### Table 3: Pairwise correlation coefficients of the relative improvements of boosting, bagging and round robin over c5.0.

2003

"... In PAGE 8: ...erformance ratios ripper-rr/ripper and c5.0-boost/c5.0 was about 0.618, which is in the same range as the correlation coefficient 0.69 that reported for the error reduction rates of bagging and boosting decision trees ( Table3 of Opitz and Maclin, 1999). Table 3 shows the pairwise correlation coefficients of the results for c5.... In PAGE 9: ... While the combination with bagging or boosting with round robin learn- ing did not produce substantially different results than bagging or boosting alone, the combination of bagging and boosting emerged as a clear winner. In the light of the results of Table3 , which showed that the predictions of round robin learning and both bagging or boosting are less correlated than the predictions of bagging and boosting, this came as a surprise. We would have expected that the classi- fiers that differ more from each other have a better chance of complementing each other into an even better classifier that exploits the different strengths of each of its constituents.... ..."

Cited by 3

### Table 1. Comparison of C4.5c, Boosting and Cost-Boosting Datasets C4.5c Boosting Cost-Boosting Cost-Boosting vs C4.5c vs C4.5c vs Boosting

"... In PAGE 5: ... Table1 shows the misclassi cation costs and the number of high cost errors of C4.5c, and the ratios for the pair-wise comparison among C4.... ..."

### Table 1: MPEG-4 anthropometric constraints and corresponding pair-wise potentials.

### Table 4. LADTree and AdaBoost.MH results

"... In PAGE 9: ...he table. On both scales the exhaustive method is the best. As the exhaus- tive method is not practical for large class datasets we chose the 1-against-1 method to compare against LADTrees, as this method is very similar in overall performance. Table4 compares the \winner quot; of Table 3 (1-against-1) to AdaBoost.MH and... In PAGE 10: ...Boost.MH outperform this method on average after 100 iterations (i.e. using trees with 100 tests). Table4 shows that LT(100) outperforms most of the early datasets (class sizes 3-13) but struggles against two of the later datasets. For soybean 1-against-1 uses a tree of size 1710, and for primary-tumor it uses a tree of size 2310.... In PAGE 10: ... There are no obvious performance di erences between these methods at 10 and 100 iterations. Table4 also compares the two logistic methods. Due to the number of trees used by LT1PC it outperforms LT both on average and on pairwise tests.... ..."

### Table 2. Comparison of the recognition accuracy and the number of features used by the Bayes classifier without feature selection (Bayes All), Bayes with pairwise-greedy feature selection (Bayes FS), AdaBoost, linear SVM (L-SVM), non-linear SVM (NL-SVM), and FSLP.

"... In PAGE 5: ...1.0%, 92.4%, and 91.9%, respectively (see Table2 ), which are comparable. Notice, however, that the average number of features selected by FSLP was 17.... In PAGE 6: ... For the AdaBoost method, peak performance was 71.9% using 80 features (see Table2 ) for each pair of classes. As shown in Figure 4, using more features slightly lowered recogni- tion accuracy.... ..."

### Table 2. Comparison of the recognition accuracy and the number of features used by the Bayes classifier without feature selection (Bayes All), Bayes with pairwise-greedy feature selection (Bayes FS), AdaBoost, linear SVM (L-SVM), non-linear SVM (NL-SVM), and FSLP.

"... In PAGE 5: ...1.0%, 92.4%, and 91.9%, respectively (see Table2 ), which are comparable. Notice, however, that the average number of features selected by FSLP was 17.... In PAGE 6: ... For the AdaBoost method, peak performance was 71.9% using 80 features (see Table2 ) for each pair of classes. As shown in Figure 4, using more features slightly lowered recogni- tion accuracy.... ..."

### Table 2 shows a comparison between these two algorithms. It reports the fraction of pairwise misrankings for both algorithms using the same experimental set-up as previously described:

"... In PAGE 19: ...Table2 . Comparison of MPRank and RankBoost for pairwise misrankings.... ..."

### Table 1 presents details of the imputed must-link and cannot-link constraint sets generated for each dataset. Note that the numbers reported do not take into account any additional cannot-link constraints that can be inferred from the im- puted must-link constraints. We compare the imputed sets to the correct pairwise relations defined by the natural classification of the datasets, using measures of pairwise precision (PP) and pairwise recall (PR). Given an imputed set Yprime, the former refers to the fraction of imputed pairs that are correctly constrained, while the latter represents the fraction of the complete set Y recovered:

2007

"... In PAGE 9: ...9 or higher were achieved for both constraint types. Table1 also lists the Table 1. Details of imputed constraint sets for text datasets.... In PAGE 9: ...9 or higher were achieved for both constraint types. Table 1 also lists the Table1 . Details of imputed constraint sets for text datasets.... In PAGE 12: ...the recall of the imputed constraints did not have a direct impact on the choice of suitable representatives for the first phase of ensemble-based selection. Also, in the case of the 3-news-similar dataset, which achieved a relatively low level of pairwise precision as shown in Table1 , both the imputed constraints and the related co-association values still proved useful when selecting real constraints. 5 Conclusion In this paper, we demonstrated that it is often possible to correctly impute sets of pairwise constraints for data by examining the co-associations in an ensemble of clusterings.... ..."

### Table 5: Test set misclassi cation errors for L2Boost, L2WCBoost (with constraints) and LogitBoost. Optimal (minimizing cross-validated test set error) number of boosts is given in parentheses; if the optimum is not unique, the minimum is given. \Larger tree quot; denotes a tree learner such that the ancestor nodes of the terminal leaves contain at most 10 observations: resulting average tree size (integer-rounded) for L2WCBoost is 12 terminal nodes.

2002

"... In PAGE 21: ... Monk 1 has Bayes error equal to zero. The estimated test set misclassi cation errors, using an average of 50 random divisions into training with 90% and test set with 10% of the data, are given in Table5 . The comparison is made when using the optimal (with respect to cross-validated test set error) number of boosting iterations for every boosting algorithm; these numbers are given in parentheses.... ..."

### Table 5: Test set misclassi cation errors for L2Boost, L2WCBoost (with constraints) and LogitBoost. Optimal (minimizing cross-validated test set error) number of boosts is given in parentheses; if the optimum is not unique, the minimum is given. \Larger tree quot; denotes a tree learner such that the ancestor nodes of the terminal leaves contain at most 10 observations: resulting average tree size (integer-rounded) for L2WCBoost is 12 terminal nodes.

2002

"... In PAGE 21: ... Monk 1 has Bayes error equal to zero. The estimated test set misclassi cation errors, using an average of 50 random divisions into training with 90% and test set with 10% of the data, are given in Table5 . The comparison is made when using the optimal (with respect to cross-validated test set error) number of boosting iterations for every boosting algorithm; these numbers are given in parentheses.... ..."