An optimization prediction method for software testing based on big data management

Abstract

This invention discloses an optimization prediction method for software testing based on big data management, specifically including: Step 1, collecting test data and performing time index alignment and numerical processing to form a test feature set; Step 2, constructing a static Markov blanket and a dynamic Markov blanket, and determining the target feature subset by switching input boundaries; Step 3, constructing a neural prediction model based on the target feature subset and outputting predicted values; Step 4, dividing the target feature subset into relevant feature clusters and updating the model parameters using the decorrelation-based isoangular minimum angle regression rule; Step 5, outputting test results, defect risk, and resource consumption prediction values; Step 6, generating test optimization configuration results based on the predicted values. This invention completes the prediction through isoangular regression using a two-layer Markov blanket.

Description

Technical Field

[0001] This invention relates to the fields of software testing and data-driven prediction technology, and in particular to an optimized prediction method for software testing based on big data management. Background Technology

[0002] As software systems continue to expand in scale and become increasingly complex in function, the types and scale of data involved in the software testing process are constantly growing. Information such as test execution records, defect records, test environment status, and test resource consumption are continuously generated during testing. Existing software test management methods largely rely on manual experience or rule-based static strategies to organize the test process. When faced with increasing test data volume, frequent switching between test phases, and limited test resources, it is difficult to effectively characterize test execution results, defect exposure, and resource consumption trends.

[0003] Some existing technical solutions attempt to incorporate statistical analysis or machine learning methods to analyze test data, but they still have shortcomings in practical applications. On the one hand, test data comes from diverse sources, and the correlations between features are complex. Directly inputting all features into the prediction model can easily introduce redundant information, affecting prediction stability. On the other hand, the testing process typically exhibits obvious phased characteristics, with the feature correlation structure changing across different testing phases. Using a fixed feature set for prediction struggles to adapt to the data distribution changes brought about by phase transitions. Furthermore, when there are strong correlations between features, conventional regression or optimization methods are prone to path instability during parameter updates, affecting the repeatability of prediction results.

[0004] Therefore, how to perform structured processing of multi-source test data in the context of big data management for software testing, reasonably determine the range of predictive features, and build a stable prediction and optimization mechanism under the condition that relevant features have dependencies remains one of the technical problems that need to be solved in the current field of software testing management. Summary of the Invention

[0005] One objective of this invention is to propose an optimization prediction method for software testing based on big data management. This invention introduces a two-layer Markov blanket constraint mechanism and a correlation-decorrelated minimum angular regression path advancement model for related feature clusters. It performs dynamic feature boundary determination and parameter path update on multi-source test data generated during the software testing process, constructs a continuously evolving test prediction and configuration optimization process, and realizes the coordinated scheduling of test execution order, resource allocation and process configuration.

[0006] An optimization prediction method for software testing based on big data management according to an embodiment of the present invention includes the following steps: Step 1: Collect test execution record data, defect record data, test environment status data, and test resource consumption data generated during the software testing process. Align the collected data with execution time index and perform numerical processing to form a test feature set. Step 2: For the target variable to be predicted, a static Markov blanket is constructed in the test feature set to determine a stable relevant feature subset, and a dynamic Markov blanket is constructed within the sliding time window to determine a stage-related feature subset. The target feature subset for the current prediction stage is determined between the stable relevant feature subset and the stage-related feature subset by the input boundary switching rule. Step 3: Construct a neural prediction model based on the target feature subset. The neural prediction model receives the feature vectors corresponding to the target feature subset and outputs the predicted values ​​for the test process. Step 4: Perform correlation measurement operation on the feature components within the target feature subset and divide them into related feature clusters. Construct an orthogonalized cluster basis in the feature subspace corresponding to each related feature cluster. In the parameter update process of the neural prediction model, adopt the decorrelated isoangular direction minimum angle regression path advancement rule oriented towards related feature clusters. Step 5: Output the predicted values ​​of test results, defect exposure risk, and test resource consumption based on the neural prediction model; Step 6: Generate test optimization configuration results based on the predicted values ​​of test results, defect exposure risk, and test resource consumption.

[0007] Optionally, step 1 specifically includes: The system collects test execution records, defect records, test environment status data, and test resource consumption data generated during software testing. Test execution records include test case identifiers and execution time information; defect records include defect identifiers and defect discovery time information; test environment status data includes test environment identifiers and a set of environment configuration parameters; and test resource consumption data includes test task identifiers and resource usage time intervals. A unified time index is constructed using the execution time information and defect discovery time information. Time index alignment is then performed on the test execution records, defect records, test environment status data, and test resource consumption data to generate a data sequence arranged according to the unified time index. Numerical processing is then performed on the time-index-aligned data sequence, including discrete encoding of identifier fields and numerical mapping of time interval fields, to generate a test feature set.

