KR20230122370A - Method and system for predicting heterogeneous defect through correlation-based selection of multiple source projects and ensemble learning - Google Patents

Abstract

A method and system for predicting heterogeneous defects through correlation-based multiple learning project selection and ensemble learning are disclosed. A heterogeneous defect prediction method performed by a defect prediction system according to an embodiment includes selecting at least one source project for ensemble learning by analyzing a correlation between a plurality of source projects and a target project; and predicting heterogeneous defects of the target project through ensemble learning of the selected at least one or more source projects.

Description

Heterogeneous defect prediction method and system through correlation-based multiple learning project selection and ensemble learning

The description below relates to software defect prediction techniques.

Software defect prediction (SDP) has been studied extensively for decades and has now become one of the major topics in software engineering. Software defect prediction aims to predict fault-prone modules (functions, classes or files depending on its granularity), and helps software engineers reduce testing effort and improve software quality.

Recently, heterogeneous defect prediction (HDP) has been proposed to predict defects even if the two projects have heterogeneous sets of metrics. Several heterogeneous defect prediction studies focus on matching metrics to improve prediction performance. This heterogeneous defect prediction method requires one source project data and one target project data as input and returns the predicted target label as output.

There are many open source project data available for model training. Therefore, in order to apply the heterogeneous defect prediction method, the developer must select an appropriate source project whose data distribution is similar to that of the target project. Nonetheless, most studies did not propose guidelines for the selection of source projects, and averaged the evaluation criteria assuming that source projects were chosen randomly.

Furthermore, the defect data is generally highly disproportionate, meaning that there are far fewer defective modules than non-faulty modules. The class imbalance problem has a significant negative impact on the performance of defect prediction models, but some techniques do not consider this problem.

As the correlation analysis is performed, a method and system for selecting at least one or more source projects for ensemble learning may be provided.

It is possible to provide a method and system for predicting heterogeneous defects in a target project through ensemble learning to solve the class imbalance problem.

The heterogeneous defect prediction method performed by the defect prediction system includes: selecting at least one source project for ensemble learning by analyzing a correlation between a plurality of source projects and a target project; and predicting heterogeneous defects of the target project through ensemble learning of the selected at least one or more source projects.

The heterogeneous defect prediction method may further include performing metric matching between a plurality of source projects and a target project.

The performing of the metric matching includes metrics of the plurality of source projects and metrics of the target project using the Kolmogorov-Smirnov test to integrate heterogeneous metric sets for the plurality of source projects and the target project. It may include measuring the similarity of the pair of metrics.

The performing of the metric matching may include selecting a metric of the source project equal to or greater than a preset standard according to a gain ratio of the source project.

The performing of the metric matching may include calculating a p-value of the Kolmogorov-Smirnov test for the metric pair, and filtering the metric pair if the calculated p-value is less than a cutoff threshold. can do.

The performing of the metric matching may include selecting a metric pair having the highest p-value by applying maximum weight bipartite matching to remaining metric pairs through filtering of the metric pair.

The selecting may include calculating a suitability score between source data for the plurality of source projects and target data for the target project as metric matching between the plurality of source projects and the target project is performed. can do.

In the selecting step, if matching between the plurality of source projects and the target project is successful, the ratio of the correlation score between the plurality of source projects and the target project to the number of modules between the plurality of source projects and the target project. It may include calculating a suitability score consisting of

The selecting may include deriving a suitability score of 0 when matching between the plurality of source projects and the target project fails.

The selecting step may include randomly sampling a module from source data for the plurality of source projects and source data for the target data, respectively, and source data for the plurality of source projects and source data for the target data. It may include a step of sorting.

The selecting may include calculating a correlation score using a Spearman correlation coefficient between source data for the plurality of source projects and target data for the target data as the sampling and sorting are completed, and the calculated correlation score and setting the p-value of to 0 when it is greater than a cutoff threshold, and deriving an average of correlation scores calculated for each of the metrics performed on the plurality of source projects and the target project.

