Model training method and device, equipment, storage medium and computer program product

Abstract

The invention relates to a model training method and device, equipment, a storage medium and a computer program product. The method comprises the following steps: determining a similar job set of a to-be-predicted job, and training an initial prediction model corresponding to the to-be-predicted job according to the similar job set to obtain a target prediction model used for predicting the execution time of the to-be-predicted job. By adopting the method, the accuracy of operation time prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and particularly to a model training method, device, equipment, storage medium, and computer program product. Background Art

[0002] In high-performance computing, due to resource limitations, resources are usually reserved for high-priority jobs for a period of time. During this period, the reserved resources are idle resources, which will cause waste of resources. Therefore, the backfilling scheduling strategy has emerged. The backfilling scheduling strategy is to execute low-priority jobs through idle resources without delaying the execution of any high-priority jobs, so as to improve resource utilization. This requires more accurate prediction of the execution time of jobs to avoid affecting the execution of high-priority jobs.

[0003] Traditional methods for predicting job time usually predict the execution time of a job to be predicted based on a neural network model that has been trained for job time prediction.

[0004] However, in the above job time prediction method, the neural network model used to predict the execution time of a job has a problem of inaccurate prediction when predicting the job time. Summary of the Invention

[0005] Based on this, it is necessary to provide a model training method, device, equipment, storage medium, and computer program product that can improve the accuracy of job time prediction for the above technical problems.

[0006] In a first aspect, this application provides a model training method, including:

[0007] Determine a set of similar jobs for the job to be predicted;

[0008] Train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0009] In the above embodiment, when determining the set of similar jobs for the job to be predicted, since the jobs in the set of similar jobs are highly similar to the job to be predicted, training the initial prediction model with the jobs similar to the job to be predicted in the set of similar jobs, the obtained target prediction model is more suitable for predicting the job time of the job to be predicted. Therefore, the target prediction model trained in the embodiment of this application for the job to be predicted has higher accuracy for job time prediction.

[0010] In one of the embodiments, the method for obtaining the initial prediction model includes:

[0011] Classify multiple historical jobs according to the types of the jobs to obtain multiple sample type sets;

[0012] Train an initial neural network model according to the multiple sample type sets to obtain an initial prediction model.

[0013] In the above embodiments, the initial neural network model is trained by different types of sample type sets to obtain initial prediction models corresponding to different types of jobs, so that the initial prediction model used for predicting the execution time of the job to be predicted is determined according to the type of the job to be predicted, making the final prediction of the execution time of the job to be predicted more accurate.

[0014] In one of the embodiments, the initial neural network model includes multiple sub-network models. Training the initial neural network model according to the multiple sample type sets to obtain an initial prediction model includes:

[0015] For each sample type set, determine the goodness of fit of each sub-network model according to the sample type set;

[0016] Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each sample type set.

[0017] In the above embodiments, for each sample type set, determine the goodness of fit of each sub-network model according to the sample type set, and integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each sample type set. Since the initial prediction model is a model integrated by multiple sub-network models, compared with a single network model, the prediction of the execution time of the job to be predicted is more accurate.

[0018] In one of the embodiments, integrating the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each sample type set includes:

[0019] Sort the multiple sub-network models according to the goodness of fit;

[0020] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each sample type set.

[0021] In the above embodiments, sorting the sub-network models according to the goodness of fit can determine the sub-network model most suitable for the sample type set, group the sub-network models in sequence according to the sorting, and integrate the sub-network models in each group. The obtained initial prediction model is more suitable for predicting the execution time of the job to be predicted corresponding to the sample type set.

[0022] In one embodiment, the sorted multiple sub-network models are grouped in sequence, and the sub-network models of each group are integrated to obtain an initial prediction model corresponding to each sample type set, including:

[0023] The sorted multiple sub-network models are grouped in sequence, and the sub-network models of each group are integrated to obtain multiple integrated models;

[0024] Determine the error value of each integrated model according to the operation characteristics of each historical operation in the sample type set;

[0025] According to the error value, an initial prediction model corresponding to the sample type set is determined from multiple integrated models.

[0026] In the above embodiment, since the smaller the error value, the better the effect of the integrated model in time prediction, each integrated model is evaluated according to its error value, and the initial prediction model corresponding to the sample type set is more accurate in time prediction.

[0027] In one embodiment, the sorted multiple sub-network models are grouped in sequence, including:

[0028] A preset number of sub-models are extracted multiple times from the sorted sub-network models to obtain groups corresponding to each extracted model; wherein the preset number is related to the number of times the sub-models are extracted.

[0029] In the above embodiment, for multiple sub-network models sorted by goodness of fit, a preset number is determined according to the number of extractions, and sub-model extraction is performed according to the preset number. The sub-network models included in the extracted groups all include the sub-network models with higher rankings. The higher the ranking of the sub-network models, the greater the goodness of fit, and the better the effect of the integrated model.

[0030] In one embodiment, determining a set of similar jobs to the job to be predicted includes:

[0031] According to the type of the job to be predicted, determining a type set corresponding to the job to be predicted; the type set includes at least one historical job;

[0032] A similar job set is determined from the type set according to the job characteristics of the job to be predicted.

[0033] In the above embodiments, a type set corresponding to the job to be predicted is determined, and a set of similar jobs corresponding to the job to be predicted is determined from the type set. Since the type set is a set obtained by classifying the types of historical jobs, in this application, a set of similar jobs is determined from the type set that matches the type of the predicted job, avoiding analyzing all historical jobs, thereby reducing the amount of data calculation and also improving the efficiency of determining the set of similar jobs.

[0034] In one embodiment, the type set includes at least one candidate job set. Determining a set of similar jobs from the type set according to the job characteristics of the job to be predicted includes:

[0035] Obtaining the similarity between the job characteristics of the job to be predicted and each candidate job set;

[0036] Determining a set of similar jobs from at least one candidate job set according to the similarity.

[0037] In the above embodiments, the server determines the similarity between the job characteristics of the job to be predicted and each candidate job set, numericalizes the similarity, and the set of similar jobs determined from the candidate job set according to the similarity is more accurate, and this method is simple and easy to implement.

[0038] In one embodiment, obtaining the similarity between the job characteristics of the job to be predicted and each candidate job set includes:

[0039] Obtaining the similarity between the job characteristics of the job to be predicted and the job characteristics of the target job in each candidate job set; the target job is any job in the candidate job set.

[0040] In the above embodiments, the server selects a target job from the candidate job set, and determines the similarity between the job characteristics of the target job and the job characteristics of the job to be predicted as the similarity between the job to be predicted and each candidate job set, and the calculation efficiency is higher.

[0041] In one embodiment, obtaining the similarity between the job characteristics of the job to be predicted and the job characteristics of the target job in each candidate job set includes:

[0042] Performing numerical processing on the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job;

[0043] Obtaining the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0044] In the above embodiments, after numericalizing the job characteristics of the job to be predicted and the target job, it is convenient to calculate the similarity between the job to be predicted and the target job. Moreover, when calculating the similarity between the job to be predicted and the target job, both text job characteristics and numerical job characteristics are considered, and various characteristics between jobs can be more comprehensively referred to, making the calculated similarity more accurate.

[0045] In one of the embodiments, the method further includes:

[0046] Classify multiple historical jobs to obtain a type set;

[0047] Determine at least one candidate job set according to the similarity between the historical jobs in the type set.

[0048] In the above embodiments, the server re-partitions the obtained historical jobs, and then divides different candidate job sets according to the similarity under different type sets, which can provide rich training data for training the initial prediction model for different jobs to be predicted. Moreover, after two partitions, the job set matching the job to be predicted can be determined more accurately and quickly.

[0049] In one of the embodiments, the method further includes:

[0050] Determine the initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0051] In the above embodiments, different initial prediction models are preset for different types of jobs to be predicted. When predicting the execution time of the job to be predicted, it saves computing resources and improves prediction efficiency. Moreover, the initial prediction model matches the type of the job to be predicted, which can improve the accuracy of job time prediction.

[0052] In a second aspect, the present application provides a method for predicting job time, and the method includes:

[0053] Input the job to be predicted into the target prediction model for time prediction to obtain the execution time of the job to be predicted;

[0054] Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the similar job set of the job to be predicted.

