### Table 1. A pair-wise comparison of di erent evaluation metrics.Pair of Metrics

2000

"... In PAGE 5: ... On each comparison we used Mathematica (Wolfram, 1999) to obtain a numeric approximation of equation 9. The results are shown in Table1 . The coordinates for R1, R2, and R3 are indicated on the rst row.... In PAGE 5: ... The area of the half space in Figure 3(b) equals (100 100)=2 = 5000, and in general the area equals (n1 n0)=2. An analysis of Table1 shows no di erence between In- formation Gain and the G statistic, and between Gini and 2 (Table 1, last row). We can prove the rst two metrics share the same bias by observing that the G statistic is simply a constant factor times Information Gain, and thus both metrics project the same isomet- rics lines over the coverage plane.... In PAGE 5: ... The area of the half space in Figure 3(b) equals (100 100)=2 = 5000, and in general the area equals (n1 n0)=2. An analysis of Table 1 shows no di erence between In- formation Gain and the G statistic, and between Gini and 2 ( Table1 , last row). We can prove the rst two metrics share the same bias by observing that the G statistic is simply a constant factor times Information Gain, and thus both metrics project the same isomet- rics lines over the coverage plane.... In PAGE 5: ... Most distances are within 5%, except for the Laplace function showing di erences up to 16%. Notice in Table1 that the di erence between any pair of metrics increases from point R1 to point R2, and then decreases from R2 to R3. We conclude a ceiling e ect (Cohen, 1995, Chapter 3, p.... In PAGE 6: ... We compute the absolute di erence in predictive accu- racy between each pair of algorithms. We also compute the average distance in bias between pairs of evalua- tion metrics in Table1 (average over the three regions). The results can be paired-up by matching average dis- tance in bias between metrics with the corresponding average accuracy di erence (e.... In PAGE 7: ... Without skewed distributions the experiment described above produces a correlation coe cient of r = 0:99. 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 ( Table1 assumes equal class proportions). In addition, Table 3 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.... ..."

Cited by 8

### Table 1: A pair-wise comparison of evaluation metrics. Pair of Metrics

"... In PAGE 8: ... On each comparison we used Mathematica (Wolfram, 1999) to obtain a numeric approximation of equation 9. The results are shown in Table1 . The coordinates for R1, R2, and R3 are indicated on the rst row.... In PAGE 8: ... The area of the half space in Figure 3(b) equals (100 100)=2 = 5000, and in general the area equals (n1 n0)=2. An analysis of Table1 shows no di erence between Information Gain and the G statis- tic. The two metrics share the same bias since the G statistics is simply a constant factor times information gain (Section 3.... In PAGE 9: ... Most distances are within 5%, except for the Laplace function showing di erences up to 16%. Notice in Table1 that the di erence between any pair of metrics increases from point R1 to point R2, and then decreases from R2 to R3. An explanation for such general behavior can be related to the ceiling and oor e ects of experimental design (Cohen, 1995, Chapter 3, p.... In PAGE 9: ...xpression E and of its complement E0 (i.e., by looking into all numbers in the contingency table of Figure 1), and partial evaluators, characterized by evaluating the quality of the coverage of E alone, ignoring the complement E0. The results in Table1 can be explained by noticing that the Laplace function belongs to the second kind of family while all others functions belong to the rst family. We could consider a version of the Laplace function L2 that belongs to the rst kind of family (i.... In PAGE 10: ... Table 2 compares the two versions of the Laplace function against Information Gain. The distance between L2 and Information Gain is similar to the distance between pairs of metrics in Table1 (excluding occurrences with the original Laplace function). L2, however, appears di erent than other metrics close to axis-line A, where we observe an increase in distance-bias from R2 to R3.... In PAGE 10: ... 3.5 The Case of Unequal Class Proportions Results on Table1 are based on problem domains with equal class proportions. But, what happens if class proportions are unequal? To answer this question we analyze the e ect of skewed class distributions over the coverage plane.... In PAGE 14: ... We compute the absolute di erence in predictive accuracy between each pair of algorithms. We also compute the average distance in bias between pairs of evaluation metrics in Table1 (average over the three regions). The results can be paired-up by matching average distance in bias between metrics with the corresponding... ..."

### TABLE 5 Distribution of (p,p) contacts as a function of monomer type and AAG(H0,HO) AAG(H4D,H4)

1996

Cited by 1

### Table 4 An interaction test suite with 11 seeded rows requires only 4 extra tests to cover all pair-wise interactions.

2006

"... In PAGE 10: ... This can arise, for example, when extending a seed; interactions covered in the seed are thereafter neutral. In Table4 , the combination (8,11) was specified as neutral and consequently is included only in the last test (since no benefit arises from its coverage). 3.... ..."

Cited by 4

### Table 4 An interaction test suite with 11 seeded rows requires only 4 extra tests to cover all pair-wise interactions.

2006

"... In PAGE 10: ... This can arise, for example, when extending a seed; interactions covered in the seed are thereafter neutral. In Table4 , the combination (8,11) was specified as neutral and consequently is included only in the last test (since no benefit arises from its coverage). 3.... ..."

Cited by 4

### Table 1: Scale of pair-wise comparisons

"... In PAGE 5: ... Example 4: A student wants to determine the relative importance of four activities in his life: A part-time job, study, personal activities and social activities. Using AHP, he will first make six pair-wise comparisons according to a pre-defined scale shown in Table1 . The results of the comparisons are shown in a matrix given in Table 2.... ..."

### Table 1: Scale of pair-wise comparisons

"... In PAGE 7: ... Example 4: A student wants to determine the relative importance of four activities in his life: A part-time job, study, personal activities and social activities. Using AHP, he will first make six pair-wise comparisons according to a pre-defined scale shown in Table1 . The results of the comparisons are shown in a matrix given in Table 2.... ..."

### Table 2 Pair-wise prediction characteristics

2006

"... In PAGE 3: ...EGASP evaluation of the submitted results and those displayed in Table 1, but do not materially change the results or the interpretation of them. Transcript predictions from individual informant sources For the set of all 44 ENCODE regions, each individual informant results in a similar number of predicted tran- scripts ( Table2 ). Informant sources at greater evolutionary distances tend to result in fewer, longer transcripts than informants within the mammalian lineage.... In PAGE 3: ... Informant sources at greater evolutionary distances tend to result in fewer, longer transcripts than informants within the mammalian lineage. However, the summary information from the ENCODE regions presented in Table2 only hints at the diversity of predicted transcripts from the various informant sources. For example, mouse and rat shared a common ancestor approximately 25 million years ago and align similar fractions of the human genome using our alignment procedure (see Materials and methods and Table 3), but using these two rodent genome sequences as informant sources leads to a significantly different set of transcripts.... In PAGE 3: ... Conversely, the opossum aligns approximately one-third the total number of bases as the rodent sequences, while retaining alignment in 76% of the coding regions (Table 3). Informative value of the pair-wise alignments The alignment characteristics for each of the six informant sequences shown in Table 3 are primarily responsible for the characteristics of the pair-wise prediction sets shown in Table2 . To asses how the alignments affect the various http://genomebiology.... ..."