The predicting of the heterogeneous defects may include constructing a logistic regression-based predictive model for each of the selected at least one or more source projects, and using the constructed logistic regression-based predictive model to predict a defect label of the target project steps may be included.

The predicting of the heterogeneous defects may include returning probabilities predicted from each logistic regression-based predictive model through soft voting, and determining an average of the returned probabilities as a final predicted label. can

A computer program stored in a computer readable storage medium to execute the heterogeneous defect prediction method performed by the defect prediction system analyzes the correlation between a plurality of source projects and a target project, thereby generating at least one source project for ensemble learning. choosing; and predicting heterogeneous defects of the target project through ensemble learning of the selected at least one or more source projects.

It may include a computer program stored in a non-transitory computer readable recording medium to execute the heterogeneous defect prediction method in the defect prediction system.

The defect prediction system includes: a correlation analyzer for selecting at least one source project as the correlation between a plurality of source projects and a target project is analyzed; and an ensemble learning unit that predicts heterogeneous defects of the target project through ensemble learning of the selected at least one or more source projects.

Guidelines can be provided for selecting source projects within heterogeneous defect prediction that do not have target data for training.

The class imbalance problem can be mitigated and heterogeneous defect prediction performance can be improved by selecting a source project suitable for the target project through correlation-based and ensemble learning methods.

Cost can be reduced by inspecting and testing modules that are expected to have defects in the software under development by predicting defects in advance by applying it to a newly developed project. In addition, since more testing costs can be concentrated on the part where defects are expected to exist, the quality of the software can be improved.

1 is a block diagram for explaining the configuration of a defect prediction system according to an embodiment.
2 is a flowchart illustrating a heterogeneous defect prediction method in a defect prediction system according to an embodiment.
3 is a diagram for explaining a heterogeneous defect prediction operation according to an embodiment.
4 is a diagram illustrating pseudocode of a heterogeneous defect prediction operation according to an embodiment.
5 is a diagram showing pseudocodes of suitability scores according to an embodiment.

Hereinafter, an embodiment will be described in detail with reference to the accompanying drawings.

In the embodiment, a novel heterogeneous defect prediction operation that further improves performance by selecting an appropriate source project for a given target project and considering imbalanced data of selecting a plurality of source projects based on correlation will be described.

1 is a block diagram illustrating a configuration of a defect prediction system according to an embodiment, and FIG. 2 is a flowchart illustrating a heterogeneous defect prediction method in a defect prediction system according to an embodiment.

The processor of the defect prediction system 100 may include a correlation analysis unit 110 and an ensemble learning unit 120 . Components of such a processor may be representations of different functions performed by the processor according to control instructions provided by program codes stored in the defect prediction system. The processor and components of the processor may control the defect prediction system to perform steps 210 to 220 included in the heterogeneous defect prediction method of FIG. 2 . In this case, the processor and components of the processor may be implemented to execute instructions according to the code of an operating system included in the memory and the code of at least one program.

The processor may load the program code stored in the file of the program for the heterogeneous defect prediction method into the memory. For example, when a program is executed in the defect prediction system, a processor may control the defect prediction system to load a program code from a program file into a memory under control of an operating system. At this time, each of the correlation analysis unit 110 and the ensemble learning unit 120 executes a command of a corresponding part of the program code loaded into the memory to perform different functional functions of the processor for executing the subsequent steps 210 to 220. can be expressions.

In advance, metric matching between a plurality of source projects and a target project may be performed by a metric matching unit (not shown). The metric matching unit may measure the similarity of metric pairs including metrics of the plurality of source projects and metrics of the target project by using the Kolmogorov-Smirnov test to integrate heterogeneous metric sets of the plurality of source projects and the target project. The metric matching unit may select a metric of a source project equal to or greater than a preset standard according to a gain ratio of the source project. The metric matcher may calculate a p-value of the Kolmogorov-Smirnov test for the metric pair, and filter the metric pair if the calculated p-value is smaller than a cutoff threshold. The metric matching unit may select a metric pair having the highest p-value by applying maximum weight bipartite matching to remaining metric pairs through metric pair filtering.