[0055] In the above embodiments, according to the type of the job to be predicted, the initial prediction model preset for the type of the job to be predicted is determined. According to the job characteristics of the job to be predicted, the similar job set with the highest similarity to the job to be predicted is determined, and the initial prediction model is trained using the similar job set. The obtained target prediction model is trained for the job to be predicted, and the execution time prediction of the job to be predicted is more accurate.

[0056] In a third aspect, the present application further provides a model training device, including:

[0057] A determination module, configured to determine a set of similar jobs for a job to be predicted;

[0058] A training module, configured to train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0059] In a fourth aspect, the present application further provides a device for predicting job time, including:

[0060] A prediction module, configured to input the job to be predicted into the target prediction model for time prediction to obtain the execution time of the job to be predicted;

[0061] wherein, the target prediction model is obtained by training an initial prediction model corresponding to the job to be predicted according to the set of similar jobs of the job to be predicted.

[0062] In a fifth aspect, an embodiment of the present application provides a computer device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the method in the above first aspect and / or second aspect are implemented.

[0063] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method in the above first aspect and / or second aspect are implemented.

[0064] In a seventh aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method in the above first aspect and / or second aspect are implemented.

[0065] For the above model training method, device, equipment, storage medium and computer program product, by determining a set of similar jobs for the job to be predicted, and training an initial prediction model corresponding to the job to be predicted according to the set of similar jobs, a target prediction model for predicting the execution time of the job to be predicted is obtained. By determining the set of similar jobs for the job to be predicted, since the jobs in the set of similar jobs are highly similar to the job to be predicted, training the initial prediction model with the jobs similar to the job to be predicted in the set of similar jobs, the obtained target prediction model is more suitable for predicting the job time of the job to be predicted. Therefore, the target prediction model trained for the job to be predicted in the embodiment of the present application has higher accuracy in job time prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 An application environment diagram of a model training method in an embodiment;

[0068] Figure 2 A schematic diagram of a flow chart of a model training method in one embodiment;

[0069] Figure 3 A schematic diagram of a process for obtaining an initial prediction model in another embodiment;

[0070] Figure 4 is a flow chart of step 401 in another embodiment;

[0071] Figure 5 A schematic diagram of an exemplary process of obtaining an initial prediction model in another embodiment;

[0072] Figure 6 is a flow chart of step 402 in another embodiment;

[0073] Figure 7 is a flow chart of step 602 in another embodiment;

[0074] Figure 8 is a flow chart of an exemplary integrated sub-network model in another embodiment;

[0075] Figure 9 is a schematic flow chart of step 201 in another embodiment;

[0076] Figure 10 is a flow chart of step 902 in another embodiment;

[0077] Figure 11 A schematic diagram of a process of obtaining similarities between job features of a job to be predicted and job features of a target job in each candidate job set in another embodiment;

[0078] Figure 12 A schematic diagram of a process for classifying historical jobs in another embodiment;

[0079] Figure 13 A schematic diagram of an exemplary process of classifying historical jobs in another embodiment;

[0080] Figure 14Schematic flowchart of an exemplary job time prediction method in an embodiment;

[0081] Figure 15 Block diagram of the structure of a model training device in an embodiment;

[0082] Figure 16 Block diagram of the structure of a job time prediction device in an embodiment;

[0083] Figure 17 Internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0084] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0085] HPC (High Performance Computing) refers to using a large number of processors to process large-scale data in a high-speed parallel manner. The typical example of an HPC system is a supercomputer.

[0086] In an HPC environment, job clusters are widely used to manage high-performance computing tasks. A job cluster is a distributed computing model that allows users to submit large-scale computing tasks, divide these computing tasks into different jobs, and process the computing tasks submitted by users at high speed by executing jobs in a high-speed parallel manner.

[0087] With the development of HPC, the high-performance computing tasks it faces, such as industrial computing, scientific computing, intelligent computing, etc., have an increasing job volume. In high-performance computing, due to resource limitations, resources are usually reserved for high-priority jobs for a period of time. During this period, the reserved resources are idle resources, which will cause waste of resources. At this time, a scheduling strategy that can reasonably allocate computing resources and efficiently manage jobs appears, the backfill scheduling strategy.

[0088] The backfill scheduling strategy is to backfill low-priority jobs to idle resources for execution without delaying the execution of any high-priority jobs, so as to improve resource utilization rate, which is the most direct and effective method to improve resource utilization rate.

[0089] However, the backfill scheduling strategy is restricted by the predicted execution time corresponding to the job at the time of submission when backfilling jobs. In order to avoid the failure of job backfilling and resulting in low resource utilization rate of the system, it is necessary to accurately predict the execution time of jobs.

[0090] In some scenario embodiments, there are a variety of job time prediction methods, which can be roughly classified into three categories:

[0091] 1. By mining historical job information, a fixed number of historical jobs similar to the characteristics of the job to be predicted are selected, and the average value of the actual execution times of each historical job is calculated, and the average value is used as the predicted execution time of the job to be predicted. However, the number of jobs to be predicted selected by this method is limited, and the method of calculating the average value has a large error, and the accuracy of the predicted time is low.

[0092] 2. Design a specific algorithm for predicting the job time for jobs of a specific application type. However, this specific method is not applicable to all types of jobs, and the application scenario is limited.

[0093] 3. Predict the execution time of the job to be predicted through a trained neural network model. Compared with the above two methods for predicting job time, this method has higher accuracy in predicting the execution time of the job to be predicted, and the application scenario is not limited. However, the neural network model is obtained by training based on ordinary jobs. Although it is applicable to most jobs, the predicted time still has the problem of low accuracy.

[0094] In view of this, the embodiments of the present application provide a model training method, device, device, storage medium, and computer program product, determine a set of similar jobs for the job to be predicted, and train the initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model for predicting the execution time of the job to be predicted. Determining the set of similar jobs for the job to be predicted, since the jobs in the set of similar jobs are highly similar to the job to be predicted, training the initial prediction model according to the jobs similar to the job to be predicted in the set of similar jobs, the obtained target prediction model is more suitable for predicting the job time of the job to be predicted. Therefore, the target prediction model trained for the job to be predicted in the embodiments of the present application has higher accuracy in predicting the job time.

[0095] The model training method provided by the embodiments of the present application can be applied to, for example Figure 1 the application environment shown. The data storage system can store the data that the server 102 needs to process. The data storage system can be integrated on the server 102, or placed in the cloud or other network servers. The server 102 can be implemented by an independent server or a server cluster composed of multiple servers.

[0096] In an exemplary embodiment, as Figure 2 shown, a model training method is provided. Taking the method applied to the Figure 1 server 102 in as an example for description, it includes the following steps 201 and 202.

[0097] Step 201, determine a set of similar jobs for the job to be predicted.

[0098] Among them, the job to be predicted, that is, the job for which time prediction is required. When the server predicts the job to be predicted, it is necessary to determine a set of similar jobs for the job to be predicted. The set of similar jobs contains at least one job, and the similarity between the jobs contained in the set of similar jobs and the job to be predicted is greater than a preset threshold. That is, the jobs contained in the set of similar jobs and the job to be predicted are similar jobs with a relatively high similarity.

[0099] In the embodiments of the present application, the server can obtain multiple historical jobs, calculate the similarity between the job to be predicted and the historical jobs, and form a set of similar jobs with the jobs whose similarity is greater than the preset threshold; optionally, the server can also pre-set multiple job sets according to a large number of historical jobs, and determine the set of similar jobs for the job to be predicted from the multiple job sets. For example, for each job set, select any job from it, calculate the similarity between the job to be predicted and this job, and use this similarity as the similarity between the job to be predicted and each job set. Sort the similarities of each job set, and use the job set with the highest similarity as the set of similar jobs for the job to be predicted. Or, for each job set, calculate the similarity between the job to be predicted and each job in the job set, add or weighted sum the similarities corresponding to each job to obtain the similarity between the job to be predicted and each job set, and use the job set with the highest similarity as the set of similar jobs.

[0100] Optionally, regarding the similarity between the job to be predicted and the job, it can be evaluated by methods such as Euclidean distance and Jaccard similarity coefficient. The server can calculate the similarity between the two based on the characteristics of the job to be predicted and the job. The higher the similarity, the higher the similarity between the job to be predicted and this job.

