Script file generation method, task processing method and electronic device
Patent Information
- Application Number
- CN202411224854.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2044-09-02
Smart Images
Figure CN118733006B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and in particular to a script file generation method, a task processing method and an electronic device. Background Art
[0002] Automatic machine learning is a special form of machine learning that focuses on the process-based and automated implementation of some or multiple links in the machine learning development process. By using a variety of automatic learning technologies and algorithms, it can effectively lower the threshold for machine learning development, better serve business experts who lack machine learning expertise, reduce the workload of algorithm engineers, and improve the efficiency of algorithm development, deployment, and application.
[0003] In existing automatic machine learning solutions, although automation has been achieved for each link of algorithm modeling, the combined construction of the overall modeling process generally still needs to be completed manually by algorithm engineers. Summary of the invention
[0004] In view of this, the present application provides a script file generation method, a task processing method, a device and an electronic device.
[0005] Specifically, the present application is implemented through the following technical solutions:
[0006] According to a first aspect of an embodiment of the present application, a script file generation method is provided, comprising:
[0007] Obtaining user input information; wherein the user input information includes data information, evaluation indicators and expected effects related to the modeling task; the data information includes description information of the input data set;
[0008] Constructing a large language model prompt word according to the user input information and the scene knowledge base;
[0009] Input the large language model prompt word into the large language model, so that the large language model determines the task flow plan according to the large language model prompt word, and selects the target module to be called from the candidate tool modules, and generates an algorithm modeling Pipeline script file according to the task flow plan and the target module; wherein the candidate tool modules include the AutoML module and the specific modeling module;
[0010] When it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements according to the evaluation index and the expected effect, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied;
[0011] When it is determined that the effect evaluation result of the algorithm modeling Pipeline does not meet the requirements based on the evaluation index and the expected effect, the construction of the large language model prompt words and the generation of the algorithm modeling Pipeline script file are iteratively executed until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied.
[0012] According to a second aspect of an embodiment of the present application, a task processing method is provided, including:
[0013] Get the data to be processed;
[0014] The data to be processed is input into a task processing model, and the task processing result is determined according to the output result of the task processing model; wherein the task processing model is generated according to the algorithm modeling Pipeline script file to be applied provided in the first aspect.
[0015] According to a third aspect of an embodiment of the present application, there is provided a task processing device, including:
[0016] An acquisition unit, used for acquiring data to be processed;
[0017] A processing unit is used to input the data to be processed into a task processing model, and determine the task processing result based on the output result of the task processing model; wherein the task processing model is generated based on the Pipeline script file of the algorithm modeling to be applied provided by the first aspect.
[0018] According to a fourth aspect of an embodiment of the present application, an electronic device is provided, comprising a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the method provided in the first aspect or the second aspect.
[0019] According to a fifth aspect of an embodiment of the present application, a machine-readable storage medium is provided, wherein the machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a processor, the method provided in the first aspect or the second aspect is implemented.
[0020] The script file generation method of the embodiment of the present application obtains user input information, constructs a large language model prompt word based on the user input information and a scenario knowledge base, and inputs the large language model prompt word into the large language model, so that the large language model determines the task flow plan based on the large language model prompt word, and selects the target module to be called from the candidate tool module, and generates an algorithm modeling Pipeline script file based on the task flow plan and the target module, and then, when it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; when it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; When the effect evaluation result of the algorithm modeling Pipeline script file does not meet the requirements, the construction of the large language model prompt words and the generation of the algorithm modeling Pipeline script file are iteratively executed until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied. By introducing the large language model into the automatic machine learning method, using the large language model as the agent, the AutoML module and the specific modeling module as tools, the good task understanding and planning capabilities of the large language model are combined with the advantages of the automatic machine learning core capability module in fast machine learning modeling, thereby realizing a more intelligent machine learning modeling method. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a flowchart of a script file generation method shown in an exemplary embodiment of the present application;
[0022] Figure 2 is a flowchart of a task processing method shown in an exemplary embodiment of the present application;
[0023] Figure 3 It is a schematic diagram of an implementation flow of a large model enhanced script file generation method shown in an exemplary embodiment of the present application;
[0024] Figure 4 It is a schematic diagram of an implementation flow of a large model enhanced script file generation method shown in an exemplary embodiment of the present application;
[0025] Figure 5 It is a structural schematic diagram of a script file generating device shown in an exemplary embodiment of the present application;
[0026] Figure 6 It is a structural schematic diagram of a script file generating device shown in an exemplary embodiment of the present application;
[0027] Figure 7 is a structural schematic diagram of a task processing device shown in an exemplary embodiment of the present application;
[0028] Figure 8 It is a schematic diagram of the hardware structure of an electronic device shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION
[0029] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application, some technical terms involved in the embodiments of the present application are briefly explained below.
[0030] 1. Automatic Machine Learning (AutoML): A technology that automates some or multiple links, or even the entire process, during the development and operation of traditional machine learning or deep learning models.
[0031] 2. Large Language Model (LLM): A deep learning model trained using large amounts of text data that can generate natural language text or understand the meaning of language text.
[0032] In order to make the above-mentioned purposes, features and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are further described in detail below with reference to the accompanying drawings.
[0033] See also Figure 1 , is a flow chart of a script file generation method provided in an embodiment of the present application, wherein the script file generation can be applied to a server, such as a cloud server or a local server, etc. Figure 1 As shown, the script file generation method may include the following steps:
[0034] Step S100, obtaining user input information; wherein the user input information includes data information, evaluation indicators and expected effects related to the modeling task; the data information includes description information of the input data set.
[0035] Exemplarily, the user input information may mainly include information (which may be referred to as data information) used to describe an input data set related to a modeling task, for example, describing what types of data the input data set includes, the role of each type of data, and the like.
[0036] Exemplarily, since the input data set is usually a data table, the above data information can also be called data table meta-information, which is used to describe the type of the data table, the relationship between each column of data in the data table, the function of each column of data, etc.
[0037] Exemplarily, the evaluation index refers to an index used to evaluate the algorithm model (which may be called a task processing model) constructed and trained in the automatic machine learning process, such as accuracy, detection rate, recall rate or F1 score.
[0038] The expected effect is the effect that the algorithm model is expected to achieve and corresponds to the evaluation index.
