Academia.eduAcademia.edu

Figure 1 – uploaded by Farshid Keynia

See full PDFdownloadDownload figure
F — number of features, F-cat — number of categorical variables, SS — number of software size variables, Num — number of projects. THE SIX SELECTED DATASETS CONSISTING OF 792 SOFTWARE PROJECT CASES TABLE I

Table 1 F — number of features, F-cat — number of categorical variables, SS — number of software size variables, Num — number of projects. THE SIX SELECTED DATASETS CONSISTING OF 792 SOFTWARE PROJECT CASES TABLE I

Related Figures (17)

Fig. 1. Generalities of the proposed method.
Fig. 1. Generalities of the proposed method.
Let us suppose that x, is a member to cluster c;. In this case, CS index can be defined according to relations (6), (7) and (8) Conceptually, DB index can be considered as intending to minimise the intra-cluster distance and maximise the inter- cluster distance, which correspond to the two principles of cohesion inside the cluster and separation between the clusters, respectively. Relations (2), (3), (4) and (5) show the index calculation method.
Let us suppose that x, is a member to cluster c;. In this case, CS index can be defined according to relations (6), (7) and (8) Conceptually, DB index can be considered as intending to minimise the intra-cluster distance and maximise the inter- cluster distance, which correspond to the two principles of cohesion inside the cluster and separation between the clusters, respectively. Relations (2), (3), (4) and (5) show the index calculation method.
THE INITIAL SETTINGS AND THE VALUES OF PARAMETERS OF GENETIC ALGORITHMS TABLE III
THE INITIAL SETTINGS AND THE VALUES OF PARAMETERS OF GENETIC ALGORITHMS TABLE III
THE INITIAL SETTINGS AND THE VALUES OF PARAMETERS OF ICA
THE INITIAL SETTINGS AND THE VALUES OF PARAMETERS OF ICA
TABLE IV Fig. 2. Comparing methods (MMRE, MdMRE, PRED) on COCOMO dataset.
TABLE IV Fig. 2. Comparing methods (MMRE, MdMRE, PRED) on COCOMO dataset.
Fig. 3. Box plot for MRE in COCOMO dataset.
Fig. 3. Box plot for MRE in COCOMO dataset.
THE RESULTS ON ALBRECHT DATASET Fig. 4. Comparing methods (MMRE, MdMRE, PRED) on Albrecht dataset. TABLE V
THE RESULTS ON ALBRECHT DATASET Fig. 4. Comparing methods (MMRE, MdMRE, PRED) on Albrecht dataset. TABLE V
Fig. 6. Comparing methods (MMRE, MdMRE, PRED) on Desharnais dataset.
Fig. 6. Comparing methods (MMRE, MdMRE, PRED) on Desharnais dataset.
Fig. 5. Box plot for MRE in Albrecht dataset.
Fig. 5. Box plot for MRE in Albrecht dataset.
Fig. 7. Box plot for MRE in Desharnais dataset.
Fig. 7. Box plot for MRE in Desharnais dataset.
Fig. 8. Comparing methods (MMRE, MdMRE, PRED) on Maxwell dataset.
Fig. 8. Comparing methods (MMRE, MdMRE, PRED) on Maxwell dataset.
Fig. 9. Box plot for MRE in Maxwell dataset.
Fig. 9. Box plot for MRE in Maxwell dataset.
Fig. 12. Comparing methods (MMRE, MdMRE, PRED) on the China dataset. Fig. 11. Box plot for MRE in ISBSG dataset.
Fig. 12. Comparing methods (MMRE, MdMRE, PRED) on the China dataset. Fig. 11. Box plot for MRE in ISBSG dataset.
Fig. 10. Comparing methods (MMRE, MdMRE, PRED) on ISBSG dataset. THE RESULT ON THE CHINA DATASET
Fig. 10. Comparing methods (MMRE, MdMRE, PRED) on ISBSG dataset. THE RESULT ON THE CHINA DATASET
COMPARISON OF PERFORMANCE RATE OF THE PROPOSED MODEL TABLE X It can be stated based on Table X that the regression-based models have had better performance in comparison to that of
COMPARISON OF PERFORMANCE RATE OF THE PROPOSED MODEL TABLE X It can be stated based on Table X that the regression-based models have had better performance in comparison to that of
