known heuristic techniques of ranking confirms
the robustness and the superiority of the EDBFM based method on complex dataset.

References

1. Alelyani, S., Liu, H., Wang, L.: The effect of the characteristics of the dataset on the selection
stability. In: Proceedings of the 23rd IEEE International Conference on Tools with Artificial
Intelligence (ICTAI), pp. 970-977 (2011)
2. Arauzo-Azofra, A., Aznarte, A.L., Benitez, J.M.: Empirical study of feature selection methods
based on individual feature evaluation for classification problems. Expert Syst. Appl. 37 (3),
8170-8177 (2011)