[0039] For example, assuming the evaluation indicator is accuracy, the expected effect can be an accuracy rate exceeding 90%.
[0040] Step S110: construct a large language model prompt word based on the user input information and the scene knowledge base.
[0041] Step S120: input the large language model prompt words into the large language model so that the large language model determines the task flow plan based on the large language model prompt words, and selects the target module to be called from the candidate tool modules, and generates an algorithm modeling Pipeline script file based on the task flow plan and the target module; wherein the candidate tool modules include the AutoML module and the specific modeling module.
[0042] In the embodiments of the present application, in order to realize a more intelligent machine learning modeling method and further evolve and upgrade the original automatic machine learning into autonomous machine learning, a large language model can be introduced into the automatic machine learning method. The large language model is used as an agent, and the AutoML module and the specific modeling module are used as tools. The good task understanding and planning capabilities of the large language model are combined with the advantages of the AutoML core capability module in fast machine learning modeling.
[0043] In addition, in order to improve the accuracy of the large language model's understanding of the task, a scenario knowledge base can also be constructed. In the process of constructing the prompt words (Prompt) of the large language model, the prompt words of the constructed large prediction model can also be enhanced based on the relevant knowledge in the scenario knowledge base.
[0044] Exemplarily, in the initial state, the scenario knowledge base may include relevant technical documents of one or more scenarios collected in advance. With the application of the technical solution provided in the embodiment of the present application, the scenario knowledge base may add new knowledge data based on the modeling task and the script files generated during the task process.
[0045] Accordingly, when user input information is obtained, a large language model prompt word can be constructed based on the obtained user input information and the scenario knowledge base. The large language model prompt word can serve as a medium for user interaction with the large language model. The large language model can automatically implement task process planning and algorithm modeling based on the large language model prompt word.
[0046] Exemplarily, the large language model prompt words may include but are not limited to an instruction part and a data part. The instruction part is mainly used to indicate the task objectives and task requirements of the large language model, and the data part may mainly include relevant descriptive information of the task, which may include but is not limited to: user input information, tool list information, etc.
[0047] Exemplarily, the above tool list information may include commonly used automatic machine learning capability modules, such as AUM (asset management scale) feature unit, general detailed list feature unit, parameter tuning feature unit, etc.
[0048] Exemplarily, the above tool list may be defined and packaged by a user.
[0049] Exemplarily, for any tool in the tool list, it may include three elements: function, input, and output of the tool.
[0050] Exemplarily, in an embodiment of the present application, the candidate tool module may include, in addition to the AutoML module provided in the above tool list, a specific modeling module (such as a general third-party modeling tool), for example, sklearn (a Python library for machine learning), Pandas (a popular open source data analysis toolkit for the Python programming language), and other modules.
[0051] In the embodiment of the present application, the constructed large language model prompt words can be input into the large language model.
[0052] The large language model can determine the task process planning based on the input large language model prompt words, and select the automatic machine learning capability module (which can be called the target module) to be called from the candidate automatic machine learning capability modules, and determine the algorithm modeling Pipeline based on the task process planning and the target module.
[0053] For example, an algorithm modeling pipeline may include:
[0054] 1. Problem definition and data collection.
[0055] For example, the specific problem and goal of modeling are determined. For example, whether the task is a classification task or a regression task, the prediction goal is determined, and the indicators for evaluating the model are determined. And relevant data sets are collected (the data sets can be provided by the user).
[0056] 2. Data preprocessing.
[0057] 2.1. Data cleaning: handling missing values, outliers and duplicate values.
[0058] 2.2. Feature engineering: Select, transform and extract features, which may include standardization, normalization, encoding categorical variables, etc.
[0059] Exemplarily, according to the task flow planning, the selected target module can be used to extract features from the corresponding data to achieve feature construction.
[0060] 2.3. Dataset segmentation: Divide the dataset into training set and test set, etc.
[0061] 3. Model selection and training.
[0062] 3.1. Choose an appropriate model based on the problem type (e.g., decision tree, logistic regression, neural network, etc.).
[0063] 3.2. Train the model on the training set and adjust the model hyperparameters to optimize performance.
[0064] 4. Model evaluation.
[0065] For example, the performance of the model is evaluated on the test set, and appropriate evaluation indicators (such as accuracy, precision, recall, root mean square error, etc.) are used to evaluate the predictive ability of the model.
[0066] Step S130: When the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied.
[0067] Step S140: When it is determined that the effect evaluation result of the algorithm modeling Pipeline script file does not meet the requirements based on the evaluation indicators and the expected effects, iteratively execute the construction of the large language model prompt words and the generation of the algorithm modeling Pipeline script file until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied.
[0068] In an embodiment of the present application, when an algorithm modeling Pipeline script file is generated in the above manner, it can be determined whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation indicators and expected effects (the evaluation indicators and expected effects included in the user input information).
[0069] For example, assuming that the evaluation indicator is the accuracy of the model, and the expected effect is that the accuracy of the model corresponding to the algorithm modeling Pipeline is not lower than the preset threshold, then when the accuracy of the model corresponding to the algorithm modeling Pipeline is not lower than the preset accuracy threshold, it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements; when the accuracy of the model corresponding to the algorithm modeling Pipeline is lower than the preset accuracy threshold, it is determined that the effect evaluation result of the algorithm modeling Pipeline script file does not meet the requirements.
[0070] Among them, the model corresponding to the algorithmic modeling Pipeline can be obtained by running the algorithmic modeling Pipeline script file.
[0071] Exemplarily, when it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation indicators and the expected effects, the algorithm modeling Pipeline script file can be determined as the algorithm modeling Pipeline script file to be applied.
[0072] Exemplarily, when it is determined that the effect evaluation result of the algorithm modeling Pipeline does not meet the requirements based on the evaluation indicators and expected effects, the construction of the large language model prompt words and the generation of the algorithm modeling Pipeline script file are iteratively executed until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied.
[0073] Exemplarily, the preset iteration stop condition may include that the number of iterations reaches a preset maximum number of iterations, or that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements.
