Prediction Method for Continuous Integration Build Results by Fusing Multidimensional Features and Temporal Dependencies

Abstract

The present application relates to a method for predicting continuous integration build results by integrating multi-dimensional features and temporal dependencies. The method includes: obtaining build data corresponding to a project repository and extracting relevant features of the predicted build results and time consumption. After standardizing the features, constructing feature vectors, generating interaction features using the Hadamard product, and combining LSTM to capture the non-linear and temporal dependencies of historical builds to form model input features. Building a prediction model including an input layer, a graph attention network module, a temporal convolutional module, and a multi-task output layer. The input layer re-standardizes the features, the graph attention network models dynamic associations, the temporal convolution extracts temporal features, and the multi-task output layer splices vectors for prediction. Training the model according to a preset loss function and using the trained model to predict build results. Using this method can meet the requirements of high-precision prediction.

Description

Technical Field

[0001] This application relates to the technical field of software development, and particularly to a method for predicting continuous integration build results by integrating multi-dimensional features and temporal dependencies. Background Art

[0002] Continuous Integration and Continuous Deployment (CI / CD) are two key processes in modern software development practices, aiming to improve the speed, quality, and frequency of software delivery. GitHub Actions (GHA) is GitHub's CI / CD platform that allows users to automate software development workflows in GitHub repositories. Although GitHub Actions provides great convenience for developers, there are still some problems. Whether there are errors in the configuration files of the workflow or bugs in the committed code, they may both lead to the failure of the final build. Build failure not only results in a waste of time; it also distracts attention, consumes a large amount of computing resources, and reduces productivity. Therefore, it is significantly important to predict the build result and duration in advance. Existing technical solutions involve tasks of predicting the success or failure of builds and the prediction task of build duration in GitHub Actions. Considering the sequential characteristics of CI / CD build data, it is essential to obtain time-related input features, such as the result of the previous build, in these two prediction tasks. Although some researchers' studies have shown that non-technical factors like team communication are also valuable in predicting build failures, the process of obtaining information like team communication is often quite cumbersome. To accurately predict the build result, a robust and comprehensive dataset is indispensable. When looking for publicly accessible CI / CD build datasets, it is found that the options are limited, especially those that meet specific requirements. Existing CI / CD datasets, such as Travis Torrent, although it is a rich resource containing a large amount of build data from Travis CI and GitHub Actions, do not fully meet the requirements. There is a need to study a model specific to GitHub Actions for predicting the build results and durations of GitHub repositories, which requires creating a customized dataset. In addition, some researchers have found that there are significant differences in the accuracy between predicting build failures and successes, which prompts them to adopt a cascade classifier to correct this imbalance. Some researchers' studies have shown that the imbalance of the dataset is the main reason for the low prediction failure accuracy, and they suggest using techniques such as oversampling and undersampling to enhance the performance of the model on imbalanced datasets. In addition, considering the time series characteristics of CI / CD data, they compared the performance differences of the models finally obtained by training models through cross-validation and time series validation. The experimental results show that cross-validation can generally exhibit more superior performance in most cases.

[0003] However, existing continuous integration and continuous deployment (CI / CD) build prediction technologies mainly rely on the TravisTorrent dataset. Travis Torrent is not a dedicated dataset for GitHub Actions (GHA), and there is currently no publicly available GHA build dataset. This results in the inability of existing technologies to meet the specific requirements for GHA build prediction in terms of data, limiting the accuracy and applicability of the model. Most current technologies only use traditional machine learning methods to predict whether a build is successful or the time it takes, and these methods have deficiencies in prediction effects, generally having large prediction biases and being unable to meet the requirements of high-precision prediction. Summary of the Invention

[0004] Based on this, it is necessary to provide a continuous integration build result prediction method that combines multi-dimensional features and temporal dependencies to meet the requirements of high-precision prediction for the above technical problems.

