CN113434401B - Software defect prediction method based on sample distribution characteristics and SPY algorithm - Google Patents

Abstract

The invention discloses a software defect prediction method based on sample distribution characteristics and SPY algorithm; based on the distribution characteristics of software defect data sets, the invention proposes a boundary k value determination formula. Choose appropriate minority class boundary samples for different datasets according to the formula. In addition, the present invention combines the SPY algorithm with the boundary sampling algorithm, optimizes the SPY algorithm through the edge samples, sets some majority class samples in the minority class boundary region as SPY samples, and sets a smaller training sample weight for the SPY samples, in the original minority class Algorithms that use boundary sampling inside the boundary region. SPY samples can guide the minority class samples on the boundary to ensure that the minority class samples in the boundary area can be correctly classified. At the same time, by setting a smaller sample weight for SPY samples, the influence of SPY samples on the classification of most samples is reduced, and a better classification effect is finally achieved.

Description

Software Defect Prediction Method Based on Sample Distribution Characteristics and SPY Algorithm

technical field

The present invention relates to a method for predicting software defects, in particular to a method for predicting software defects based on sample distribution characteristics and the SPY algorithm; The defect data set improves the classification effect of the model, and ultimately helps testers find defect files and allocate test resources more effectively, thereby reducing the cost of software testing.

Background technique

For datasets with balanced category distribution, traditional classification algorithms can achieve better classification results. But in actual application scenarios, the distribution of data is usually unbalanced, such as financial fraud, medical diagnosis, software failure and other scenarios. In these scenarios, the data are mainly divided into two categories, most of the samples belong to the majority class data, and the rest belong to the minority class samples. When the traditional classification algorithm classifies the unbalanced data, the result tends to be the majority class, but the recognition rate of the minority class samples is low. But in actual scenarios, minority class samples have more practical value. Therefore, the classification of imbalanced data has high research value.

Existing approaches to deal with the class imbalanced data classification problem mainly start from two aspects: 1) Data sampling. Balance the original dataset by adding minority class data or subtracting majority class data. The method of increasing minority class samples is called oversampling method, and the method of reducing majority class samples is called undersampling method. There is also a hybrid sampling method that mixes oversampling and oversampling. The above methods directly balance the amount of data, but will change the distribution of the original data. 2) The classification algorithm mainly includes two parts: cost-sensitive learning and ensemble learning. The cost of misclassification of majority and minority samples is different. Cost-sensitive learning balances the classification tendency of the classifier by setting different misclassification penalty factors for the two types of samples and increasing the misclassification penalty factor of minority samples. Make the minority class samples be classified as correctly as possible. Cost-sensitive learning does not change the distribution of the original samples, but it needs to determine the penalty factors for the two types of samples. The integrated learning method is to combine several weak classifiers, assign different weights and integrate them into a strong classifier according to the classification performance of each weak classifier. In the SPY algorithm, the majority class samples around some minority class samples are regarded as SPY samples, and their labels are modified to balance the data set. However, due to the large number of SPY samples required for the balanced data set, this will affect the correct classification of the majority class samples. Therefore, the present invention combines sample edge sampling with the SPY algorithm, optimizes the SPY sample selection method, and adds training weight control to it, thereby improving the overall prediction performance.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the invention proposes a software defect prediction method based on sample distribution characteristics and SPY algorithm.

The invention proposes a formula for determining the boundary k value based on the distribution characteristics of the software defect data set. Select appropriate minority class boundary samples for different data sets according to the formula. In addition, the present invention combines the SPY algorithm with the boundary sampling algorithm, optimizes the SPY algorithm through edge samples, and proposes a new boundary oversampling method BSGSMOTE, which mainly sets part of the majority class samples in the minority class boundary area as SPY samples, and Set a smaller training sample weight for the SPY sample, and use a boundary sampling algorithm inside the original minority class boundary area. The SPY samples can guide the minority class samples on the boundary to ensure that the minority class samples in the border area can be correctly classified. At the same time, by setting a smaller sample weight for SPY samples, the influence of SPY samples on the classification of majority class samples is reduced, and a better classification effect is finally achieved.