[0008] Optionally, step 2, constructing a static Markov blanket, includes the following steps: In the test feature set, select the target node identifier corresponding to the predicted target variable and use the target node identifier as the center of conditional correlation analysis. Along each feature component identifier within the test feature set, construct conditional correlation pairs between the feature component identifier and the target node identifier one by one. Within a data interval of a fixed time span, perform a conditional independence test on each conditional correlation pair. The conditional independence test calculates the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of other feature component identifiers unchanged. Perform a threshold comparison operation on the conditional correlation statistics and write the feature component identifiers that meet the conditional correlation retention rules into the candidate correlation set. For the feature component identifiers in the candidate correlation set, perform a pairwise conditional correlation consistency check operation. The check operation includes calculating the conditional correlation statistic between any two feature component identifiers in the candidate correlation set while keeping the target node identifier unchanged. Perform a removal operation on the feature component identifiers that fail the conditional correlation consistency check to form an intermediate stable correlation set. For the feature component identifiers in the intermediate stable correlation set, perform a reverse conditional correlation check operation on each one. The reverse conditional correlation check operation includes calculating the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of the remaining feature component identifiers in the intermediate stable correlation set unchanged. The feature component identifiers that pass the inverse conditional correlation check are written into the stable correlation feature subset. The stable correlation feature subset is used as the component of the static Markov blanket and recorded in the feature boundary buffer.

[0009] Optionally, step 2, constructing the dynamic Markov blanket, specifically includes: Set sliding time window parameters in the test feature set, including the window start time index, window end time index, and window span length. Move along the sliding time window on a unified time index axis, and extract a subset of feature data within the window coverage area at each window position. Select the target node identifier corresponding to the predicted target variable within the feature data subset, and select the target node identifier corresponding to the predicted target variable to participate in the stage conditional correlation calculation. Construct stage conditional correlation calculation pairs between each feature component identifier and the target node identifier one by one along the feature component identifier within the feature data subset. Within the current sliding time window, perform a stage conditional independence test operation on the stage conditional correlation calculation pairs. The stage conditional independence test operation includes calculating the conditional correlation statistics under the condition that the values ​​of the other feature component identifiers are within the control window. Perform an in-window threshold comparison operation on the stage conditional correlation statistics, and write the feature component identifiers that meet the stage correlation retention rules into the stage candidate correlation set. For the feature component identifiers in the stage candidate correlation set, perform a pairwise stage conditional correlation consistency check operation within the window. The check operation includes calculating the stage conditional correlation statistics between any two feature component identifiers under the condition that the target node identifiers are controlled. Feature component identifiers that fail the phase condition correlation consistency check within the window are removed to form an intermediate phase correlation set. Window sequence stability determination is performed on the feature component identifiers within the intermediate phase correlation set. This window sequence stability determination includes comparing the changes in condition correlation statistics within the current window and adjacent windows. Feature component identifiers that pass the window sequence stability determination are written into the phase-related feature subset. The phase-related feature subset is used as a component of the dynamic Markov blanket and recorded in the feature boundary buffer.

[0010] Optionally, step 3 includes the following steps: The algorithm reads the identifiers of each feature component in the target feature subset, generates an input node set according to the order of the feature component identifiers, and assigns a numerical input interface with the same dimension as the corresponding feature component to each input node in the input node set. At least one intermediate mapping layer is configured after the input node set. The intermediate mapping layer consists of several computation nodes, each of which establishes a connection relationship with the input node set. The connection weight parameters and bias parameters corresponding to the connection relationships are recorded in the model parameter storage area. An output node is configured after the intermediate mapping layer. The output node establishes a connection relationship with the computation nodes in the intermediate mapping layer. The connection weight parameters and bias parameters corresponding to the output node are recorded in the model parameter storage area, thus forming a neural prediction model structure that includes the input node set, the intermediate mapping layer, and the output node.

[0011] Optionally, step 4, constructing the orthogonalized cluster basis, specifically includes: Read the feature component identifier set corresponding to each relevant feature cluster within the target feature subset; for each relevant feature cluster, extract the numerical vectors of each feature component within the relevant feature cluster in the test feature set to form the feature matrix of the relevant feature cluster; perform a linearity independence judgment operation on the feature matrix of the relevant feature cluster, filter out linearly dependent feature component vectors, and generate an initial basis vector set within the cluster; perform an orthogonalization operation on the initial basis vector set within the cluster to generate a set of cluster basis vectors that satisfy the orthogonal constraint relationship; write the set of cluster basis vectors into the feature cluster basis storage area as the orthogonalized cluster basis of the corresponding relevant feature cluster.

