Warning classification method based on cost-sensitive neural network with threshold operation

Abstract

The present invention relates to a warning classification method of a cost-sensitive neural network based on a threshold operation, comprising: S1 analyzing jar files of a series of versions of target software using a FindBugs tool to obtain static warnings of the target software; S2 marking the static warnings obtained by S1; S3 adopts the cost-sensitive BP neural network, uses the samples in the sample set to train the classifier, uses the classifier to classify all the samples in the sample set, and calculates the true category probability value for predicting effective warnings or false positive warnings. Threshold The mode of operation is to adjust the real category probability value to obtain a new category probability value, and use the new category probability value to predict and classify all samples in the sample set. The method of the invention has an average increase of 44.07% in the effective warning recall rate Recall, can quickly achieve a higher and stable recall rate, and can achieve a lower classification cost than the traditional neural network method.

Description

technical field [0001] The invention relates to software static analysis, in particular to the classification of software static warnings, in particular to a warning classification method based on a cost-sensitive neural network based on threshold operation. Background technique [0002] During the software development process, a lot of resources are consumed in finding and resolving bugs in software. Except for some large companies that use automated testing methods during the development process, in most small and medium-sized software companies or project teams, software testing is still in the manual processing stage. Its testing tasks are often heavy and inefficient. Smaller functional changes require regression testing of other functions. In the case of limited resources and time constraints, software testing work is usually not completed smoothly, resulting in many undiscovered errors remaining in the software system. The lack of testing methods has become a bottlen...

AI Technical Summary

Problems solved by technology

[0006] However, there are two types of problems in the warning classification process: the first type is the problem of unequal costs caused by misclassification, classifying a false positive warning as a valid warning (misclassification I), and classifying a valid warning as a false positive warning. Report a warning (misclassification II), and the resulting cost is not equal. It is different from the software defect prediction category. Misclassification II may cause software crashes, while misclassification I only requires developers to spend 5 minutes for review, that is, misclassification The cost of II will be much higher than that of misclassification I; the second type of problem is a class imbalance problem. As mentioned above, false positive warnings may account for the vast majority of overall warnings, and the experimental data of the present invention is also extremely unbalanced

The purpose of previous warning classification techniques is often to reduce the error rate of classification, while ignoring the different costs caused by the above-mentioned different classification errors and the imbalance of data

Images