"... In PAGE 5: ... LERILS outperformed unordered RIPPER on only hepatitis and xd6, but was not significantly outperformed by RIPPER on any of the datasets. Table5 contains hypothesis sizes for a single run of the algorithms. These are indicative of comprehensibility, although comprehensibility involves more than the rule size and there is no widely accepted measure.... ..."

"... In PAGE 7: ... Gain Gini Gain Ratio Laplace G Statistic 2 voting 1:0 0:38 95:0 (0:26) 95:13 (0:38) 94:95 (0:39) 92:53 (1:35) cancer 1:01 0:65 94:16 (0:64) 94:64 (0:61) 94:51 (0:36) 90:34 (1:08) new-thyroid-hyper 1:01 0:17 96:67 (1:33) 97:10 (1:05) 96:43 (0:75) 90:42 (1:79) new-thyroid-hypo 1:0 0:15 96:19 (1:04) 96:10 (0:90) 95:62 (0:87) 94:19 (1:98) star cluster 1:04 0:67 96:95 (0:71) 96:81 (0:80) 96:86 (0:83) 83:90 (2:68) promoters 1:18 0:50 76:2 (4:07) 77:8 (2:71) 78:7 (3:82) 74:2 (4:02) mushroom 1:25 0:45 99:73 (0:02) 99:53 (0:32) 99:73 (0:02) 78:39 (0:01) ionosphere 1:25 0:66 89:63 (1:64) 89:74 (1:71) 90:31 (1:24) 66:06 (3:42) crx 1:38 0:45 86:01 (0:34) 86:12 (0:26) 85:94 (0:32) 62:05 (2:39) mean region 1 92:28 92:55 92:56 81:34 hepatitis 1:58 0:18 70:2 (2:93) 78:0 (1:90) 85:9 (2:34) 64:5 (3:11) lymphography-2 1:6 0:52 81:71 (2:22) 81:93 (2:0) 80:21 (3:13) 72:29 (3:09) lymphography-3 1:67 0:44 79:21 (3:03) 79:43 (2:85) 76:79 (1:87) 64:86 (4:54) zoo 1:92 0:05 85:33 (2:27) 85:44 (3:20) 85:0 (3:55) 47:0 (10:27) credit 1:95 0:67 67:36 (3:69) 69:81 (2:50) 73:09 (3:18) 54:90 (4:25) chess-end game 2:01 0:50 79:41 (0:86) 80:73 (0:33) 75:44 (0:09) 72:80 (0:94) heart 2:06 0:45 71:31 (1:03) 71:82 (1:35) 71:52 (1:27) 62:38 (4:77) mean region 2 76:36 78:17 78:28 62:68 diabetes 2:51 0:32 67:62 (2:13) 70:86 (1:80) 69:89 (0:84) 43:32 (0:95) bupa 2:61 0:45 59:18 (2:56) 60:63 (2:57) 57:72 (2:18) 47:12 (1:35) tic-tac-toe 2:69 0:67 58:41 (2:37) 67:88 (4:15) 73:68 (0:88) 50:77 (1:97) mean region 3 61:74 66:46 67:10 47:07 mean overall 77:16 78:94 79:30 65:02 e cient) of r = 0:97, which points to a strong vari- able interdependence between the distance-bias mea- sure and di erences in accuracy performance. Results for other similar experiments are all summarized on Table3 . Each entry shows the correlation coe cient obtained from tting a linear regression model to the data.... In PAGE 7: ... Such improvement indicates that the relation between the distance-bias measure and di erences in predictive ac- curacy is more evident when the computation of equa- tion 9 is done assuming a class distribution similar to that of the domain under analysis (Table 1 assumes equal class proportions). In addition, Table3 shows results of experiments similar to the two above, ex- cept we group domains on three regions according to the coverage of the expression in the nal rule hypoth- esis. The highest correlation is observed when domains are grouped into regions and skewed distributions are eliminated (bottom entries on column 3, Table 3).... In PAGE 7: ... The rst experiment uses a simple algorithm that outputs a single-feature as the nal hypothesis. After repeating the experiments reported above, correlation coe cients comparable to the rst row of results on Table3 (average over all do- mains with and without eliminating skewed distribu- tions) take values of r = 0:98 and r = 0:98 respectively. The second experiment increases the complexity of the algorithm by using a decision tree as the hypothesis.... ..."