A task processing method and device

Abstract

The present application discloses a task processing method and device, which relate to the field of model selection technology, including converting the sample data set characteristics of the target task into computable quantitative indicators by constructing a feature quantitative evaluation matrix containing the structural features and dynamic features of the sample data set, and then obtaining the target task information in combination with the task type and task constraints. The target task information is further matched and screened with the feature quantitative evaluation matrix, task type and task constraints in historical task information to determine a set of candidate models, so as to reuse the model adaptation experience of historical tasks without trying a large number of models one by one, significantly reducing the redundant calculations caused by manual trial and error in large-scale data sets or complex tasks, quickly narrowing the range of candidate models, and improving the efficiency of machine learning model selection.

Description

Technical Field

[0001] The present application relates to the field of model selection technology, and in particular to a task processing method and device. Background Art

[0002] In the application scenarios of machine learning models, faced with diverse task requirements and a rich selection of models, accurately matching and task-adapting models is the core prerequisite for efficiently completing tasks.

[0003] Currently, model selection primarily relies on manual experience and trial: based on the task type (such as classification or regression) and preliminary data analysis, several potentially applicable models are manually selected for training and evaluation. However, for large datasets or complex tasks, this approach is time-consuming and prone to missing superior candidate models. Furthermore, it struggles to quickly respond to diverse task requirements and lacks efficient automation mechanisms, failing to meet the efficiency requirements of practical applications.

[0004] Therefore, how to improve the efficiency of model selection becomes an urgent problem to be solved. Summary of the Invention

[0005] The present application provides a task processing method and apparatus to at least solve the problem of low efficiency of model selection in related technologies.

[0006] This application provides a task processing method, including:

[0007] Obtain target task information; the target task information includes a feature quantification evaluation matrix corresponding to the target task, a task type, and task constraints; the feature quantification evaluation matrix includes structural features and dynamic features corresponding to a sample data set of the target task;

[0008] Based on the target task information, combined with multiple historical task information and historical machine learning models corresponding to the historical tasks, a set of machine learning models to be used is matched and screened; the historical task information set includes feature quantitative evaluation matrices, task types, and task constraints corresponding to the multiple historical tasks respectively;

[0009] Determine a target machine learning model from the set of machine learning models to be used.

[0010] The present application also provides a task processing device, comprising:

[0011] An acquisition unit is used to acquire target task information; the target task information includes a feature quantification evaluation matrix corresponding to the target task, a task type, and task constraints; the feature quantification evaluation matrix includes structural features and dynamic features corresponding to a sample data set of the target task;

[0012] A matching unit is configured to match and select a set of machine learning models to be used based on the target task information, in combination with multiple historical task information and historical machine learning models corresponding to the historical tasks; the historical task information set includes feature quantification evaluation matrices, task types, and task constraints corresponding to the multiple historical tasks;

[0013] A determination unit is used to determine a target machine learning model from the set of machine learning models to be used.

[0014] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned task processing methods when executing the computer program.

[0015] The present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned task processing methods are implemented.

[0016] The present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned task processing methods when executed by a processor.

[0017] This application converts the sample dataset characteristics of the target task into computable quantitative indicators by constructing a feature quantitative evaluation matrix that includes the structural features and dynamic features of the sample dataset, and then obtains the target task information in combination with the task type and task constraints. The target task information is further matched and screened with the feature quantitative evaluation matrix, task type and task constraints in the historical task information to determine a set of candidate models to reuse the model adaptation experience of historical tasks. There is no need to try a large number of models one by one, which significantly reduces the redundant calculations caused by manual trial and error in large-scale datasets or complex tasks, quickly narrows the range of candidate models, and improves the efficiency of machine learning model selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0019] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. 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 any creative work.

[0020] Figure 1 One of the flowcharts of a task processing method provided in an embodiment of the present application;

[0021] Figure 2 The second flowchart of a task processing method provided in an embodiment of the present application;

[0022] Figure 3 The third flowchart of a task processing method provided in an embodiment of the present application;

[0023] Figure 4 A schematic diagram of the structure of a task processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] It should be noted that, in the description of this application, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence.

[0026] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0027] The embodiment of the present application provides a task processing method, referring to Figure 1 As shown, the specific steps include:

[0028] S11. Obtain target task information.

[0029] The target task information includes the feature quantitative evaluation matrix, task type and task constraints corresponding to the target task; the feature quantitative evaluation matrix includes the structural features and dynamic features corresponding to the sample data set of the target task.

[0030] In some embodiments, the target task is a specific business or technical task that the user currently needs to solve by training a machine learning model; for example, identifying the type of goods on a conveyor belt or determining whether a user-uploaded image is a cat. The target task information is a multi-dimensional description of the target task, facilitating its use in subsequent model matching and selection.

[0031] Specifically, the feature quantification evaluation matrix in the target task information is a quantitative summary of the features reflected by the sample data set on which the target task depends, and the essential characteristics and change patterns of the sample data set are reflected through the feature quantification evaluation matrix; the task type is used to clarify the essential requirements for the machine learning model and helps to determine the range of candidate models; for example, task types include but are not limited to: classification tasks, regression tasks, time series prediction tasks, clustering tasks; task constraints can be understood as the restrictions on the model during actual deployment to ensure that the optimization results meet the engineering feasibility; task constraints include but are not limited to: accuracy constraints, resource constraints, video memory / memory constraints, computing resource constraints, and time constraints.

[0032] Furthermore, the feature quantification evaluation matrix includes the structural features and dynamic features corresponding to the sample data set of the target task; among them, the structural features represent the static properties of the sample data set itself, reflecting its inherent statistical characteristics, and are directly related to the data distribution and composition; the dynamic features represent the dynamic properties of the sample data set in a specific task that are related to the model interaction, and their values ​​may vary with the task type or the model training process.

[0033] The structural features in the feature quantification evaluation matrix may include the following features:

[0034] Sample-feature balance δ: This is the logarithm of the ratio of the sample data size to the number of features. This characteristic reflects the relative relationship between the sample dataset and the number of features. A large logarithm of this ratio indicates that the sample data size is sufficient relative to the number of features, leading to more stable model learning. Conversely, a low logarithm indicates overfitting or insufficient learning. For example, in image recognition tasks, a large number of samples but few features results in a large δ value, which helps the model learn common image characteristics. A small number of samples but many features results in a small δ value, making the model susceptible to the influence of individual sample characteristics.

