### TABLE 1. Dimensions of re-authoring techniques.

### Table 9. Desirable qualities of RE techniques (McPhee and Eberlein 2002).

in ABSTRACT

"... In PAGE 6: ...able 8. Reasons for choosing RE techniques (McPhee and Eberlein 2002)................ 20 Table9 .... In PAGE 27: ...able 7. Top 10 RE techniques (McPhee and Eberlein 2002). Rank Familiarity Usefulness (TTM) Usefulness (Non-TTM) 1 Scenarios/Use Cases Requirements Prioritization Requirements Change Management 2 Semi-Formal Modeling Requirements Change Management Semi-Formal Modeling 3 Informal Modeling Scenarios/Use Cases Requirements Reviews 4 Requirements Change Management Semi-Formal Modeling Scenarios/Use Cases 5 Evolutionary Prototyping Requirements Testing Requirements Checklists 6 Interviews Evolutionary Prototyping Requirements Testing 7 Requirements Prioritization Requirements Reuse Requirements Tracing 8 Requirements Reviews Interviews Viewpoint-Oriented Techniques 9 Throw-away Prototyping Requirements Reviews Interviews 10 Requirements Checklists Requirements Checklists Designer as Apprentice The reasons for choosing RE techniques and their desirable qualities were also studied in the survey. These are shown in Table 8 and Table9 . Facilitation of good communication is the second most important item in both the lists so importance of good communication is evident.... ..."

### Table 1: Rules extracted from network \Iris-Bin quot; by BIO-RE technique.

"... In PAGE 16: ... Also, it shows the remarkable performance of the rules extracted from network \Iris-Cont quot; by Full-RE. Note that: (i) The numeric values compared with input features Iis in the rules extracted by both BIO-RE and Partial- RE represent the mean ( is) of these input feature (see the rule bodies in Table1 and Table 2). This coarse thresholding is largely responsible for the (relatively) poor performance of the two networks and subsequently of the extracted rules.... In PAGE 21: ... Also, they were extracted to cover all training and testing data sets and hence increase the completeness of the extracted set of rules. Examples of such rules are: R1 and R4 of Table1 , R4 of Table 2, and R2-R5 of Table 5. 5 Comparative Performance Evaluation Since both iris and breast cancer problems have continuous input features, Full-RE is naturally suited for them.... In PAGE 25: ...CONCLUSIONS 25 Table1 0: A qualitative comparison of di erent rule extraction techniques. BIO-RE Partial-RE Full-RE NeuroRule KT or Subset MofN Provides CF No Yes Yes No No No May need a default rule Yes Yes No Yes Yes Yes Works for 1.... ..."

### Table 1: A Comparison Among Dimensionality Re- duction Techniques

"... In PAGE 1: ... The proposed dimension- ality reduction techniques include Singular Value Decompo- sition (SVD) [13, 18], Discrete Fourier Transform (DFT) [24], Discrete Wavelet Transform (DWT) [5, 23, 12, 27], Piecewise Linear Approximation (PLA) [20, 17], Piecewise Aggregate Approximation (PAA) [15, 29], Adaptive Piece- wise Constant Approximation (APCA) [16] and Chebyshev Polynomials (CP) [4]. We list seven popular dimensionality reduction techniques in Table1 , in terms of the time complexity, space complex- ity, and capability to be indexed in the reduced space, where n is the length of each time series, N is the total number of time series in the database, and (2m) is the reduced dimen- sionality. In terms of the time complexity, CP is more costly than PLA, DWT, PAA and APCA (Table 1); and PLA is... In PAGE 1: ... We list seven popular dimensionality reduction techniques in Table 1, in terms of the time complexity, space complex- ity, and capability to be indexed in the reduced space, where n is the length of each time series, N is the total number of time series in the database, and (2m) is the reduced dimen- sionality. In terms of the time complexity, CP is more costly than PLA, DWT, PAA and APCA ( Table1 ); and PLA is... ..."

### Table 1: A Comparison Among Dimensionality Re- duction Techniques

"... In PAGE 1: ... The proposed dimension- ality reduction techniques include Singular Value Decompo- sition (SVD) [13, 18], Discrete Fourier Transform (DFT) [24], Discrete Wavelet Transform (DWT) [5, 23, 12, 27], Piecewise Linear Approximation (PLA) [20, 17], Piecewise Aggregate Approximation (PAA) [15, 29], Adaptive Piece- wise Constant Approximation (APCA) [16] and Chebyshev Polynomials (CP) [4]. We list seven popular dimensionality reduction techniques in Table1 , in terms of the time complexity, space complex- ity, and capability to be indexed in the reduced space, where n is the length of each time series, N is the total number of time series in the database, and (2m) is the reduced dimen- sionality. In terms of the time complexity, CP is more costly... In PAGE 1: ... We list seven popular dimensionality reduction techniques in Table 1, in terms of the time complexity, space complex- ity, and capability to be indexed in the reduced space, where n is the length of each time series, N is the total number of time series in the database, and (2m) is the reduced dimen- sionality. In terms of the time complexity, CP is more costly than PLA, DWT, PAA and APCA ( Table1... ..."

### Table 5: Results for uniform grid and multiresolution data re- duction techniques

2001

"... In PAGE 7: ... code automatically determined the best decomposition to min- imize one of the two different error criteria defined in Section 2: the standard deviation, n, in the leaf octants, and the maxi- mum deviation, en, in the leaf octants. In Table5 , we report the the average n and maximum en over all octants for each case. The first value gives a measure of the overall fidelity of the re- duced data set to the original data set; the latter value gives a worst-case measure of fidelity.... ..."

Cited by 3

### Table 2. Rules extracted by Partial-RE technique. R# Rule Body Class CF Correct Mis red

1996

"... In PAGE 3: ...e a hidden unit) is replaced by the corresponding set of conjuncted input features that causes it to be active. Final rules are written in this format: \If Xi i And Xg g cf ?! Consequentj quot; (see Table2 for examples). Experimental results show that Partial-RE algorithm is suitable for large size problems, since extracting all possible rules is NP-hard and extracting only the most e ective rules is a practical alternative.... ..."

Cited by 8

### Table 3. Rules extracted by Full-RE technique. R# Rule Body Class CF Times Correct Times Mis red

1996

"... In PAGE 4: ... In each rule extracted between a hidden and output node, Full-RE replaces each hidden node (hj) by the left hand side of the rule(s) whose right hand side is hj. Final rules extracted by Full-RE are represented in the same format of Partial-RE expect that each i is replaced by one of the discretization boundaries (say di;l) selected by Full-RE (see Table3 ). Note that there is no restriction on the number of premises in the nal rules extracted by the Full-RE.... ..."

Cited by 8

### Table 1: Rules extracted from network \Cancer-Bin quot; by BIO-RE technique. Cancer Performance Measures

1996

"... In PAGE 13: ... 5.3 Rule Extraction and Ordering Results Table1 , 2, and 3 present three sets of ordered rules extracted by the three rule extraction techniques, along with the corresponding performance measures. Table 4 provides an overall comparison between the extracted rules and their corresponding trained networks.... In PAGE 15: ... This is due to the fact that the original input features of the breast cancer problem have the same range (1,10). Note that: The numeric values compared with input features Xis in the rules extracted by both BIO- RE and Partial-RE represent the mean ( is) of these input feature (see the rule bodies in Table1 and Table 2). Since the original inputs of the breast cancer problem are all normalized integer values between (1,10) therefore, binarizing and discretizing them did not degrade the performance of neither BIO-RE nor Partial-RE.... In PAGE 15: ... Also, they were extracted to cover all training and testing data sets and hence increase the completeness of the extracted set of rules. Examples of such rules are: R8 and R10 of Table1... In PAGE 31: ... An important issue that needs to be addressed here is: \Should one discard rules with low soundness, completeness and/or high false-alarm measure(s)? quot;. An example of such a rule is R10 in Table1 . For small data sets we might still retain such rules at the bottom of the application 7IF 9 more than one rule with zero false-alarm THEN select the one with the highest completeness measure... ..."

Cited by 4

### Table 3. Rules extracted from network \Cancer-Cont quot; by Full-RE technique. Cancer Performance Measures

1997

"... In PAGE 2: ... Final rules extracted by Full-RE are represented in the same format of Partial-RE expect that each i is replaced by one of the discretization boundaries (say di;l) selected by Full-RE as described earlier. See Table3 for examples. 3 Rule Evaluation and Ordering Proce- dure To evaluate the performance of rules extracted from trained networks by any of the three presented tech- niques (or by any other rule extraction approach), a simple rule evaluation procedure which attaches three performance measures to each extracted rule is devel- oped.... ..."

Cited by 13