In this paper, we propose two weighted learning methods for the construction of single hidden layer feedforward neural networks. Both methods incorporate weighted least squares. Our idea is to allow the training insta...In this paper, we propose two weighted learning methods for the construction of single hidden layer feedforward neural networks. Both methods incorporate weighted least squares. Our idea is to allow the training instances nearer to the query to offer bigger contributions to the estimated output. By minimizing the weighted mean square error function, optimal networks can be obtained. The results of a number of experiments demonstrate the effectiveness of our proposed methods.展开更多
Software defect prediction （SDP） is an active research field in software engineering to identify defect-prone modules. Thanks to SDP, limited testing resources can be effectively allocated to defect-prone modules. A...Software defect prediction （SDP） is an active research field in software engineering to identify defect-prone modules. Thanks to SDP, limited testing resources can be effectively allocated to defect-prone modules. Although SDP requires sufficient local data within a company, there are cases where local data are not available, e.g., pilot projects. Companies without local data can employ cross-project defect prediction （CPDP） using external data to build classifiers. The major challenge of CPDP is different distributions between training and test data. To tackle this, instances of source data similar to target data are selected to build classifiers. Software datasets have a class imbalance problem meaning the ratio of defective class to clean class is far low. It usually lowers the performance of classifiers. We propose a Hybrid Instance Selection Using Nearest-Neighbor （HISNN） method that performs a hybrid classification selectively learning local knowledge （via k-nearest neighbor） and global knowledge （via na/ve Bayes）. Instances having strong local knowledge are identified via nearest-neighbors with the same class label. Previous studies showed low PD （probability of detection） or high PF （probability of false alarm） which is impractical to overall performance as well as high PD and low PF. use. The experimental results show that HISNN produces high overall performance as well as high PD and low PF.展开更多
Personal credit risk assessment is an important part of the development of financial enterprises. Big data credit investigation is an inevitable trend of personal credit risk assessment, but some data are missing and ...Personal credit risk assessment is an important part of the development of financial enterprises. Big data credit investigation is an inevitable trend of personal credit risk assessment, but some data are missing and the amount of data is small, so it is difficult to train. At the same time, for different financial platforms, we need to use different models to train according to the characteristics of the current samples, which is time-consuming. <span style="font-family:Verdana;">In view of</span><span style="font-family:Verdana;"> these two problems, this paper uses the idea of transfer learning to build a transferable personal credit risk model based on Instance-based Transfer Learning (Instance-based TL). The model balances the weight of the samples in the source domain, and migrates the existing large dataset samples to the target domain of small samples, and finds out the commonness between them. At the same time, we have done a lot of experiments on the selection of base learners, including traditional machine learning algorithms and ensemble learning algorithms, such as decision tree, logistic regression, </span><span style="font-family:Verdana;">xgboost</span> <span style="font-family:Verdana;">and</span><span style="font-family:Verdana;"> so on. The datasets are from P2P platform and bank, the results show that the AUC value of Instance-based TL is 24% higher than that of the traditional machine learning model, which fully proves that the model in this paper has good application value. The model’s evaluation uses AUC, prediction, recall, F1. These criteria prove that this model has good application value from many aspects. At present, we are trying to apply this model to more fields to improve the robustness and applicability of the model;on the other hand, we are trying to do more in-depth research on domain adaptation to enrich the model.</span>展开更多
基金supported by the NSC under Grant No.NSC-100-2221-E-110-083-MY3 and NSC-101-2622-E-110-011-CC3"Aim for the Top University Plan"of the National Sun-Yat-Sen University and Ministry of Education
文摘In this paper, we propose two weighted learning methods for the construction of single hidden layer feedforward neural networks. Both methods incorporate weighted least squares. Our idea is to allow the training instances nearer to the query to offer bigger contributions to the estimated output. By minimizing the weighted mean square error function, optimal networks can be obtained. The results of a number of experiments demonstrate the effectiveness of our proposed methods.
文摘Software defect prediction （SDP） is an active research field in software engineering to identify defect-prone modules. Thanks to SDP, limited testing resources can be effectively allocated to defect-prone modules. Although SDP requires sufficient local data within a company, there are cases where local data are not available, e.g., pilot projects. Companies without local data can employ cross-project defect prediction （CPDP） using external data to build classifiers. The major challenge of CPDP is different distributions between training and test data. To tackle this, instances of source data similar to target data are selected to build classifiers. Software datasets have a class imbalance problem meaning the ratio of defective class to clean class is far low. It usually lowers the performance of classifiers. We propose a Hybrid Instance Selection Using Nearest-Neighbor （HISNN） method that performs a hybrid classification selectively learning local knowledge （via k-nearest neighbor） and global knowledge （via na/ve Bayes）. Instances having strong local knowledge are identified via nearest-neighbors with the same class label. Previous studies showed low PD （probability of detection） or high PF （probability of false alarm） which is impractical to overall performance as well as high PD and low PF. use. The experimental results show that HISNN produces high overall performance as well as high PD and low PF.
文摘Personal credit risk assessment is an important part of the development of financial enterprises. Big data credit investigation is an inevitable trend of personal credit risk assessment, but some data are missing and the amount of data is small, so it is difficult to train. At the same time, for different financial platforms, we need to use different models to train according to the characteristics of the current samples, which is time-consuming. <span style="font-family:Verdana;">In view of</span><span style="font-family:Verdana;"> these two problems, this paper uses the idea of transfer learning to build a transferable personal credit risk model based on Instance-based Transfer Learning (Instance-based TL). The model balances the weight of the samples in the source domain, and migrates the existing large dataset samples to the target domain of small samples, and finds out the commonness between them. At the same time, we have done a lot of experiments on the selection of base learners, including traditional machine learning algorithms and ensemble learning algorithms, such as decision tree, logistic regression, </span><span style="font-family:Verdana;">xgboost</span> <span style="font-family:Verdana;">and</span><span style="font-family:Verdana;"> so on. The datasets are from P2P platform and bank, the results show that the AUC value of Instance-based TL is 24% higher than that of the traditional machine learning model, which fully proves that the model in this paper has good application value. The model’s evaluation uses AUC, prediction, recall, F1. These criteria prove that this model has good application value from many aspects. At present, we are trying to apply this model to more fields to improve the robustness and applicability of the model;on the other hand, we are trying to do more in-depth research on domain adaptation to enrich the model.</span>