In step 210, the correlation analysis unit 110 may select at least one source project by analyzing the correlation between a plurality of source projects and a target project. The correlation analyzer 110 may calculate a suitability score between the source data of the plurality of source projects and the target data of the target project as metric matching is performed between the plurality of source projects and the target project. When the matching between the plurality of source projects and the target project is successful, the correlation analysis unit 110 determines the suitability consisting of the correlation score between the plurality of source projects and the target project and the ratio of the number of modules between the plurality of source projects and the target project. score can be calculated. The correlation analysis unit 110 may derive a suitability score of 0 when matching between a plurality of source projects and a target project fails. The correlation analysis unit 110 randomly samples a module from each of the source data for the source data and the target data for a plurality of source projects, and sorts the source data for the plurality of source projects and the source data for the target data. can do. As the sampling and sorting are completed, the correlation analysis unit 110 calculates a correlation score using the Spearman correlation coefficient between source data for a plurality of source projects and target data for target data, and calculates p- of the calculated correlation score. If the value is greater than the cutoff threshold, it may be set to 0, and an average of correlation scores calculated for each of the metrics performed on a plurality of source projects and target projects may be derived.

In step 220, the ensemble learning unit 120 may predict heterogeneous defects of the target project through ensemble learning of at least one selected source project. The ensemble learning unit 120 may build a logistic regression-based prediction model for each of the selected at least one or more source projects, and predict a defect label of the target project using the logistic regression-based prediction model. The ensemble learning unit 120 may return probabilities predicted from each logistic regression-based prediction model through soft voting, and may determine an average of the returned probabilities as a final predicted label.

3 is a diagram for explaining a heterogeneous defect prediction operation according to an embodiment.

The defect prediction system may perform metric matching through the Kolmogorov-Smirnov test (310). The defect prediction system may utilize the Kolmogorov-Smirnov test 310 to integrate heterogeneous metric sets between the source project and the target project. For example, 15% of the source project's metrics are chosen as the gain ratio score, and the metrics between the source and target projects are p- You can match based on value. If the p-value for which all metric matching has been performed is less than a cutoff threshold (eg, 0.05), the source project may be filtered out.

More specifically, if the source project is a project including 34 metrics, the top 6 metrics ranked according to the gain ratio may be selected. The similarity of all pairs of metrics can be measured using the Kolmogorov-Smirnov test 310 to integrate the metric sets of the two different projects (the metric sets of the source project and the target project). As an example, 6 (after metric selection) and 29 metrics may be included in the project including the source project and the target project, respectively. The p-values of the Kolmogorov-Smirnov test (310) for the possible 174 metric pairs have to be calculated. At this time, if the p-value is smaller than the cutoff threshold (eg, 0.05), the match is discarded and not considered. Then, the matching metric set with the highest sum of scores may be selected by applying maximum weight bipartite matching. For example, it is assumed that there are two source metrics (S ₁ and S ₂ ) and two target metrics (T ₁ and T ₂ ). Kolmogorov-Smirnov test for metric pairs (S ₁ , T ₁ ), (S ₁ , T ₂ ), (S ₂ , T ₁ ), (S ₂ , T ₂ ) because the score sum is the highest at 1.3 (310) When the p-values of are 0.9, 0.7, 0.6, and 0.1, respectively, the (S1, T2) and (S2, T1) metric pairs can be selected.

The defect prediction system may select at least one or more source projects from a plurality of source projects through correlation analysis 320 . The defect prediction system can calculate a suitability score for each source project to select a source project for ensemble learning. For example, a higher suitability score means that the source project to learn from is appropriate. Learning may be performed by selecting at least one source project having the highest suitability score. The number of selected source projects may vary depending on the selection ratio, and the default value may be 0.6.