Fig. 13. Box plot for MRE in the China dataset. We OWN! MOadels ala We CVOlUUONALy aleOMUl MIOGCIS Nave shown the weakest performance. Moreover, the results obtained for each model largely depend on the used dataset. As ar example, PSO model has been proved strong for some datasets, weak for some others and yet intermediately performed for the third group of datasets. Although parameter setting plays an effective role in the performance of these algorithms, it can be observed for the proposed method that it has exhibited the best performance for all of the datasets. The issue is reflective of the reality that the proposed model is least dependent on the usec set of data. One factor largely influential on the model performance can be the number of the training samples because the models, particularly the proposed model, have displayed the best performance for the China dataset whose records are a lot higher than those of the other data collections. According to the fact that the proposed model makes use of two phases, i.e.. feature selection and clustering, two factors, namely the selection of the most effective features and localisation, have been able to corroborate the model performance. The enhancement rates are more distinct in regard of the China dataset; in addition, the localisation operation has been found having no significant effect on the improvement of the proposed model performance due to a low number of samples. Table X] presents the mean percentage of performance improvement of the proposed model for various scales on the datasets used herein. The results indicate that the efficiency of the majority of the models is undeniably dependent on the dataset used. Table X summarises the model efficiency rates according to three levels: good, medium and weak. The table indicators have been computed corresponding to the following algorithm:
Fig. 13. Box plot for MRE in the China dataset. We OWN! MOadels ala We CVOlUUONALy aleOMUl MIOGCIS Nave shown the weakest performance. Moreover, the results obtained for each model largely depend on the used dataset. As ar example, PSO model has been proved strong for some datasets, weak for some others and yet intermediately performed for the third group of datasets. Although parameter setting plays an effective role in the performance of these algorithms, it can be observed for the proposed method that it has exhibited the best performance for all of the datasets. The issue is reflective of the reality that the proposed model is least dependent on the usec set of data. One factor largely influential on the model performance can be the number of the training samples because the models, particularly the proposed model, have displayed the best performance for the China dataset whose records are a lot higher than those of the other data collections. According to the fact that the proposed model makes use of two phases, i.e.. feature selection and clustering, two factors, namely the selection of the most effective features and localisation, have been able to corroborate the model performance. The enhancement rates are more distinct in regard of the China dataset; in addition, the localisation operation has been found having no significant effect on the improvement of the proposed model performance due to a low number of samples. Table X] presents the mean percentage of performance improvement of the proposed model for various scales on the datasets used herein. The results indicate that the efficiency of the majority of the models is undeniably dependent on the dataset used. Table X summarises the model efficiency rates according to three levels: good, medium and weak. The table indicators have been computed corresponding to the following algorithm:
PERFORMANCE IMPROVEMENT OF THE PROPOSED MODEL According to Table XI, it can be stated that the highest improvement percentage has been documented for MMRE and MdMRE scales on the China dataset. One possible reason behind such an improvement can be the large number of the records in this set of data that, per se, causes the localisation phase of the proposed method exhibit more appropriate performance. The improvement results are also considerable for the other datasets. TABLE XI
PERFORMANCE IMPROVEMENT OF THE PROPOSED MODEL According to Table XI, it can be stated that the highest improvement percentage has been documented for MMRE and MdMRE scales on the China dataset. One possible reason behind such an improvement can be the large number of the records in this set of data that, per se, causes the localisation phase of the proposed method exhibit more appropriate performance. The improvement results are also considerable for the other datasets. TABLE XI
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved