18.11.2012 Views

anytime algorithms for learning anytime classifiers saher ... - Technion

anytime algorithms for learning anytime classifiers saher ... - Technion

anytime algorithms for learning anytime classifiers saher ... - Technion

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />

Table 4.6: Average cost of classification as a percentage of the standard cost of<br />

classification <strong>for</strong> mc = 1000. The first 25 rows list <strong>for</strong> each dataset the average cost<br />

over the different 4 cost-assignments, while the last 5 rows give the results <strong>for</strong> the<br />

datasets with costs from (Turney, 1995).<br />

Dataset C4.5 LSID3 IDX CSID3 EG2 DTMC ICET ACT<br />

Breast-cancer 54.3 56.7 48.0 55.4 48.6 43.6 45.2 45.5<br />

Bupa 91.5 84.8 78.4 85.5 78.2 67.2 69.8 72.8<br />

Car 41.4 44.3 53.8 42.8 51.4 57.2 41.2 42.1<br />

Flare 26.0 26.0 27.1 26.3 27.1 26.0 25.9 26.0<br />

Glass 53.9 47.2 55.0 50.9 50.2 49.2 46.1 39.7<br />

Heart 45.0 49.5 43.0 44.2 42.8 40.3 41.2 41.4<br />

Hepatitis 44.2 38.1 31.9 38.1 32.2 28.6 29.3 29.7<br />

Iris 22.3 20.8 24.1 21.3 22.8 18.9 16.8 16.3<br />

KRK 57.8 63.7 65.6 63.6 64.8 65.2 55.1 55.6<br />

Multi-ANDOR 47.9 14.1 31.5 32.2 32.4 49.0 16.2 11.8<br />

Monks1 16.3 12.2 15.8 16.2 16.3 52.3 15.6 12.2<br />

Monks2 65.4 77.4 71.3 69.4 70.1 57.5 57.5 57.5<br />

Monks3 23.8 22.9 22.5 22.3 22.3 25.5 22.1 22.0<br />

Multiplexer 59.3 16.2 51.4 54.7 52.0 57.9 18.5 19.9<br />

MultiXOR 62.5 18.5 51.0 56.9 52.3 56.7 49.5 21.3<br />

Nursery 20.3 21.6 21.4 20.5 20.8 24.7 18.6 20.2<br />

Pima 74.6 80.3 65.6 71.3 68.4 62.3 60.8 64.0<br />

Tae 82.1 73.5 72.3 73.6 74.3 64.2 76.8 67.4<br />

Tic-tac-toe 53.3 50.2 45.6 48.6 45.9 60.9 44.8 40.2<br />

Titanic 72.5 62.3 65.9 67.8 66.2 64.3 60.5 61.2<br />

Thyroid 24.3 27.3 25.4 24.1 24.1 17.7 24.1 17.7<br />

Voting 12.7 14.1 14.8 13.0 14.3 11.8 12.0 11.8<br />

Wine 20.3 21.9 30.0 20.7 26.8 20.6 18.8 19.1<br />

XOR3d 89.0 49.6 75.4 77.8 75.3 84.4 70.1 66.7<br />

XOR5 77.9 31.6 79.7 80.7 80.6 58.4 58.4 67.4<br />

Bupa 93.0 91.2 92.8 88.5 92.8 80.8 84.3 91.2<br />

Heart 56.0 61.1 37.4 38.9 38.3 38.6 35.4 42.3<br />

Hepatitis 93.1 86.5 68.4 74.2 73.8 77.3 68.1 76.8<br />

Pima 75.8 84.4 96.0 85.4 96.0 73.4 70.5 72.1<br />

Thyroid 23.8 34.8 24.7 23.6 22.7 22.4 24.6 22.9<br />

requires that the learner carefully balance the two types of cost. In these cases<br />

ACT has the lowest average cost and the largest number of t-test wins. Moreover,<br />

the Wilcoxon test indicates that it is superior. ICET is the second best method.<br />

As reported by Turney (1995), ICET is clearly better than the greedy methods<br />

EG2, IDX, and CSID3.<br />

Note that EG2, IDX, and CSID3, which are insensitive to misclassification<br />

cost, produced the same trees <strong>for</strong> all values of mc. These trees, however, are<br />

89

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!