### Table 7-12 Mean and Standard deviation of best results from each type of run

"... In PAGE 72: ... The exception to this is the experiment where 60% of the operations are merge operations. Table7 -1 Number of runs with best result gt;= to a given % of examples correctly classified for each experiment Best gt;= 87% 88% 89% 90% 91% 92% 93% 94% 95% 0MaxYes 73 69 61 46 36 19 5 0 0 25MaxNo 75 74 74 67 55 36 16 7 1 25MaxYes 75 74 69 58 42 26 9 2 0 25MeanNo 75 73 70 60 47 31 7 4 0 50MaxNo 75 74 68 54 34 18 9 1 0 50MaxYes 73 70 56 42 28 10 4 3 0 50MeanNo 74 71 64 51 39 21 10 2 0 60MaxNo 73 71 62 50 35 26 6 1 0 60MeanNo 74 72 60 47 29 15 6 1 0 66MaxYes 75 72 60 44 24 12 4 0 0 Table 7-1 shows the number of runs which had a score greater or equal to the given percentage value of examples correctly classified. For example, 25MaxNo is the only experiment that had a run with a score greater or equal to 95% correctly classified.... In PAGE 72: ... The exception to this is the experiment where 60% of the operations are merge operations. Table 7-1 Number of runs with best result gt;= to a given % of examples correctly classified for each experiment Best gt;= 87% 88% 89% 90% 91% 92% 93% 94% 95% 0MaxYes 73 69 61 46 36 19 5 0 0 25MaxNo 75 74 74 67 55 36 16 7 1 25MaxYes 75 74 69 58 42 26 9 2 0 25MeanNo 75 73 70 60 47 31 7 4 0 50MaxNo 75 74 68 54 34 18 9 1 0 50MaxYes 73 70 56 42 28 10 4 3 0 50MeanNo 74 71 64 51 39 21 10 2 0 60MaxNo 73 71 62 50 35 26 6 1 0 60MeanNo 74 72 60 47 29 15 6 1 0 66MaxYes 75 72 60 44 24 12 4 0 0 Table7 -1 shows the number of runs which had a score greater or equal to the given percentage value of examples correctly classified. For example, 25MaxNo is the only experiment that had a run with a score greater or equal to 95% correctly classified.... In PAGE 74: ... The standard Holm method was used as supplied with the R statistical system. Table7 -2 shows the mean and standard deviation of the 75 runs for each experiment. Here, the 25MaxNo experiment has the highest mean value.... In PAGE 74: ...xperiment. Here, the 25MaxNo experiment has the highest mean value. There are two main questions. Firstly, is this better than the base case with zero merging and secondly, is it better than the other experiments? Table7 -2 Mean and Standard deviation of best results from each run for a particular set of parameters Best correct value Mean Std Deviation 0MaxYes 90.11 1.... In PAGE 75: ... 7-16 Examining the adjusted probability values from the multiple comparisons of means in Table7 -3 a probability of less than 0.05 is required to show that there is a significant difference between the mean values.... In PAGE 75: ... Therefore the methods that use merging for 25 percent of the rule discovery operations are not significantly different from each other and that 25MaxNo is significantly better than the remainder of the results. Now, examining the differences for the various merge percentages in Table7 -4 it can be seen that merging 25% of the time is significantly different to all other methods with the worst probability being 0.0013.... In PAGE 75: ...0013. Finally, comparing the means for the different types of run shown in Table7 -5, it can be seen that the only significant difference is between using a special pass and using the maximum value for the new rule strength. Table 7-3 probability values from a pairwise comparison of means (note 0.... In PAGE 75: ...ith the worst probability being 0.0013. Finally, comparing the means for the different types of run shown in Table 7-5, it can be seen that the only significant difference is between using a special pass and using the maximum value for the new rule strength. Table7 -3 probability values from a pairwise comparison of means (note 0.000 represents a value of less than 0.... In PAGE 75: ...000 1.000 Table7 -4 Comparison of percent merged (note 0.000 represents a value of less than 0.... In PAGE 76: ... 7-17 Table7 -5 comparison of run type (note 0.000 represents a value of less than 0.... In PAGE 76: ... The results are compared below with the previously generated numbers from the merge experiment. Table7 -6 comparisons for best results with and without duplicates Duplicates Mean Std Deviation Yes 90.11 1.... In PAGE 78: ... 7-19 Table7 -7 Number of runs with best result gt;= to a given % of examples correctly classified for no duplicate experiments Best = gt; 94% 95% 96% 97% 98% 99% 100% 0MaxYes 72 64 54 28 7 0 0 25MaxNo 100 100 99 98 94 73 30 25MaxYes 100 99 98 95 89 71 23 25MeanNo 100 100 99 93 84 69 31 50MaxNo 100 100 100 100 96 82 25 50MaxYes 99 98 97 86 75 63 43 50MeanNo 100 100 100 99 93 81 38 60MaxNo 100 100 100 100 97 78 23 60MeanNo 100 99 99 98 94 85 34 66MaxYes 98 96 91 91 79 63 31 Table 7-7 shows the number of runs with a score greater or equal to the given percentage value of examples correctly classified. All results for 50MaxNo and 60MaxNo were better or equal to 97% correctly classified.... In PAGE 78: ... 