More specifically, in order to calculate a suitability score between source data for a plurality of source projects and target data for a target project, the suitability score may be defined as follows. Referring to FIG. 4, pseudocode of a heterogeneous defect prediction operation is shown, and referring to FIG. 5, an example of pseudocode for suitability score is shown.

Equation 1:

where X _S and X _T denote integrated source data and target data, respectively, and module(X) denotes the number of modules of data X, am. These formulas are applied only if the metric match succeeds (lines 3-5), and if the metric match fails, the suitability score is zero (line 6). The suitability score may include two terms including a correlation score between the plurality of source projects and the target project and a ratio of the number of modules between the plurality of source projects and the target project. At this time, through the experiment may be set to 0.95 by default.

A correlation score between two-dimensional data (integrated source data and target data) X _S and X _T can be calculated by the corScore function provided in the pseudo code of FIG. 5 . The correlation coefficient can be calculated between one-dimensional data with the same number of instances. So, for each metric data, corScore can align the two data by randomly sampling a module from the data with more modules to match (lines 1-9). As the sampling and alignment procedure is completed, a correlation score can be obtained (line 10). A correlation score can be calculated using Spearman's correlation coefficient. If the p-value of the result of calculating the correlation score is greater than a preset value (eg, 0.05), it means that the correlation is not statistically significant. Accordingly, the correlation score may be set to 0 (lines 11-13). After correlation scores for all metrics have been calculated, an average can be derived (lines 14-16). This entire process can be performed repeatedly (line 17). In other words, the entire process can be repeated 100 times.

The defect prediction system may predict heterogeneous defects of a target project through ensemble learning 330 for at least one selected source project. Ensemble learning is used to improve the existing method of class imbalance learning. Ensemble classifiers have stronger generalization ability and better insights into unseen data than the basic learners by combining the prediction results of the basic learners. For example, multi-kernel classifiers can be integrated with ensemble learning to improve prediction performance by mitigating the effects of class imbalance learning. The defect prediction system can utilize ensemble learning to mitigate the class imbalance problem, train a predictive model using a logistic regression classifier for each selected source project, and predict defect labels for the target project. The results (predicted defect labels) can then be combined to reduce the misclassification of defect modules via soft voting classifiers.

More specifically, logistic regression is the most widely used classifier in heterogeneous defect prediction. Referring to FIG. 4 , a logistic regression-based predictive model for each selected at least one or more source projects may be built, and defect labels of the target project may be predicted (lines 13 to 15). Voting can be applied to aggregate prediction results and mitigate class imbalance problems. Generally, there are two voting strategies including hard voting and soft voting. In hard voting, the labels predicted by each basic classifier are voted on, and the majority of labels becomes the final result. In soft voting, predicted probabilities from each basic classifier are returned, and the average of the returned probabilities determines the final result. In the embodiment, the final predicted label L (row 16) through soft voting may be defined as follows.

Here, N' is the number of selected source projects, and Z _i represents the predicted probability of a single source project.

According to an embodiment, a source project suitable for a given target project may be selected and the class imbalance problem may be mitigated through an ensemble learning method for correlation-based selection of a plurality of source projects and prediction of heterogeneous defects. As an example, experiments on 37 projects can be performed, and the Wilcoxon signed-rank test and Cliff's Effect size tests may be performed. According to the experimental results, the proposed method in the embodiment shows much better performance than the conventional heterogeneous defect prediction method without selecting an appropriate source project, the performance of both correlation-based selection and ensemble learning approaches is improved, and the learning project It can be seen that selecting a plurality of source projects with 0.6 produces the optimal result. In conclusion, even without target data for learning, heterogeneous defect prediction provides selection guidelines for source projects to improve prediction performance and apply them in the early stages of a project. In other words, we can confirm that choosing an appropriate number of source data is important for predictive performance.