[0012] Optionally, step 4, which adopts the decorrelation-oriented isoangular minimum angle regression path advancement rule for relevant feature clusters, includes the following steps: Read the set of orthogonalized cluster basis vectors corresponding to each relevant feature cluster, and initialize the active cluster basis set to an empty set; calculate the correlation between the prediction residual vector and each orthogonalized cluster basis vector, and perform a sorting operation on the correlation values; select the orthogonalized cluster basis vector with the largest correlation value from the sorting results and write it into the active cluster basis set; construct a parameter update direction vector in the vector space corresponding to the active cluster basis set, and maintain a symmetrical inner product relationship between the parameter update direction vector and each orthogonalized cluster basis vector in the active cluster basis set; perform parameter step-size advancement operation along the parameter update direction vector, update the parameter components in the neural prediction model corresponding to the active cluster basis set, and record the corresponding parameter step-size position; during the parameter step-size advancement process, periodically calculate the correlation change results between the orthogonalized cluster basis vectors not written into the active cluster basis set and the prediction residual vector; when the correlation change result corresponding to any orthogonalized cluster basis vector not written into the active cluster basis set meets the entry judgment condition, write the corresponding orthogonalized cluster basis vector into the active cluster basis set, reconstruct the parameter update direction vector in the vector space corresponding to the updated active cluster basis set, and continue to perform path advancement operation.

[0013] Optionally, step 6 specifically includes: The system reads predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values, and establishes a prediction value association table according to the test case identifier. For each test case identifier in the prediction value association table, it performs numerical normalization processing on the predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values ​​to generate a standardized prediction value vector. Based on a preset weight vector, it performs a weighted summation operation on the standardized prediction value vector to generate a comprehensive test case evaluation value. It then performs a sorting operation on the comprehensive test case evaluation values ​​to form a test case priority sequence. Based on the test case priority sequence, it performs a scheduling order mapping operation on the test cases, an allocation adjustment operation on the test resource occupancy time interval, and a configuration mapping operation on the test execution process nodes to generate a test optimization configuration result.

[0014] The beneficial effects of this invention are: (1) This invention constructs a static Markov blanket and a dynamic Markov blanket in the test feature set and performs input boundary switching in different prediction stages, so that the prediction model can adapt to the feature-related structure differences brought about by the changes in the test stage while maintaining a stable related feature structure, thereby avoiding the problem of redundant or mismatched features introduced by the fixed feature set in the stage switching scenario.

[0015] (2) Under the condition that there is a dependency relationship between related features, the present invention introduces the construction process of related feature clusters and orthogonalized clusters, and adopts the decorrelated minimum angular regression path advancement method oriented towards related feature clusters, so that the model parameter update process maintains structural consistency in the related feature space, reduces the interference of related features on the parameter path, and enhances the controllability of the prediction process.

[0016] (3) The present invention maps the predicted values ​​of test results, the predicted values ​​of defect exposure risks and the predicted values ​​of test resource consumption to the test optimization configuration result generation process in a unified manner, so that the test case scheduling order, resource allocation arrangement and process configuration adjustment have a unified quantitative basis, which is conducive to the continuous optimization of the test management process in large-scale test scenarios. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an optimized prediction method for software testing based on big data management proposed in this invention. Figure 2 This is a schematic diagram of the switchable input boundary mechanism of a double-layer Markov blanket for an optimized prediction method for software testing based on big data management proposed in this invention. Figure 3 This is a schematic diagram of the decorrelation-oriented isoangular minimum angle regression path advancement rule for an optimized prediction method for software testing based on big data management proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 An optimization prediction method for software testing based on big data management includes the following steps: Step 1: Collect test execution record data, defect record data, test environment status data, and test resource consumption data generated during the software testing process. Align the collected data with execution time index and perform numerical processing to form a test feature set. Step 2: For the target variable to be predicted, a static Markov blanket is constructed in the test feature set to determine a stable relevant feature subset, and a dynamic Markov blanket is constructed within the sliding time window to determine a stage-related feature subset. The target feature subset for the current prediction stage is determined between the stable relevant feature subset and the stage-related feature subset by the input boundary switching rule. Step 3: Construct a neural prediction model based on the target feature subset. The neural prediction model receives the feature vectors corresponding to the target feature subset and outputs the predicted values ​​for the test process. Step 4: Perform correlation measurement operation on the feature components within the target feature subset and divide them into related feature clusters. Construct an orthogonalized cluster basis in the feature subspace corresponding to each related feature cluster. In the parameter update process of the neural prediction model, adopt the decorrelated isoangular direction minimum angle regression path advancement rule oriented towards related feature clusters. Step 5: Output the predicted values ​​of test results, defect exposure risk, and test resource consumption based on the neural prediction model; Step 6: Generate test optimization configuration results based on the predicted values ​​of test results, defect exposure risk, and test resource consumption.

[0020] In this embodiment, step 1 specifically includes: The system collects test execution records, defect records, test environment status data, and test resource consumption data generated during software testing. Test execution records include test case identifiers and execution time information; defect records include defect identifiers and defect discovery time information; test environment status data includes test environment identifiers and a set of environment configuration parameters; and test resource consumption data includes test task identifiers and resource usage time intervals. A unified time index is constructed using the execution time information and defect discovery time information. Time index alignment is then performed on the test execution records, defect records, test environment status data, and test resource consumption data to generate a data sequence arranged according to the unified time index. Numerical processing is then performed on the time-index-aligned data sequence, including discrete encoding of identifier fields and numerical mapping of time interval fields, to generate a test feature set.