[0101] Step 202: Train the initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model.

[0102] Among them, the initial prediction model can be a basic prediction model trained using a large number of historical jobs. This initial prediction model can be a single regression model. Or, in the embodiments of the present application, since the ensemble model has better accuracy, is not prone to overfitting, is insensitive to outliers, and has a fast training speed compared to a single model, the initial prediction model can also be an ensemble model composed of multiple regression models. The selection of the multiple regression models that make up the ensemble model can be set according to the application environment. For example, appropriate regression models can be selected using algorithms such as Adaptive Boosting, Support Vector Regression, Random Forest Regression, Bayesian Ridge Regression, Elastic Net, Decision Tree Regression, Gradient Boosting Regression Tree, and K-Nearest Neighbor Regression.

[0103] Optionally, the initial prediction model is a pre-trained model, and the server can directly train the initial prediction model according to the set of similar jobs of the job to be predicted; optionally, multiple candidate models can be pre-trained, and the server can determine the candidate model corresponding to the job to be predicted from the multiple candidate models as the initial prediction model according to the characteristics or type of the job to be predicted.

[0104] In the embodiments of the present application, the server can train the initial prediction model according to the set of similar jobs. Since the jobs in the set of similar jobs are all jobs determined to be similar to the job to be predicted for the job to be predicted, therefore, training the initial prediction model based on the set of similar jobs can obtain a target prediction model for the job to be predicted, and this target prediction model is specifically used to predict the execution time of the job to be predicted.

[0105] In a possible implementation manner, for each job in the set of similar jobs, the server extracts the job characteristics of each job, and iteratively trains the initial prediction model according to the characteristic values of each job. For the process of one iterative training, the server randomly selects a job from the set of similar jobs, and inputs the job characteristics of this job into the initial prediction model to obtain the predicted execution time output. According to a preset loss function, calculate the loss value between the predicted execution time and the actual execution time of this job, and adjust the parameters of the initial prediction model according to this loss value to obtain an intermediate prediction model. Traverse all the jobs in the similar jobs according to the above process, and train the initial prediction model multiple times. When the number of training times reaches a preset number threshold, and / or the loss value between the predicted execution time output by the intermediate prediction model and the actual execution time of the job is less than a certain loss threshold, the training ends, and this intermediate prediction model is used as the target prediction model. The above is an exemplary introduction to the process of the server training the initial prediction model, and the embodiments of the present application do not limit the specific training process of the server for the initial prediction model.

[0106] The model training method provided by the embodiments of the present application determines a set of similar jobs for the job to be predicted, and trains the initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model for predicting the execution time of the job to be predicted. Determining the set of similar jobs for the job to be predicted, since the jobs in the set of similar jobs are highly similar to the job to be predicted, training the initial prediction model according to the jobs similar to the job to be predicted in the set of similar jobs, the obtained target prediction model is more suitable for predicting the job time of the job to be predicted. Therefore, the target prediction model trained for the job to be predicted in the embodiments of the present application has higher accuracy in predicting the job time.

[0107] In one embodiment, based on Figure 2For the illustrated embodiment, refer to Figure 3 , this embodiment relates to the process of obtaining an initial prediction model. As Figure 3 shown, this process may include step 301 and step 302.

[0108] Step 301, classify multiple historical jobs according to the type of the jobs to obtain multiple sample type sets.

[0109] In the embodiments of the present application, the server presets different initial prediction models according to different types. Therefore, before determining the initial prediction models corresponding to each type, it is necessary to first determine the sample data used to determine the initial prediction models corresponding to each type, that is, the sample type sets. In the present application, each sample type set includes multiple historical jobs, and the types of each historical job are the same.

[0110] Step 302, train an initial neural network model according to multiple sample type sets to obtain an initial prediction model.

[0111] Different types correspond to different initial prediction models. Regarding the process of obtaining the initial prediction model, in the embodiments of the present application, the server trains the initial neural network model according to the sample sets corresponding to each type.

[0112] In the embodiments of the present application, the initial neural network model may include various regression models, such as an adaptive boosting model, a support vector regression model, a random forest regression model, a Bayesian ridge regression model, an elastic net model, a decision tree regression model, a gradient boosting regression tree model, a K-nearest neighbor regression model, etc. The server may train different neural network models according to the sample type sets corresponding to each type, so as to determine the initial prediction models corresponding to each type. The obtained initial prediction model may be a single model or an integrated model integrated by multiple models.

[0113] It can be understood that the initial prediction models corresponding to each type are not fixed. As time goes by, more and more jobs are executed, the scale of historical jobs expands, and the characteristic attributes of the jobs to be predicted will also be updated. In order to enable the initial prediction model to adapt to the specific application scenario, it is necessary to continuously update the initial prediction model. For example, periodically obtain the recently executed historical jobs, re-divide the sample type sets according to the newly obtained historical jobs, and thus update the initial prediction model according to the newly divided sample type sets to obtain an initial prediction model that is more suitable for the current application scenario.

[0114] In the above embodiments, the initial neural network model is trained with different types of sample type sets to obtain initial prediction models corresponding to different types of jobs, so that the initial prediction model used for predicting the execution time of the job to be predicted is determined according to the type of the job to be predicted, making the final prediction of the execution time of the job to be predicted more accurate.

[0115] In one embodiment, based on the above Figure 3 illustrated embodiment, refer to Figure 4 , this embodiment relates to the process of training the initial neural network model according to multiple sample type sets to obtain the initial prediction model. As Figure 4 shown, step 301 may include step 401 and step 402.

[0116] Step 401, for each sample type set, determine the goodness of fit of each sub-network model according to the sample type set.

[0117] In the embodiments of the present application, the initial neural network model includes multiple sub-network models, and the sub-network models may be regression models, including adaptive boosting models, support vector regression models, random forest regression models, Bayesian ridge regression models, elastic net models, decision tree regression models, gradient boosting regression tree models, K-nearest neighbor regression models, etc.

[0118] In the embodiments of the present application, the server may determine the adaptation degree of each sub-network model to different sample type sets according to the goodness of fit. The specific calculation process is as follows:

[0119] Suppose there are a total of n sample type sets {type1, type2,..., type n} and m sub-network models {model1, model2,..., model m}.

[0120] For type1 in a sample type set, calculate the goodness of fit corresponding to a sub-network model model1 based on type1. The server inputs the numerical features of each historical job in type1 (including the number of CPUs requested by the job, the amount of memory requested by the job, the number of nodes requested by the job, and the execution time limit requested by the job) into model1 respectively. Model1 outputs the predicted execution time of each historical job. Substitute the predicted execution time and the actual execution time corresponding to each historical job into formula (1) to obtain the goodness of fit corresponding to the sub-network model model1 calculated based on the sample type set type1.

[0121]

[0122] Among them, ESS is the sum of squared residuals, and the calculation formula is shown in Formula (2):

[0123]

[0124] Among them, there are n historical jobs in this sample type set, and y 预测执行时间 is the predicted execution time corresponding to the i-th historical job, and y 实际执行时间 is the actual execution time corresponding to the i-th historical job.

[0125] TSS is the total sum of squares, and the calculation formula is shown in Formula (3):

[0126]

[0127] Among them, is the mean of the actual execution times of each historical executed job.

[0128] RSS is the residual sum of squares, and the calculation formula is shown in Formula (4):

[0129]

[0130] Among them, the larger the goodness of fit, the more suitable the sub-network model model1 is for this sample type set type1.

[0131] According to the above embodiments, the server can obtain the goodness of fit of each sub-network model corresponding to each sample type set, which can be

[0132]

[0133] Step 402: Integrate multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each sample type set.

[0134] Compared with a single sub-network model, an integrated model that combines multiple sub-network models has advantages such as better accuracy, being less prone to overfitting, being insensitive to outliers, and having a fast training speed. Therefore, in the embodiments of the present application, for a type of sample type set, the server can select at least two sub-network models according to the goodness of fit for integration to obtain an initial prediction model corresponding to this type of sample type set.