[0035] Data completeness rate ρ: This is the ratio of the total number of missing features in all samples to the total number of features. ρ provides a direct indicator of the degree of data missingness. Values ​​closer to 0 indicate greater data completeness, indicating that the model learns based on complete information, ensuring accuracy. Values ​​closer to 1 indicate a high degree of missing features, potentially preventing the model from effectively learning data patterns and impacting predictive performance. For example, in medical datasets, excessively high ρ values ​​can make it difficult for disease diagnosis models to accurately diagnose conditions.

[0036] Distribution anomaly coefficient γ: This is the sum of the absolute values ​​of skewness and kurtosis. Skewness measures the degree of asymmetry in the data distribution, while kurtosis reflects the degree of concentration of the data distribution around the mean. A larger γ value indicates a greater deviation from normality, increasing the likelihood of outliers or extreme distributions that could interfere with model training. A γ value close to 0 indicates a relatively normal data distribution, which promotes stable model learning. In financial data, a large γ value indicates abnormal data fluctuations, which can affect the accuracy of risk prediction models.

[0037] Feature type entropy τ: Calculated by summing the negative logarithms of the proportions of each type of feature. It measures the diversity of feature types in a dataset. A high τ value indicates a richer set of features, allowing the model to learn more diverse information and capture complex data patterns. A low τ value indicates a single set of features, limiting the information learned by the model and making it difficult to handle complex tasks. In text classification tasks, a high τ value can improve the performance of the classification model due to the presence of multiple feature types. A low τ value, however, can negatively impact the performance of the model if the number of feature types is small.

[0038] Furthermore, the dynamic features in the feature quantification evaluation matrix may include the following features:

[0039] Intrinsic dimensionality ω: This refers to the minimum number of principal components required to achieve a cumulative variance contribution of 95% or more during principal component analysis. ω reflects the effective dimensionality of the data. A small value indicates that the data can be explained by a small number of principal components, resulting in lower dimensionality and a lighter learning burden for the model. A large value indicates higher dimensionality, requiring more principal components, increasing model complexity, and making training more challenging. For high-resolution image data, a large ω value means that more factors must be considered when processing the data.

[0040] Collinearity strength ξ: This is the maximum Pearson correlation coefficient between any two features in the dataset. This measure measures the degree of linear correlation between features. A ξ value close to 1 indicates strong linear correlation, which can lead to unstable model parameter estimation and reduced model generalization. A ξ value close to 0 indicates weak correlation between features, which facilitates the model to learn each feature independently. In a housing price prediction model, a high correlation between house size and number of bedrooms (a large ξ value) can interfere with the model's ability to accurately predict housing prices.

[0041] Feature discrimination ϕ: Calculated through variance analysis, it quantifies the strength of the association between a feature and the target variable. Larger ϕ values ​​indicate a stronger discriminative power of the feature for the target variable and a greater importance for model prediction. Smaller ϕ values ​​indicate weaker discriminative power and a smaller contribution to prediction. In disease diagnosis, symptom features with high ϕ values ​​are crucial for the diagnostic model. Unrelated features, however, have low ϕ values ​​and should be considered for removal.

[0042] Feature Importance Variance κ: This value, pre-calculated based on XGBoost (eXtreme Gradient Boosting, a highly efficient ensemble learning algorithm), measures the dispersion of feature importance. A large κ value indicates a large variance in feature importance, requiring a focus on important features during model training. A small κ value indicates a more balanced feature importance, allowing for a more balanced consideration of all features during training. In a customer churn prediction model, a large κ value can be used to focus on key features in model construction, improving prediction accuracy.

[0043] Based on the structural features and dynamic features of the above example, the feature quantification evaluation matrix shown in the following formula is generated:

[0044]

[0045] Where D represents the sample data set, Represents the feature quantization evaluation matrix.

[0046] S12. Match the target task information with each historical task information in the historical task information set, and filter out a preset number of historical task information to generate a set of machine learning models to be used based on the historical machine learning models corresponding to the preset number of historical task information.

[0047] Among them, the historical task information set includes feature quantitative evaluation matrices, task types and task constraints corresponding to multiple historical tasks.

[0048] In an embodiment of the present application, after obtaining the target task information, a historical task information set can be obtained, and then by reusing the model application experience accumulated in historical tasks, a set of candidate models suitable for the current task can be quickly screened out, avoiding blindness in the model selection process and improving model selection efficiency and adaptability.

[0049] Specifically, the historical task information set includes historical task information corresponding to multiple historical tasks. Each historical task information also includes its corresponding feature quantitative evaluation matrix, task type and task constraints. Therefore, when matching with the target task information, corresponding analysis can be performed from three dimensions to obtain a machine learning model that is more suitable for the target task.

[0050] Furthermore, when matching the target task information with the historical task information, the similarity between the target task information and the historical task information can be calculated, and then a preset number of historical task information can be filtered out based on the similarity. For example, the preset number can be set to 5, and then the historical machine learning models corresponding to these 5 historical task information can be used to generate a set of machine learning models to be used.

[0051] S13. Determine the target machine learning model from the set of machine learning models to be used.

[0052] In an embodiment of the present application, after obtaining the set of machine learning models to be used through the above steps, the target machine learning model can be determined from the set.

[0053] Specifically, the target machine learning model can be determined from the set of machine learning models to be used by trial and error. This is because the set of machine learning models to be used is selected based on matching the target task information with the historical task information set. The set of machine learning models to be used only includes a preset number of models, and all of them have a high degree of compatibility with the target task in terms of task type, feature quantization evaluation matrix, and task constraints. They are not randomly selected models. The target machine learning model is then determined through trial and error.

[0054] The method of determining the target machine learning model from the set of machine learning models to be used can be further carried out by performing performance evaluation on each machine learning model in the set of machine learning models to be used to determine the optimal machine learning model, and use the model as the target machine learning model.

[0055] S14. Process the target task based on the target machine learning model.

[0056] Furthermore, after obtaining the target machine learning model, the corresponding target task can be processed.

[0057] This application converts the sample dataset characteristics of the target task into computable quantitative indicators by constructing a feature quantitative evaluation matrix that includes the structural features and dynamic features of the sample dataset, and then obtains the target task information in combination with the task type and task constraints. The target task information is further matched and screened with the feature quantitative evaluation matrix, task type and task constraints in the historical task information to determine a set of candidate models to reuse the model adaptation experience of historical tasks. There is no need to try a large number of models one by one, which significantly reduces the redundant calculations caused by manual trial and error in large-scale datasets or complex tasks, quickly narrows the range of candidate models, and improves the efficiency of machine learning model selection.

[0058] As an extension and refinement of the above embodiment, refer to Figure 2 As shown, the embodiment of the present application also provides another task processing method, including the following steps:

[0059] S21. Obtain target task information.

[0060] The target task information includes the feature quantitative evaluation matrix, task type and task constraints corresponding to the target task; the feature quantitative evaluation matrix includes the structural features and dynamic features corresponding to the sample data set of the target task.

[0061] Specifically, the target task information expression generated based on the feature quantification evaluation matrix, task type and task constraints is as follows:

[0062]

[0063] in, It is the feature quantitative evaluation matrix obtained in the multidimensional feature perception stage, which fully describes the structure and dynamic characteristics of the data; Indicates the task type, such as classification, regression, clustering, etc., and clarifies the nature and objectives of the task; Represents resource constraints, such as computing time and memory, reflecting the resource limitations of the task. Task feature fingerprints integrate data features, task types, and resource constraints, providing comprehensive reference information for model selection.

[0064] S22: Calculate the task similarity between the target task information and each historical task information in the historical task information set.

[0065] Specifically, referring to the expression of the task information in S21 above, an expression of each historical task information is constructed, thereby preparing for the subsequent calculation of task similarity.

[0066] Furthermore, the target task information and the historical task information set may be combined according to a similarity calculation formula to obtain the task similarity between the target task information and each historical task information in the historical task information set.

[0067] The similarity calculation formula is:

[0068]

[0069] in, represents the target task, represents the said historical mission; Indicates the task similarity between the target task and the historical task; Indicates the target task information, Indicates the historical task information; Represents the time attenuation coefficient, which is used to reduce the impact of time interval on similarity; Indicates the execution timestamp of the target task, Indicates the execution timestamp of the historical task.

[0070] In the formula It is the cosine similarity formula. If the information expression of two tasks is The more similar they are (for example, features and constraints are highly overlapped), the smaller the angle between their vectors, and the closer the cosine value is to 1 (high similarity); conversely, it is closer to 0 (low similarity), which measures how similar the features and constraints between tasks are.

[0071] in the formula is a time decay function, which is used to correct the impact of task execution time interval on similarity. The calculation results of this part range from (1, 2]. The smaller the time interval, the closer the result is to 2, indicating a higher gain in similarity due to time. The larger the time interval, the closer the result is to 1, indicating a decrease in similarity. This is because historical tasks with closer time periods are more valuable as a reference, which can improve the timeliness and environmental adaptability of model selection.

[0072] in, Used to reduce the impact of time interval on similarity. The larger the time interval, the more significant the effect (e.g. When it is very large, the time correction item for the task two years ago and the current task will be very small).

[0073] Through cosine similarity and time decay correction, we ultimately arrive at task similarity. This result simultaneously considers the correlation between task content matching and execution time, avoiding invalid experience due to similar content but distant pasts, and preventing the neglect of invalid references due to recent but significantly different content, thus ensuring scientific and scenario-appropriate similarity assessment.

[0074] It should be noted that the method for obtaining the historical task information set may include the following steps:

[0075] Step 1: Build a knowledge graph that includes multiple machine learning models and the task adaptation relationships corresponding to each machine learning model.

[0076] Among them, the nodes in the knowledge graph are machine learning models, and the edges are composed of adaptation conditions, confidence levels, and historical model call times.

[0077] In some embodiments, the knowledge graph is composed of three types of elements: nodes, edges, and attribute nodes; machine learning models are used as nodes, including predefined models (such as classic algorithms XGBoost, ResNet and other system-preset general models) and user-defined models (personalized models developed by users based on business needs). The basic attributes of the model (such as model name, type, and core parameter range) will be recorded in the node as the basic entity unit of the knowledge graph.

[0078] The edge connects the model node and the task adaptation relationship, and is composed of a triplet of adaptation conditions, model confidence, and model historical call count. Among them, the adaptation conditions describe the characteristic constraints of the adaptation task corresponding to the model, such as task type (classification, regression, etc.), feature quantification evaluation matrix, business scenario restrictions (such as applicable image classification tasks), and clarify the matching rules between the model and the task; the model confidence is the historical performance of the quantitative model on the adaptation task, which can be calculated through the actual execution performance of the model, that is, the verification set accuracy, business indicators, etc. (value range [0, 1]), to reflect the reliability of the model in the corresponding task; the model historical call count is the frequency of use of the statistical model in similar adaptation tasks, reflecting the application preference of the model in engineering practice and assisting in judging the practical value of the model.

[0079] Step 2: Filter out multiple historical tasks based on the knowledge graph.

[0080] Furthermore, through the constructed knowledge graph, the target task information of the target task is used as the retrieval condition, and the relationship edges that are adapted to the target task are first matched in the knowledge graph, and then the historical tasks that meet the adaptation rules are obtained.

[0081] Step 3: Obtain task information corresponding to multiple historical tasks to generate a historical task information set.

[0082] Furthermore, after screening out multiple historical tasks that preliminarily meet the target task based on the knowledge graph, task information of the multiple historical tasks is obtained to generate a historical task information set.

[0083] In this application, the knowledge graph transforms the scattered adaptation experience between models and tasks into a structured knowledge network, enabling accurate reuse of adaptation rules. Furthermore, based on the multi-dimensional search of the knowledge graph, historical tasks that are compatible with the target task are initially screened out. This lays the foundation for subsequent task similarity calculations and reduces interference from invalid historical data.

[0084] S23. Based on task similarity, a preset number of historical tasks are filtered out.