[0021] In this embodiment, step 2, constructing a static Markov blanket, includes the following steps: In the test feature set, select the target node identifier corresponding to the predicted target variable and use the target node identifier as the center of conditional correlation analysis. Along each feature component identifier within the test feature set, construct conditional correlation pairs between the feature component identifier and the target node identifier one by one. Within a data interval of a fixed time span, perform a conditional independence test on each conditional correlation pair. The conditional independence test calculates the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of other feature component identifiers unchanged. Perform a threshold comparison operation on the conditional correlation statistics and write the feature component identifiers that meet the conditional correlation retention rules into the candidate correlation set. For the feature component identifiers in the candidate correlation set, perform a pairwise conditional correlation consistency check operation. The check operation includes calculating the conditional correlation statistic between any two feature component identifiers in the candidate correlation set while keeping the target node identifier unchanged. Perform a removal operation on the feature component identifiers that fail the conditional correlation consistency check to form an intermediate stable correlation set. For the feature component identifiers in the intermediate stable correlation set, perform a reverse conditional correlation check operation on each one. The reverse conditional correlation check operation includes calculating the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of the remaining feature component identifiers in the intermediate stable correlation set unchanged. The feature component identifiers that pass the inverse conditional correlation check are written into the stable correlation feature subset. The stable correlation feature subset is used as the component of the static Markov blanket and recorded in the feature boundary buffer.

[0022] In this embodiment, step 1, which involves constructing conditional correlation calculation pairs between feature component identifiers and target node identifiers one by one, specifically includes: Read the time series of the feature component corresponding to the current feature component identifier, and read the time series of the target variable corresponding to the target node identifier; retrieve the column index positions corresponding to the current feature component identifier and the target node identifier in the field index table of the test feature set, and record the column indexes of the feature component and the target variable; perform alignment and truncation on the time series of the feature component and the time series of the target variable on a unified time index axis to generate a sequence of sample pairs of equal length; extract the set of feature component identifiers other than the current feature component identifier and the target node identifier from the test feature set to generate a set of condition variable identifiers; perform column index retrieval and alignment and truncation on each identifier in the set of condition variable identifiers to generate a condition variable matrix; write the feature component identifier, the target node identifier, the feature component column index, the target variable column index, the sequence of sample pairs of equal length, the set of condition variable identifiers, and the condition variable matrix into the condition correlation calculation pair record to complete the construction of the condition correlation calculation pair between the current feature component identifier and the target node identifier.

[0023] In this embodiment, step 1, performing a threshold comparison operation on the conditional correlation statistics, specifically includes: After calculating the conditional correlation statistic for each conditional correlation calculation pair, the current conditional correlation statistic value is read. The preset conditional correlation threshold parameters are read from the parameter configuration table, including the absolute value threshold and the significance probability threshold. The absolute value of the current conditional correlation statistic is taken to generate the statistic amplitude value. The statistic amplitude value and the absolute value threshold are compared; when the statistic amplitude value is greater than or equal to the absolute value threshold, a first judgment flag is generated. Simultaneously, the significance probability value corresponding to the current conditional correlation statistic is read, and the significance probability value is compared with the significance probability threshold; when the significance probability value is less than or equal to the significance probability threshold, a second judgment flag is generated. When both the first and second judgment flags are valid, the current feature component identifier is written to the candidate correlation set. If either judgment flag does not meet the conditions, no writing operation is performed.

[0024] In this embodiment, step 1 involves sequentially performing reverse conditional correlation verification operations, reading the feature component time series corresponding to the current feature component identifier, and reading the target variable time series corresponding to the target node identifier; removing the current feature component identifier from the intermediate stable correlation set to generate a set of remaining feature component identifiers, and constructing a condition variable matrix using the feature component time series corresponding to the set of remaining feature component identifiers; performing alignment and truncation on the current feature component time series, target variable time series, and condition variable matrix on a unified time index axis to generate a set of sample data of equal length; calculating the conditional correlation statistic between the current feature component time series and the target variable time series under the condition of controlling the value of the condition variable matrix; reading the preset reverse conditional correlation threshold parameter, performing absolute value calculation on the conditional correlation statistic value, and performing a size comparison operation with the reverse conditional correlation threshold parameter; when the absolute value of the statistic is greater than or equal to the reverse conditional correlation threshold parameter, retaining the current feature component identifier in the stable correlation feature subset; when the absolute value of the statistic is less than the reverse conditional correlation threshold parameter, removing the current feature component identifier from the stable correlation feature subset.

