CN113204481B - Class imbalance software defect prediction method based on data resampling - Google Patents

Abstract

The invention provides a class imbalance software defect prediction method based on data resampling. According to the method, the Euclidean distance between a minority class data set and a majority class element and between the minority class data set and the minority class element is calculated, the minority class data and the majority class data which are closest to the minority class data are screened out, and the distance parameter of the minority class data is obtained through the Euclidean distance; marking the minority data in the minority data set according to the distance parameters, and obtaining minority data point types; calculating each K near-point set with few elements in the minority data sets, and counting the number of majority data and minority data in the K near-point sets to obtain the number of newly generated minority data; and respectively selecting two classifiers, performing confidence evaluation on the newly generated software defect prediction minority class data to obtain a training data set, training the selected classifiers, and obtaining a final prediction result through weighted voting. The invention can well solve the class imbalance problem in the software defect prediction process.

Description

A Class Imbalanced Software Defect Prediction Method Based on Data Resampling

technical field

The invention belongs to the field of software defect prediction, in particular to a class unbalanced software defect prediction method based on data resampling.

Background technique

With the development of society and the improvement of science and technology, the Internet has been deeply integrated into all aspects of our life, whether it is online shopping, going out to take a car, smart home, restaurant ordering and other various activities in our daily life can be done through The software is completed, and the usage scenarios of the software have penetrated into all aspects of our food, clothing, housing and transportation. In the process of software development, the demand for software functions continues to increase, the number of people served by software is constantly increasing, and the software development time is constantly being compressed. Various problems lead to software defects in the process of software development. The occurrence of defects will make the software unable to provide normal functions, cause huge production and economic losses, and have a huge impact on people's normal life. Therefore, it is important and necessary to try to avoid the occurrence of software defects. Therefore, in the software development process Predicting software defects can help developers find defects that occur in the software development process as soon as possible, and can timely modify the code of software defects, thereby avoiding various production economic losses.

However, in the real development environment, the data with software defects is far smaller than the data without software defects. The software defect prediction model constructed at this time is more difficult to find code modules with software defects. However, the ideal software defect prediction model The model needs to be more sensitive to defective data and can more accurately predict whether the code module has defects. Therefore, it is very important to solve the class imbalance problem of software defect prediction. In view of the above shortcomings, the present invention proposes a class-imbalanced software defect prediction method.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to solve the class imbalance problem in software defect prediction and propose a software defect prediction method for class imbalance problem, which is generally applicable to software defect prediction. In order to achieve the above object, the present invention comprises the following steps:

Step 1: Select any minority class data in the minority class data set to perform Euclidean distance calculation with each minority class data in the minority class data set in turn, and select the minority class with the closest distance to the selected minority class data in the minority class data set. Data, select any minority class data in the minority class data set and perform Euclidean distance calculation with each majority class data in the majority class data set in turn, and filter out the majority class data in the majority class data set with the closest distance to the selected minority class data. , calculate the distance of the selected minority data according to the nearest Euclidean distance between the selected minority data and each minority data in the minority data set, and according to the nearest Euclidean distance between the selected minority data and each minority data in the majority data set parameter; mark the minority class data according to the distance parameter of the minority class data in the minority class data set, and obtain the data point type of the minority class data; calculate the K nearest neighbor point set of each minority class data in the minority class data set, and further The K-nearest neighbor point set of each minority class data is divided into the K-nearest neighbor point majority class data set and the K-nearest neighbor point minority class data set. The number of minority class data in the data set, calculate the number of newly generated minority class data for each minority class data in the minority class data set;

Step 2: Select the first classifier and the second classifier respectively, and perform confidence evaluation on the newly generated software defect prediction minority data to obtain a training data set;

Step 3, use the first classifier, the second classifier and the obtained training set S' selected in step 2 to obtain the final prediction result through weighted voting;

Preferably, the software defect data in step 1 is: S={S _min , S _max };

The minority class data set described in step 1 is:

The majority class data set described in step 1 is:

Among them, S _min represents the minority class data set, S _max represents the majority class data set, pi represents the _i -th minority class data in the minority class data set, i∈[1, N], N represents the minority class data set in the minority class data set The number of class data, d _k represents the kth majority class data in the majority class data set, k∈[1,K], K represents the number of majority class data in the majority class data set;

i∈[1，N]， min_i∈[1，N]；The minority class data that is closest to the selected minority class data described in step 1 is:

i∈[1,N], min _i∈ [1,N];

表示少数类数据集合中与选取的第i个少数类数据距离最近的少数类数据，N表示少数类数据集合中少数类数据的数量；in,

Represents the minority class data in the minority class data set that is closest to the selected i-th minority class data, and N represents the number of minority class data in the minority class data set;

i∈[1，N], max_i∈[1，K]；The majority class data that is closest to the selected minority class data described in step 1 is:

i∈[1,N], max _i∈ [1,K];

表示多数类数据集合中与选取的第i个少数类数据距离最近的多数类数据，K表示少数类数据集合中少数类数据的数量；in,

Indicates the majority class data in the majority class data set that is closest to the selected i-th minority class data, and K represents the number of minority class data in the minority class data set;

The nearest Euclidean distance between the minority class data selected in step 1 and each minority class data in the minority class data set is:

The nearest Euclidean distance between the minority class data selected in step 1 and each majority class data in the majority class data set is:

The distance parameter of the selected minority data in step 1 is:

Among them, ∝ _i is the distance parameter of the i-th minority class data in the minority class data set;

In step 1, the minority class data is marked in the minority class data set according to the distance parameter of the minority class data as:

If ∝ _i < 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a safe point, flag _i =1;

If ∝ _i = 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a confusion point, flag _i = 2;

If ∝ _i > 1, the data point type in the minority class data set and the selected i-th minority class data are marked as dangerous points, and flag _i =3;

Calculate the K-nearest neighbor point set of each minority class data in the minority class data set described in step 1:

In step 1, the K nearest neighbor point set of each minority class data is divided into K nearest neighbor point majority class data set and K nearest neighbor point minority class data set, specifically:

The number of majority class data in the K nearest neighbor majority class data set described in step 1, denoted as

The number of minority class data in the minority class data set of K nearest neighbors described in step 1, denoted as

Step 1: Calculate the number of newly generated minority class data for each minority class data in the minority class data set, specifically:

Among them, ∝ _i is the distance parameter of the ith minority class data in the minority class data set, and n _i is the number of newly generated minority class data of each i th minority class data in the minority class data set;

Described in step 1, calculate the newly generated software defect prediction data;

As described in step 1, the i-th minority data in the minority data set will generate n _i new minority data, so the newly generated minority data is represented by p ^new _{i, j} , where j ∈ [1, n _i ]

The deviation of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 deviates from the majority class, denoted as ε _i,j ;

Among them, the deviation ε _i,j of the jth newly generated data of the ith minority class data in the minority class data set deviates from the majority class, and its calculation formula is:

为偏离多数类程度参数，取值为0-1的随机数,

为其最近的多数类数据。in,

is the degree of deviation from the majority class parameter, a random number of 0-1,

is its most recent majority class data.

The bias of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 is biased towards the majority class, denoted as σ _i,j ;

The bias σ _i,j of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 is biased towards the majority class, and its calculation formula is:

为偏向少数类层度参数，其取值为0-1.5的随机数，

为其最近的少数类数据。in,

is a level parameter that is biased towards the minority class, and its value is a random number from 0 to 1.5.

for its nearest minority class data.

The newly generated software defect prediction data described in step 1 is a minority class data, which is denoted as p ^new _i,j ;

The calculation formula of the jth newly generated data of the ith minority class data of the newly generated software defect prediction data is:

p ^new _i,j = p _i +ε _i,j +σ _i,j

Obtaining a new minority data set described in step 1, denoted as S _new ;

The number n _i of defect data newly generated by the minority class point p _i in step 1 is obtained according to the method of the minority class data p ^new generated above to obtain a new minority class data set S _new .