[0074] It can be seen that in Figure 1In the method flow shown, by obtaining user input information, a large language model prompt word is constructed based on the user input information and the scenario knowledge base, and the large language model prompt word is input into the large language model, so that the large language model determines the task flow plan based on the large language model prompt word, and selects the target module to be called from the candidate tool modules, and generates an algorithm modeling Pipeline script file based on the task flow plan and the target module, and then, when it is determined based on the evaluation indicators and the expected effect that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; when it is determined based on the evaluation indicators and the expected effect that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; When the effect evaluation result of a certain algorithm modeling Pipeline script file does not meet the requirements, the construction of the large language model prompt words and the processing of the algorithm modeling Pipeline script file generation are iteratively executed until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied. By introducing a large language model into the automatic machine learning method, using the large language model as the agent, and the AutoML module and the specific modeling module as tools, the good task understanding and planning capabilities of the large language model are combined with the advantages of the automatic machine learning core capability module in fast machine learning modeling, thereby realizing a more intelligent machine learning modeling method.
[0075] In some embodiments, determining whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect may include:
[0076] According to the input data set, the feature construction module in the target module is used to construct features, and the feature enhancement module in the target module is used to enhance the constructed features, and a sample set is generated according to the enhanced features;
[0077] Use the XGB model and the LGB model to build the algorithm model, and use the HPO module in the target module and the sample set to determine the optimal hyperparameters of the algorithm model;
[0078] Based on the evaluation indicators, expected effects, and optimal hyperparameters, determine whether the effect evaluation results of the algorithm modeling Pipeline script file meet the requirements.
[0079] Exemplarily, for any algorithm modeling Pipeline script file generated in the above manner, in the process of evaluating the effect of the algorithm modeling Pipeline script file, feature construction can be performed based on the input data set using the feature construction module in the target module, and the constructed features can be enhanced using the feature enhancement module in the target module, and a sample set can be generated based on the enhanced features.
[0080] Exemplarily, the feature building blocks may include dedicated feature building blocks and / or general feature building blocks.
[0081] Accordingly, for the data in the input data set (such as a data table), if there is a dedicated feature construction module, the dedicated feature construction module can be used to construct features for the data; if there is no dedicated feature construction module, the general feature construction module can be used to construct features for the data.
[0082] Exemplarily, for the features obtained by constructing the input data set, a feature enhancement module can be used to enhance the constructed features, and a sample set can be generated based on the enhanced features to optimize the model performance of the trained algorithm model.
[0083] For example, in the process of algorithm model building, the XGB (an efficient open source machine learning framework) model and the LGB (Light Gradient Boosting Machine) model can be used to build the algorithm model, and the HPO (Hyperparameter Optimization) module can be used to obtain the optimal hyperparameters of the algorithm model.
[0084] Exemplarily, hyperparameters may include but are not limited to tree depth and learning rate.
[0085] Exemplarily, the HPO module can determine multiple groups of candidate hyperparameters within the value range of each hyperparameter, and use the sample set to evaluate the effect of each group of candidate hyperparameters respectively, and then determine the optimal hyperparameters based on the effect of each group of candidate hyperparameters.
[0086] For example, for any group of candidate hyperparameters, the training set included in the sample set can be used to train the algorithm model under this group of hyperparameters, and the trained algorithm model can be tested using the test set, thereby obtaining the evaluation index values corresponding to the training set (such as precision value, accuracy value, etc.) and the evaluation index values corresponding to the test set, and then determining the optimal hyperparameters from each group of candidate hyperparameters.
[0087] For example, when the optimal hyperparameters of the algorithm model are determined, it can be determined whether the effect evaluation results of the algorithm modeling Pipeline script file meet the requirements based on the evaluation indicators, expected effects, and the optimal hyperparameters.
[0088] For example, based on the evaluation index value of the algorithm model corresponding to the test set under the optimal hyperparameters and the evaluation index value of the corresponding test set, if both meet the expected effect requirements, it can be determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements; otherwise, it is determined that the effect evaluation result of the algorithm modeling Pipeline script file does not meet the requirements.
[0089] In one example, the above-mentioned use of the HPO module in the target module and the sample set to determine the optimal hyperparameters of the algorithm model includes:
[0090] Divide the sample set into K folds; K ≥ 2;
[0091] For any set of hyperparameters determined by the HPO module, each of the K folds is used as a test set, and the remaining K-1 folds are used as training sets. The algorithm model is trained for K rounds, and the evaluation index values of the K algorithm models on the training set and the evaluation index values of the K algorithm models on the test set are obtained.
[0092] The average value of the evaluation index values of the K algorithm models on the training set is determined as the evaluation index value on the training set corresponding to the set of hyperparameters; and the average value of the evaluation index values of the K algorithm models on the test set is determined as the evaluation index value on the test set corresponding to the set of hyperparameters;
[0093] The optimal hyperparameters are determined based on the evaluation index values on the training set and the evaluation index values on the test set corresponding to each group of hyperparameters determined by the HPO module.
[0094] For example, in order to avoid overfitting of the algorithm model, the performance of the algorithm model can be evaluated by K-fold cross-validation.
[0095] Exemplarily, when a sample set is obtained in the above manner, the sample set may be divided into K folds (ie, the sample set may be divided into K parts).
[0096] Among them, the difference between the number of samples in each sample does not exceed the preset number threshold.
[0097] Exemplarily, for any set of hyperparameters determined using the HPO module, the algorithm model can be trained for K rounds by taking each of the K folds as a test set and the remaining K-1 folds as a training set.
[0098] For example, assuming that the sample set is divided into three folds (assuming they are Z1, Z2 and Z3), the algorithm model can be trained for one round with Z1 as the test set, Z2 and Z3 as the training sets, and the algorithm model can be trained for one round with Z2 as the test set, Z1 and Z3 as the training sets, and the algorithm model can be trained for one round with Z3 as the test set, Z1 and Z2 as the training sets (a total of 3 rounds of training).
[0099] By performing K rounds of training on the algorithm model in the above manner, the evaluation index values of the K algorithm models on the training set and the evaluation index values of the K algorithm models on the test set can be obtained. Furthermore, the average value of the evaluation index values of the K algorithm models on the training set can be determined as the evaluation index value on the training set corresponding to the group of hyperparameters; and the average value of the evaluation index values of the K algorithm models on the test set can be determined as the evaluation index value on the test set corresponding to the group of hyperparameters.