[0025] In this embodiment, step 2, constructing the dynamic Markov blanket, specifically includes: Set sliding time window parameters in the test feature set, including the window start time index, window end time index, and window span length. Move along the sliding time window on a unified time index axis, and extract a subset of feature data within the window coverage area at each window position. Select the target node identifier corresponding to the predicted target variable within the feature data subset, and select the target node identifier corresponding to the predicted target variable to participate in the stage conditional correlation calculation. Construct stage conditional correlation calculation pairs between each feature component identifier and the target node identifier one by one along the feature component identifier within the feature data subset. Within the current sliding time window, perform a stage conditional independence test operation on the stage conditional correlation calculation pairs. The stage conditional independence test operation includes calculating the conditional correlation statistics under the condition that the values ​​of the other feature component identifiers are within the control window. Perform an in-window threshold comparison operation on the stage conditional correlation statistics, and write the feature component identifiers that meet the stage correlation retention rules into the stage candidate correlation set. For the feature component identifiers in the stage candidate correlation set, perform a pairwise stage conditional correlation consistency check operation within the window. The check operation includes calculating the stage conditional correlation statistics between any two feature component identifiers under the condition that the target node identifiers are controlled. Feature component identifiers that fail the phase condition correlation consistency check within the window are removed to form an intermediate phase correlation set. Window sequence stability determination is performed on the feature component identifiers within the intermediate phase correlation set. This window sequence stability determination includes comparing the changes in condition correlation statistics within the current window and adjacent windows. Feature component identifiers that pass the window sequence stability determination are written into the phase-related feature subset. The phase-related feature subset is used as a component of the dynamic Markov blanket and recorded in the feature boundary buffer.

[0026] In this embodiment, step 2, which involves performing a window sequence stability determination operation on the feature component identifiers within the intermediate relevant set of the stage, specifically includes: Following the time index order of the sliding time window, read the stage condition-related statistics values ​​corresponding to the current window and its adjacent previous window; perform difference calculation on the stage condition-related statistics values ​​of the current window and the previous window to generate a statistic change value; perform absolute value calculation on the statistic change value to generate a statistic amplitude change value; read the preset window stability threshold parameter and perform a size comparison operation on the statistic amplitude change value and the window stability threshold parameter; when the statistic amplitude change value is less than or equal to the window stability threshold parameter, generate a window stability pass mark; repeat the above operation along several consecutive windows, recording the number of times the current feature component identifier passes window stability within consecutive windows; when the number of window stability passes reaches the preset consecutive window count parameter, write the current feature component identifier into the stage-related feature subset; when the number of window stability passes does not reach the preset consecutive window count parameter, no writing operation is performed.

[0027] In this embodiment, step 3 includes the following steps: The algorithm reads the identifiers of each feature component in the target feature subset, generates an input node set according to the order of the feature component identifiers, and assigns a numerical input interface with the same dimension as the corresponding feature component to each input node in the input node set. At least one intermediate mapping layer is configured after the input node set. The intermediate mapping layer consists of several computation nodes, each of which establishes a connection relationship with the input node set. The connection weight parameters and bias parameters corresponding to the connection relationships are recorded in the model parameter storage area. An output node is configured after the intermediate mapping layer. The output node establishes a connection relationship with the computation nodes in the intermediate mapping layer. The connection weight parameters and bias parameters corresponding to the output node are recorded in the model parameter storage area, thus forming a neural prediction model structure that includes the input node set, the intermediate mapping layer, and the output node.

[0028] In this embodiment, step 4, constructing the orthogonalized cluster basis, specifically includes: Read the feature component identifier set corresponding to each relevant feature cluster within the target feature subset; for each relevant feature cluster, extract the numerical vectors of each feature component within the relevant feature cluster in the test feature set to form the feature matrix of the relevant feature cluster; perform a linearity independence judgment operation on the feature matrix of the relevant feature cluster, filter out linearly dependent feature component vectors, and generate an initial basis vector set within the cluster; perform an orthogonalization operation on the initial basis vector set within the cluster to generate a set of cluster basis vectors that satisfy the orthogonal constraint relationship; write the set of cluster basis vectors into the feature cluster basis storage area as the orthogonalized cluster basis of the corresponding relevant feature cluster.

[0029] In this embodiment, step 4, which adopts the decorrelation-oriented isoangular minimum angle regression path advancement rule for relevant feature clusters, includes the following steps: Read the set of orthogonalized cluster basis vectors corresponding to each relevant feature cluster, and initialize the active cluster basis set to an empty set; calculate the correlation between the prediction residual vector and each orthogonalized cluster basis vector, and perform a sorting operation on the correlation values; select the orthogonalized cluster basis vector with the largest correlation value from the sorting results and write it into the active cluster basis set; construct a parameter update direction vector in the vector space corresponding to the active cluster basis set, and maintain a symmetrical inner product relationship between the parameter update direction vector and each orthogonalized cluster basis vector in the active cluster basis set; perform parameter step-size advancement operation along the parameter update direction vector, update the parameter components in the neural prediction model corresponding to the active cluster basis set, and record the corresponding parameter step-size position; during the parameter step-size advancement process, periodically calculate the correlation change results between the orthogonalized cluster basis vectors not written into the active cluster basis set and the prediction residual vector; when the correlation change result corresponding to any orthogonalized cluster basis vector not written into the active cluster basis set meets the entry judgment condition, write the corresponding orthogonalized cluster basis vector into the active cluster basis set, reconstruct the parameter update direction vector in the vector space corresponding to the updated active cluster basis set, and continue to perform path advancement operation.