[0005] A continuous integration build result prediction method that combines multi-dimensional features and temporal dependencies, the method includes:

[0006] Screen projects in GitHub according to the project screening tool and obtain the build data of the corresponding project repositories according to the GitHub REST API and web crawler technology, and extract the relevant features required for predicting the build result and time consumption in the build data;

[0007] Perform standardization processing on the relevant features, and construct a feature vector for predicting success / failure and time consumption according to the processed relevant features; generate interaction features using the Hadamard product to capture the non-linear interaction relationship between the feature vectors for predicting success / failure and time consumption, and use LSTM to extract sequence dependencies to capture the temporal dependency relationship of the historical build-related features in the feature vectors for predicting success / failure and time consumption, obtaining the model input features;

[0008] Construct a build result and time consumption prediction model; the result and time consumption prediction model includes an input layer, a graph attention network module, a temporal convolutional module, and a multi-task output layer; standardize the model input features in the input layer to obtain the standardized features; model the dynamic association between the standardized features in the graph attention network module to obtain a dynamic association vector; perform feature extraction on the standardized features in the temporal convolution to obtain a temporal feature vector;

[0009] In the multi-task output layer, perform prediction of build success / failure and time consumption on the vector obtained by concatenating the dynamic association vector and the temporal feature vector;

[0010] Train the result and time-consuming prediction model according to a preset loss function to obtain a trained result and time-consuming prediction model; use the trained result and time-consuming prediction model to predict the construction result.

[0011] Beneficial effects: In view of the lack of dedicated datasets in GitHub Actions, this application uses a project screening tool, GitHub REST API, and web crawlers to obtain construction data and build a dedicated dataset. By extracting relevant features for predicting construction results and time consumption, specific requirements are met, laying a solid data foundation for accurate prediction and making up for the defect that existing data cannot fit the construction prediction of GitHub Actions. In the feature processing link, relevant features are standardized, Hadamard products are used to generate interaction features to capture non-linear relationships, and LSTM is used to extract sequence dependencies and mine time series features, making the input features of the model more representative. The constructed prediction model includes an input layer, a graph attention network module, a temporal convolutional module, and a multi-task output layer. Each module works together. The input layer is secondarily standardized, the graph attention network models dynamic associations, the temporal convolution extracts time series features, and the multi-task output layer uses task correlations to improve the overall prediction accuracy. At the same time, the first loss function dynamically balances the data distribution, overcomes the impact of data imbalance on accuracy, greatly improves the prediction accuracy, and solves the problem of large prediction deviation of traditional methods. At the same time, the multi-task temporal graph network MT-TGN proposed in this application supports incremental training. In the real-time scenario of CI / CD, it can quickly update the model based on newly constructed data and adapt to changes in the software development process in a timely manner. It is more suitable for actual applications than traditional models, effectively reducing the waste of time, manpower, and computing costs caused by construction failures, and improving software development efficiency. At the same time, by fully analyzing various influencing factors of construction results and time consumption, covering dimensions such as current construction, historical construction, configuration files, and repositories, it is possible to more accurately identify potentially failed constructions and long-time-consuming constructions, significantly improving the efficiency of software continuous integration and continuous deployment construction, and helping developers identify potential problems in advance and optimize the development process. Description of the Drawings

[0012] Figure 1 It is a schematic flowchart of a continuous integration construction result prediction method that integrates multi-dimensional features and temporal dependencies in an embodiment;

[0013] Figure 2 It is an internal structure diagram of a computer device in an embodiment. Detailed Implementation Manner

[0014] In order to make the purpose, technical solutions, and advantages of this application clearer, the following further elaborates on this application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0015] In one embodiment, as Figure 1 shown, a method for predicting continuous integration build results by integrating multi-dimensional features and temporal dependencies is provided, including the following steps:

[0016] Step 102, screen the projects in GitHub according to the project screening tool, obtain the build data of the corresponding project repositories according to the GitHub REST API and web crawler technology, and extract the relevant features required for predicting the build results and time consumption from the build data.

[0017] As pointed out in the background art, the existing continuous integration and continuous deployment (CI / CD) build prediction technologies mainly rely on the Travis Torrent dataset, which is not a dedicated dataset for GitHub Actions (GHA), and there is a lack of publicly available GHA build datasets, resulting in the inability of the existing technologies to meet the specific requirements for GHA build prediction in terms of data.

[0018] This application screens the projects in GitHub according to the project screening tool, obtains the build data of the corresponding project repositories according to the GitHub REST API and web crawler technology, and in this way constructs a customized dataset specifically for GitHub Actions, solving the problem of the lack of publicly available build datasets suitable for GitHub Actions and providing basic data support for subsequent accurate prediction. Extracting the relevant features required for predicting the build results and time consumption from the build data ensures that the features included in the dataset are closely related to predicting the GitHub Actions build results and time consumption, meeting the specific requirements and making up for the defect that the existing datasets cannot fully meet the requirements.

[0019] Step 104, perform standardization processing on the relevant features, construct feature vectors for predicting success / failure and time consumption based on the processed relevant features; generate interaction features using the Hadamard product to capture the non-linear interaction relationship between the feature vectors for predicting success / failure and time consumption, and use LSTM to extract sequence dependencies to capture the temporal dependency relationship of the historical build-related features in the feature vectors for predicting success / failure and time consumption, obtaining the model input features.

[0020] Performing standardization processing on the relevant features can eliminate the influence of the dimension and scale between different features, enabling the model to treat each feature more fairly, contributing to the convergence of the model, and improving the stability and accuracy of the prediction.

[0021] In the actual construction process, there may be complex non-linear relationships between various features, and traditional machine learning methods often struggle to capture these relationships. Generating interaction features through the Hadamard product can uncover the potential connections between features, provide richer information for the model, and thus improve the prediction accuracy. Considering that CI / CD construction data has sequential characteristics, input features related to time, such as the result of the previous construction, have an important impact on the current construction result. LSTM (Long Short-Term Memory network) can handle sequential data well and capture the temporal information in the data, which traditional machine learning methods lack, thereby improving the accuracy of predicting the construction result and duration.

[0022] Step 106, construct a construction result and duration prediction model; the result and duration prediction model includes an input layer, a graph attention network module, a temporal convolutional module, and a multi-task output layer; standardize the model input features in the input layer to obtain standardized features; model the dynamic associations between the standardized features in the graph attention network module to obtain a dynamic association vector; extract features from the standardized features in the temporal convolution to obtain a temporal feature vector.

[0023] Standardizing the model input features in the input layer to obtain standardized features further ensures the standardization of the input data and is beneficial to the processing of subsequent modules. The graph attention network can dynamically allocate attention weights according to the importance between features, better capture the dynamic associations between features, which is very important for understanding the complex relationships between various factors in the construction process and helps improve the prediction accuracy. The temporal convolutional module (TCN) consists of 3 layers of dilated convolution (dilation factor is 2 l , l = 1, 2, 3), each layer contains a residual connection, and the output temporal features extract features from the standardized features in the temporal convolution to obtain a temporal feature vector. Temporal convolution can further extract the temporal features in the data, combined with the temporal dependencies captured by the previous LSTM, to more comprehensively mine the temporal information in the data and improve the prediction accuracy.

[0024] Step 108, in the multi-task output layer, perform predictions on whether the construction is successful and the duration using the vector obtained by concatenating the dynamic association vector and the temporal feature vector.

[0025] In the multi-task output layer, perform predictions on whether the construction is successful and the duration using the vector obtained by concatenating the dynamic association vector and the temporal feature vector. By means of multi-task learning, predicting both whether the construction is successful and the duration simultaneously can utilize the correlation between the two tasks to promote each other and improve the overall prediction accuracy.

[0026] Step 110: Train the result and time-consuming prediction model according to a pre-set loss function to obtain a trained result and time-consuming prediction model; use the trained result and time-consuming prediction model to perform construction result prediction.