7-19 Table 7-7 Number of runs with best result gt;= to a given % of examples correctly classified for no duplicate experiments Best = gt; 94% 95% 96% 97% 98% 99% 100% 0MaxYes 72 64 54 28 7 0 0 25MaxNo 100 100 99 98 94 73 30 25MaxYes 100 99 98 95 89 71 23 25MeanNo 100 100 99 93 84 69 31 50MaxNo 100 100 100 100 96 82 25 50MaxYes 99 98 97 86 75 63 43 50MeanNo 100 100 100 99 93 81 38 60MaxNo 100 100 100 100 97 78 23 60MeanNo 100 99 99 98 94 85 34 66MaxYes 98 96 91 91 79 63 31 Table7 -7 shows the number of runs with a score greater or equal to the given percentage value of examples correctly classified. All results for 50MaxNo and 60MaxNo were better or equal to 97% correctly classified.... In PAGE 79: ...8.2 Statistical comparisons Table7 -8 Mean and Standard deviation of best results Best correct value Mean Std Deviation 0MaxYes 95.40 2.... In PAGE 79: ...85 1.6269 Table7 -8 shows the mean and standard deviation of the 100 runs for each experiment excluding all duplicate examples. Using the statistical comparisons of means method for large samples (greater than 30 samples), some conclusions can be drawn as described below.... In PAGE 79: ... Using the statistical comparisons of means method for large samples (greater than 30 samples), some conclusions can be drawn as described below. Table7 -9 shows the results of these comparisons of means. The result for 0MaxYes, the base case, where no merging takes place, is significantly different to all the other experiments.... In PAGE 79: ...elow 0.05. Therefore, little can be deduced from these. Examining the percentage merging statistics in Table7 -10 we can again conclude that no merging is inferior to using merging in any form as already discovered in the paragraph above. The type of run statistics in Table 7-11 are more interesting in that they show that a special run is also different from the other types of runs.... In PAGE 79: ...elow 0.05. Therefore, little can be deduced from these. Examining the percentage merging statistics in Table 7-10 we can again conclude that no merging is inferior to using merging in any form as already discovered in the paragraph above. The type of run statistics in Table7 -11 are more interesting in that they show that a special run is also different from the other types of runs. Table 7-12 presents the summary statistics for classifying the results by run type.... In PAGE 79: ... The type of run statistics in Table 7-11 are more interesting in that they show that a special run is also different from the other types of runs. Table7 -12 presents the summary statistics for classifying the results by run type. ... In PAGE 80: ... 7-21 proportion of the normal rule discovery pass is significantly superior to replacing a proportion of rule discovery passes with a special merge only pass. Table7 -9 probability values from a pair-wise comparison of means for no duplicate experiments (note 0.000 represents a value of less than 0.... In PAGE 80: ...052 0.046 Table7 -10 comparison of percent merged (note 0.000 represents a value of less than 0.... In PAGE 80: ...069 0.006 Table7 -11 comparison of run type (note 0.000 represents a value of less than 0.... In PAGE 81: ...Once the merge operator is combined with removing the duplicate examples the classifier system is able to classify all the examples correctly. Using the merge operator at a level of 50 to 60% gives a high probability of perfect scores of all examples being correctly classified, as well as all results being 97% correct or better as shown for 50MaxNo and 60MaxNo in Table7 -7. 7.... In PAGE 81: ...9 Using the test data Figure 7-8 shows how the modified classifier system performed in comparison to the original algorithm as used in chapter 5. Table7 -13 shows the detailed numbers used in Figure 7-8.The modified algorithm included both the removal of duplicates from the training set and the use of the merging operation during the discovery cycle.... In PAGE 83: ... These differences would require the settings of these parameters to be chosen specifically for different data sets to be used. Table7 -13 list of results comparing the classifier system before and after modifications for the 7Pdata classifier system Average correct StdDev Count Standard Error Noise00N 91.97 2.... In PAGE 83: ...582 100 0.711 Table7 -14 shows the number of runs which had a score greater or equal to the given percentage value of examples correctly classified for the 7Cdata. This shows that all experiments except for 0MaxYes, the base case and 25MeanNo were better or equal to 96% correctly classified.... In PAGE 84: ... 7-25 Table7 -15 shows the mean and standard deviation of the 100 runs for each experiment for the 7Cdata. Using the statistical comparisons of means method for large samples (greater than 30 samples) produced the results shown in Table 7-16 through Table 7-18.... In PAGE 84: ... 