[0100] Still taking the previous example, assuming that the evaluation index values of the algorithm model on the test set in the three rounds of training are P11, P12 and P13 respectively, then (P11+P12+P13) / 3 can be used as the evaluation index value of the algorithm model on the training set corresponding to this set of hyperparameters.
[0101] Exemplarily, the optimal hyper-parameters may be determined based on the evaluation index values on the training set and the evaluation index values on the test set corresponding to each group of hyper-parameters determined by the HPO module.
[0102] In some embodiments, the obtaining of user input information may include:
[0103] When a modeling instruction sent by the terminal device is detected, an information input interface is displayed on a display screen of the terminal device;
[0104] Obtain user input information entered through the information input interface;
[0105] After determining the current algorithm modeling Pipeline script file as the algorithm modeling Pipeline script file to be applied, the following may also be included:
[0106] The algorithm modeling Pipeline script file to be applied is displayed on the display screen of the terminal device.
[0107] For example, a user may initiate a modeling task to a server through a terminal device (such as a PC or a mobile terminal).
[0108] When the server detects a modeling instruction sent by the terminal device, it displays an information input interface through the display screen of the terminal device. The user can input information related to the modeling task through the information input interface. Then, the server can obtain the user input information entered through the information input interface.
[0109] In addition, when the server determines the algorithm modeling Pipeline script file to be applied in the above manner, the algorithm modeling Pipeline script file to be applied can be displayed on the display screen of the terminal device.
[0110] In some embodiments, constructing a large language model prompt word based on user input information may include:
[0111] When it is determined that prompt word enhancement is needed, similar texts are searched from the knowledge base based on the task description information;
[0112] Based on user input information and similar texts, a large language model prompt word is constructed.
[0113] Exemplarily, in order to improve the accuracy of the large language model's understanding of the task, in the process of constructing the large language model prompt words based on user input information, task-related knowledge can also be retrieved as needed, and the large language model prompt words can be constructed based on the user input information and the retrieved task-related information.
[0114] In one example, whether prompt word enhancement is needed can be determined according to task complexity.
[0115] Exemplarily, the complexity of a task can be determined based on task description information, data information, evaluation indicators, and expected effects.
[0116] For example, the more difficult it is to achieve the expected effect, the higher the task complexity can be considered.
[0117] Taking accuracy as the evaluation indicator, and the expected effect requiring the model accuracy to be no less than a preset accuracy threshold as an example, the task complexity can be determined based on the preset accuracy threshold.
[0118] For example, when the preset accuracy threshold exceeds 0.9 (which can be called a specified threshold and the value can be adjusted according to actual needs), it is determined that the task complexity is high and prompt word enhancement is required; when the preset accuracy threshold does not exceed 0.9, it is determined that the task complexity is low and prompt word enhancement is not required.
[0119] In another example, whether prompt word enhancement is needed may be determined according to the number of iterations.
[0120] For example, when the large language model prompt word construction process is performed for the first time, prompt word enhancement may not be performed; when the large language model prompt word construction process is not performed for the first time, prompt word enhancement may be performed.
[0121] In yet another example, whether prompt word enhancement is needed may be determined based on the number of iterations and the complexity of the task.
[0122] For example, when the large language model prompt word construction process is performed for the first time, it can be determined whether prompt word enhancement is needed according to the task complexity; when the large language model prompt word construction process is not performed for the first time, prompt word enhancement can be performed.
[0123] For example, when it is determined that prompt word enhancement is required, similar text can be searched from a knowledge base (also referred to as a scenario knowledge base) based on the task description information, that is, task-related information can be searched so that the large language model can better understand the task.
[0124] Accordingly, a large language model prompt word can be constructed based on user input information and similar texts.
[0125] In one example, searching for similar texts from a knowledge base based on the task description information may include:
[0126] According to the task description information, determine the vectorized representation corresponding to the task description information;
[0127] Based on the similarity between the vectorized representation corresponding to the task description information and the vectorized representation of the text fragment, a text fragment whose similarity with the task description information meets the requirements is determined; wherein the text fragment is obtained by segmenting the text in the knowledge base, and the vectorized representation of the text fragment is the vectorized representation corresponding to the text fragment.
[0128] Exemplarily, when prompt word enhancement is required, vectorized features corresponding to the task description information may be determined based on the task description information.
[0129] For example, feature extraction may be performed on the task description information to obtain vectorized features corresponding to the task description information.
[0130] When the vectorized features corresponding to the task description information are obtained, the text segment having a similarity with the task description information that meets the requirements can be determined based on the similarity between the vectorized representation corresponding to the task description information and the vectorized representation of the text segment.
[0131] In one example, the number of text segments whose similarity with the task description information meets the requirement is positively correlated with the number of iterations.
[0132] Exemplarily, the number of text segments whose similarity with the task description information meets the requirement and is determined in the above process may increase as the number of iterations increases.
[0133] For example, in the first iteration, similar text fragments may not be retrieved. In the second iteration, the number of text fragments that need to be determined to meet the similarity requirements may be N1. In the third iteration, the number of text fragments that need to be determined to meet the similarity requirements may be N2, N2>N1, and so on.
[0134] Exemplarily, the text segment is obtained by segmenting the text in the knowledge base, and the vectorized representation of the text segment is the vectorized representation corresponding to the text segment.
[0135] It should be noted that segmenting the text in the knowledge base and determining the vectorized features of the text segments can be implemented offline, without the need to be performed in real time every time a similar text search is performed.
[0136] In some embodiments, the above-mentioned construction of the large language model prompt word based on the user input information and the scene knowledge base may include:
[0137] For the first large language model prompt word construction process, the large language model prompt word is constructed based on the user input information and the scene knowledge base;
[0138] For non-first large language model prompt word construction processing, the large language model prompt words are constructed based on user input information, the scenario knowledge base, and historical iteration record information; wherein the historical iteration record information includes the construction of the large language model prompt words before this large language model prompt word construction processing, and the relevant record information during the processing of generating the algorithm modeling Pipeline script file.
[0139] Exemplarily, in order to determine the algorithm modeling Pipeline faster and better for the large language model, in the iterative process of generating the algorithm modeling Pipeline script file, for the non-first large language model prompt word construction process, in addition to being able to be constructed based on user input information and the scenario knowledge base, historical iteration record information can also be introduced, that is, for the non-first large language model prompt word construction process, the large language model prompt word is constructed based on user input information, the scenario knowledge base, and historical iteration record information.