[0027] Train the result and time-consuming prediction model according to a pre-set loss function. The first loss function can dynamically balance the data distribution. In the actual construction dataset, there may be an imbalance in the number of successful and failed construction samples. The weight of easily classified samples can be reduced, and the weight of difficult-to-classify samples can be increased, so as to improve the performance of the model on the imbalanced dataset, reduce prediction bias, and improve prediction accuracy.

[0028] The model needs to predict both whether the construction is successful and the time-consuming. The multi-task joint loss function integrates the losses of these two tasks, enabling the model to comprehensively consider the performance of the two tasks during training. By optimizing the joint loss to adjust the parameters, the model can achieve better performance on both tasks, avoiding the situation where one task performs well but the other is affected. There is a certain correlation between the two tasks of whether the construction is successful and the time-consuming. For example, complex code may be both prone to construction failure and take a long time. The multi-task joint loss function can utilize this correlation to achieve the transfer and sharing of knowledge between the two tasks, enabling the model to learn more comprehensive information, thereby improving the prediction accuracy of each task. By designing an appropriate form of the multi-task joint loss function, such as setting weight parameters according to task importance or data characteristics, the proportion of each task in the joint loss can be flexibly adjusted. In the training of this model, the attention degree for construction result prediction and time-consuming prediction can be dynamically adjusted according to actual needs, ensuring the balanced development of the model on different tasks and improving the overall prediction effect.

[0029] Use the trained result and time-consuming prediction model to perform construction result prediction. At the beginning of each construction, the features of the current construction are extracted in real time and input into the trained model. The model outputs a prediction of whether the construction is successful (failure probability) and a prediction of the time-consuming (estimated time), and the results are fed back to the developer or the automation system. When there is new construction data, the model can be quickly updated and adjusted to adapt to the new situation and meet the real-time requirements of CI / CD. This is an advantage that traditional models (such as random forests) do not have and can be better applied to the actual software development process.

[0030] The above-mentioned continuous integration build result prediction method that integrates multi-dimensional features and temporal dependencies. In view of the lack of dedicated datasets in GitHub Actions in this application, a project screening tool, GitHub REST API, and web crawlers are used to obtain build data and construct an exclusive dataset. By extracting relevant features of the predicted build results and elapsed time to meet specific requirements, a solid data foundation for accurate prediction is laid, making up for the defect that existing data cannot fit the build prediction of GitHub Actions. In the feature processing stage, relevant features are standardized, Hadamard product is used to generate interaction features to capture non-linear relationships, and LSTM is used to extract sequence dependencies and mine temporal features, making the input features of the model more representative. The constructed prediction model includes an input layer, a graph attention network module, a temporal convolutional module, and a multi-task output layer. Each module works together. The input layer performs secondary standardization, the graph attention network models dynamic associations, the temporal convolutional layer extracts temporal features, and the multi-task output layer uses task correlations to improve the overall prediction accuracy. At the same time, the first loss function dynamically balances the data distribution, overcomes the impact of data imbalance on accuracy, greatly improves the prediction accuracy, and solves the problem of large prediction deviation of traditional methods.

[0031] The proposed multi-task temporal graph network MT-TGN supports incremental training. In the real-time CI / CD scenario, it can quickly update the model based on new build data and adapt to changes in the software development process in a timely manner. It is more suitable for practical applications than traditional models, effectively reducing the waste of time, manpower, and computing costs caused by build failures, and improving software development efficiency. At the same time, it fully analyzes various influencing factors of build results and elapsed time, covering dimensions such as the current build, historical builds, configuration files, and repositories, and can more accurately identify potentially failed builds and long-elapsed builds, significantly improving the build efficiency of software continuous integration and continuous deployment, and helping developers identify potential problems in advance and optimize the development process.