[0030] In this embodiment, step 4, performing parameter step size advancement calculation along the parameter update direction vector, specifically includes: Read the parameter update direction vector corresponding to the current active cluster basis set and read the current prediction residual vector; calculate the correlation between each orthogonalized cluster basis vector and the prediction residual vector along each orthogonalized cluster basis vector not written into the active cluster basis set; calculate the inner product between each orthogonalized cluster basis vector and the parameter update direction vector in the current active cluster basis set; construct a step size candidate formula, which includes the ratio between the correlation value corresponding to the unwritten active cluster basis vector and the inner product value of the active cluster basis vector; perform a filtering operation on each step size candidate value, retaining the candidate step size values ​​that meet the condition that the step size is positive; select the smallest step size value from the retained candidate step size values ​​as the parameter advancement step size for this time; multiply the parameter update direction vector by the current parameter advancement step size to perform numerical updates on the parameter components in the neural prediction model corresponding to the active cluster basis set; after the update is completed, recalculate the prediction residual vector and record the current parameter step size position.

[0031] In this embodiment, step 6 specifically includes: The system reads predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values, and establishes a prediction value association table according to the test case identifier. For each test case identifier in the prediction value association table, it performs numerical normalization processing on the predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values ​​to generate a standardized prediction value vector. Based on a preset weight vector, it performs a weighted summation operation on the standardized prediction value vector to generate a comprehensive test case evaluation value. It then performs a sorting operation on the comprehensive test case evaluation values ​​to form a test case priority sequence. Based on the test case priority sequence, it performs a scheduling order mapping operation on the test cases, an allocation adjustment operation on the test resource occupancy time interval, and a configuration mapping operation on the test execution process nodes to generate a test optimization configuration result.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a software system test management scenario characterized by large scale, frequent testing activities, and limited testing resources. In this scenario, a large amount of test execution records, defect records, environmental state change information, and test resource consumption information are continuously generated during the testing process. As the testing phases progress, the key factors affecting test results and defect exposure differ significantly at different stages. Traditional test management methods based on fixed features or empirical rules struggle to reflect changes in test status in a timely manner, easily leading to problems such as unreasonable test case sorting, delayed concentrated defect exposure, and unbalanced test resource allocation. In this scenario, firstly, multi-source data generated during test execution is uniformly collected, and data from different sources is mapped to a unified timeline using a time index alignment method. Then, the identifier and status data are numerically processed to form a test feature set with a consistent structure. Subsequently, a two-layer feature filtering is performed on the test feature set, focusing on test results and defect exposure as prediction targets. On the one hand, a static Markov blanket is constructed to identify feature components that maintain stable correlations throughout the entire testing cycle. On the other hand, a dynamic Markov blanket is constructed within a continuously sliding time interval to characterize stage-related features that change significantly in different testing phases. An input boundary switching mechanism automatically selects appropriate feature subsets as model inputs for different prediction phases. Based on this, a neural prediction model is constructed to model the target feature subset, outputting predicted test results, defect exposure risk, and test resource consumption. Addressing the correlation issues within the target feature subset, the feature components are correlated and divided into several related feature clusters. An orthogonalized cluster basis is constructed within each feature cluster, and a decorrelation-oriented isoangular minimum angle regression path is adopted during model parameter updates to ensure stability within the related feature space. During test execution, the system continuously outputs prediction results and adjusts the test case execution order, test resource allocation ratio, and execution strategy based on the predicted test results, defect risk, and resource consumption information. This approach allows testing resources to be allocated more to high-risk, high-reward test cases, while avoiding excessive resource consumption in low-risk testing phases.

[0033] Table 1: Comparison of Test Optimization Prediction Results

[0034] As shown in Table 1, the comparison of test optimization prediction results reveals that, under the same test scale and conditions, the method of this invention exhibits significant differences in several key test management indicators. Firstly, regarding the average defect exposure time, the traditional method yields a value of 42, while the method of this invention reduces this value to 27, a reduction of approximately 35%. This result demonstrates that by differentiating test characteristics into stages and introducing a prediction mechanism, high-risk defects can be identified at an earlier stage of the testing process, thereby reducing the repair costs and test rework risks associated with delayed defect exposure.

[0035] Regarding the proportion of defects found in concentrated areas, the traditional method yields a statistical result of 58%, while the method of this invention increases the corresponding proportion to 76%, an improvement of approximately 18%. This change indicates that, guided by the prediction results, testing resources can more effectively cover areas with high defect density, resulting in a more concentrated distribution of defect findings and avoiding fluctuations in testing rhythm caused by scattered defect exposure.

[0036] In terms of test resource utilization, the traditional method achieves a utilization rate of 61%, while the method of this invention improves it to 82%, an increase of 21%. This result reflects that by introducing test resource consumption prediction and incorporating it into test configuration adjustments, test resource allocation is more aligned with actual risk levels, reducing the resource consumption of low-value test cases.

