Linear classifier-oriented software testing method based on metamorphic test

Abstract

The invention discloses a linear classifier-oriented software testing method based on metamorphic testing, which comprises the following steps of: 1, randomly generating a plurality of groups of training data sets and testing data sets as source data sets, and testing a to-be-tested program by utilizing the source data sets to obtain a training result; 2, generating a new data set according to a mode in a metamorphic relation according to the hyperplane model obtained by the source data set, and training on the new data set to obtain a group of new hyperplane models; and 3, calculating whether the hyperplane model obtained by training the hyperplane model and the new data set meets a metamorphic relation, and if the metamorphic relation is not met on any training data set, judging that the program has bug. According to the stability of the linear classifier, two new metamorphic relations are provided, the test result is more accurate, the training result on the source data set is applied when the source data set generates the new data set, and the test effect is improved.

Description

technical field [0001] Invention patents relate to technical fields such as software engineering, testing of machine learning programs and software testing. Background technique [0002] Machine learning technology is widely used in various fields of daily life, such as image processing, object recognition, autonomous driving, etc. However, the technology related to the quality assurance of machine learning software is not mature enough, because the software quality assurance of machine learning faces two difficulties. First, the statistical characteristics of machine learning cannot be carried out according to the effect of the model trained by the machine learning program. logical analysis. Second, machine learning programs often do not have test predictions. Even if the model trained by the machine learning program can achieve a prediction accuracy rate of more than 90% on a certain data set, it does not mean that there are no bugs in the program. Therefore, research on...

AI Technical Summary

Problems solved by technology

However, the technology related to the quality assurance of machine learning software is not mature enough, because the software quality assurance of machine learning faces two major difficulties. logical analysis

2. Machine learning programs often do not have test predictions. Even if the model trained by the machine learning program can achieve a prediction accuracy rate of more than 90% on a certain data set, it does not mean that there are no bugs in the program.

The complete machine learning software includes three parts: machine learning training program, machine learning model and data set. Although it is very important for the quality assurance of the machine learning model, the model is trained by the training program. If there is a problem with the training program, then the training The model is also prone to problems

However, there are not many works on testing machine learning training programs at present. Most of these works use the method of metamorphosis testing for testing. However, the metamorphosis relationship selected in these works is often very intuitive, such as changing the order of training data. The training result does not change

Due to being too intuitive and simple, the current method often has the problem that the test effect is not good enough. In addition, due to the lack of rigorous theoretical analysis, the current method also has the problem of misjudgment of program bugs, that is, it is wrong to believe that the program exists. bug

In summary, the current method still has a lot of room for improvement and improvement.

Images