[0135] For a set of sample types of a certain type, regarding the method by which the server selects multiple sub-network models for integrating the initial prediction model corresponding to the set of sample types according to the goodness of fit. Optionally, since the larger the goodness of fit, the more suitable the sub-network model is for the set of sample types, a goodness-of-fit threshold can be preset in advance, and multiple sub-network models with a goodness of fit greater than the threshold are selected for integration. Optionally, the server sorts the sub-network models from largest to smallest according to their goodness of fit. If the preset number of sub-network models for integration is 3, the server selects the top 3 sub-network models for integration.

[0136] Exemplarily, after the server determines multiple sub-network models for integrating the initial prediction model corresponding to a set of sample types, the sub-network models can be integrated by various methods. Optionally, the server integrates the multiple sub-network models based on the Stacking hierarchical model integration framework, that is, the output of one sub-network model is used as the input of another sub-network model, nested layer by layer, and the output of the last-layer sub-network model is used as the prediction execution time. Optionally, the server integrates the sub-network models based on the Blending hybrid strategy, that is, the outputs of multiple sub-network models are weighted and averaged to obtain the final output result as the prediction execution time. Optionally, the server integrates the sub-network models based on Bagging (Bootstrap aggregation). Regarding the specific method of integrating each sub-network model, no specific limitation is made here.

[0137] Refer to Figure 5 , and an exemplary introduction to the process of obtaining the initial prediction model is given:

[0138] Suppose there are a total of n sets of sample types {type1, type2, ……, typen} and m sub-network models {model1, model2, ……, modelm}; for different sets of sample types, calculate the goodness of fit corresponding to each sub-network model to obtain the goodness of fit of each sub-network model corresponding to type1 The goodness of fit of each sub-network model corresponding to type2 ……, the goodness of fit of each sub-network model corresponding to type n Integrate the sub-network models according to the goodness of fit to obtain the initial prediction models {IM1, IM2, ……, IMn} corresponding to each set of sample types.

[0139] In the above embodiments, for each set of sample types, the goodness of fit of each sub-network model is determined according to the set of sample types, and multiple sub-network models are integrated according to the goodness of fit to obtain an initial prediction model corresponding to each set of sample types. Since the initial prediction model is a model integrated from multiple sub-network models, compared with a single network model, the prediction of the execution time of the job to be predicted is more accurate.

[0140] In one embodiment, based on the above Figure 4 illustrated embodiment, refer to Figure 6 , this embodiment relates to the process of integrating multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each set of sample types. As Figure 6 illustrated, step 402 may include step 601 and step 602.

[0141] Step 601, sort multiple sub-network models according to the goodness of fit.

[0142] Step 602, group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each set of sample types.

[0143] Among them, the larger the goodness of fit, the more suitable the corresponding sub-network model is for the set of sample types. In the embodiments of the present application, the server may sort multiple sub-network models in descending order of the goodness of fit. The higher the ranking, the more suitable the sub-network model is for the set of sample types.

[0144] In a possible implementation manner, the preset number of sub-network models for integration is 3. Then the server selects the top 3 sub-network models ranked in the sorting for grouping and integrates them to obtain an initial prediction model.

[0145] In another possible implementation manner, the server groups the sorted multiple sub-network models in sequence, and integrates the sub-network models in each group. For example, the first time, the top two sub-network models are selected for grouping, the second time, the top three sub-network models are selected for grouping, and so on. Finally, all sub-network models are grouped, and the sub-network models in each group are integrated to obtain multiple candidate network models. The server evaluates each candidate network model and selects an initial prediction model from each candidate network model according to the evaluation results.

[0146] In this embodiment, regarding the method for evaluating candidate network models, optionally, a plurality of historical jobs are randomly selected from the sample type set, the job characteristics of each historical job are input into each candidate network model, and according to the output results of each candidate network and the actual execution time of each historical job, the accuracy of the output results of each candidate network model is calculated, and the candidate network model with the highest accuracy is selected as the initial prediction model; optionally, according to the historical jobs in the sample type set, the error value of each candidate network model is calculated. The smaller the error value, the more accurate the prediction result of the candidate network model. The candidate network model with the smallest error value is selected as the initial prediction model. The error value can be prediction error (PE), mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), weighted absolute error (WAE), mean absolute percentage error (MAPE), etc. Here, the method for evaluating candidate network models is not specifically limited.

[0147] In the above embodiment, the sub-network models are sorted according to the goodness of fit, the sub-network model most suitable for the sample type set is determined, the sub-network models are grouped in sequence according to the sorting, and the sub-network models in each group are integrated. The obtained initial prediction model is more suitable for predicting the execution time of the to-be-predicted jobs corresponding to the sample type set.

[0148] In one embodiment, based on the above Figure 6 illustrated embodiment, refer to Figure 7 , this embodiment relates to the process of sequentially grouping the sorted multiple sub-network models and integrating the sub-network models in each group to obtain the initial prediction model corresponding to each sample type set. As Figure 7 shown, step 602 may include steps 701 to 703.

[0149] Step 701, sequentially group the sorted multiple sub-network models, and integrate the sub-network models in each group to obtain multiple integrated models.

[0150] Regarding the method for the server to sequentially group the sorted multiple sub-network models, optionally, the preset number of sub-network models for integration is 3, then the server selects the top 3 sub-network models ranked in the sorting for grouping; optionally, the server sequentially groups the sorted multiple sub-network models. For example, the first time, the first two sub-network models are selected for grouping, the second time, the first three sub-network models are selected for grouping, and so on, until all sub-network models are grouped.

[0151] In the embodiment of the present application, the server can integrate the sub-network models in each group based on Bagging to obtain an integrated model. Refer to Figure 8 , and an exemplary introduction is made to the integration process:

[0152] For the set of sample types type1, a certain grouping corresponding to type1 is integrated. Suppose this grouping includes m sub-network models; for a sub-network model m in the grouping, the server randomly selects multiple historical jobs from the sample job set to form a sub-training set m corresponding to the sub-network model m. Based on the numerical feature X and actual execution time Y of each historical job in the sub-training set m, the sub-network model m is trained to obtain the base model m. According to this method, all sub-network models are trained to obtain multiple base models that make up the initial prediction model. The multiple base models are grouped into an integrated model. When the numerical features of a job to be predicted are input into each base model respectively, each base model will output a corresponding predicted value. The mean value of all predicted values is the predicted execution time corresponding to the job to be predicted.

[0153] Step 702: Determine the error value of each integrated model according to the job characteristics of each historical job in the sample type set.

[0154] Among them, the error value can be prediction error PE, mean absolute error MAE, mean square error MSE, root mean square error RMSE, weighted absolute error WAE, mean absolute percentage error MAPE, etc. Taking the mean absolute error value MAE as an example of the error value, the formula for calculating the MAE corresponding to the integrated model IM is shown in formula (5):

[0155]

[0156] MAE of the integrated model IM IM is the mean value of the MAEs corresponding to the m base models.

[0157] The formula for calculating the MAE of each base model is shown in formula (6):

[0158]

[0159] Exemplarily, assume that the sub-training set corresponding to the base model m is the sub-training set m, and the sub-training set m includes j historical jobs randomly selected from the sample job set. The numerical features of each historical job are input into the base model m to obtain the predicted value corresponding to the historical job. The mean value of the difference between the predicted value corresponding to each historical job and the historical execution time is the MAE corresponding to the base model m. m .

[0160] Step 703: Determine the initial prediction model corresponding to the sample type set from multiple integrated models according to the error value.

[0161] Among them, the smaller the error value, the better the corresponding integrated model performs in time prediction. In the embodiments of the present application, for a set of sample types, the integrated model with the smallest error value among the integrated models is selected as the initial prediction model corresponding to the set of sample types.

[0162] In the above embodiments, since the smaller the error value, the better the integrated model performs in time prediction. Therefore, by evaluating each integrated model according to its error value, the initial prediction model corresponding to the set of sample types determined is more accurate in time prediction.

[0163] In one embodiment, based on the above Figure 7 shown embodiment, this embodiment relates to the process of grouping multiple sorted sub-network models in sequence. Step 701 may include: performing sub-model extraction on the multiple sorted sub-network models for a preset number of times to obtain the grouping corresponding to each extracted model.

[0164] In the embodiments of the present application, sub-model extraction is performed multiple times on the sorted sub-network models, and a group of sub-network models is extracted for integration into an integrated model. The number of sub-network models included in each extracted grouping is a preset number, and the preset number is related to the number of times of sub-model extraction.