[0140] Exemplarily, the historical iteration record information includes relevant record information of executing the construction of the large language model prompt words before the current large language model prompt word construction process, and the process of generating the algorithm modeling Pipeline script file.
[0141] For example, for the kth (k≥2) large language model prompt word construction process, the historical iteration record information may include the construction of the large language model prompt words from the 1st to k-1th executions, as well as the relevant record information during the processing of the algorithm modeling Pipeline script file generation.
[0142] Exemplarily, for any execution of the construction of the large language model prompt words and the processing of generating the algorithm modeling Pipeline script file, the constructed large language model prompt words, the obtained algorithm modeling Pipeline script file, and the evaluation results can be recorded.
[0143] In some embodiments, when the algorithm modeling Pipeline script file to be applied is determined, the following may also be included:
[0144] The generated algorithm modeling Pipeline script file and user input information are saved to the scene knowledge base.
[0145] Exemplarily, when an algorithm modeling Pipeline script file is generated in the above manner, the generated algorithm modeling Pipeline script file and the user input information in this task can be deposited into the scenario knowledge base for subsequent use.
[0146] That is, in an embodiment of the present application, in addition to considering relevant technical information of different scenarios, the scenario knowledge base may also include the algorithm modeling Pipeline script files generated during the historical task process and the corresponding user input information.
[0147] In some embodiments, the scenario knowledge base includes knowledge documents of traffic control scenarios, and the candidate tool module includes a dedicated feature construction module for traffic control scenarios;
[0148] and / or,
[0149] The scenario knowledge base includes knowledge documents of industrial material supply and demand forecasting scenarios, and the candidate tool modules include a dedicated feature building module of the industrial material supply and demand forecasting scenarios;
[0150] and / or,
[0151] The scenario knowledge base includes knowledge documents of traffic overload prediction scenarios, and the candidate tool modules include dedicated feature construction modules for traffic overload prediction scenarios;
[0152] Among them, the candidate tool module also includes a general feature construction module.
[0153] Exemplarily, the scenarios to which the embodiments of the present application are applicable may include, but are not limited to, part or all of the following scenarios: traffic control scenarios, industrial material supply and demand prediction scenarios, and flow limit overload prediction scenarios.
[0154] In addition to the dedicated feature building modules for each scenario, the candidate tool modules may also include general feature building modules.
[0155] See also Figure 2 , is a flowchart of a task processing method provided in an embodiment of the present application, such as Figure 2 As shown, the task processing method may include the following steps:
[0156] Step S200: Obtain data to be processed.
[0157] Exemplarily, the data to be processed may include, but is not limited to, one or more of text data to be processed, image data to be processed, and audio data to be processed.
[0158] Step S210: input the data to be processed into the task processing model, and determine the task processing result according to the output result of the task processing model.
[0159] Exemplarily, the task processing model can be generated based on the algorithm modeling Pipeline script file to be applied obtained in the above method embodiment.
[0160] For example, the task processing model can be generated by executing the algorithm modeling Pipeline script file to be applied obtained in the above method embodiment.
[0161] Exemplarily, the acquired data to be processed may be input into a task processing model, and the task processing result may be determined based on the output result of the task processing model.
[0162] For example, assuming that the task is to predict traffic excess, the task processing model may be a traffic excess prediction model, and the data to be processed may include basic information of the user and information related to traffic usage (which may be in the form of a data table).
[0163] The acquired data to be processed can be input into the task processing model, and the traffic over-limit prediction result (i.e., predicting whether the user's traffic usage will exceed the preset traffic limit) can be determined based on the output result of the task processing model.
[0164] In order to enable those skilled in the art to better understand the technical solutions provided in the embodiments of the present application, the technical solutions provided in the embodiments of the present application are described below with reference to specific examples.
[0165] In this embodiment, for automatic machine learning, it is necessary to selectively determine which modules to use to build the modeling process based on scenario characteristics, data distribution, and task objectives.
[0166] In view of this technical problem, the embodiment of the present application can use the large language model as the agent and each machine learning module as the candidate tool, and autonomously generate the algorithm modeling process based on the information about the task, data and evaluation indicators provided by the user, so as to realize the adaptive generation and iteration of the algorithm modeling pipeline based on the scene characteristics, data distribution, task objectives and other information. The specific implementation process can be as follows: Figure 3 shown.
[0167] Below Figure 3 The relevant technical details in the implementation process shown are explained.
[0168] 1. User input information.
[0169] Exemplarily, the user input information may include a description of the task to be modeled, data information (which may be referred to as data table meta-information), evaluation indicators, and expected effects, etc.
[0170] Exemplarily, a large language model prompt word (hereinafter referred to as Prompt) may be constructed based on user input information, and the Prompt may serve as a medium for the user to interact with the large language model.
[0171] For example, a typical Prompt template might look like this:
[0172]
[0173] 2. Similar scene retrieval.
[0174] For example, based on the user input information, similar scenarios can be retrieved in the scenario knowledge base to enhance the subsequent Pipeline construction.
[0175] For example, in similar scene retrieval, vector retrieval is mainly used to obtain similar document fragments based on the document knowledge represented by vectorization, and Prompt is expanded to improve the effect.
[0176] Exemplarily, similar scene retrieval may mainly include the following processing:
[0177] 1) Text (i.e. documents in the scene knowledge base) fragment segmentation:
[0178] For example, for any text:
[0179] chunk1, chunk2, …, chunk n = TextSplit(document)
[0180] Among them, chunk is a text segment, and TextSplit() is a text segmentation algorithm.
[0181] 2) Task description and text fragment representation (i.e. vectorized representation):
[0182] vector task = EmbeddingModel(task description)
[0183] vector i = EmbeddingModel(chunk i ) for i = 1, 2, …, n
[0184] Among them, EmbeddingModel is a representation model used to obtain vectorized representation.
[0185] 3) Similar text retrieval:
[0186] {chunk i , …, chunk j SimilaritySearch(vector task )
[0187] Wherein, SimilaritySearch() is a similarity search, and in this embodiment, a similarity search is performed based on vectorized representation.