The subject of the present invention comprises the following steps:

Step 1) Extracting sample features based on the software defect data set; including obtaining sample imbalance, obtaining the average distance between similar samples, and obtaining sample variance;

a) Calculate the sample imbalance

The ratio of the number of samples in the majority class to the number of samples in the minority class in the statistical data set. Calculation sample

The formula for this unbalance degree is:

imblance= _numN / _numP .

In the formula, num _N is the number of samples in the majority class, and num _P is the number of samples in the minority class.

b) Calculate the average distance between similar samples

The average distance between samples of the same class describes how close the samples are to each other. Taking the minority class sample as an example, for each minority class sample P _i in the minority class sample set S _p , calculate the distance from _Pi to the surrounding k nearest neighbor samples of the same type, and obtain the distance d ₁ , d ₂ ...d _k , put Calculate the average of the k distances obtained, and the average distance dp _i from the P _i sample to the surrounding neighbor samples can be obtained, the formula is:

dp _i =Avg(d ₁ ,d ₂ . . . d _k ).

Calculate each dp _i , and finally get dp=[dp ₁ ,dp ₂ ...dp _nump ], take the average value of dp as the average distance measure index dp _average between minority class samples. Similarly, we can also get the average distance dn _average between samples of the majority class.

c) Calculate the sample variance

The variance of the sample describes the degree of dispersion of the sample. The sample variance can be obtained by summing the average of the squares of the differences between each sample value and the overall sample mean. For the same type of samples, the larger the variance of the sample, the more dispersed the distribution of the sample; on the contrary, the smaller the variance of the sample, the more concentrated the distribution of the sample. The overall data set has two types of samples, the majority class and the minority class, and the variances of obtaining the two types of samples are S ₁ and S ₂ . The formula for calculating the sample variance is:

In the formula, ^σ2 represents the overall variance, X represents a single sample, μ represents the mean of the overall sample, and N is the number of samples in the data set.

Step 2) Adaptive boundary k value calculation, select appropriate boundary samples according to the k-nearest neighbor algorithm, and prepare for boundary sample resampling;

a) Calculation of adaptive boundary k value

The distribution characteristics of different data sets are different, and the value of k needs to be adjusted adaptively according to the distribution characteristics of the data sets. According to the distribution of the overall data set, two calculation formulas for the k value are proposed from the perspectives of distance and variance. In order to prevent the boundary k from being too large or too small, the value of k is limited to a range between 5 and 15.

Starting from the distance of the sample individual, combined with the overall imbalance rate of the sample, the following formula is obtained.

stk ₁ ∈ [5,15]

In the formula, imbalance is the sample imbalance rate, dp _average is the average distance between minority class samples, and dn _average is the average distance between majority class samples.

From the perspective of the variance of the two types of sample populations, combined with the imbalance rate of the sample as a whole, another calculation formula for the k value is obtained:

stk ₂ ∈ [5,15]

In the formula, imbalance is the sample imbalance rate, S _P is the overall variance of the minority class samples, and S _N is the overall variance of the majority class samples.

b) Boundary sample selection

According to the boundary k value obtained, the K nearest neighbor algorithm is used to calculate the k nearest neighbor samples around each minority class sample. Among the k nearest neighbor samples, if the number of majority class samples is more than the number of minority class samples, and the number of neighboring minority class samples is not 0, it will be selected as the minority class boundary sample.

Step 3) Select the SPY samples around the minority class samples to help the two types of samples in the border area to be better classified, so as to improve the overall software defect prediction level;

According to the adaptive boundary k value obtained in step 2, select boundary SPY samples. SPY samples refer to the majority class samples that are close to the boundary of the minority class samples, and the corresponding SPY samples can be found by calculating the neighbors around the minority class samples. Carry out the nearest neighbor sample analysis on the boundary of the minority class, and select the appropriate majority class sample as the SPY sample. The specific method is to perform k-nearest neighbor analysis on the minority class samples, and calculate the nearest neighbor samples around the minority class samples. For each minority class sample, if the number of minority class samples in its neighbor samples is greater than the number of majority class samples, it means that the sample is in a relatively safe area. At this time, the majority class samples in these neighbor samples Can be regarded as a SPY sample.

Step 4) Perform oversampling in the marginal minority samples to balance the dataset;

Minority samples are oversampled by linear interpolation. The present invention uses a k-nearest neighbor algorithm with k=5 to obtain boundary neighbor samples of a minority class on the boundary. And random linear interpolation is performed between two minority class samples, so that the newly generated minority class samples can be distributed in the minority class boundary area, and the randomness makes the newly generated samples more diverse.

Step 5) Set training weights for SPY samples and other samples respectively, to reduce the impact of SPY samples on majority class samples, so as to improve the overall effect;

Set the training weight of the SPY sample. Since the essence of the SPY sample is the majority class sample, setting the label of the SPY sample as the minority class sample label will inevitably affect the classification decision of the surrounding majority class samples. By reducing the training weight of SPY samples, its impact on the decision boundary can be reduced, and it can guide the classification decision of minority class samples. In the present invention, the training weight of the SPY sample is set to 0.5, and the weights of other samples are all set to 1.

Step 6) use the machine learning model to train and predict the data set;

Put the obtained category-balanced software defect data set into the training model for training, and obtain the trained model. The model uses logistic regression model, decision tree model, k-nearest neighbor model and Bayesian model. After the model training is completed, the test set samples are preprocessed and input into the model to obtain the predicted labels of the model.

Beneficial effects of the present invention:

1. This technology adaptively determines the value of k according to the distribution characteristics of the original samples, and uses the k-nearest neighbor algorithm to find suitable minority class boundary samples, making preparations for generating new samples in the minority class boundary area.

2. This technology combines the SPY algorithm with the boundary sampling algorithm. By setting the majority class samples in the specified area as SPY samples, decision-making guidance is made on the minority class samples, and the control of training weight is added, so that more minority class samples can be correctly classified and the probability of defect class samples being identified is improved.

Description of drawings

Figure 1 Definition diagram of minority class boundary samples.

Figure 2 Definition of average distance between similar samples.

Fig. 3 Definition map of SPY sample.

Figure 4 The overall flowchart of the algorithm model.

Detailed ways

The present invention will be described in detail below in conjunction with the software defect prediction data set according to the accompanying drawings. Overall process of the present invention is as shown in accompanying drawing 4, and concrete steps are as follows:

Step 1. Perform 5-fold cross-validation on the original software defect data set, select 80% of it as a training set, and the remaining 20% as a test set. Feature extraction is performed on the samples in the training set to obtain the three features of sample imbalance, average distance between similar samples and sample variance.

1) Obtain the sample imbalance degree

The ratio of the number of samples in the majority class to the number of samples in the minority class in the software defect data set, the formula for calculating the imbalance degree of the samples is:

imblance= _numN / _numP ;

In the formula, num _N is the number of samples in the majority class, and num _P is the number of samples in the minority class; as shown in Figure 1;

2) Obtain the average distance between similar samples

As shown in Figure 2, the average distance between samples of the same type describes the closeness between samples; taking minority samples as an example, for each minority sample P _i in the minority sample set S _p , calculate P _i to The distances of the surrounding k neighbor samples of the same type can be obtained as distances d ₁ , d ₂ ...d _k , and the average value of the obtained k distances can be obtained to obtain the average distance dp _i from the _Pi sample to the surrounding neighbor samples. The formula is:

dp _i =Avg(d ₁ ,d ₂ ...d _k );

Calculate each dp _i , and finally get dp=[dp ₁ ,dp ₂ ...dp _nump ], take the average value of dp as the average distance measurement index dp _average between minority class samples; similarly, we can also get majority class samples The average distance dn _average between;

3) Get the sample variance

The variance of the sample describes the degree of dispersion of the sample; the variance of the sample can be obtained by the sum of the square values of the difference between each sample value and the mean of the total sample; Scattered; on the contrary, the smaller the variance of the sample, the more concentrated the distribution of the sample; the overall data set has two types of samples, the majority class and the minority class, and the variances of the two types of samples obtained are S ₁ and S ₂ ; the formula for calculating the sample variance is:

In the formula, ^σ2 represents the overall variance, X represents a single sample, μ represents the mean of the overall sample, and N is the number of samples in the data set.

Step 2. Substituting the three features obtained in step 1 into the proposed calculation formulas for the two boundary k values to obtain two k values. The two formulas are as follows:

Starting from the distance of the sample individual, the formula of combining the imbalance rate of the sample is as follows:

In the formula, imbalance is the sample imbalance rate, dp _average is the average distance between minority class samples, and dn _average is the average distance between majority class samples.

Starting from the variance of the sample population, the formula combined with the imbalance rate of the sample is as follows:

In the formula, imbalance is the sample imbalance rate, S _P is the overall variance of the minority class samples, and S _N is the overall variance of the majority class samples.

According to the boundary k value calculated in step 2, a better one is selected as the K-nearest neighbor parameter, and this patent uses k2 as the neighbor parameter. For the minority class samples, use the K nearest neighbor algorithm to calculate the k nearest neighbor samples around each minority class sample. If it is 0, the analyzed minority class samples will be regarded as minority class boundary samples.

Step 3: Carry out k-nearest neighbor analysis on the minority class samples, and calculate the nearest neighbor samples around the minority class samples. For each minority class sample, if the number of minority class samples in its neighbor samples is greater than the number of majority class samples, it means that the sample is in a relatively safe area, and the majority class samples in these neighbor samples are selected as SPY sample.

Step 4. Perform linear interpolation sampling on the minority class samples in the boundary area. The present invention adopts the k-nearest neighbor algorithm of k=5 to obtain the boundary neighbor samples of the minority class on the boundary. And random linear interpolation is performed between two minority class samples, so that the newly generated minority class samples can be distributed in the minority class boundary area, and the randomness makes the newly generated samples more diverse. The formula for linear interpolation is as follows:

n _i =(p _i -p _j )*δ+p _j

Step 5. Set the SPY sample label as the minority sample label, modify its training weight to 0.5, and set the training weight of other samples to 1.

Step 6. Use the training set to train on classic classification models such as logistic regression models, naive Bayesian models, logistic regression models, support vector machine models and decision tree models. Then normalize the sample data on the test set, put the normalized data into the trained model for classification prediction, and calculate the evaluation indicators Recall, F1, AUC and G-Mean.

Analysis of the final data distribution results: the present invention mainly generates new samples in the boundary area of the minority class samples, and the new samples are in the boundary area. In the process of selecting the boundary samples of the minority class, the noise samples in the samples can be found and removed by the k-nearest neighbor algorithm. At the same time, part of the majority class samples can be regarded as SPY samples. In the case of not adding too many minority class samples, the decision-making of the two types of samples can be moved to the minority class sample area through SPY samples to ensure that the minority class samples can be correctly classified. At the same time, by controlling the training weights of the two types of samples, the influence of SPY samples on the classification of most samples is reduced, and the overall classification performance is finally improved.

Claims (4)

1. The software defect prediction method based on the sample distribution characteristics and the SPY algorithm is characterized by comprising the following steps of:

step 1) extracting sample characteristics based on a software defect data set; acquiring sample unbalance, acquiring average distance between similar samples and acquiring sample variance;

step 2), calculating a self-adaptive boundary k value, and selecting a boundary sample according to a k nearest neighbor algorithm;

a) Adaptive boundary k value calculation

The distribution characteristics of different data sets are different, and the value of k needs to be adaptively adjusted according to the distribution characteristics of the data sets; according to the distribution condition of the whole data set, two calculation formulas of a k value are provided from the aspects of distance and variance; to prevent the boundary k from being too large or too small, the k value is constrained to range between 5 and 15;

starting from the distance of the individual sample, combining the integral unbalance rate of the sample to obtain the following formula;

s.t.k ₁ ∈[5,15]

where the average is the sample imbalance, dp _average Is a small numberAverage distance between class samples, dn _average Is the average distance between most classes of samples;

from the aspect of the variance of the two types of sample populations, the imbalance rate of the sample population is combined to obtain another calculation formula of the k value:

s.t.k ₂ ∈[5,15]