[0165] Exemplarily, there are m sub-network models in total. The preset number corresponding to the first sub-model extraction is 1, the preset number corresponding to the second sub-model extraction is 2, and so on. The preset number corresponding to the k-th sub-model extraction is k, where k is an integer greater than 1 and not greater than m. Then, every k times, the server extracts the first k sub-network models from the sorted sub-network models according to the preset number k corresponding to the sub-model extraction for grouping.

[0166] In the above embodiments, for the multiple sorted sub-network models sorted according to the goodness of fit, the preset number is determined according to the number of extractions, and sub-model extraction is performed according to the preset number. The groupings of sub-network models extracted include sub-network models that are ranked higher. The higher the ranking of the sub-network model, the greater the goodness of fit, and the better the performance of the integrated integrated model.

[0167] In the above embodiments, since the target prediction model trained using the similarity job set can more accurately predict the job time, the method for determining the similarity job set is particularly important. In one embodiment, based on the above Figure 2 shown embodiment, refer to Figure 9 , this embodiment relates to the process of determining the set of similar jobs for the job to be predicted. As Figure 9 shown, step 201 may include step 901 and step 902.

[0168] Step 901: Determine the type set corresponding to the job to be predicted according to the type of the job to be predicted.

[0169] In the embodiments of the present application, different jobs to be predicted may correspond to different types. The type is one of the basic attributes of the job to be predicted, and the type of the job to be predicted can be divided according to the application field, the function of the job, etc. When the server obtains the job to be predicted, it can determine the type corresponding to the job to be predicted. In the embodiments of the present application, the types of the prediction jobs may include molecular dynamics, meteorological image research, electronic structure and quantum mechanics, communication, artificial intelligence, which are not limited in the embodiments of the present application.

[0170] In a possible implementation manner, a large number of historical jobs can be divided into multiple type sets according to types. Each type set includes at least one historical job, and the type of each historical job corresponds to the type of the type set. For example, when the type corresponding to the type set is molecular dynamics, the types of the historical jobs included in the type set are all molecular dynamics.

[0171] Step 902: Determine the similar job set from the type set according to the job characteristics of the job to be predicted.

[0172] Among them, the job characteristics may be the attribute characteristics, function characteristics, etc. of the job to be predicted. In the embodiments of the present application, optionally, the job characteristics may include text characteristics such as user name, user group name, partition applied for by the job, node applied for by the job, working path of the job, type of the job, etc., and may also include numerical characteristics such as the number of CPUs applied for by the job, the amount of memory applied for by the job, the number of nodes applied for by the job, the execution time limit applied for by the job, which are not limited in the embodiments of the present application.

[0173] In the embodiments of the present application, the server can determine the corresponding type set according to the type of the job to be predicted. In a possible implementation manner, the server can calculate the similarity between the job to be predicted and the historical jobs according to the job characteristics of the job to be predicted, and group the historical jobs with a similarity greater than the preset similarity threshold into a similar job set; in another possible implementation manner, each type set includes at least one candidate job set, and each candidate job set includes at least one historical job. The server can calculate the similarity with the candidate job set according to the job characteristics of the job to be predicted, so as to determine the similar job set corresponding to the job to be predicted.

[0174] Exemplarily, for a set of candidate jobs, there can be multiple methods for the server to calculate the similarity between the job to be predicted and the set of candidate jobs. Optionally, the server can default that the similarity between each historical job in each candidate job set is the same. Therefore, the server can randomly select a historical job from the candidate job set and calculate the similarity between the features of the job to be predicted and the features of the randomly selected historical job. The server can use this similarity as the similarity between the job to be predicted and the candidate job set. Optionally, the server can also calculate the similarity between the job to be predicted and each historical job in the candidate job set, and calculate the average value of each similarity, and use the average value as the similarity between the job to be predicted and the candidate job set.

[0175] In the embodiments of the present application, to determine the set of similar jobs corresponding to the job to be predicted and determine the set of similar jobs corresponding to the job to be predicted from the set of types. Since the set of types is a set that has been divided according to the types of historical jobs, in the present application, the set of similar jobs is determined from the set of types that matches the type of the predicted job, avoiding analyzing all historical jobs, thereby reducing the amount of data calculation and also improving the efficiency of determining the set of similar jobs.

[0176] In one embodiment, based on the above Figure 9 illustrated embodiment, refer to Figure 10 , this embodiment relates to the process of determining the set of similar jobs from the set of types according to the job features of the job to be predicted. As Figure 10 shown, step 902 may include step 1001 and step 1002.

[0177] Step 1001, obtain the similarity between the job features of the job to be predicted and each candidate job set.

[0178] In the embodiments of the present application, the set of types includes at least one candidate job set. The candidate job set includes multiple historical jobs. The similarity between each historical job in a candidate job set is extremely high. For example, the similarity between each historical job is greater than a preset threshold. The preset threshold can be 95%, 90%, etc. Those skilled in the art can determine the preset threshold according to actual needs, and the embodiments of the present application do not limit it.

[0179] In the embodiments of the present application, for a candidate job set, there are various methods for the server to calculate the similarity between the job to be predicted and the candidate job set. Optionally, the server may default that the similarity between each historical job in each candidate job set is the same. Therefore, the server may randomly select a historical job from the candidate job set and calculate the similarity between the features of the job to be predicted and the randomly selected historical job. The server may use this similarity as the similarity between the job to be predicted and the candidate job set. Optionally, the server may also calculate the similarity between the job to be predicted and each historical job in the candidate job set, and calculate the average value of the similarities, and use the average value as the similarity between the job to be predicted and the candidate job set.

[0180] Optionally, step 1001 may include: obtaining the similarity between the job features of the job to be predicted and the job features of the target job in each candidate job set.

[0181] Wherein, the target job is any job in the candidate job set. In the embodiments of the present application, the server may default that the similarity between each historical job in each candidate job set is the same. Therefore, the server may randomly select a historical job from the candidate job set as the target job, and calculate the similarity according to the features of the job to be predicted and the target job. The server may use this similarity as the similarity between the job to be predicted and the candidate job set.

[0182] In the above embodiments, the server selects the target job from the candidate job set, and determines the similarity between the job features of the target job and the job features of the job to be predicted as the similarity between the job to be predicted and each candidate job set, and the calculation efficiency is higher.

[0183] Step 1002, determine the similar job set from at least one candidate job set according to the similarity.

[0184] In the embodiments of the present application, after obtaining the similarity between the job features of the job to be predicted and each candidate job set, optionally, the server may use the candidate job set with the largest similarity as the similar job set corresponding to the job to be predicted. Optionally, the server may also select candidate job sets with similarities greater than a preset similarity threshold from at least one candidate job set, and use the union of the candidate job sets with similarities greater than the similarity threshold as the similar job set.

[0185] In the above embodiments, the server determines the similarity between the job features of the job to be predicted and each candidate job set, numerically represents the similarity, and the similar job set determined from the candidate job set according to the similarity is more accurate, and the method is simple and easy to implement.

[0186] Since in some scenarios, the job characteristics of a job may be non - numerical features such as text, in order to facilitate the calculation of the similarity between the job to be predicted and the jobs in the candidate set, it is necessary to numerically process the non - numerical features. In one embodiment, based on the above Figure 10 shown embodiment, refer to Figure 11 , this embodiment is related to the process of obtaining the similarity between the job characteristics of the job to be predicted and the job characteristics of the target jobs in each candidate job set. As Figure 11 shown, this process may include step 1101 and step 1102.

[0187] Step 1101, numerically process the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0188] The job characteristics of the job to be predicted may include text - type features or numerical - type features. For text - type features, they cannot be directly used for similarity calculation. Therefore, in the embodiments of the present application, the server can numerically process the text - type features. Regarding the numerical processing method, optionally, the server can use a Label Encoder to convert the text - type features into integer values; optionally, the server can use One - Hot Encoding to convert the text - type features into binary values. Here, no specific limitation is made on the specific numerical processing method.

[0189] In the embodiments of the present application, after the server numerically processes the text - type features in the job characteristics of the job to be predicted, the numerically processed text features and numerical features are used as the numerical characteristics of the job to be predicted. The method of obtaining the numerical characteristics of the target job is the same as the method of obtaining the numerical characteristics of the job to be predicted, and will not be elaborated here.