N’为新成少数类数据集S_new包含元素的个数，对新数据的类别标记为缺陷数据，记为弱标记L_w，对于第i个新成少数类数据集用符号p’_i表示，其标记为

in,

N' is the number of elements contained in the newly formed minority data set S _new , and the new data category is marked as defect data, denoted as weak label L _w , and the i-th newly formed minority data set is represented by the symbol p' _i , which is marked as

Preferably, the step 2 is as follows:

In step 2, the influence degree of the first classifier and the influence degree of the second classifier are calculated respectively;

H₁预测类别为

In step 2, the first classifier H ₁ is trained by using the newly formed minority data set S _new , and the newly formed minority data set S _new is used to bring into the first classifier H ₁ in turn to obtain the predicted class L _p1 , for S The i-th point p' _i in _new is weakly marked as

H1 _predicts the class to be

H₂预测类别为

The second classifier H ₂ is trained by using the newly formed minority data set S _new , and the second classifier H 2 is sequentially brought into the second classifier H ₂ by using the newly formed minority data set S _new to obtain the _predicted class L _p2 . points p' _i , whose weak labels are

H ₂ predicts the class to be

The degree of influence of the first classifier is:

取值为1，否则取值为0，第二分类器H₂预测类别与弱标记L_w类别相同

取值为1，否则取值为0。Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0.

The degree of influence of the second classifier is:

取值为1，否则取值为0，第二分类器H₂预测类别与弱标记L_w类别相同

取值为1，否则取值为0。Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0.

In step 2, the label of the minority data is updated according to the influence degree of the first classifier and the influence degree of the second classifier, and the label of the minority data is updated to construct the updated original software defect data;

的置信度，用符号γ_i表示。As described in step 2, compute weak markers

The confidence level of , denoted by the symbol γ _i .

进行判断，依据分类器的影响程度来判断，计算公式为

当置信度γ_i＞β，将这个少数类数据

加入训练数据，当γ_i≤β的时候，直接删除，不把这个少数类数据加入到新训练集。Weak labeling of the new minority dataset as described in step 2

Judgment is made according to the influence degree of the classifier, and the calculation formula is

When the confidence γ _i > β, the minority class data

Add training data, when γ _i ≤ β, delete it directly, and do not add this minority class data to the new training set.

As described in step 2, the newly generated minority data, that is, S _new , is re-screened to obtain newly generated minority data S _new ′, and S _new ′ is added to the original software defect data S to obtain a new training set S ′;

Preferably, the step 3 specifically includes the following steps:

After obtaining the new training data set S', train the first classifier H ₁ and the second classifier H ₂ , and obtain the first classifier respectively by predicting the data v through the trained first classifier H ₁ and the second classifier H ₂ Predict the result L ₁ and the second classifier L ₂ , continue to use the influence degree o ₁ of the first classifier and the influence degree o ₂ of the second classifier, and use the calculation formula L _pre =L ₁ *o ₁ +L ₂ *o ₂ to get the prediction result;

As described in step 3, when the value of L _pre is greater than β, the category of v is predicted to be a minority category;

As described in step 3, when the value of L _pre is less than or equal to β, the category of v is predicted to be the majority category;

Compared with the prior art, the advantages and positive effects of the present invention are:

The present invention can well solve the class imbalance problem.

This paper adds a screening process for newly generated minority data, removes data that deviates from reality, and retains data that can show the true characteristics of minority classes.

This paper proposes a software defect prediction method that can solve the class imbalance, which can be widely applied to various software defect data and solve the class imbalance problem.

Description of drawings

FIG. 1 is a diagram of a software defect prediction method for class imbalance according to the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings and specific implementations. Here, the present invention is explained only by the description of the suitability of the present invention, but is not regarded as the present invention. limit.

The overall implementation flow chart of the present invention is shown in Figure 1, and the specific implementation is as follows:

Step 1: Select any minority class data in the minority class data set to perform Euclidean distance calculation with each minority class data in the minority class data set in turn, and select the minority class with the closest distance to the selected minority class data in the minority class data set. Data, select any minority class data in the minority class data set and perform Euclidean distance calculation with each majority class data in the majority class data set in turn, and filter out the majority class data in the majority class data set with the closest distance to the selected minority class data. , calculate the distance of the selected minority data according to the nearest Euclidean distance between the selected minority data and each minority data in the minority data set, and according to the nearest Euclidean distance between the selected minority data and each minority data in the majority data set parameter; mark the minority class data according to the distance parameter of the minority class data in the minority class data set, and obtain the data point type of the minority class data; calculate the K nearest neighbor point set of each minority class data in the minority class data set, and further The K-nearest neighbor point set of each minority class data is divided into the K-nearest neighbor point majority class data set and the K-nearest neighbor point minority class data set. The number of minority class data in the data set, calculate the number of newly generated minority class data for each minority class data in the minority class data set.