[0188] Exemplarily, in the above manner, the N (Top N) text segments with the highest similarity to the vectorized representation corresponding to the task description information can be determined, and the text segment with the highest similarity can be selected, or a text segment can be selected from the N text segments, and the text corresponding to the selected text segment can be determined as the similar text.
[0189] It should be noted that the specific implementation of the above-mentioned text segmentation, representation model and vector retrieval method is not limited in the embodiments of the present application.
[0190] 3. Generate algorithm modeling Pipeline script file.
[0191] Exemplarily, the large language model can determine the task flow plan based on the input prompt, and select a suitable module (i.e., the target module mentioned above) from the candidate tool modules. Then, the algorithm modeling Pipeline script file can be generated based on the task flow plan and the target module.
[0192] Exemplarily, the prompt construction and algorithm modeling pipeline script file generation can be iteratively executed multiple times to improve the effectiveness of the algorithm modeling process.
[0193] 4. Algorithm modeling Pipeline output.
[0194] Exemplarily, the generated executable algorithm modeling Pipeline script file is output to the user.
[0195] Exemplarily, based on the generated algorithm modeling Pipeline script file, a task processing model can be obtained, and the task processing model can be used for task processing.
[0196] 5. New knowledge is stored in the database.
[0197] For example, when the algorithm modeling Pipeline script file is obtained, it can be deposited into the scenario knowledge base together with the user input information for subsequent use.
[0198] Exemplarily, the scenario knowledge base can be organized in a semi-structured form, which can record the description information of historical tasks, data learning, and generated algorithm modeling Pipeline script files, etc., to facilitate the subsequent knowledge reuse and promotion acceleration of similar scenarios.
[0199] For example, the structure of the knowledge base can be as follows:
[0200]
[0201] In this embodiment, based on the above process, an automatic machine learning process enhanced by a large language model can be implemented.
[0202] For example, the large language model is used as an agent to mainly understand tasks, plan processes, and select tools, while the original automatic machine learning module and related public open source modules are used as candidate tools for algorithm modeling. At the same time, the external scene knowledge base can also improve the generation of algorithm modeling processes and the reuse of knowledge between similar scenes. The specific implementation process can be found in Figure 4 .
[0203] Exemplarily, during the first iteration, the construction of the Prompt may include user input information, and may automatically decide whether to retrieve relevant information from the knowledge base to enhance result generation based on the complexity of the task; for non-first iterations, the construction of the Prompt may include not only user input information, but also previous iteration records, etc.
[0204] Based on the constructed Prompt as the input of the large language model, an algorithm modeling Pipeline script file can be generated, and then the algorithm evaluation module is called to get the result, and whether to stop or continue the iteration is decided based on whether the result meets expectations.
[0205] The technical solution provided in the embodiments of the present application is described below with reference to specific examples.
[0206] Assuming that a user wants to build a traffic overlimit prediction model based on customer traffic data, the available data tables mainly include: ① customer basic information table table1; ② customer monthly average traffic table table2; ③ customer traffic usage information table table3; ④ customer traffic usage log details table table4; algorithm evaluation indicator selection precision (accuracy); ⑤ customer traffic overlimit identification table table5. Given the above input information, the execution process of the method provided in the embodiment of the present application is generally as follows:
[0207] First iteration:
[0208] 1-1. According to the provided data table metadata, it is determined that the corresponding feature unit module needs to be selected from the feature library first.
[0209] For example, it is assumed that table1 and table2 have dedicated feature construction modules available for calling; while table3 and table4 do not have dedicated feature construction modules, but can call the general detailed table feature construction module and complete effective feature construction.
[0210] The customer traffic excess identification table table5 can be used as sample label information.
[0211] 1-2. After completing feature construction, all features are connected into a wide table based on the sample label table, and the training set and test set are divided to facilitate verification of the algorithm effect.
[0212] 1-3. Based on the description information entered by the user, it can be inferred that this is a binary classification task. Therefore, the commonly used LGB model can be selected to complete the algorithm modeling on the training set, and then evaluated on the test set to calculate the Precision index.
[0213] Assume that based on the above basic process, the training set Precision is 0.84 and the test set Precision is 0.8, which is lower than the user's expected target (assuming 0.9), thus triggering the second round of iteration.
[0214] 2-1. Reuse the above feature construction process to obtain the feature table and sample label table.
[0215] 2-2. Reuse the above table connection process to obtain the feature wide table and complete the division of training set and test set.
[0216] 2-3. Retrieve similar knowledge from the scene knowledge base and find out that the final algorithm solution in the existing similar historical scenes called the feature enhancement AutoFE (Automatic Feature Engineering) module, and used the XGB and LGB models respectively in the modeling stage, and called the HPO module for hyper-parameter optimization of the two.
[0217] 2-4. Verify the fusion scheme again on the test set. Assume that the Precision on the training set is 0.98 and the Precision on the test set is 0.85.
[0218] The above process uses a more refined modeling method, and the test set indicators have also improved, but they are still lower than expected. At the same time, there is a large deviation between the training set and the test set, and it is inferred that there may be overfitting. Therefore, the third round of iteration is triggered:
[0219] 3-1. Reuse the above feature construction process to obtain the feature table and sample label table.
[0220] 3-2. When retrieving from the scene knowledge base to another scene, a K-fold cross-validation scheme is used to avoid overfitting. Therefore, the constructed feature wide table is randomly divided into 5 folds.
[0221] 3-3. Reuse the algorithm modeling process of the second iteration, but use 5-fold cross-validation to determine the optimal hyperparameters and fusion strategy when adjusting algorithm parameters and fusion models.
[0222] 3-4. Based on the above process, verify again and assume that the training set Precision is 0.92 and the test set Precision is 0.91.
[0223] Since the algorithm indicators have met the user's target requirements at this time, the iteration process is ended, and the complete algorithm modeling Pipeline is solidified into a code file, and the input and production code files related to this task are solidified into the scenario knowledge base for subsequent use.
[0224] It can be seen that in the technical solution provided in the embodiment of the present application, by introducing a large model as an intelligent agent in automatic machine learning and giving full play to its general understanding, task planning and tool calling capabilities, a more intelligent machine learning modeling method is realized, and the original automatic machine learning is further evolved and upgraded to autonomous machine learning; in addition, through the text retrieval enhancement technology of the large language model, the technical empowerment of algorithm modeling with existing scenario knowledge is realized, and the knowledge reuse and promotion acceleration between similar scenarios are improved.