More specifically, three questions may be set for the experiment. Question 1 (does the method proposed in the embodiment achieve better prediction performance than the existing heterogeneous defect prediction methods?), question 2 (two key methods of the method proposed in the embodiment (correlation-based multi-source selection and ensemble How effective is learning) for predictive performance?), Question 3 (What are the appropriate parameter values (i.e., alpha of the fit function, correlation coefficient, and selection ratio) for the method proposed in the example?), three questions are set. It can be.

Additionally, publicly available datasets widely used for heterogeneous fault prediction can be collected, including NASA, SOFTLAB, ReLink, AIEEM, and PROMISE. The NASA data set contains metrics for various codes of NASA software systems. NASA's static code metrics can include size and complexity. Therefore, only 5 project data such as CM1, MW1, PC1, PC3, and PC4 with 37 common metrics will be used. The SOFTLAB data set is controller software for white goods. The SOFTLAB data set consists of 29 metrics, including Halstead and McCabe's circulant matrices. The ReLink data set was collected to obtain high-quality defect data and improve predictive performance, and includes 26 metrics, including file-level static code complexity measures such as lines of code, cyclical complexity, and more. The AEEEM data set consists of change, bug, and version information for five systems. AEEEM data contains 61 metrics, including source code metric, entropy of change metric, conversion of source code metric, etc. PROMISE data sets can be composed of open source projects such as ant, camel, and ivy. The PROMISE data set contains 20 metrics including CK metrics and OO metrics. To mitigate the threat of external validity, we decide to additionally use the JIRA dataset with 65 metrics for the experiment. A total of 37 projects were used in 6 data sets, and these projects are described in Table 1. Each of these data sets has a different number of metrics.

Table 1:

In addition, the method proposed in the embodiment can be compared with four other heterogeneous defect prediction methods of HDP-KS, EMKCA, CTKCCA, and MSMDA. An empirical study on heterogeneous defect prediction can be evaluated that CTKCCA is best performed considering linear indivisibility and class imbalance problems, followed by EMKCA. Among the metric selection-based approaches, HDP-KS showed the highest performance. HDP-KS, EMKCA, and CTKCCA measured the mean or median of the evaluation criteria assuming that source projects were randomly selected. MSMDA as a mixed project model requiring target data for training is a multi-source selection method based on manifold discriminant sorting.

Two commonly used performance metrics can be used to evaluate these heterogeneous defect prediction techniques: AUC for effort-unrecognized performance metric (also called NPI) and ACC for effort-perceived performance metric (EPI).

In software fault prediction, one can generally treat faulty modules as positive instances and non-faulty modules as negative instances. Depending on the actual flaws and prediction flaws of the module, prediction results can be classified into four types: true positive (tp), false positive (fp), false negative (fn), and true negative (tn). The error matrix in software defect prediction is described in Table 2.

Table 2:

AUC represents the area under the receiver operating characteristic curve. The curve plots the false positive rate (probability of false alarm, also called pf) as the x-coordinate and the true positive rate (probability of detection, also called pd) as the y-coordinate. pf and pd are defined as fp/(fp+tn) and tp/(tp+fn), respectively. Threshold-dependent measures such as precision and recall can be affected by setting cutoff thresholds and class imbalance issues. On the other hand, AUC is widely used in defect prediction studies as it is threshold independent and less sensitive to class imbalance problems. AUC falls between 0 and 1, with higher AUC indicating better performance. A complete model has an AUC of 1.0 and random prediction yields an AUC of 0.5.

Unlike NPIs such as AUC, EPI considers effort to examine program modules. Most existing HDP studies have generally evaluated performance only in NPI, but this can lead to biased results. Therefore, we further consider ACC for EPI used for just-in-time (JIT) defect prediction.

ACC represents the recall of a faulty module when it expends 20% of its total effort. We will use the number of modules tested to measure effort. For example, a project may have 1,000 modules, and a bug may occur in 100 of them. When 20% of the total effort is spent, the 200 modules with the highest prediction scores can be tested. Let's assume that 35 of the 200 checked modules have bugs. Then 35 bad modules out of 100 can be found. So ACC is 35/100 = 0.35.