[0032] In one of the embodiments, projects in GitHub are screened according to a project screening tool, and build data of the corresponding project repositories are obtained according to GitHub REST API and web crawler technology, including:

[0033] Set project screening criteria, use the number of stars of the project as an indicator of project popularity; consider projects on the main branch as an indicator of project originality, and exclude project repositories on the fork branch; use the number of builds of the project as a measurement standard, and only project repositories with the number of builds reaching a certain standard are required; consider the collaborative development of the project, and the project needs to have a certain number of contributors and commits. According to the project screening criteria, the build data of the corresponding project repositories are obtained by means of GitHub REST API and web crawler technology.

[0034] In a specific embodiment, suitable projects are screened out from a large number of open-source software projects. According to the research content of the present invention and with reference to relevant literature, several screening criteria are initially determined, including: 1) Considering the popularity of the project, using the number of stars of the project as an indicator of project popularity; 2) Considering the originality of the project, only considering the projects in the main branch and excluding the project repositories in the forked branches; 3) Considering the development history of the project, using the number of builds of the project as a measurement criterion and only considering the project repositories with the number of builds reaching a certain standard; 4) Considering the collaborative development of the project, the project needs to have a certain number of contributors and commits. According to the above screening criteria, the build data of the corresponding project repositories is obtained by means of the GitHub REST API and web crawler technology.

[0035] In one of the embodiments, the relevant features include historical build-related features, current build-related features, configuration file-related features, and repository-related features.

[0036] In a specific embodiment, the data in the dataset obtained in the previous step is analyzed. The dataset contains a large amount of build information. First, the entries with failed build results should be selected, and then the faults existing therein are analyzed. Then, according to the characteristics of the build, the feature information required for model prediction should be extracted. According to the research content of the present invention and with reference to relevant documents, four types of build features are initially determined: 1) Historical build-related features; 2) Current build-related features; 3) Build configuration-related features; 4) Repository-related features where the build belongs. The relevant features are shown in Table 1.

[0037]

[0038] In one of the embodiments, the graph attention network module models the dynamic association between the standardized features to obtain a dynamic association vector, including:

[0039] The attention weight calculation formula for the graph attention network module to model the dynamic association between the standardized features to obtain a dynamic association vector is:

[0040] ;

[0041] Among them, represents the importance degree of node j to node i , represents a learnable weight vector, represents a linear transformation matrix, represents the standardized feature, represents the neighbor node, represents node i 's neighbor node set, denotes the activation function, and || denotes the concatenation operation.

[0042] In one embodiment, the multi-task output layer includes a success / failure prediction module and a time-consuming prediction module; the success / failure prediction module includes a fully connected layer and a Softmax layer; the time-consuming prediction module includes a fully connected layer and a ReLU.

[0043] In a specific embodiment, the success / failure prediction module includes a fully connected layer (192→64→2) + Softmax, which outputs the failure probability.

[0044] The time-consuming prediction module includes a fully connected layer (192→64→1) + ReLU, which outputs the estimated time consumption value.

[0045] In one embodiment, in the multi-task output layer, success / failure prediction and time-consuming prediction are performed on the vector obtained by concatenating the dynamic association vector and the time series feature vector, including:

[0046] In the multi-task output layer, success / failure prediction and time-consuming prediction are performed on the vector obtained by concatenating the dynamic association vector and the time series feature vector. The failure probability is output by processing the concatenated vector according to the success / failure prediction module, and the estimated time consumption value is output by processing the concatenated vector according to the time-consuming prediction module.

[0047] In one embodiment, the first loss function is:

[0048] ;

[0049] where represents the predicted probability of the model for the s-th sample, represents the sample set, is used to adjust the weight of the loss function, is the focal parameter.

[0050] In one embodiment, the multi-task joint loss function is:

[0051] ;

[0052] where represents the loss function value of the regression task, represents the loss function value of the classification task, represents the control parameter, with a value of 0.1, used to control the speed at which the weight coefficient λ tilts towards the classification task, represents the number of training epochs.

[0053] It should be understood that although Figure 1The steps in the flowchart are sequentially shown according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least some of the other steps or sub-steps or stages of other steps.

[0054] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in Figure 2 . The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for predicting the continuous integration build results by fusing multi-dimensional features and temporal dependencies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device may be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0055] Those skilled in the art can understand that Figure 2 The structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component layout.