The software defect data in step 1 is: S={S _min , S _max };

The minority class data set described in step 1 is:

The majority class data set described in step 1 is:

Among them, S _min represents the minority class data set, S _max represents the majority class data set, pi represents the _i -th minority class data in the minority class data set, i∈[1, N], N represents the minority class data set in the minority class data set The number of class data, d _k represents the kth majority class data in the majority class data set, k∈[1,K], K represents the number of majority class data in the majority class data set;

i∈[1，N]， min_i∈[1，N]；The minority class data that is closest to the selected minority class data described in step 1 is:

i∈[1,N], min _i∈ [1,N];

表示少数类数据集合中与选取的第i个少数类数据距离最近的少数类数据，N表示少数类数据集合中少数类数据的数量；in,

Represents the minority class data in the minority class data set that is closest to the selected i-th minority class data, and N represents the number of minority class data in the minority class data set;

i∈[1，N], max_i∈[1，K]；The majority class data that is closest to the selected minority class data described in step 1 is:

i∈[1,N], max _i∈ [1,K];

表示多数类数据集合中与选取的第i个少数类数据距离最近的多数类数据，K表示少数类数据集合中少数类数据的数量；in,

Indicates the majority class data in the majority class data set that is closest to the selected i-th minority class data, and K represents the number of minority class data in the minority class data set;

The nearest Euclidean distance between the minority class data selected in step 1 and each minority class data in the minority class data set is:

The nearest Euclidean distance between the minority class data selected in step 1 and each majority class data in the majority class data set is:

The distance parameter of the selected minority data in step 1 is:

Among them, ∝ _i is the distance parameter of the i-th minority class data in the minority class data set;

In step 1, the minority class data is marked in the minority class data set according to the distance parameter of the minority class data as:

If ∝ _i < 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a safe point, flag _i =1;

If ∝ _i = 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a confusion point, flag _i = 2;

If ∝ _i > 1, the data point type in the minority class data set and the selected i-th minority class data are marked as dangerous points, and flag _i =3;

In step 1, the set of K nearest neighbors of each minority class data in the minority class data set is calculated, and the experimental setting K=5:

In step 1, the K nearest neighbor point set of each minority class data is divided into K nearest neighbor point majority class data set and K nearest neighbor point minority class data set, specifically:

The number of majority class data in the K nearest neighbor majority class data set described in step 1, denoted as

The number of minority class data in the minority class data set of K nearest neighbors described in step 1, denoted as

Step 1: Calculate the number of newly generated minority class data for each minority class data in the minority class data set, specifically:

Among them, ∝ _i is the distance parameter of the ith minority class data in the minority class data set, and n _i is the number of newly generated minority class data of each i th minority class data in the minority class data set;

Described in step 1, calculate the newly generated software defect prediction data;

As described in step 1, the i-th minority data in the minority data set will generate n _i new minority data, so the newly generated minority data is represented by p ^new _{i, j} , where j ∈ [1, n _i ]

The deviation of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 deviates from the majority class, denoted as ε _i,j ;

Among them, the deviation ε _i,j of the jth newly generated data of the ith minority class data in the minority class data set deviates from the majority class, and its calculation formula is:

为偏离多数类程度参数，取值为0-1的随机数,

为其最近的多数类数据。in,

is the degree of deviation from the majority class parameter, a random number of 0-1,

is its most recent majority class data.

The bias of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 is biased towards the majority class, denoted as σ _i,j ;

The bias σ _i,j of the j-th newly generated data of the i-th minority-class data in the minority-class data set described in step 1 is biased towards the majority class, and its calculation formula is:

为偏向少数类层度参数，其取值为0-1.5的随机数，

为其最近的少数类数据。in,

is a level parameter that is biased towards the minority class, and its value is a random number from 0 to 1.5.

for its nearest minority class data.