To solve question 1, the predictive performance of the method proposed in the embodiment can be compared with four criteria. HDP-KS, EMKCA, and CTKCCA are one-to-one HDPs, meaning they require one source project and one target project as input. Therefore, you can select one of the 37 projects as a target project and use each project in a different data set as a source project. For example, if CM1 is the target project, there are 32 prediction combinations excluding 5 projects in the NASA dataset (see Table 1). In total, EMKCA and CTKCCA have 1,092 possible prediction combinations. HDP-KS tries the same 1,092 combinations, but only 797 combinations are valid because 295 combinations fail to match the metric. The method proposed in MSMDA and the embodiment is a many-to-one heterogeneous defect prediction, in which several source projects can be selected and learned for each target project. Thus, there are 37 possible prediction combinations in MSMDA and the method proposed in the examples. Unlike the other four methods, MSMDA is a mixed project heterogeneous defect prediction that uses 10% of the labeled target data for training, so the average can be obtained after 30 iterations of the MSMDA process. All methods can use LIBLINEAR's logistic regression classifier with default parameters for fair comparison. Thus, reference MATLAB code can be utilized. However, we found a serious bug in the preprocessing of target data by label information. That is, all defective modules are placed first, and non-defective modules come after. Obviously, such a process is problematic because the label information of the target data can only be accessed by evaluating the predicted outcome. In experiments, the code can be modified to remove the sorting process that can degrade the performance of the baseline.

In order to solve question 2, some processes may be omitted from the method proposed in the embodiment and two new methods for comparing results may be created. First, HDP-KS-EL (HDPKS with ensemble learning) can apply ensemble learning to all prediction results without going through a correlation-based selection process. Second, CASS (Correlation-based Single Source Selection) can select one source project by a correlation-based selection process without applying ensemble learning. CASS can select only one project with the highest suitability score and train the model with the selected project. In the embodiment, AUC and ACC of the three methods of HDP-KS-EL, CASS, and the method proposed in the embodiment (CAMEL) can be compared. Table 3 summarizes whether each method applies a selection approach and an ensemble learning approach.

Table 3:

For questions 1 and 2, statistical tests may be performed. A non-parametric Wilcoxon signed-rank test can be performed at the 95% confidence level to assess statistical significance. The Wilcoxon single rank test is used to compare two related samples and the results between the method proposed in the examples (CAMEL) and each baseline can be compared. In addition, to measure the size of the difference in performance between the method (CAMEL) proposed in the embodiment and each reference value, Cliff's non-parametric effect size test can be calculated. Cliff's belongs to [-1, 1], and Table 4 shows various Indicates the efficiency level for the value.

Table 4:

To solve question 3, we want to find the optimal alpha, correlation coefficient and selection ratio. It has been found that it is necessary to determine the parameters empirically (e.g. Gaussian kernel parameters of EMKCA) as the performance of most HDP approaches is affected by parameter settings. First, the optimal alpha value can be found in the fit function. from 0.05 to 1.00 Values are used at intervals of 0.05. Considering that the difference in results between 0.95 and 1.00 is large, we try an additional value of 0.99. As a result, we can calculate the AUC value with 21 different alpha values and find the best one. Note that when is equal to 1.00, the fit function includes only the correlation score between the two projects and not the ratio of the number of modules. After finding the optimal alpha value, you can try three types of correlation coefficients and nine selection ratios. To calculate the correlation score between matching metrics, Pearson's correlation coefficient, Spearman's rank correlation coefficient, and Kendall's rank correlation coefficient are used, respectively. , Spearman's r and Kendall's It is said. Coefficients can be computed using Python's Scipy module. To select multiple source projects, use a selection ratio of 0.1 to 0.9 with an interval of 0.1. That is, according to the selection ratio, the matching source project with the highest suitability score can be selected as 10%, 20%, ??, 80%, or 90%. Therefore, it is possible to calculate the AUC value with 27 parameter settings and find the best value.