[0085] Specifically, based on the task similarity calculated in step S22 (i.e., the similarity between the target task and each historical task), the tasks can be sorted by similarity, and a preset number of historical tasks most similar to the target task can be selected from the historical task information set, thereby ensuring that the screening results meet both quality requirements and the quantity requirements for subsequent processing. For example, if the preset number is 5, the five most similar historical tasks are selected based on similarity.

[0086] S24. Generate a set of machine learning models to be used based on the historical machine learning models corresponding to a preset number of historical tasks.

[0087] Furthermore, the preset number of historical tasks screened out in step S23 can be used to obtain the historical machine learning models associated with them from the knowledge graph to form an initial candidate set of models to be used.

[0088] Preferably, the models in the initial candidate set can also be checked for compatibility with the operating environment to exclude models that cannot run in the target task environment due to hardware limitations, software dependencies, etc. At the same time, if the model type of the initial candidate set is single (for example, all tree models), other types of high-confidence models (such as neural network models) can be supplemented from the knowledge graph to ensure sufficient diversity in the set of models to be used and increase the probability of finding the optimal model.

[0089] S25. Determine the target machine learning model from the set of machine learning models to be used.

[0090] It should be noted that, in order to determine the target machine learning model from the set of machine learning models to be used, genetic algorithms and Bayesian optimization methods can be used in the embodiments of the present application to determine the target machine learning model.

[0091] Specifically, refer to Figure 3 As shown, the steps of determining the target machine learning model from the set of machine learning models to be used include the following:

[0092] S251. Obtain a model parameter set of each machine learning model in the machine learning model set to be used, and encode the model parameter set of each machine learning model into a parameter sequence set.

[0093] Specifically, the model parameter set of the machine learning model (such as the learning rate and number of layers of the neural network, the depth and number of leaf nodes of the tree model) is converted into a computable parameter sequence.

[0094] Extract all parameters of each model from the set of machine learning models to be used (such as the learning rate of a classification model, the number of leaf nodes of a tree model, the hidden layer dimensions of a neural network, etc.). Because the genetic algorithm needs to convert the parameters into a unified format for subsequent operations such as crossover and mutation, these parameters need to be encoded. Continuous parameters (such as the learning rate) can be converted to a fixed range of values ​​through normalization, and discrete parameters (such as the number of network layers) can be mapped to specific integers, ultimately forming a uniformly structured parameter sequence. These sequences generated by the model parameter sets of each machine learning model together constitute a parameter sequence set, preparing for the evolutionary iterative operations of the genetic algorithm and allowing the parameters of different models to be processed uniformly.

[0095] S252: Construct an initial parameter population based on the parameter sequence set.

[0096] Then, based on the encoded parameter sequence set, an initial parameter population is constructed. The initial population contains the parameter sequence corresponding to each model in the set of machine learning models to be used; and subsequent evolutionary iterations are carried out based on the genetic algorithm.

[0097] S253. Perform multiple rounds of evolutionary iterations on the initial parameter population, and after each round of evolutionary iterations, calculate the model performance index corresponding to each parameter sequence in the current parameter population, and select the target parameter sequence based on the model performance index.

[0098] Specifically, multiple rounds of evolutionary iterations are performed on the initial parameter population, and each round of iteration includes two operations: crossover and mutation.

[0099] The crossover operation is to select two parameter sequences from the current population, exchange some parameter fragments according to certain rules (for example, exchange the first half and the second half of the sequence), and generate a new parameter sequence.

[0100] The mutation operation randomly adjusts some parameter values ​​in the parameter sequence (such as slightly modifying the learning rate value or fine-tuning the number of network layers) to introduce new parameter features. These two operations simulate gene recombination and mutation in biological evolution, thereby obtaining a new parameter sequence.

[0101] After each round of iteration, the comprehensive score of model performance indicators (such as prediction accuracy, inference speed, and video memory usage) corresponding to each new parameter sequence is calculated. Multiple parameter sequences with better model performance indicators can then be selected to form a target parameter sequence set. Crossover and mutation operations are then re-performed until the target parameter sequence set is re-determined. This process is repeated over multiple rounds, terminating at a preset number of iterations or when the population performance improvement falls below a threshold. Ultimately, the optimized population, or optimal parameter sequence set, is obtained. The parameter sequence with the best comprehensive score for model performance indicators in the final round is then used as the target parameter sequence.

[0102] Preferably, the target parameter sequence set obtained in the last round can be combined with the logic of Bayesian optimization, that is, using the existing model performance indicators to predict which parameter sequences have a greater probability of performing better in subsequent iterations, and then determine the target parameter sequence based on this probability.

[0103] S254. Determine the machine learning model corresponding to the target parameter sequence as the target machine learning model.

[0104] After multiple rounds of evolutionary iterations and screening, a target parameter sequence is obtained, perhaps through extensive exploration using a genetic algorithm and / or precise screening using Bayesian optimization. The machine learning model corresponding to this target parameter sequence is then determined as the target machine learning model, which is able to better meet the performance and efficiency requirements of the task.

[0105] In an embodiment of the present application, the model parameters are encoded into an evolvable sequence and an initial population is constructed. The current population is obtained by combining multiple rounds of evolutionary iterations. The performance indicators of each machine learning model are then obtained based on the parameter sequence in the current population. The target parameter sequence is screened out using the performance indicators as the measurement standard, and the target machine learning model is further determined based on the target parameter sequence, thereby solving the problem of low efficiency of traditional manual methods.

[0106] S26. According to the optimization objective function corresponding to the target task, determine the target hyperparameters corresponding to the target machine learning model from the preset hyperparameter space.

[0107] After determining the target machine learning model in the above S25, in order to improve task processing efficiency, it is necessary to further determine the corresponding target hyperparameters for the target machine learning model.

[0108] Specifically, the hyperparameter space refers to setting a set of value ranges of adjustable hyperparameters (such as learning rate [0.001, 0.1], tree depth [3, 10], batch size [16, 256], etc.) for different machine learning models (such as XGBoost, ResNet), and then determining the target hyperparameters corresponding to the target machine learning model from the preset hyperparameter space based on the optimization objective function corresponding to the target task.

[0109] For the target machine learning model, based on the task constraints of the target task (such as latency, accuracy, resource usage, etc.), select an adaptive hyperparameter combination from the hyperparameter space to avoid blind hyperparameter selection and ensure that the model achieves optimal performance while meeting the constraints.