7-25 Table 7-15 shows the mean and standard deviation of the 100 runs for each experiment for the 7Cdata. Using the statistical comparisons of means method for large samples (greater than 30 samples) produced the results shown in Table7 -16 through Table 7-18. The statistical inferences are detailed below.... In PAGE 84: ... 7-25 Table 7-15 shows the mean and standard deviation of the 100 runs for each experiment for the 7Cdata. Using the statistical comparisons of means method for large samples (greater than 30 samples) produced the results shown in Table 7-16 through Table7 -18. The statistical inferences are detailed below.... In PAGE 84: ...able 7-18. The statistical inferences are detailed below. 0 5 10 15 66MaxYes Average 0 5 10 15 50MaxYes average 0 5 10 15 60MaxNo Average 0 5 10 15 50MaxNo Average 0 5 10 15 200000 600000 1000000 1400000 1800000 2200000 2600000 3000000 3400000 3800000 4200000 4600000 60MeanNo Average 0 5 10 15 50MeanNo Average 0 5 10 15 25MaxNo Average 0 5 10 15 25MeanNo Average 0 5 10 15 25MaxYes Average 0 5 10 15 200000 600000 1000000 1400000 1800000 2200000 2600000 3000000 3400000 3800000 4200000 4600000 0MaxYes Average count and location where best values occur merge + nodup. Alternate data Figure 7-9 Location of best values for merge with no duplicates experiments for the 7Cdata The type of run statistics in Table7 -18 shows the results of comparing each run to every other run. The result for 0MaxYes, the base case, where no merging takes place was significantly different to all the other experiments.... In PAGE 85: ... 7-26 in Table7 -15 shows that the 25% merge runs are worse than using a higher level of merging. Table 7-14 Number of runs with best result gt;= to a given % of examples correctly classified for merge with no duplicates on 7Cdata Best = gt; 94% 95% 96% 97% 98% 99% 100% 0MaxYes 62 41 30 19 6 0 0 25MaxNo 100 100 99 96 92 84 57 25MaxYes 100 100 100 100 98 89 34 25MeanNo 100 100 100 98 93 77 56 50MaxNo 100 100 100 100 99 95 73 50MaxYes 100 100 100 100 99 98 92 50MeanNo 100 100 100 100 97 92 75 60MaxNo 100 100 100 100 99 94 76 60MeanNo 100 100 100 99 95 92 81 66MaxYes 100 100 100 100 97 95 84 The summary statistics for classifying the results by merge percentage in Table 7-17 support the conclusion that using a low merge rate of 25% is not as effective as using a rate of 50%, 60% or 66%.... In PAGE 85: ... 7-26 in Table 7-15 shows that the 25% merge runs are worse than using a higher level of merging. Table7 -14 Number of runs with best result gt;= to a given % of examples correctly classified for merge with no duplicates on 7Cdata Best = gt; 94% 95% 96% 97% 98% 99% 100% 0MaxYes 62 41 30 19 6 0 0 25MaxNo 100 100 99 96 92 84 57 25MaxYes 100 100 100 100 98 89 34 25MeanNo 100 100 100 98 93 77 56 50MaxNo 100 100 100 100 99 95 73 50MaxYes 100 100 100 100 99 98 92 50MeanNo 100 100 100 100 97 92 75 60MaxNo 100 100 100 100 99 94 76 60MeanNo 100 100 100 99 95 92 81 66MaxYes 100 100 100 100 97 95 84 The summary statistics for classifying the results by merge percentage in Table 7-17 support the conclusion that using a low merge rate of 25% is not as effective as using a rate of 50%, 60% or 66%. The type of run statistics in Table 7-18 show that using the mean value for the new rule strength has a p-value of only 0.... In PAGE 85: ... Table 7-14 Number of runs with best result gt;= to a given % of examples correctly classified for merge with no duplicates on 7Cdata Best = gt; 94% 95% 96% 97% 98% 99% 100% 0MaxYes 62 41 30 19 6 0 0 25MaxNo 100 100 99 96 92 84 57 25MaxYes 100 100 100 100 98 89 34 25MeanNo 100 100 100 98 93 77 56 50MaxNo 100 100 100 100 99 95 73 50MaxYes 100 100 100 100 99 98 92 50MeanNo 100 100 100 100 97 92 75 60MaxNo 100 100 100 100 99 94 76 60MeanNo 100 100 100 99 95 92 81 66MaxYes 100 100 100 100 97 95 84 The summary statistics for classifying the results by merge percentage in Table 7-17 support the conclusion that using a low merge rate of 25% is not as effective as using a rate of 50%, 60% or 66%. The type of run statistics in Table7 -18 show that using the mean value for the new rule strength has a p-value of only 0.031 and is therefore significantly different.... In PAGE 86: ... The results also improved on the test sets at very low levels of noise at the 1% and 2% level. At higher levels the original algorithm gave better performance Table7 -15 Mean and Standard deviation of best results for 7Cdata Best correct value Mean Std Deviation 0MaxYes 94.63 2.... In PAGE 86: ...86 0.4637 Table7 -16 probability values from a pair-wise comparison of means for the experiments on 7Cdata (note 0.000 represents a value of less than 0.... In PAGE 87: ... 7-28 Table7 -17 comparison of percent merged in 7Cdata (note 0.000 represents a value of less than 0.... In PAGE 87: ...780 0.730 Table7 -18 comparison of run types with 7Cdata (note 0.000 represents a value of less than 0.... ..."