The newly generated software defect prediction data described in step 1 is a minority class data, which is denoted as p ^new _i,j ;

The calculation formula of the jth newly generated data of the ith minority class data of the newly generated software defect prediction data is:

p ^new _i,j = p _i +ε _i,j +σ _i,j

Obtaining a new minority data set described in step 1, denoted as S _new ;

The number n _i of defect data newly generated by the minority class point p _i in step 1 is obtained according to the method of the minority class data p ^new generated above to obtain a new minority class data set S _new .

N’为新成少数类数据集S_new包含元素的个数，对新数据的类别标记为缺陷数据，记为弱标记L_w，对于第i个新成少数类数据集用符号p’_i表示，其标记为

in,

N' is the number of elements contained in the newly formed minority data set S _new , and the new data category is marked as defect data, denoted as weak label L _w , and the i-th newly formed minority data set is represented by the symbol p' _i , which is marked as

Step 2: Select the first classifier and the second classifier respectively, and perform confidence evaluation on the newly generated software defect prediction minority data to obtain a training data set;

The step 2 is as follows:

In step 2, the influence degree of the first classifier and the influence degree of the second classifier are calculated respectively;

H₁预测类别为

In step 2, the first classifier H ₁ is trained by using the newly formed minority data set S _new , and the newly formed minority data set S _new is used to bring into the first classifier H ₁ in turn to obtain the predicted class L _p1 , for S The i-th point p' _i in _new is weakly marked as

H1 _predicts the class to be

H₂预测类别为

The second classifier H ₂ is trained by using the newly formed minority data set S _new , and the second classifier H 2 is sequentially brought into the second classifier H ₂ by using the newly formed minority data set S _new to obtain the _predicted class L _p2 . points p' _i , whose weak labels are

H ₂ predicts the class to be

The degree of influence of the first classifier is:

取值为1，否则取值为0，第二分类器H₂预测类别与弱标记L_w类别相同

取值为1，否则取值为0。Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0.

The degree of influence of the second classifier is:

取值为1，否则取值为0，第二分类器H₂预测类别与弱标记L_w类别相同

取值为1，否则取值为0。Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0.

In step 2, the label of the minority data is updated according to the influence degree of the first classifier and the influence degree of the second classifier, and the label of the minority data is updated to construct the updated original software defect data;

的置信度，用符号γ_i表示。As described in step 2, compute weak markers

The confidence level of , denoted by the symbol γ _i .

进行判断，依据分类器的影响程度来判断，计算公式为

当置信度γ_i＞β＝0.5，将这个少数类数据

加入训练数据，当γ_i≤β＝0.5的时候，直接删除，不把这个少数类数据加入到新训练集。Weak labeling of the new minority dataset as described in step 2

Judgment is made according to the influence degree of the classifier, and the calculation formula is

When the confidence γ _i > β = 0.5, this minority class data

Add training data, when γ _i ≤ β = 0.5, delete it directly, and do not add this minority data to the new training set.

As described in step 2, the newly generated minority data, that is, S _new , is re-screened to obtain newly generated minority data S _new ′, and S _new ′ is added to the original software defect data S to obtain a new training set S ′;

Step 3, use the first classifier, the second classifier and the obtained training set S' selected in step 2 to obtain the final prediction result through weighted voting;

The step 3 specifically includes the following steps:

After obtaining the new training data set S', train the first classifier H ₁ and the second classifier H ₂ , and obtain the first classifier respectively by predicting the data v through the trained first classifier H ₁ and the second classifier H ₂ Predict the result L ₁ and the second classifier L ₂ , continue to use the influence degree o ₁ of the first classifier and the influence degree o ₂ of the second classifier, and use the calculation formula L _pre =L ₁ *o ₁ +L ₂ *o ₂ to get the prediction result;

As described in step 3, when the L _pre value is greater than β=0.5, the category of v is predicted to be a minority category;

As described in step 3, when the value of L _pre is less than or equal to β=0.5, the class of v is predicted to be the majority class.