[0056] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0057] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0058] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for predicting the results of continuous integration builds by fusing multi-dimensional features and temporal dependencies, characterized in that, The method comprises: Filter projects in GitHub using the project screening tool, obtain the build data of the corresponding project repository using the GitHub REST API and web crawler technology, and extract relevant features from the build data required to predict build results and time consumption; The related features are standardized, and feature vectors of success or failure and time-consuming prediction are constructed according to the processed related features; interactive features are generated using Hadamard product to capture the nonlinear interactive relationship between the feature vectors of success or failure and time-consuming prediction, and sequence dependency is extracted using LSTM to capture the temporal dependency of historically constructed related features in the feature vectors of success or failure and time-consuming prediction, so as to obtain model input features; Constructing a result and time-consuming prediction model; the result and time-consuming prediction model includes an input layer, a graph attention network module, a temporal convolution module and a multi-task output layer; standardizing the model input features at the input layer to obtain standardized features; modeling the dynamic associations between the standardized features in the graph attention network module to obtain a dynamic association vector; extracting features from the standardized features in the temporal convolution to obtain a temporal feature vector; In the multi-task output layer, the vector obtained by splicing the dynamic association vector and the time series feature vector is used to predict the success or failure of the construction and the time consumption; The result and time-consuming prediction model is trained according to a preset loss function to obtain a trained result and time-consuming prediction model; and the trained result and time-consuming prediction model is used to perform result prediction; The preset loss function includes a first loss function and a multi-task joint loss function; the first loss function is: Among them, represents the predicted probability of the model for the s-th sample, represents the sample set, represents the node j for the node i importance, used to adjust the weight of the loss function, is the focus parameter; The multi-task joint loss function is: Among them, represents the loss function value of the regression task, represents the loss function value of the classification task, represents a control parameter with a value of 0.1, used to control the λ speed at which the weight coefficient tilts towards the classification task, represents the number of training epochs.

2. The method according to claim 1, characterized in that Use the project screening tool to filter projects in GitHub and obtain the build data of the corresponding project repository based on the GitHub REST API and web crawler technology, including: Set project screening criteria and use the number of stars of the project as an indicator of project popularity; consider the main branch project as an indicator of project originality and exclude the project repositories of forked branches; use the number of project builds as a measurement standard and only require project repositories with a certain number of builds that meet certain standards; consider the collaborative development of the project, and the project needs to have a certain number of contributors and commits. According to the project screening criteria, use the GitHub REST API and web crawler technology to obtain the build data of the corresponding project repository.

3. The method according to claim 1, wherein The related features include historical build related features, current build related features, configuration file related features and belonging warehouse related features.

4. The method according to claim 1, wherein Modeling the dynamic association between the standardized features in the graph attention network module to obtain a dynamic association vector includes: The dynamic association between the standardized features is modeled in the graph attention network module, and the attention weight calculation formula of the dynamic association vector is obtained as follows: Among them, represents a node j for the node i importance degree of represents a learnable weight vector represents a linear transformation matrix represents the normalized feature represents a neighbor node represents a node i set of neighbor nodes of represents an activation function, || represents a concatenation operation 5. The method according to claim 1, characterized in that The multi-task output layer includes a success prediction module and a time-consuming prediction module; the success prediction module includes a fully connected layer and a Softmax layer; the time-consuming prediction module includes a fully connected layer and a ReLU.

6. The method according to claim 5, wherein In the multi-task output layer, the vector obtained by concatenating the dynamic association vector and the temporal feature vector is used to perform success prediction and time-consuming prediction, including: In the multi-task output layer, the vector obtained by concatenating the dynamic association vector and the temporal feature vector is used to perform success prediction and time-consuming prediction. The failure probability is output according to the processing of the concatenated vector by the success prediction module, and the time-consuming estimation value is output according to the processing of the concatenated vector by the time-consuming prediction module.

Images