[0110] In the embodiment of the present application, after obtaining the target machine learning model, the target hyperparameters corresponding to the target machine learning model are determined from the preset hyperparameter space through the optimization objective function corresponding to the target task. On the basis of determining the target machine learning model, the hyperparameters are obtained in a targeted manner to form a complete process of model selection and hyperparameter adaptation, thereby improving the overall task processing efficiency.

[0111] Specifically, according to the optimization objective function corresponding to the target task, determining the target hyperparameters corresponding to the target machine learning model from the preset hyperparameter space includes the following steps 1 and 2:

[0112] Step 1: Construct the optimization objective function and optimization boundary conditions.

[0113] In the embodiment of this application, the core requirements of the target task are decomposed into quantifiable optimization objectives (such as prediction accuracy, resource consumption, and operating efficiency), and the direction of hyperparameter optimization is clarified through the form of mathematical functions. Specifically, the optimization objective function includes three-dimensional objective functions, which are constructed by comprehensively considering the prediction performance, resource consumption, and operating efficiency of the model, namely: prediction accuracy objective function , used to measure the prediction accuracy of the model, reflecting the accuracy of the model's prediction of the target task (such as classification, regression); resource consumption objective function , used to quantify the resource consumption of the model, and the natural logarithmic transformation can smooth the numerical range of video memory occupancy, which is convenient for optimizing calculations; the operating efficiency objective function , which is used to measure the operating efficiency of the model and directly reflects the time cost during model training or inference.

[0114] Furthermore, combined with the core requirements of the three-dimensional objective function (prediction accuracy, resource consumption, and operational efficiency), the optimization boundary conditions of each dimension are clarified. That is, for each objective function, corresponding constraints are set to limit the optimization of the hyperparameter group to a practically feasible range. Specifically, corresponding thresholds can be set as constraints. The specifics include the following:

[0115] (1) Set optimization boundary conditions for prediction accuracy, that is, set the maximum allowable value of the weighted prediction error (i.e., the accuracy threshold). For example, if the task requires a classification accuracy of no less than 90%, the output of the prediction accuracy objective function (weighted prediction error) is constrained to no more than 10% to avoid substandard core performance due to excessive pursuit of efficiency or resource conservation.

[0116] (2) Set optimization boundary conditions for resource consumption, that is, set an upper limit for video memory usage. Based on the actual video memory capacity of the hardware device (e.g., 4GB of video memory on the edge device), constrain the video memory usage (raw value before natural logarithm transformation) corresponding to the resource consumption objective function to not exceed the available video memory of the device, to prevent the model from failing due to insufficient video memory.

[0117] (3) Set optimization boundary conditions for operation efficiency, that is, set the maximum allowable time for a single iteration (for example, the real-time inference scenario requires that a single iteration take no more than 100 milliseconds), and constrain the output of the operation efficiency objective function to not exceed this time threshold to ensure that the model can complete training or inference within the specified time.

[0118] By constructing a multi-objective optimization function that integrates prediction performance, resource efficiency, and operating cost, and setting optimization boundary conditions based on task constraints, we can then select the optimal hyperparameter set from the pre-set hyperparameter space. By using the objective function to guide the direction and the boundary conditions to constrain the space, we achieve a deep balance between model performance and resource consumption, obtaining a more precise hyperparameter set. This also avoids the situation where a single objective (such as pursuing only accuracy) leads to overall unavailability (e.g., insufficient graphics memory to run).

[0119] It's important to note that in real-world tasks, there's often a trade-off between prediction accuracy, resource consumption, and operational efficiency (for example, high-precision models may consume more video memory and take longer to run). By combining the objective function with boundary conditions, the resulting hyperparameter combination can meet core requirements while adapting to the resource and efficiency constraints of real-world scenarios.

[0120] Step 2: Combine the optimization objective function and optimization boundary conditions to determine the target hyperparameter group from the preset hyperparameter space.

[0121] In some embodiments, the optimization objective function and the optimization boundary conditions may be combined to search for a hyperparameter group from a preset hyperparameter space to determine a target hyperparameter group.

[0122] Specifically, combining the optimization objective function and the optimization boundary conditions, determining the target hyperparameter group from the preset hyperparameter space includes the following steps:

[0123] Step 201: Obtain a set of hyperparameter groups from a preset hyperparameter space.

[0124] The hyperparameter groups in the hyperparameter group set are multiple hyperparameter groups sampled from a preset hyperparameter space.

[0125] Specifically, a method for obtaining a set of hyperparameter groups from a preset hyperparameter space can be to use a Sobol low-discrepancy sequence to uniformly and efficiently sample the hyperparameter space, generate a set of hyperparameter groups covering key areas of the parameter space, ensure that the candidate solutions subsequently screened are globally representative, and avoid the problems of local concentration and omission of high-quality areas caused by random sampling.

[0126] The Sobol sequence is a low-discrepancy sequence mathematically constructed to ensure uniform coverage of the hyperparameter space with sample points (compared to random sampling, it can cover more area with fewer samples). For a space with m hyperparameter dimensions, setting the number of samples N to 200 (adjustable based on parameter complexity and computational resources) generates 200 hyperparameter sets, which serve as the basis for subsequent screening.

[0127] Each dimension of the hyperparameter group (such as learning rate, batch size, and number of network layers) takes values ​​within the preset hyperparameter space (such as learning rate [0.001, 0.1], batch size [16, 256]), ensuring that the sampling does not exceed the potentially feasible area of ​​the task constraints.

[0128] This ensures that the generated set of hyperparameter groups not only covers the diversity of the hyperparameter space (avoiding missing high-quality solutions), but also reduces invalid sampling through low-difference sequences to improve subsequent screening efficiency.

[0129] Step 202: Based on the optimization objective function and the optimization boundary conditions, filter the set of hyperparameter groups to obtain a target hyperparameter group.

[0130] Furthermore, guided by the optimization objective function constructed above and taking the optimization boundary conditions as the bottom line, a set of hyperparameter groups to be used that both meets the constraints and has optimization potential is further screened out from the sampled hyperparameter group sets, that is, based on the optimization objective function and the optimization boundary conditions, the hyperparameter group sets are screened to obtain the set of hyperparameter groups to be used; and then the hyperparameter group sets to be used are gradient optimized to obtain the target hyperparameter group.