### Table 7-12:

1994

"... In PAGE 60: ...Table7 -20showssignificantcorrelationsonlybetweenthevisualdimensionandrt andbetween thevisualdimensionandvisualworkloadreport.Theseshowedfor algorithms1,2, 5, and6, with algorithm6 recordingthehighestcorrelationsagain.... ..."

### Table 7.12: Features in combination experiments.

1997

Cited by 12

### Table 3: The resolvable 2-(14; 7; 12) designs of Type 1.

"... In PAGE 8: ... Figure 1), combined with an isomorphism check using the graph automorphism program nauty [6]. A total of 541,192 nonisomorphic resolvable 2-(14; 7; 12) designs of Type 1 is found in this search; the order of the automorphism groups of these are presented in Table3 . The CPU time employed in the search was about 500 hours on a 800 MHz PC computer.... In PAGE 8: ... The CPU time employed in the search was about 500 hours on a 800 MHz PC computer. The numbers in Table3 agree with the numbers obtained by counting the Type 1 designs in the complete classi cation, which gives an indepen- dent partial veri cation of the complete classi cation.... ..."

Cited by 1

### Tables 7 12 37

2000

Cited by 28

### Table 5.1: The separation in pixels (pixel edge length 7.12mm) between intensity peaks in the left and right walls of left ventricle measured on a 1 dimensional hor- izontal slice located centrally in the short axis view and in a similar slice in each patient.

1998

### Table 7.12: Count of faults isolated by chance in repl. 1

1995

Cited by 32

### Table 7.12: Count of faults isolated by chance in repl. 1

1995

Cited by 32