[0037] Meanwhile, the proportion of invalid test cases decreased from 29% under the traditional method to 14%, a reduction of approximately 15%. This change indicates that, guided by the prediction results, the test case sorting and execution strategy is more targeted, reducing the repeated execution of test cases that contribute little to defects.

[0038] Regarding the mean prediction error, the traditional method corresponds to a value of 0.31, while the method of this invention reduces it to 0.18, a reduction of approximately 42%. This result indicates that, under the orthogonalization of relevant feature clusters and the equiangular path advancement mechanism, the prediction process is more stable, and the volatility of the model output is effectively controlled. In summary, the above data shows that this invention exhibits consistent improvements in defect exposure timeliness, resource utilization efficiency, and prediction stability.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An optimization prediction method for software testing based on big data management, characterized in that, Includes the following steps: Step 1: Collect test execution record data, defect record data, test environment status data, and test resource consumption data generated during the software testing process. Align the collected data with execution time index and perform numerical processing to form a test feature set. Step 2: For the target variable to be predicted, a static Markov blanket is constructed in the test feature set to determine a stable relevant feature subset, and a dynamic Markov blanket is constructed within the sliding time window to determine a stage-related feature subset. The target feature subset for the current prediction stage is determined between the stable relevant feature subset and the stage-related feature subset by the input boundary switching rule. Step 3: Construct a neural prediction model based on the target feature subset. The neural prediction model receives the feature vectors corresponding to the target feature subset and outputs the predicted values ​​for the test process. Step 4: Perform correlation measurement operation on the feature components within the target feature subset and divide them into related feature clusters. Construct an orthogonalized cluster basis in the feature subspace corresponding to each related feature cluster. In the parameter update process of the neural prediction model, adopt the decorrelated isoangular direction minimum angle regression path advancement rule oriented towards related feature clusters. Step 5: Output the predicted values ​​of test results, defect exposure risk, and test resource consumption based on the neural prediction model; Step 6: Generate test optimization configuration results based on the predicted values ​​of test results, defect exposure risk, and test resource consumption.

2. The optimization and prediction method for software testing based on big data management according to claim 1, characterized in that, Step 1 specifically includes: The system collects test execution records, defect records, test environment status data, and test resource consumption data generated during software testing. Test execution records include test case identifiers and execution time information; defect records include defect identifiers and defect discovery time information; test environment status data includes test environment identifiers and a set of environment configuration parameters; and test resource consumption data includes test task identifiers and resource usage time intervals. A unified time index is constructed using the execution time information and defect discovery time information. Time index alignment is then performed on the test execution records, defect records, test environment status data, and test resource consumption data to generate a data sequence arranged according to the unified time index. Numerical processing is then performed on the time-index-aligned data sequence. This numerical processing includes discrete encoding of identifier fields and numerical mapping of time interval fields to generate a test feature set.

3. The optimization prediction method for software testing based on big data management according to claim 2, characterized in that, Step 2, constructing a static Markov blanket, includes the following steps: In the test feature set, select the target node identifier corresponding to the predicted target variable and use the target node identifier as the center of conditional correlation analysis. Along each feature component identifier within the test feature set, construct conditional correlation pairs between the feature component identifier and the target node identifier one by one. Within a data interval of a fixed time span, perform a conditional independence test on each conditional correlation pair. The conditional independence test calculates the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of other feature component identifiers unchanged. Perform a threshold comparison operation on the conditional correlation statistics and write the feature component identifiers that meet the conditional correlation retention rules into the candidate correlation set. For the feature component identifiers in the candidate correlation set, perform a pairwise conditional correlation consistency check operation. The check operation includes calculating the conditional correlation statistic between any two feature component identifiers in the candidate correlation set while keeping the target node identifier unchanged. Perform a removal operation on the feature component identifiers that fail the conditional correlation consistency check to form an intermediate stable correlation set. For the feature component identifiers in the intermediate stable correlation set, perform a reverse conditional correlation check operation on each one. The reverse conditional correlation check operation includes calculating the conditional correlation statistic between the current feature component identifier and the target node identifier while keeping the values ​​of the remaining feature component identifiers in the intermediate stable correlation set unchanged. The feature component identifiers that pass the inverse conditional correlation check are written into the stable correlation feature subset. The stable correlation feature subset is used as the component of the static Markov blanket and recorded in the feature boundary buffer.