[0131] Specifically, the hyperparameter group to be used is first obtained based on the optimization objective function and boundary conditions. The specific method can be through constructing an initial Pareto frontier; wherein, the Pareto frontier refers to the set of hyperparameter groups to be used composed of Pareto optimal solutions screened out from the set of hyperparameter groups. Furthermore, the Pareto optimal solution means that there is no other hyperparameter group that is better than the hyperparameter group in all optimization objective functions (prediction accuracy, resource consumption, and operating efficiency).

[0132] Then, for the 200 sets of hyperparameters obtained by sampling, we first filter out the hyperparameter groups that violate the constraints by optimizing boundary conditions (such as accuracy threshold and memory limit); then we select the Pareto optimal solution from the remaining valid groups to form the initial Pareto frontier, that is, the hyperparameter group to be used, which serves as the starting point for subsequent local optimization to improve optimization efficiency.

[0133] The initial Pareto frontier, that is, the hyperparameter group in the hyperparameter group to be used, is further optimized through multi-objective gradient descent and dynamic adjustment strategy to finally determine the target hyperparameter group.

[0134] Multi-objective gradient descent starts with the hyperparameter set of the initial Pareto front and adjusts the hyperparameters through the gradient descent algorithm to iterate the three-dimensional objective function value in a more optimal direction (such as reducing prediction error, reducing video memory usage, and shortening time).

[0135] At the same time, dynamic weight adjustment can be performed, that is, the weight of the three-dimensional objective function can be adjusted in real time according to the resource surplus rate. For example, when resources are tight, the weight of the resource consumption objective function can be increased to prioritize video memory adaptation. A constraint processing mechanism can also be used to impose penalties on hyperparameter groups that may violate constraints during iteration through an adaptive penalty function (the penalty coefficient increases with the number of iterations) to ensure that the final result meets the boundary conditions.

[0136] Finally, after multiple rounds of iterations, the group with the best overall performance is selected from the optimized hyperparameter groups as the target hyperparameter group.

[0137] S27. Process the target task based on the target machine learning model and target hyperparameters.

[0138] Furthermore, the determined target machine learning model is used as the basic framework, and is combined with the selected target hyperparameters to be applied to the processing process of the target task. The model and hyperparameters can be used to train the input data of the target task and perform other operations to achieve the expected effect of the target task.

[0139] It should be noted that after processing the target task based on the target machine learning model and target hyperparameters, the actual performance of the target machine learning model in processing the target task can also be collected, and then the knowledge graph can be dynamically optimized in reverse to improve the accuracy of the knowledge graph.

[0140] This is because the knowledge graph stores the matching relationship between models and tasks. However, the performance of models in real-world scenarios changes, and new tasks are constantly emerging. If the matching relationship in the knowledge graph remains fixed, it will become increasingly difficult to adapt to actual needs. Therefore, the knowledge graph needs to be dynamically adjusted according to actual conditions to adapt to various scenarios.

[0141] Specifically, in order to enable the knowledge graph to adjust the adaptation relationship between the model and the task in real time based on the execution results of the task, the edge of the knowledge graph stores the model confidence corresponding to the matching relationship between the model and the task, and the model confidence is adjusted according to the corresponding weight. When the target task is executed, the dual-path reinforcement learning logic can be triggered based on the performance of the previously determined target machine learning model. That is, if the task is successful, the weight corresponding to the matching relationship between the corresponding model and the task is increased according to the set rules. As the weight increases, the knowledge graph has more sufficient trust in the model's adaptation to this task, and the model confidence accumulates positively. If the task fails, the associated weight is reduced. The reduced weight weakens the knowledge graph's recognition of the model's adaptability, and the model confidence is simultaneously reduced.

[0142] Combined with the methods of the above embodiments, a task processing system is built. The system can receive tasks and task information uploaded by users, and then obtain the corresponding machine learning model and hyperparameter group based on the task information; the system can simultaneously receive tasks and task information uploaded by multiple users. In order to ensure the efficient and stable operation of the entire system, a priority queue management module is also created in the processing system to give priority to scheduling tasks that are highly urgent and have sufficient resources. Through the priority queue management module, resources can be reasonably allocated according to the urgency of the task and resource availability, thereby improving the overall efficiency of the system.

[0143] The processing system also includes a performance-resource balancing controller, which can adjust the system's performance and resource allocation in real time to ensure system stability and performance.

[0144] It should be noted that different problems may arise when the processing system is running. To deal with different problems (for example, system fluctuations, resource shortages, and serious failures), the following steps are adopted in the embodiments of the present application:

[0145] Step a: Monitor the status of the task processing system.

[0146] In the embodiment of the present application, the method for monitoring the status of the task processing system is not limited in any way and is set according to actual conditions.

[0147] Step b: When the execution status of the task processing system is faulty, the target machine learning model is switched to a backup machine learning model, or the task processing system is reset.

[0148] Among them, the backup machine learning model is any one in the set of machine learning models to be used except the target machine learning model.

[0149] When the system detects a minor fault or performance fluctuation, it quickly switches to a pre-configured backup model, ensuring service continuity and availability while buying time for troubleshooting and repair. The backup machine learning model is any one of the machine learning models in the set to be used, except the target machine learning model.

[0150] When a serious system failure occurs and switching to a backup machine learning model or a lightweight model fails to solve the problem, the cold start protocol is triggered, which resets the entire system and reinitializes all components and configurations to restore to the initial state. Although this involves a full system restart, it can completely eliminate the root cause of the failure and ensure that the system returns to normal operation. It is suitable for fault recovery in extreme situations.

[0151] Step c: When the execution status of the task processing system is insufficient resources, the target machine learning model is lightweight processed.

[0152] When the system faces resource constraints (such as insufficient computing resources or high memory pressure), lightweight mode is automatically enabled. In this mode, the system optimizes resource allocation and appropriately reduces model accuracy (accuracy reduction is controlled within 3%) to ensure stable system operation under limited resources. This strategy is particularly effective in resource-constrained environments, maximizing the use of existing resources while ensuring basic functionality.

