Skip to main content

Table 2 LOOCV experiments on robustness of the feature weighting on the spiral with irrelevant noisy feature

From: Feature weight estimation for gene selection: a local hyperlinear learning approach

Spiral data LHR I-RELIEF LOGO
Dimension SVM LDA NB KNN HKNN Aver. SVM LDA NB KNN HKNN Aver. SVM LDA NB KNN HKNN Aver.
0 84.0 53.3 50.0 87.0 87.0 72.3 84.0 51.7 50.0 87.0 87.0 72.0 51.0 51.5 52.0 84.3 84.3 64.6
1000 86.0 61.2 61.8 92.0 87.0 77.6 59.3 54.3 54.8 78.3 60.5 61.4 56.8 57.7 59.3 82.3 64.3 64.1
2000 90.0 69.0 67.0 92.0 89.0 81.4 56.8 57.0 55.8 72.3 57.8 59.9 57.3 57.8 60.3 88.5 83.8 69.5
3000 87.5 67.0 64.0 91.8 86.8 79.4 56.8 55.3 52.8 74.3 56.3 59.1 60.0 54.8 54.5 85.3 76.5 66.2
4000 88.5 64.0 66.3 92.3 88.5 79.9 55.5 58.8 57.3 71.3 55.3 59.6 59.0 60.5 61.8 86.5 79.5 69.5
5000 89.0 67.8 66.8 92.8 87.8 80.8 81.3 59.5 57.0 85.3 77.8 72.2 61.8 60.8 63.7 88.8 81.0 71.2
6000 88.8 66.3 67.5 92.0 88.3 80.5 64.3 57.3 59.5 74.0 61.0 63.2 54.8 54.5 57.0 87.5 82.5 67.3
7000 89.3 69.5 70.0 92.0 89.0 81.9 83.5 61.0 54.8 88.0 79.5 73.3 63.0 65.5 67.0 87.8 83.8 73.4
8000 86.8 65.0 66.8 93.8 88.3 80.1 0.0 56.5 59.3 69.8 55.0 48.1 66.5 67.5 69.3 89.3 82.5 75.0
9000 88.8 68.5 70.8 92.8 87.0 81.6 51.5 49.0 51.2 68.0 53.5 54.6 62.0 57.5 61.0 73.3 57.0 62.1
10000 88.8 68.5 70.8 92.8 87.0 81.6 51.5 49.0 51.2 68.0 53.5 54.6 57.0 57.5 56.0 86.5 81.8 67.8
  1. When using LOOCV criteria, the LHR outperforms both I-RELIEF and LOGO in terms of accuracy after classical classifier of SVM, LDA, NB, KNN and HKNN. The better averaged value after the two methods are highlighted in bold. With the increase of dimension of the irrelevant features, the performance of both LOGO and I-RELIEF are degraded while LHR keeps stable.