[0225] The method provided by the present application is described above. The device provided by the present application is described below:
[0226] See also Figure 5 , is a schematic diagram of the structure of a script file generating device provided in an embodiment of the present application, such as Figure 5 As shown, the script file generating device may include:
[0227] The acquisition unit 510 is used to acquire user input information; wherein the user input information includes data information and evaluation indicators related to the modeling task; the data information includes description information of the input data set;
[0228] A construction unit 520, configured to construct a large language model prompt word according to the user input information and a scene knowledge base;
[0229] A generating unit 530 is used to input the large language model prompt word into the large language model, so that the large language model determines the task flow plan according to the large language model prompt word, and selects the target module to be called from the candidate tool modules, and generates an algorithm modeling Pipeline script file according to the task flow plan and the target module; wherein the candidate tool modules include the AutoML module and the specific modeling module;
[0230] An evaluation unit 540 is used to determine whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements according to the evaluation index and the expected effect;
[0231] The generating unit 530 is further configured to determine the algorithmic modeling Pipeline script file as the algorithmic modeling Pipeline script file to be applied if the effect evaluation result of the algorithmic modeling Pipeline script file meets the requirements;
[0232] The generation unit 530 is also used to iteratively execute the construction of the large language model prompt words and the processing of algorithm modeling Pipeline script file generation when the effect evaluation result of the algorithm modeling Pipeline does not meet the requirements, until the preset iteration stop condition is reached, and the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied.
[0233] In some embodiments, the evaluation unit 540 determines whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect, including:
[0234] According to the input data set, the feature construction module in the target module is used to construct features, and the feature enhancement module in the target module is used to enhance the constructed features, and a sample set is generated according to the enhanced features;
[0235] The algorithm model is constructed by using the XGB model and the LGB model, and the optimal hyperparameters of the algorithm model are determined by using the HPO module in the target module and the sample set;
[0236] Based on the evaluation index, the expected effect, and the optimal hyperparameter, determine whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements.
[0237] In some embodiments, the evaluation unit 540 uses the HPO module in the target module and the sample set to determine the optimal hyperparameters of the algorithm model, including:
[0238] Divide the sample set into K folds; K≥2;
[0239] For any set of hyperparameters determined by the HPO module, each of the K folds is used as a test set, and the remaining K-1 folds are used as training sets, and the algorithm model is trained for K rounds, and K evaluation index values of the algorithm model on the training set and K evaluation index values of the algorithm model on the test set are obtained;
[0240] The average value of the evaluation index values of the K algorithm models on the training set is determined as the evaluation index value on the training set corresponding to the group of hyperparameters; and the average value of the evaluation index values of the K algorithm models on the test set is determined as the evaluation index value on the test set corresponding to the group of hyperparameters;
[0241] The optimal hyperparameters are determined based on the evaluation index values on the training set and the evaluation index values on the test set corresponding to each group of hyperparameters determined by the HPO module.
[0242] In some embodiments, the obtaining unit 510 user input information includes:
[0243] When a modeling instruction sent by a terminal device is detected, an information input interface is displayed on a display screen of the terminal device;
[0244] Acquire user input information input through the information input interface;
[0245] After the generating unit 530 determines the current algorithm modeling Pipeline script file as the algorithm modeling Pipeline script file to be applied, it also includes:
[0246] The algorithm modeling Pipeline script file to be applied is displayed on the display screen of the terminal device.
[0247] In some embodiments, the construction unit 520 constructs a large language model prompt word according to the user input information and the scene knowledge base, including:
[0248] For the first large language model prompt word construction process, the large language model prompt word is constructed according to the user input information and the scene knowledge base;
[0249] For non-first large language model prompt word construction processing, the large language model prompt words are constructed based on the user input information, the scenario knowledge base, and the historical iteration record information; wherein the historical iteration record information includes the construction of the large language model prompt words before the current large language model prompt word construction processing, and the relevant record information during the processing of generating the algorithm modeling Pipeline script file.
[0250] In some embodiments, Figure 6 As shown, the device also includes:
[0251] The saving unit 550 is used to save the generated algorithm modeling Pipeline script file and the user input information to the scenario knowledge base.
[0252] In some embodiments, the construction unit 520 constructs a large language model prompt word according to the user input information and the scene knowledge base, including:
[0253] When it is determined that prompt word enhancement is required, searching for similar text from the knowledge base according to the task description information;
[0254] A large language model prompt word is constructed based on the user input information and the similar text.
[0255] In some embodiments, the construction unit 520 searches for similar text from a knowledge base according to the task description information, including:
[0256] Determine, based on the task description information, a vectorized representation corresponding to the task description information;
[0257] Based on the similarity between the vectorized representation corresponding to the task description information and the vectorized representation of the text fragment, determine the text fragment whose similarity with the task description information meets the requirements; wherein, the text fragment is obtained by segmenting the text in the knowledge base, and the vectorized representation of the text fragment is the vectorized representation corresponding to the text fragment.
[0258] In some embodiments, the number of text segments determined to have a similarity with the task description information that meets the requirement is positively correlated with the number of iterations.
[0259] See also Figure 7 , is a structural diagram of a task processing device provided in an embodiment of the present application, such as Figure 7 As shown, the task processing device may include:
[0260] An acquisition unit 710 is used to acquire data to be processed;
[0261] The processing unit 720 is used to input the data to be processed into the task processing model, and determine the task processing result according to the output result of the task processing model.
[0262] Exemplarily, the task processing model is generated according to the algorithm modeling Pipeline script file to be applied as described in the above method embodiment.
[0263] An embodiment of the present application provides an electronic device, including a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the script file generation method or task processing method described above.
[0264] See also Figure 8 , is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. The electronic device may include a processor 801 and a memory 802 storing machine executable instructions. The processor 801 and the memory 802 may communicate via a system bus 803. In addition, by reading and executing the machine executable instructions corresponding to the script file generation or task processing logic in the memory 802, the processor 801 may execute the script file generation method or task processing method described above.