[0153] Furthermore, through the above method, the system can respond flexibly to different failure scenarios, maximizing the stability and availability of the system while taking into account resource utilization efficiency and performance.

[0154] As an extension and refinement of the above embodiment, the processing system also includes a performance monitoring model (e.g., a Long Short-Term Memory (LSTM) anomaly detector) that inputs continuous data generated during the process of determining the target machine learning model and target hyperparameters into the LSTM neural network. The LSTM neural network learns the normal changes in this data and outputs a value representing whether the current optimization state is healthy.

[0155] The current state is then compared to the previous state. If the difference is particularly large (exceeding the threshold of 3 standard deviations), it is determined that there is a problem in the model selection process and the target hyperparameter determination process. At this time, the re-determination process can be triggered immediately and adjustment strategies can be implemented to ensure the stable progress of the system.

[0156] At the same time, the data involved in the process of determining the target model and target hyperparameters each time are recorded and saved, and stored in the historical experience pool, thereby providing reference data for subsequent tasks.

[0157] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0158] The embodiments of the present application further provide a task processing device. For descriptions of features in the embodiments corresponding to the task processing device, reference can be made to the relevant descriptions of the embodiments corresponding to the task processing method, which will not be repeated here.

[0159] In some embodiments, reference Figure 4 As shown, an embodiment of the present disclosure provides a task processing device 400, which includes:

[0160] An acquisition unit 410 is configured to acquire target task information, wherein the target task information includes a feature quantification evaluation matrix corresponding to the target task, a task type, and task constraints; the feature quantification evaluation matrix includes structural features and dynamic features corresponding to a sample data set of the target task;

[0161] A matching unit 420 is configured to match the target task information with each piece of historical task information in a historical task information set, and select a preset number of pieces of historical task information to generate a set of machine learning models to be used based on the historical machine learning models corresponding to the preset number of pieces of historical task information; the historical task information set includes feature quantification evaluation matrices, task types, and task constraints corresponding to the plurality of historical tasks, respectively.

[0162] A determining unit 430 is configured to determine a target machine learning model from the set of machine learning models to be used;

[0163] The processing unit 440 is used to process the target task based on the target machine learning model.

[0164] As an optional implementation of the embodiment of the present disclosure, the matching unit 420 is specifically used to calculate the task similarity between the target task information and each historical task information in the historical task information set; based on the task similarity, screen out the preset number of historical tasks; and generate the set of machine learning models to be used according to the historical machine learning models corresponding to the preset number of historical tasks.

[0165] As an optional implementation of the embodiment of the present disclosure, the matching unit 420 is specifically configured to obtain the task similarity between the target task information and each historical task information in the historical task information set based on a similarity calculation formula, combining the target task information and the historical task information set; wherein the similarity calculation formula is:

[0166]

[0167] in, represents the target task, represents the said historical mission; Indicates the task similarity between the target task and the historical task; Indicates the target task information, Indicates the historical task information; Represents the time attenuation coefficient, which is used to reduce the impact of time interval on similarity; Indicates the execution timestamp of the target task, Indicates the execution timestamp of the historical task.

[0168] As an optional implementation of the embodiment of the present disclosure, the determination unit 430 is specifically used to obtain a model parameter set of each machine learning model in the set of machine learning models to be used, and encode the model parameter set of each machine learning model into a parameter sequence set; based on the parameter sequence set, construct an initial parameter population; perform multiple rounds of evolutionary iterations on the initial parameter population, and after each round of evolutionary iteration, calculate the model performance indicators corresponding to each parameter sequence in the current parameter population, and screen the target parameter sequence based on the model performance indicators; determine the machine learning model corresponding to the target parameter sequence as the target machine learning model.

[0169] As an optional implementation of the embodiment of the present disclosure, the processing unit 440 is specifically used to determine the target hyperparameters corresponding to the target machine learning model from the preset hyperparameter space according to the optimization objective function corresponding to the target task; and process the target task based on the target machine learning model and the target hyperparameters.

[0170] As an optional implementation of the embodiment of the present disclosure, the processing unit 440 is specifically used to construct an optimization objective function and optimization boundary conditions: combining the optimization objective function and the optimization boundary conditions, and determining a target hyperparameter group from a preset hyperparameter space.

[0171] As an optional implementation of the embodiment of the present disclosure, the processing unit 440 is specifically used to obtain a set of hyperparameter groups from the preset hyperparameter space; the hyperparameter groups in the hyperparameter group set are multiple hyperparameter groups sampled from the preset hyperparameter space; based on the optimization objective function and the optimization boundary conditions, the hyperparameter group set is screened to obtain the target hyperparameter group.

[0172] As an optional implementation of the embodiment of the present disclosure, the processing unit 440 is specifically used to screen from the set of hyperparameter groups based on the optimization objective function and the optimization boundary conditions to obtain a set of hyperparameter groups to be used; and perform gradient optimization on the set of hyperparameter groups to be used to obtain a target hyperparameter group.

[0173] As an optional implementation of the embodiment of the present disclosure, the acquisition unit 410 is also used to construct a knowledge graph including multiple machine learning models and task adaptation relationships corresponding to each machine learning model; the nodes in the knowledge graph are machine learning models, and the edges are composed of adaptation conditions, model confidence, and the number of historical model calls; based on the knowledge graph, multiple historical tasks are screened out; and task information corresponding to the multiple historical tasks is obtained to generate the historical task information set.

[0174] As an optional implementation of the embodiment of the present disclosure, the acquisition unit 410 is also used to update the model confidence in the knowledge graph according to the task execution status of the target machine learning model.

[0175] As an optional implementation of the embodiment of the present disclosure, the processing unit 440 is also used to monitor the status of the task processing system; when the execution status of the task processing system is faulty, the target machine learning model is switched to a backup machine learning model, or the task processing system is reset; the backup machine learning model is any one of the set of machine learning models to be used except the target machine learning model; when the execution status of the task processing system is insufficient resources, the target machine learning model is lightweight processed.

[0176] An embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned task processing method embodiments.

[0177] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the steps of any of the above-mentioned task processing method embodiments when running.

[0178] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0179] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the above-mentioned task processing method embodiments are implemented.

[0180] An embodiment of the present application further provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above-mentioned task processing method embodiments are implemented.