[0190] Step 1102, obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0191] For a target job, the server can determine the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of the target job by calculating methods such as the Euclidean distance and Jaccard similarity coefficient between the target job and the job to be predicted.

[0192] In the embodiments of the present application, if the server determines the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of the target job according to the Jaccard similarity coefficient, then the calculation formula for the Jaccard similarity coefficient between the job to be predicted and the target job is shown in formula (7):

[0193]

[0194] Among them, A includes the numerical features of the job to be predicted, and B includes the numerical features of the target job. Exemplarily, the numerical features of the job to be predicted include a, b, and c, so A = {a, b, c}, and the numerical features of the target job include a, c, and d, so B = {a, c, d}. Substitute A and B into Formula 1, |A ∩ B| = |{a, c}| = 2, |A ∪ B| = |{a, b, c, d}| = 4, and the Jaccard similarity coefficient J(A, B) between the job to be predicted and the target job is calculated as 1 / 2.

[0195] In the above embodiment, after numericalizing the job features of the job to be predicted and the target job, it is convenient to calculate the similarity between the job to be predicted and the target job. Moreover, when calculating the similarity between the job to be predicted and the target job, both the text job features and the numerical job features are considered, so that various features between jobs can be more comprehensively referred to, making the calculated similarity more accurate.

[0196] In one embodiment, in order to quickly match a set of similar jobs for the job to be predicted, it is necessary to classify the historical jobs. Based on the above Figure 10 illustrated embodiment, refer to Figure 12 , this embodiment relates to the process of classifying historical jobs. As Figure 12 shown, this process may include step 1201 and step 1202.

[0197] Step 1201, classify multiple historical jobs to obtain a set of types.

[0198] Among them, the historical jobs are the jobs that have been executed in the historical period. Optionally, the server can obtain all the historical jobs that have been executed; optionally, the historical jobs a long time ago are no longer referential for predicting the execution time of the job to be predicted. Therefore, the server can also obtain the historical jobs that have been executed within a period of time according to a preset time period. Each historical job has multiple attribute features, such as text-based user name, user group name, partition applied for by the job, node applied for by the job, working path of the job, type of the job, etc., and numerical-based CPU quantity applied for by the job, memory quantity applied for by the job, node quantity applied for by the job, execution time limit applied for by the job, etc.

[0199] In the embodiments of the present application, the server may classify historical jobs according to the attribute characteristics of each historical job. For example, the server may classify historical jobs according to the types in terms of applications. Exemplarily, historical jobs may be classified into five categories: molecular dynamics, meteorological image research, electronic structure and quantum mechanics, communication, and artificial intelligence. Alternatively, the server may also divide historical jobs into different type sets according to the types in terms of functions, and the historical jobs included in each type set have the same functional type.

[0200] Step 1202: Determine at least one candidate job set according to the similarity between each historical job in the type set.

[0201] In the embodiments of the present application, in order to ensure that historical jobs with higher similarity are classified into the same set, the type set may be further divided in detail. For example, the server may calculate the similarity between each historical job in the type set (calculated between every two historical jobs), and divide each historical job according to the similarity. For example, for two historical jobs, when the similarity between these two historical jobs is greater than a preset threshold, these two historical jobs are classified into a candidate job set. Finally, each obtained candidate job set includes multiple historical jobs with a similarity greater than the preset threshold.

[0202] Refer to Figure 13 , and an exemplary introduction is made to the process of classifying historical jobs to obtain a type set:

[0203] Obtain multiple historical jobs, and the multiple historical jobs form a historical job set; according to the types of each historical job, divide the historical jobs into multiple type sets. Suppose there are n types in total, then the historical jobs can be divided into n type sets, including type1, type2,..., type n; for each type set, calculate the similarity between each historical job, and divide each historical job according to the similarity to obtain multiple candidate job sets. For example, type set type1 is divided into multiple candidate job sets type1.1, type1.2,..., type1.x, and so on. Each type set is divided again to determine the candidate job sets included in each type set.

[0204] In the above embodiments, the server performs a secondary division on the obtained historical jobs, and then divides different candidate job sets according to the similarity under different type sets, which can provide rich training data for training the initial prediction model for different jobs to be predicted. Moreover, after two divisions, it can more accurately and quickly determine the job set matching the job to be predicted.

[0205] In one embodiment, based on the above Figure 2In the illustrated embodiment, this embodiment relates to the process of determining an initial prediction model, which may include: determining an initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0206] In the embodiments of the present application, in order to save computing resources and improve prediction efficiency when predicting the execution time of the job to be predicted, different initial prediction models are preset for different types of jobs to be predicted. When the server predicts the execution time of the job to be predicted, there is no need to set the initial prediction model by integrating various conditions. The server only needs to determine the initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0207] In the above embodiment, different initial prediction models are preset for different types of jobs to be predicted. When predicting the execution time of the job to be predicted, it saves computing resources and improves prediction efficiency. Moreover, the initial prediction model matches the type of the job to be predicted, which can improve the accuracy of job time prediction.

[0208] In one embodiment, an exemplary model training method is provided, which can be applied to Figure 1 the illustrated implementation environment.

[0209] Step a: Classify multiple historical jobs according to the type of the job to obtain multiple sample type sets.

[0210] Step b: For each sample type set, determine the goodness of fit of each sub-network model according to the sample type set.

[0211] Step c: Sort multiple sub-network models according to the goodness of fit.

[0212] Step d: Extract sub-models a preset number of times from the sorted multiple sub-network models to obtain the corresponding groups for each extracted model. Wherein, the preset number is related to the number of times of sub-model extraction.

[0213] Step e: Integrate the sub-network models of each group to obtain multiple integrated models.

[0214] Step f: Determine the error value of each integrated model according to the job characteristics of each historical job in the sample type set.

[0215] Step g: Determine the initial prediction model corresponding to the sample type set from multiple integrated models according to the error value.

[0216] Wherein, the initial neural network model includes multiple sub-network models.

[0217] Step h: Determine the initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0218] Step i, classify multiple historical jobs to obtain a type set.

[0219] Step j, determine at least one candidate job set according to the similarity between the historical jobs in the type set.

[0220] Step k, determine the type set corresponding to the job to be predicted according to the type of the job to be predicted.

[0221] Among them, the type set includes at least one historical job.

[0222] Step l, numerically process the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0223] Step m, obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0224] Among them, the target job is any job in the candidate job set.

[0225] Step n, determine a similar job set from at least one candidate job set according to the similarity.

[0226] Among them, the type set includes at least one candidate job set.

[0227] Step o, train the initial prediction model corresponding to the job to be predicted according to the similar job set to obtain a target prediction model.

[0228] Among them, the target prediction model is used to predict the execution time of the job to be predicted.

[0229] In an exemplary embodiment, a method for predicting job time is provided, and the method includes:

[0230] Input the job to be predicted into the target prediction model for time prediction to obtain the execution time of the job to be predicted. Among them, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the similar job set of the job to be predicted.

[0231] In the embodiment of the present application, the job to be predicted is a job that needs time prediction, and the target prediction model is used to predict the execution time of the job to be predicted. The target prediction model is trained based on the model training method provided in the above embodiment. Among them, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the similar job set of the job to be predicted, and the similarity between the jobs in the similar job set and the job to be predicted is greater than a preset threshold.

[0232] Refer to Figure 14, an exemplary introduction to the process of the prediction method for operation time is given:

[0233] The server obtains the operation to be predicted; the server determines the type set corresponding to the operation to be predicted according to the type of the operation to be predicted; the server determines the similar operation set corresponding to the operation to be predicted from the type set according to the operation characteristics of the operation to be predicted. The method for determining the similar operation set is specifically described in the above embodiments and will not be elaborated here; the server determines the initial prediction model to be trained according to the type set corresponding to the operation to be predicted. The method for determining the initial prediction model is specifically described in the above embodiments and will not be elaborated here; the server uses the similar operation set to train the initial prediction model to obtain the target prediction model. The specific training process is described in the above embodiments and will not be elaborated here; the operation characteristics of the operation to be predicted are input into the target prediction model, and the target prediction model is used to predict the execution time of the operation to be predicted. The result output by the target prediction model is the predicted execution time of the operation to be predicted.