where the imbalance is the sample imbalance ratio, S _P Is the overall variance, S, of the minority samples _N Is the overall variance of the majority class samples;

b) Boundary sample selection

Calculating K neighbor samples around each minority sample by using a K neighbor algorithm according to the obtained boundary K value; if the number of the majority class samples is more than that of the minority class samples and the number of the neighbor minority class samples is not 0, selecting the k neighbor class samples as the minority class boundary samples;

step 3) performing k neighbor operation on the minority samples, and calculating to obtain neighbor samples around the minority samples; for each minority sample, if the number of minority samples in the neighbor samples is greater than that of majority samples, the sample is in a relatively safe area, and at this time, the majority samples in the neighbor samples are regarded as SPY samples; selecting SPY samples around the few samples, and using the SPY samples to guide the few samples in the boundary region to be better classified so as to improve the overall software defect prediction level;

step 4) oversampling in a few boundary samples to balance the data set;

step 5) respectively setting training weights for the SPY sample and other samples;

setting the training weight of the SPY sample to be 0.5, and setting the weights of other samples to be 1; the control of the weight leads SPY samples to guide the accurate classification of few samples of the boundary, and simultaneously reduces the classification influence on most samples of the boundary area, thereby integrally improving the overall classification prediction effect;

and 6) training and predicting the data set by using a machine learning model.

2. The method for predicting software defects based on sample distribution characteristics and SPY algorithm according to claim 1, wherein the step 1 of extracting sample characteristics based on the software defect data set specifically comprises the following steps:

1) Obtaining sample imbalance

The ratio of the number of the majority samples to the number of the minority samples in the software defect data set is calculated by the following formula:

imblance＝num _N /num _P ；

num in the formula _N Number of majority sample, num _P The number of the samples is a minority number;

2) Obtaining average distance between homogeneous samples

The average distance between homogeneous samples describes the proximity between samples; taking the minority class sample as an example, for each minority class sample P in the minority class sample set _i Calculating P _i The distances to the surrounding k neighboring homogeneous samples are obtained to obtain the distance d ₁ ,d ₂ …d _k The obtained k distances are averaged to obtain P _i Average distance dp of a sample to surrounding neighbor samples _i The formula is as follows:

dp _i ＝Avg(d ₁ ,d ₂ …d _k )；

calculating each dp _i Finally obtaining dp = [ dp ] ₁ ,dp ₂ …dp _nump ]Taking the average value of dp as the average distance measure index dp between the minority class samples _average (ii) a Similarly, we can also get the average distance dn between most kinds of samples _average ；

3) Obtaining a sample variance

The variance of a sample describes the degree of dispersion of the sample; the sample variance may be obtained from the sum average of the squared values of the difference between each sample value and the total sample mean; for the same type of sample, the larger the variance of the sample, the more dispersed the distribution of the sample; conversely, the smaller the variance of a sample, the more concentrated the distribution of the sample; the total data set comprises a majority type sample and a minority type sample, and the variances of the two types of samples are calculated respectively; the sample variance is calculated as:

in the formula sigma ² Represents the global variance, X represents the individual samples, μ represents the mean of the global samples, and N is the number of data set samples.

3. The method of claim 1, wherein the oversampling is performed in the boundary minority sample class according to the step 4, specifically as follows:

a few types of samples are subjected to oversampling in a linear interpolation mode; acquiring boundary neighbor samples of a few classes on a boundary by using a k neighbor algorithm with k = 5; and random linear interpolation is carried out between the two minority samples, so that the newly generated minority samples can be distributed in a minority boundary region, and meanwhile, the randomness causes the newly generated samples to have more diversity.

4. The method of claim 1, wherein the training and prediction of the data set using the machine learning model in step 6 is performed by using a sample distribution feature and an SPY algorithm, specifically as follows:

training and predicting in a classical decision model such as a logistic regression model, a decision tree model, a k nearest neighbor model and a Bayesian model through the obtained class-balanced software defect data set to obtain a trained model; after the model training is finished, preprocessing a test set sample and inputting the preprocessed test set sample into the model, so that a label predicted by the model can be obtained.

Images