[0181] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0182] The above is a detailed introduction to a task processing method and device provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A task processing method, characterized in that: Applied to a task processing system, the method includes: Obtain target task information; the target task information includes a feature quantification evaluation matrix corresponding to the target task, a task type, and task constraints; the feature quantification evaluation matrix includes structural features and dynamic features corresponding to a sample data set of the target task; Matching the target task information with each piece of historical task information in a historical task information set, screening out a preset number of pieces of historical task information, and generating a set of machine learning models to be used based on the historical machine learning models corresponding to the preset number of pieces of historical task information; the historical task information set includes feature quantification evaluation matrices, task types, and task constraints corresponding to the plurality of historical tasks, respectively; Determine a target machine learning model from the set of machine learning models to be used; Processing the target task based on the target machine learning model; The target task information is matched with each historical task information in the historical task information set to screen out a preset number of historical task information, and a set of machine learning models to be used is generated based on the historical machine learning models corresponding to the preset number of historical task information, including: Calculating task similarity between the target task information and each piece of historical task information in the historical task information set; Based on the task similarity, screening out the preset number of historical tasks; The set of machine learning models to be used is generated based on the historical machine learning models corresponding to the preset number of historical tasks.

2. The method according to claim 1, characterized in that The calculating the task similarity between the target task information and each piece of historical task information in the historical task information set includes: According to the similarity calculation formula, the target task information and the historical task information set are combined to obtain the task similarity between the target task information and each historical task information in the historical task information set; wherein the similarity calculation formula is: in, represents the target task, represents the said historical mission; Indicates the task similarity between the target task and the historical task; Indicates the target task information, Indicates the historical task information; Represents the time attenuation coefficient, which is used to reduce the impact of time interval on similarity; Indicates the execution timestamp of the target task, Indicates the execution timestamp of the historical task.

3. The method according to claim 1, characterized in that The step of determining a target machine learning model from the set of machine learning models to be used comprises: Obtain a model parameter set for each machine learning model in the set of machine learning models to be used, and encode the model parameter set for each machine learning model into a parameter sequence set; Based on the parameter sequence set, construct an initial parameter population; Performing multiple rounds of evolutionary iterations on the initial parameter population, and after each round of evolutionary iterations, calculating the model performance index corresponding to each parameter sequence in the current parameter population, and selecting the target parameter sequence based on the model performance index; The machine learning model corresponding to the target parameter sequence is determined as the target machine learning model.

4. The method according to claim 1, wherein Processing the target task based on the target machine learning model includes: Determining target hyperparameters corresponding to the target machine learning model from a preset hyperparameter space according to the optimization objective function corresponding to the target task; Process the target task based on the target machine learning model and the target hyperparameters.

5. The method according to claim 4, characterized in that Determining target hyperparameters corresponding to the target machine learning model from a preset hyperparameter space based on the optimization objective function corresponding to the target task includes: Construct the optimization objective function and optimization boundary conditions: In combination with the optimization objective function and the optimization boundary conditions, a target hyperparameter group is determined from a preset hyperparameter space.

6. The method according to claim 5, characterized in that The step of combining the optimization objective function and the optimization boundary condition to determine a target hyperparameter group from a preset hyperparameter space includes: Acquire a set of hyperparameter groups from the preset hyperparameter space; the hyperparameter groups in the set of hyperparameter groups are multiple hyperparameter groups sampled from the preset hyperparameter space; Based on the optimization objective function and the optimization boundary conditions, the set of hyperparameter groups is screened to obtain the target hyperparameter group.

7. The method according to claim 6, characterized in that The step of screening the set of hyperparameter groups based on the optimization objective function and the optimization boundary conditions to obtain the target hyperparameter group includes: Based on the optimization objective function and the optimization boundary conditions, screening from the set of hyperparameter groups to obtain a set of hyperparameter groups to be used; Gradient optimization is performed on the set of hyperparameter groups to be used to obtain a target hyperparameter group.

8. The method according to claim 2, characterized in that Before calculating the task similarity between the target task information and each piece of historical task information in the historical task information set, the method further includes: Construct a knowledge graph containing multiple machine learning models and the task adaptation relationships corresponding to each machine learning model; the nodes in the knowledge graph are machine learning models, and the edges are composed of adaptation conditions, model confidence, and the number of historical model calls; Filtering out a plurality of the historical tasks based on the knowledge graph; The task information corresponding to the plurality of historical tasks is obtained to generate the historical task information set.

9. The method according to claim 8, characterized in that The method further comprises: After the target task is executed, the model confidence in the knowledge graph is updated according to the task execution status of the target machine learning model.

10. The method according to claim 1, characterized in that The method further comprises: monitoring the status of the task processing system; When the execution state of the task processing system is faulty, the target machine learning model is switched to a backup machine learning model, or the task processing system is reset; the backup machine learning model is any one of the set of machine learning models to be used except the target machine learning model; When the execution status of the task processing system is insufficient resources, the target machine learning model is lightweight processed.

11. A task processing device, characterized in that: include: An acquisition unit, used to acquire target task information; The target task information includes a feature quantification evaluation matrix corresponding to the target task, a task type, and task constraints; The feature quantification evaluation matrix includes structural features and dynamic features corresponding to the sample data set of the target task; a matching unit, configured to match the target task information with each piece of historical task information in a historical task information set, and screen out a preset number of pieces of historical task information, so as to generate a set of machine learning models to be used based on the historical machine learning models corresponding to the preset number of pieces of historical task information; the historical task information set includes feature quantification evaluation matrices, task types, and task constraints corresponding to the plurality of historical tasks; a determination unit, configured to determine a target machine learning model from the set of machine learning models to be used; a processing unit, configured to process the target task based on the target machine learning model; The matching unit is specifically used to calculate the task similarity between the target task information and each historical task information in the historical task information set; based on the task similarity, screen out the preset number of historical tasks; and generate the set of machine learning models to be used based on the historical machine learning models corresponding to the preset number of historical tasks.

12. An electronic device, characterized in that: include: memory for storing computer programs; A processor, configured to implement the steps of the task processing method according to any one of claims 1 to 10 when executing the computer program.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the task processing method according to any one of claims 1 to 10.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the task processing method according to any one of claims 1 to 10 are implemented.

Images