This embodiment compares the method of the present invention with some existing mainstream SMOTE+SVM, SMOTE+decision tree, SMOTE+k-nearest neighbor, SMOTE+Naive Bayes methods, and selects the accuracy, F-measure, balance, AUC Metric comparison results. Among all the methods compared, the method of the present invention has the highest accuracy, and the recognition accuracy has reached the advanced level in the field.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. The equivalent changes made by the method and the implementation method thereof should all be included within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (2)

1. a class unbalanced software defect prediction method based on data resampling, is characterized in that, Step 1: Select any minority class data in the minority class data set to perform Euclidean distance calculation with each minority class data in the minority class data set in turn, and select the minority class with the closest distance to the selected minority class data in the minority class data set. Data, select any minority class data in the minority class data set and perform Euclidean distance calculation with each majority class data in the majority class data set in turn, and filter out the majority class data in the majority class data set with the closest distance to the selected minority class data. , calculate the distance of the selected minority data according to the nearest Euclidean distance between the selected minority data and each minority data in the minority data set, and according to the nearest Euclidean distance between the selected minority data and each minority data in the majority data set parameter; mark the minority class data according to the distance parameter of the minority class data in the minority class data set, and obtain the data point type of the minority class data; calculate the K nearest neighbor point set of each minority class data in the minority class data set, and further The K-nearest neighbor point set of each minority class data is divided into the K-nearest neighbor point majority class data set and the K-nearest neighbor point minority class data set. The number of minority class data in the data set, calculate the number of newly generated minority class data for each minority class data in the minority class data set; Step 2: Select the first classifier and the second classifier respectively, and perform confidence evaluation on the newly generated software defect prediction minority data to obtain a training data set; Step 3, use the first classifier, the second classifier and the obtained training set S' selected in step 2 to obtain the final prediction result through weighted voting; The software defect data is: S={S _min , S _max }; Step 1 The minority class data set is:

Step 1 The majority class data set is:

Among them, S _min represents the minority class data set, S _max represents the majority class data set, pi represents the _i -th minority class data in the minority class data set, i∈[1, N], N represents the minority class data set in the minority class data set The number of class data, d _k represents the kth majority class data in the majority class data set, k∈[1,K], K represents the number of majority class data in the majority class data set; The minority class data closest to the selected minority class data in step 1 is:

in,

Represents the minority class data in the minority class data set that is closest to the selected i-th minority class data, and N represents the number of minority class data in the minority class data set; The majority class data closest to the selected minority class data in step 1 is:

in,

Indicates the majority class data in the majority class data set that is closest to the selected i-th minority class data, and K represents the number of minority class data in the minority class data set; The nearest Euclidean distance between the minority class data selected in step 1 and each minority class data in the minority class data set is:

The nearest Euclidean distance between the minority class data selected in step 1 and each majority class data set in the majority class data set is:

Step 1 calculates the distance parameter of the selected minority class data as:

Among them, ∝ _i is the distance parameter of the i-th minority class data in the minority class data set; Step 1 Label the minority class data according to the distance parameter of the minority class data in the minority class data set as: If ∝ _i < 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a safe point, flag _i =1; If ∝ _i = 1, the data point type in the minority class data set and the selected i-th minority class data is marked as a confusion point, flag _i = 2; If ∝ _i > 1, the data point type in the minority class data set and the selected i-th minority class data are marked as dangerous points, and flag _i =3; Step 1 Calculate the set of K-nearest neighbors for each minority class in the minority class data set: In step 1, the K-nearest neighbor point set of each minority class data is divided into the K-nearest neighbor point majority class data set and the K-nearest neighbor point minority class data set, specifically: Step 1 The number of majority class data in the majority class data set of K nearest neighbors, denoted as

Step 1 The number of minority class data in the minority class data set of K nearest neighbors, denoted as

Step 1: Calculate the number of newly generated minority class data for each minority class data in the minority class data set, specifically:

Among them, ∝ _i is the distance parameter of the ith minority class data in the minority class data set, and n _i is the number of newly generated minority class data of each i th minority class data in the minority class data set; Step 1, calculate the newly generated software defect prediction data; Step 1, the i-th minority data in the minority data set will generate n _i new minority data, so the newly generated minority data is represented by p ^new _{i, j} , where j ∈ [1, n _i ] Step 1: The deviation of the j-th newly generated data of the i-th minority-class data in the minority-class data set deviates from the majority class, denoted as ε _i,j ; Among them, the deviation εi _,j of the jth newly generated data of the ith minority class data in the minority class data set deviates from the majority class, and its calculation formula is:

in,

is the degree of deviation from the majority class parameter, a random number of 0-1,

its most recent majority class data; Step 1: The bias of the j-th newly generated data of the i-th minority-class data in the minority-class data set is biased towards the majority class, denoted as σ _i,j ; Step 1 The bias σ _i,j of the j-th newly generated data of the i-th minority class data in the minority class data set is biased towards the majority class, and its calculation formula is:

in,

is a level parameter that is biased towards the minority class, and its value is a random number from 0 to 1.5.

for its most recent minority class data; The software defect prediction data newly generated in step 1 is the minority data, which is denoted as p ^new _i,j ; The calculation formula of the jth newly generated data of the ith minority class data of the newly generated software defect prediction data is: p ^new _i,j = p _i +ε _i,j +σ _i,j Step 1: Obtain a newly generated minority data set, denoted as S _new ; Step 1: The number n _i of newly generated defect data by the minority class point p _i , according to the method of the minority class data p ^new generated above, obtain the newly generated minority class data set S _new ; in,

N, is the number of elements included in the newly formed minority data set S _new , the category of the new data is marked as defect data, denoted as weak label L _w , and the i-th newly formed minority data set is represented by the symbol p' _i , which is marked as

The step 2 is as follows: Step 2: Calculate the influence degree of the first classifier and the influence degree of the second classifier respectively; Step 2: Train the first classifier H ₁ by using the newly formed minority data set S _new , use the newly formed minority data set S _new , and bring it into the first classifier H ₁ in turn to obtain the predicted class L _p1 , for S _new The ith point p' _i of , its weak mark is

H1 _predicts the class to be

The second classifier H ₂ is trained by using the newly formed minority data set S _new , and the second classifier H 2 is sequentially brought into the second classifier H ₂ by using the newly formed minority data set S _new to obtain the _predicted class L _p2 . points p' _i , whose weak labels are

H ₂ predicts the class to be

The degree of influence of the first classifier is:

Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0; The degree of influence of the second classifier is:

Among them, N is the number of elements in the minority data set S _min , and the first classifier H ₁ predicts the same category as the weak label L _w category

The value is 1, otherwise the value is 0, the second classifier _H2 predicts the same category as the weak label _Lw category

The value is 1, otherwise the value is 0; Step 2, updating the label of the minority class data according to the influence degree of the first classifier and the influence degree of the second classifier, and updating the label of the minority class data to construct the updated original software defect data; Step 2, Compute Weak Markers

The confidence level of , which is represented by the symbol γ _i ; Step 2, Weak labeling of the new minority dataset

Judgment is made according to the influence degree of the classifier, and the calculation formula is

When the confidence γ _i > β, the minority class data

Add training data, when γ _i ≤ β, delete it directly, and do not add this minority data to the new training set; In step 2, the newly generated minority data S _new is re-screened to obtain newly generated minority data S _new ′, and S _new ′ is added to the original software defect data S to obtain a new training set S ′. 2. the class unbalanced software defect prediction method based on data resampling according to claim 1, is characterized in that, Step 3 specifically includes the following steps: After obtaining the new training data set S', train the first classifier H ₁ and the second classifier H ₂ , and obtain the first classifier respectively by predicting the data v through the trained first classifier H ₁ and the second classifier H ₂ The prediction result L ₁ and the second classifier prediction result L ₂ , continue to use the influence degree o ₁ of the first classifier and the influence degree o ₂ of the second classifier, and use the calculation formula L _pre =L ₁ *o ₁ +L ₂ The value of *o ₂ to get the prediction result; Step 3, when the L _pre value is greater than β, the category of the predicted data v is a minority category; Step 3, when the value of L _pre is less than or equal to β, the category of the predicted data v is the majority category.

Images