[0234] In the above embodiments, according to the type of the operation to be predicted, the initial prediction model preset for the type of the operation to be predicted is determined. According to the operation characteristics of the operation to be predicted, the similar operation set with the highest similarity to the operation to be predicted is determined. The target prediction model obtained by training the initial prediction model with the similar operation set is trained for the operation to be predicted and can predict the execution time of the operation to be predicted more accurately.

[0235] It should be understood that although each step in the flowcharts involved in the above embodiments is displayed in sequence according to the indication of the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily need to be executed at the same moment, but can be executed at different moments. The execution order of these steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0236] Based on the same inventive concept, an embodiment of the present application further provides a model training device for implementing the above-mentioned model training method. The implementation solution provided by this device to solve the problem is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the model training device provided below can refer to the limitations on model training in the above text and will not be elaborated here.

[0237] In an exemplary embodiment, as Figure 15As shown, a model training device 1500 is provided, including: a determination module 1501 and a training module 1502, where:

[0238] The determination module 1501 is configured to determine a set of similar jobs for the job to be predicted;

[0239] The training module 1502 is configured to train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0240] In one embodiment, the model training device 1500 further includes an acquisition module for the initial prediction model, and this module includes:

[0241] A classification unit, configured to classify multiple historical jobs according to the types of the jobs to obtain multiple sample type sets;

[0242] An initial prediction model obtaining unit, configured to train an initial neural network model according to the multiple sample type sets to obtain the initial prediction model.

[0243] In one embodiment, the initial neural network model includes multiple sub-network models, and the initial prediction model obtaining unit is further configured to execute:

[0244] For each of the sample type sets, determine the goodness of fit of each of the sub-network models according to the sample type set;

[0245] Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each of the sample type sets.

[0246] In one embodiment, the initial prediction model obtaining unit is further configured to execute:

[0247] Sort the multiple sub-network models according to the goodness of fit;

[0248] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each of the sample type sets.

[0249] In one embodiment, the initial prediction model obtaining unit is further configured to execute:

[0250] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain multiple integrated models;

[0251] Determine the error value of each of the integrated models according to the job characteristics of the historical jobs in the sample type set;

[0252] Determine the initial prediction model corresponding to the sample type set from the multiple integrated models according to the error value.

[0253] In one embodiment, the initial prediction model obtaining unit is further configured to perform:

[0254] Perform sub-model extraction on the sorted multiple sub-network models for a preset number of times to obtain a group corresponding to each extracted model; wherein, the preset number is related to the number of times of sub-model extraction.

[0255] In one embodiment, the determination module 1501 includes:

[0256] A type set determination unit, configured to determine a type set corresponding to the job to be predicted according to the type of the job to be predicted; at least one historical job is included in the type set;

[0257] A similar job set determination unit, configured to determine the similar job set from the type set according to the job characteristics of the job to be predicted.

[0258] In one embodiment, at least one candidate job set is included in the type set, and the similar job set determination unit is further configured to perform:

[0259] Obtain the similarity between the job characteristics of the job to be predicted and each candidate job set;

[0260] Determine the similar job set from the at least one candidate job set according to the similarity.

[0261] In one embodiment, the similar job set determination unit is further configured to perform:

[0262] Obtain the similarity between the job characteristics of the job to be predicted and the job characteristics of the target job in each candidate job set; the target job is any job in the candidate job set.

[0263] In one embodiment, the similar job set determination unit is further configured to perform:

[0264] Perform numerical processing on the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job;

[0265] Obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0266] In one embodiment, the similar job set determination unit is further configured to perform:

[0267] Classify multiple historical jobs to obtain the type set;

[0268] Determine the at least one candidate job set according to the similarity between historical jobs in the type set.

[0269] In one embodiment, the model training device 1500 further includes:

[0270] An initial prediction model determination module, configured to determine an initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0271] Based on the same inventive concept, an embodiment of the present application further provides a prediction device for job time for implementing the prediction method of job time involved above. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the prediction device for job time provided below can refer to the limitations on the prediction of job time in the above text, and will not be repeated here.

[0272] In an exemplary embodiment, as Figure 16 shown, a prediction device 1600 for job time is provided, including: a prediction module 1601, where:

[0273] The prediction module 1601 is configured to input the job to be predicted into a target prediction model for time prediction, and obtain the execution time of the job to be predicted;

[0274] Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the set of similar jobs of the job to be predicted.

[0275] Each module in the above model training device and the prediction device for job time can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0276] In an exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 17As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. 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, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store prediction data for model training and / or job time. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it realizes a model method and / or a prediction method for job time.

[0277] Those skilled in the art can understand that Figure 17 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 different component arrangements.

[0278] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0279] Determine a set of similar jobs for the job to be predicted;

[0280] Train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0281] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0282] According to the type of the job to be predicted, determine a set of types corresponding to the job to be predicted; at least one historical job is included in the set of types;

[0283] Determine the set of similar jobs from the set of types according to the job characteristics of the job to be predicted.

[0284] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0285] Classify multiple historical jobs according to the type of the jobs to obtain multiple sample type sets;

[0286] Train an initial neural network model according to the multiple sample type sets to obtain the initial prediction model.

[0287] In one embodiment, the initial neural network model includes multiple sub-network models, and when the processor executes the computer program, the following steps are further implemented:

[0288] For each of the sample type sets, determine the goodness of fit of each of the sub-network models according to the sample type set;

[0289] Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each of the sample type sets.

[0290] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0291] Sort the multiple sub-network models according to the goodness of fit;

[0292] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each of the sample type sets.

[0293] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0294] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain multiple integrated models;

[0295] Determine the error value of each of the integrated models according to the job characteristics of the historical jobs in the sample type set;

[0296] Determine the initial prediction model corresponding to the sample type set from the multiple integrated models according to the error value.

[0297] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0298] Perform multiple extractions of sub-models with a preset number on the sorted multiple sub-network models to obtain a group corresponding to each extracted model; wherein, the preset number is related to the number of extractions of the sub-models.

[0299] In one embodiment, at least one candidate job set is included in the type set, and when the processor executes the computer program, the following steps are further implemented:

[0300] Obtain the similarity between the job characteristics of the job to be predicted and each of the candidate job sets;

[0301] Determine a similar job set from the at least one candidate job set according to the similarity.

[0302] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0303] Obtain the similarity between the job characteristics of the job to be predicted and the job characteristics of the target jobs in each of the candidate job sets; the target job is any job in the candidate job set.

[0304] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0305] Perform numerical processing on the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job;

[0306] Obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0307] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0308] Classify multiple historical jobs to obtain the type set;

[0309] Determine the at least one candidate job set according to the similarity between the historical jobs in the type set.

[0310] In one embodiment, when the processor executes the computer program, the following steps are further implemented:

[0311] Determine an initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0312] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0313] Input the job to be predicted into a target prediction model for time prediction to obtain the execution time of the job to be predicted;

[0314] Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the similar job set of the job to be predicted.

[0315] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0316] Determine a set of similar jobs for the job to be predicted;

[0317] Train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0318] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0319] Classify multiple historical jobs according to the types of jobs to obtain multiple sample type sets;

[0320] Train an initial neural network model according to the multiple sample type sets to obtain the initial prediction model.

[0321] In one embodiment, the initial neural network model includes multiple sub-network models. When the computer program is executed by a processor, the following steps are further implemented:

[0322] For each of the sample type sets, determine the goodness of fit of each of the sub-network models according to the sample type set;

[0323] Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each of the sample type sets.

[0324] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0325] Sort the multiple sub-network models according to the goodness of fit;

[0326] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each of the sample type sets.

[0327] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0328] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain multiple integrated models;

[0329] Determine the error value of each of the integrated models according to the job characteristics of the historical jobs in the sample type set;

[0330] Determine the initial prediction model corresponding to the sample type set from the multiple integrated models according to the error value.

[0331] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0332] Perform sub-model extraction on the sorted multiple sub-network models for a preset number of times to obtain the grouping corresponding to each extracted model; wherein, the preset number is related to the number of sub-model extraction times.

[0333] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0334] Determine the type set corresponding to the job to be predicted according to the type of the job to be predicted; the type set includes at least one historical job;

[0335] Determine the set of similar jobs from the type set according to the job characteristics of the job to be predicted.