4. The optimization and prediction method for software testing based on big data management according to claim 3, characterized in that, Step 2, constructing the dynamic Markov blanket, specifically includes: Set sliding time window parameters in the test feature set, including the window start time index, window end time index, and window span length. Move along the sliding time window on a unified time index axis, and extract a subset of feature data within the window coverage area at each window position. Select the target node identifier corresponding to the predicted target variable within the feature data subset, and select the target node identifier corresponding to the predicted target variable to participate in the stage conditional correlation calculation. Construct stage conditional correlation calculation pairs between each feature component identifier and the target node identifier one by one along the feature component identifiers within the feature data subset. Within the current sliding time window, perform a stage conditional independence test operation on the stage conditional correlation calculation pairs. The stage conditional independence test operation includes calculating the conditional correlation statistics under the condition that the values ​​of the remaining feature component identifiers are controlled within the window. Perform an in-window threshold comparison operation on the stage conditional correlation statistics, and write the feature component identifiers that meet the stage correlation retention rules into the stage candidate correlation set. For the feature component identifiers in the stage candidate correlation set, perform a pairwise stage conditional correlation consistency check operation within the window. The check operation includes calculating the stage conditional correlation statistics between any two feature component identifiers under the condition that the target node identifiers are controlled. Feature component identifiers that fail the phase condition correlation consistency check within the window are removed to form an intermediate phase correlation set. Window sequence stability determination is performed on the feature component identifiers within the intermediate phase correlation set. This window sequence stability determination includes comparing the changes in condition correlation statistics within the current window and adjacent windows. Feature component identifiers that pass the window sequence stability determination are written into the phase-related feature subset. The phase-related feature subset is used as a component of the dynamic Markov blanket and recorded in the feature boundary buffer.

5. The optimization prediction method for software testing based on big data management according to claim 4, characterized in that, Step 3 includes the following steps: The algorithm reads the identifiers of each feature component in the target feature subset, generates an input node set according to the order of the feature component identifiers, and assigns a numerical input interface with the same dimension as the corresponding feature component to each input node in the input node set. At least one intermediate mapping layer is configured after the input node set. The intermediate mapping layer consists of several computation nodes, each of which establishes a connection relationship with the input node set. The connection weight parameters and bias parameters corresponding to the connection relationships are recorded in the model parameter storage area. An output node is configured after the intermediate mapping layer. The output node establishes a connection relationship with the computation nodes in the intermediate mapping layer. The connection weight parameters and bias parameters corresponding to the output node are recorded in the model parameter storage area, thus forming a neural prediction model structure that includes the input node set, the intermediate mapping layer, and the output node.

6. The optimization prediction method for software testing based on big data management according to claim 5, characterized in that, Step 4, constructing the orthogonalized cluster basis, specifically includes: Read the feature component identifier set corresponding to each relevant feature cluster within the target feature subset; for each relevant feature cluster, extract the numerical vectors of each feature component within the relevant feature cluster in the test feature set to form the feature matrix of the relevant feature cluster; perform a linearity independence judgment operation on the feature matrix of the relevant feature cluster, filter out linearly related feature component vectors, and generate an initial basis vector set within the cluster; perform an orthogonalization operation on the initial basis vector set within the cluster to generate a set of cluster basis vectors that satisfy the orthogonal constraint relationship; write the set of cluster basis vectors into the feature cluster basis storage area as the orthogonalized cluster basis of the corresponding relevant feature cluster.

7. The optimization prediction method for software testing based on big data management according to claim 6, characterized in that, Step 4, which adopts the decorrelation-oriented isoangular minimum angle regression path advancement rule for relevant feature clusters, includes the following steps: Read the set of orthogonalized cluster basis vectors corresponding to each relevant feature cluster, and initialize the active cluster basis set to an empty set; calculate the correlation between the prediction residual vector and each orthogonalized cluster basis vector, and perform a sorting operation on the correlation values; select the orthogonalized cluster basis vector with the largest correlation value from the sorting results and write it into the active cluster basis set; construct a parameter update direction vector in the vector space corresponding to the active cluster basis set, and maintain a symmetrical inner product relationship between the parameter update direction vector and each orthogonalized cluster basis vector in the active cluster basis set; perform parameter step-size advancement operation along the parameter update direction vector, update the parameter components in the neural prediction model corresponding to the active cluster basis set, and record the corresponding parameter step-size position; during the parameter step-size advancement process, periodically calculate the correlation change results between the orthogonalized cluster basis vectors not written into the active cluster basis set and the prediction residual vector; when the correlation change result corresponding to any orthogonalized cluster basis vector not written into the active cluster basis set meets the entry judgment condition, write the corresponding orthogonalized cluster basis vector into the active cluster basis set, reconstruct the parameter update direction vector in the vector space corresponding to the updated active cluster basis set, and continue to perform path advancement operation.

8. The optimization prediction method for software testing based on big data management according to claim 7, characterized in that, Step 6 specifically includes: The system reads predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values, and establishes a prediction value association table according to the test case identifier. For each test case identifier in the prediction value association table, it performs numerical normalization processing on the predicted test result values, defect exposure risk prediction values, and test resource consumption prediction values ​​to generate a standardized prediction value vector. Based on a preset weight vector, it performs a weighted summation operation on the standardized prediction value vector to generate a comprehensive test case evaluation value. It then performs a sorting operation on the comprehensive test case evaluation values ​​to form a test case priority sequence. Based on the test case priority sequence, it performs a scheduling order mapping operation on the test cases, an allocation adjustment operation on the test resource occupancy time interval, and a configuration mapping operation on the test execution process nodes to generate a test optimization configuration result.

Images