The devices described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA) , a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. The device can be commanded. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. can be embodied in Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler.

As described above, although the embodiments have been described with limited examples and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

Claims (15)

In the heterogeneous defect prediction method performed by the defect prediction system,
selecting at least one source project for ensemble learning by analyzing a correlation between a plurality of source projects and a target project; and
Predicting heterogeneous defects of a target project through ensemble learning of the selected at least one or more source projects.
Heterogeneous defect prediction method comprising a. According to claim 1,
A step of performing metric matching of multiple source projects and target projects
Heterogeneous defect prediction method further comprising. According to claim 2,
The step of performing the metric matching,
Measuring a similarity between a metric pair including a metric of the plurality of source projects and a metric of the target project using a Kolmogorov-Smirnov test to integrate heterogeneous metric sets of the plurality of source projects and the target project.
Heterogeneous defect prediction method comprising a. According to claim 3,
The step of performing the metric matching,
Selecting a metric of a source project equal to or higher than a preset standard according to a gain ratio of the source project
Heterogeneous defect prediction method comprising a. According to claim 3,
The step of performing the metric matching,
calculating a p-value of the Kolmogorov-Smirnov test for the pair of metrics, and filtering the pair of metrics if the calculated p-value is less than a cutoff threshold;
Heterogeneous defect prediction method comprising a. According to claim 5,
The step of performing the metric matching,
Selecting a metric pair having the highest p-value by applying maximum weight bipartite matching to remaining metric pairs through filtering of the metric pair;
Heterogeneous defect prediction method comprising a. According to claim 1,
The selection step is
Calculating a suitability score between source data for the plurality of source projects and target data for the target project as metric matching between the plurality of source projects and the target project is performed
Heterogeneous defect prediction method comprising a. According to claim 7,
The selection step is
When matching between the plurality of source projects and the target project is successful, a suitability score composed of a correlation score between the plurality of source projects and the target project and a ratio of the number of modules between the plurality of source projects and the target project steps to calculate
Heterogeneous defect prediction method comprising a. According to claim 7,
The selection step is
Deriving a suitability score of 0 when matching between the plurality of source projects and the target project fails.
Heterogeneous defect prediction method comprising a. According to claim 7,
The selection step is
Randomly sampling modules from each of the source data for the plurality of source projects and the source data for the target data, and aligning the source data for the plurality of source projects and the source data for the target data.
Heterogeneous defect prediction method comprising a. According to claim 10,
The selection step is
Upon completion of the sampling and the sorting, a correlation score is calculated using a Spearman correlation coefficient between source data for the plurality of source projects and target data for the target data, and the p-value of the calculated correlation score is a cutoff Deriving an average of correlation scores calculated for each of the metrics performed on the plurality of source projects and the target project by setting it to 0 if it is greater than the threshold value
Heterogeneous defect prediction method comprising a. According to claim 1,
The step of predicting the heterogeneous defect,
Constructing a logistic regression-based predictive model for each of the selected at least one or more source projects, and predicting a defect label of the target project using the constructed logistic regression-based predictive model.
Heterogeneous defect prediction method comprising a. According to claim 12,
The step of predicting the heterogeneous defect,
Returning predicted probabilities from each logistic regression-based prediction model through soft voting, and determining an average of the returned probabilities as a final predicted label.
Heterogeneous defect prediction method comprising a. A computer program stored in a computer readable storage medium for executing a heterogeneous defect prediction method performed by a defect prediction system,
selecting at least one source project for ensemble learning by analyzing a correlation between a plurality of source projects and a target project; and
Predicting heterogeneous defects of a target project through ensemble learning of the selected at least one or more source projects.
A computer program stored in a computer readable storage medium comprising a. In the defect prediction system,
A correlation analyzer for selecting at least one source project for ensemble learning by analyzing the correlation between a plurality of source projects and a target project; and
An ensemble learning unit that predicts heterogeneous defects of a target project through ensemble learning of the selected at least one source project.
Defect prediction system comprising a.

Images