[0336] In one embodiment, the type set includes at least one candidate job set. When the computer program is executed by the processor, the following steps are further implemented:

[0337] Obtain the similarity between the job characteristics of the job to be predicted and each candidate job set;

[0338] Determine the set of similar jobs from the at least one candidate job set according to the similarity.

[0339] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0340] Obtain the similarity between the job characteristics of the job to be predicted and the job characteristics of the target job in each candidate job set; the target job is any job in the candidate job set.

[0341] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0342] Perform numerical processing on the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job;

[0343] Obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0344] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:

[0345] Classify the multiple historical jobs to obtain the type set;

[0346] Determine the at least one candidate job set according to the similarity between the historical jobs in the type set.

[0347] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0348] Determine the initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0349] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0350] Input the job to be predicted into the target prediction model for time prediction to obtain the execution time of the job to be predicted;

[0351] Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the set of similar jobs of the job to be predicted.

[0352] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0353] Determine the set of similar jobs of the job to be predicted;

[0354] Train the initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

[0355] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0356] Classify multiple historical jobs according to the type of the jobs to obtain multiple sample type sets;

[0357] Train the initial neural network model according to the multiple sample type sets to obtain the initial prediction model.

[0358] In one embodiment, the initial neural network model includes multiple sub-network models. When the computer program is executed by a processor, the following steps are further implemented:

[0359] For each of the sample type sets, determine the goodness of fit of each of the sub-network models according to the sample type set;

[0360] Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each sample type set.

[0361] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0362] Sort the multiple sub-network models according to the goodness of fit;

[0363] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each sample type set.

[0364] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0365] Group the sorted multiple sub-network models in sequence, and integrate the sub-network models in each group to obtain multiple integrated models;

[0366] Determine the error value of each integrated model according to the job characteristics of each historical job in the sample type set;

[0367] Determine the initial prediction model corresponding to the sample type set from the multiple integrated models according to the error value.

[0368] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0369] Perform multiple extractions of sub-models with a preset number on the sorted multiple sub-network models to obtain the group corresponding to each extracted model; wherein, the preset number is related to the number of extractions of sub-models.

[0370] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0371] Determine the type set corresponding to the to-be-predicted job according to the type of the to-be-predicted job; the type set includes at least one historical job;

[0372] Determine the similar job set from the type set according to the job characteristics of the to-be-predicted job.

[0373] In one embodiment, the type set includes at least one candidate job set. When the computer program is executed by a processor, the following steps are further implemented:

[0374] Obtain the similarity between the job characteristics of the to-be-predicted job and each candidate job set;

[0375] Determine a set of similar jobs from the at least one set of candidate jobs according to the similarity.

[0376] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0377] Obtain the similarity between the job characteristics of the job to be predicted and the job characteristics of the target jobs in each set of candidate jobs; the target job is any job in the set of candidate jobs.

[0378] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0379] Perform numerical processing on the job characteristics of the job to be predicted and the job characteristics of each target job to obtain the numerical characteristics of the job to be predicted and the numerical characteristics of each target job;

[0380] Obtain the similarity between the numerical characteristics of the job to be predicted and the numerical characteristics of each target job.

[0381] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0382] Classify the multiple historical jobs to obtain the set of types;

[0383] Determine the at least one set of candidate jobs according to the similarity between the historical jobs in the set of types.

[0384] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:

[0385] Determine the initial prediction model corresponding to the job to be predicted according to the type of the job to be predicted.

[0386] In one embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0387] Input the job to be predicted into the target prediction model for time prediction to obtain the execution time of the job to be predicted;

[0388] Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the job to be predicted according to the set of similar jobs of the job to be predicted.

[0389] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0390] 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, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in this application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in this application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0391] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of 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.

[0392] 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 patent scope of the present application. 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 fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A model training method, characterized in that, The method includes: Determine a set of similar jobs for the job to be predicted; Train an initial prediction model corresponding to the job to be predicted according to the set of similar jobs to obtain a target prediction model; the target prediction model is used to predict the execution time of the job to be predicted.

2. The method according to claim 1, characterized in that, The method for obtaining the initial prediction model includes: Classify multiple historical jobs according to the types of jobs to obtain multiple sample type sets; Train an initial neural network model according to the multiple sample type sets to obtain the initial prediction model.

3. The method according to claim 2, characterized in that The initial neural network model includes multiple sub-network models. Training the initial neural network model according to the multiple sample type sets to obtain the initial prediction model includes: For each of the sample type sets, determine the goodness of fit of each of the sub-network models according to the sample type set; Integrate the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each of the sample type sets.

4. The method according to claim 3, wherein Integrating the multiple sub-network models according to the goodness of fit to obtain an initial prediction model corresponding to each of the sample type sets includes: Sort the multiple sub-network models according to the goodness of fit; Group the sorted multiple sub-network models in sequence and integrate the sub-network models in each group to obtain an initial prediction model corresponding to each of the sample type sets.

5. The method according to claim 4, characterized in that Grouping the sorted multiple sub-network models in sequence and integrating the sub-network models in each group to obtain an initial prediction model corresponding to each of the sample type sets includes: Group the sorted multiple sub-network models in sequence and integrate the sub-network models in each group to obtain multiple integrated models; Determine the error value of each of the integrated models according to the job characteristics of the historical jobs in the sample type set; Determine the initial prediction model corresponding to the sample type set from the multiple integrated models according to the error value.

6. The method according to claim 5, wherein Grouping the sorted multiple sub-network models in sequence includes: Perform multiple extractions of a preset number of sub-models on the sorted multiple sub-network models to obtain a group corresponding to each extraction model; where the preset number is related to the number of extractions of the sub-models.

7. The method according to claim 1, characterized in that, Determining the set of similar jobs for the job to be predicted includes: Determine a type set corresponding to the job to be predicted according to the type of the job to be predicted; the type set includes at least one historical job; Determine the set of similar jobs from the type set according to the job characteristics of the job to be predicted.

8. The method according to claim 7, wherein The type set includes at least one candidate job set. Determining the set of similar jobs from the type set according to the job characteristics of the job to be predicted includes: Obtain the similarity between the job characteristics of the job to be predicted and each of the candidate job sets; Determine the set of similar jobs from the at least one candidate job set according to the similarity.

9. The method according to claim 8, wherein Obtaining the similarity between the job characteristics of the job to be predicted and each of the candidate job sets includes: Obtain the similarity between the job characteristics of the to-be-predicted job and the job characteristics of the target job in each of the candidate job sets; the target job is any job in the candidate job set.

10. The method according to claim 9, characterized in that, The obtaining the similarity between the job characteristics of the to-be-predicted job and the job characteristics of the target job in each of the candidate job sets includes: Perform numerical processing on the job characteristics of the to-be-predicted job and the job characteristics of each target job to obtain the numerical characteristics of the to-be-predicted job and the numerical characteristics of each target job; Obtain the similarity between the numerical characteristics of the to-be-predicted job and the numerical characteristics of each target job.

11. The method according to claim 8, characterized in that, The method further includes: Classify the multiple historical jobs to obtain the type set; Determine the at least one candidate job set according to the similarity between the historical jobs in the type set.

12. The method according to claim 1, characterized in that, The method further includes: Determine the initial prediction model corresponding to the to-be-predicted job according to the type of the to-be-predicted job.

13. A method for predicting operation time, characterized in that, The method includes: Input the to-be-predicted job into the target prediction model for time prediction to obtain the execution time of the to-be-predicted job; Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the to-be-predicted job according to the similar job set of the to-be-predicted job.

14. A model training device, characterized in that, The device includes: A determination module, configured to determine a similar job set of the to-be-predicted job; A training module, configured to train the initial prediction model corresponding to the to-be-predicted job according to the similar job set to obtain a target prediction model; the target prediction model is used to predict the execution time of the to-be-predicted job.

15. A prediction device for operation time, characterized in that, The device includes: A prediction module, configured to input the to-be-predicted job into the target prediction model for time prediction to obtain the execution time of the to-be-predicted job; Wherein, the target prediction model is obtained by training the initial prediction model corresponding to the to-be-predicted job according to the similar job set of the to-be-predicted job.

16. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 13 are implemented.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 13 are implemented.

18. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 13 are implemented.

Images