[0265] The memory 802 mentioned herein may be any electronic, magnetic, optical or other physical storage device that may contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium may be: RAM (RadomAccess Memory, random access memory), volatile memory, non-volatile memory, flash memory, storage drive (such as hard disk drive), solid state drive, any type of storage disk (such as optical disk, DVD, etc.), or similar storage medium, or a combination thereof.
[0266] In some embodiments, a machine-readable storage medium is also provided. Figure 8 The memory 802 in the machine readable storage medium stores machine executable instructions, and when the machine executable instructions are executed by the processor, the script file generation method or task processing method described above is implemented. For example, the storage medium can be ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
Claims
1. A script file generation method, characterized in that: include: Obtaining user input information; wherein the user input information includes data information, evaluation indicators and expected effects related to the modeling task; the data information includes description information of the input data set; Constructing a large language model prompt word according to the user input information and the scene knowledge base; Input the large language model prompt word into the large language model, so that the large language model determines the task flow plan according to the large language model prompt word, and selects the target module to be called from the candidate tool modules, and generates an algorithm modeling Pipeline script file according to the task flow plan and the target module; wherein the candidate tool modules include the AutoML module and the specific modeling module; When it is determined that the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements according to the evaluation index and the expected effect, the algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; In the case where it is determined according to the evaluation index and the expected effect that the effect evaluation result of the algorithm modeling Pipeline script file does not meet the requirements, iteratively executing the construction of the large language model prompt words and the processing of the algorithm modeling Pipeline script file generation until a preset iteration stop condition is reached, the current algorithm modeling Pipeline script file is determined as the algorithm modeling Pipeline script file to be applied; Wherein, the target module includes a feature construction module, a feature enhancement module and an HPO module; the feature construction module includes a dedicated feature construction module and / or a general feature construction module; For the data in the input data set, if there is a dedicated feature construction module, the dedicated feature construction module is used to construct the features of the data; if there is no dedicated feature construction module, the general feature construction module is used to construct the features of the data; Determining whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements based on the evaluation index and the expected effect includes: According to the input data set, the feature construction module in the target module is used to construct features, and the feature enhancement module in the target module is used to enhance the constructed features, and a sample set is generated according to the enhanced features; The algorithm model is constructed by using the XGB model and the LGB model, and the optimal hyperparameters of the algorithm model are determined by using the HPO module in the target module and the sample set; Based on the evaluation index, the expected effect, and the optimal hyperparameter, determine whether the effect evaluation result of the algorithm modeling Pipeline script file meets the requirements.
2. The method according to claim 1, characterized in that The method of using the HPO module in the target module and the sample set to determine the optimal hyperparameter of the algorithm model includes: Divide the sample set into K folds; K≥2; For any set of hyperparameters determined by the HPO module, each of the K folds is used as a test set, and the remaining K-1 folds are used as training sets, and the algorithm model is trained for K rounds, and K evaluation index values of the algorithm model on the training set and K evaluation index values of the algorithm model on the test set are obtained; The average value of the evaluation index values of the K algorithm models on the training set is determined as the evaluation index value on the training set corresponding to the group of hyperparameters; and the average value of the evaluation index values of the K algorithm models on the test set is determined as the evaluation index value on the test set corresponding to the group of hyperparameters; The optimal hyperparameters are determined based on the evaluation index values on the training set and the evaluation index values on the test set corresponding to each group of hyperparameters determined by the HPO module.
3. The method according to claim 1, characterized in that The obtaining of user input information includes: When a modeling instruction sent by a terminal device is detected, an information input interface is displayed on a display screen of the terminal device; Acquire user input information input through the information input interface; After determining the current algorithm modeling Pipeline script file as the algorithm modeling Pipeline script file to be applied, the method further includes: The algorithm modeling Pipeline script file to be applied is displayed on the display screen of the terminal device.
4. The method according to claim 1, characterized in that: The step of constructing a large language model prompt word based on the user input information and the scene knowledge base includes: For the first large language model prompt word construction process, the large language model prompt word is constructed according to the user input information and the scene knowledge base; For non-first large language model prompt word construction processing, the large language model prompt words are constructed based on the user input information, the scenario knowledge base, and the historical iteration record information; wherein the historical iteration record information includes the construction of the large language model prompt words before the current large language model prompt word construction processing, and the relevant record information during the processing of generating the algorithm modeling Pipeline script file.
5. The method according to claim 1, characterized in that When the algorithm modeling Pipeline script file to be applied is determined, the method further includes: The generated algorithm modeling Pipeline script file and the user input information are saved to the scenario knowledge base.
6. The method according to claim 1, characterized in that The user input information also includes task description information; The step of constructing a large language model prompt word based on the user input information and the scene knowledge base includes: In the case where it is determined that prompt word enhancement is required, determining a vectorized representation corresponding to the task description information according to the task description information; According to the similarity between the vectorized representation corresponding to the task description information and the vectorized representation of the text segment, determine the text segment whose similarity with the task description information meets the requirement; wherein the text segment is obtained by segmenting the text in the knowledge base, and the vectorized representation of the text segment is the vectorized representation corresponding to the text segment; the number of text segments whose similarity with the task description information meets the requirement is positively correlated with the number of iterations; A large language model prompt word is constructed based on the user input information and the text segment whose similarity with the task description information meets the requirements.
7. The method according to any one of claims 1 to 6, characterized in that: The scenario knowledge base includes knowledge documents of traffic control scenarios, and the candidate tool modules include dedicated feature construction modules for traffic control scenarios; and / or, The scenario knowledge base includes knowledge documents of industrial material supply and demand forecasting scenarios, and the candidate tool module includes a dedicated feature construction module for industrial material supply and demand forecasting scenarios; and / or, The scenario knowledge base includes knowledge documents of traffic overload prediction scenarios, and the candidate tool module includes a dedicated feature construction module for traffic overload prediction scenarios; Wherein, the candidate tool module also includes a general feature construction module.
8. A task processing method, characterized in that: include: Get the data to be processed; The data to be processed is input into a task processing model, and the task processing result is determined according to the output result of the task processing model; wherein the task processing model is generated according to the algorithm modeling Pipeline script file to be applied in the method described in any one of claims 1-7.
9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the method according to any one of claims 1-7 or 8.