Analysis model training data generation method and electronic equipment

Abstract

The invention provides an analysis model training data generation method and electronic equipment. The analysis model training data generation method comprises the steps that basic training data of a target analysis model is acquired through a first data acquisition instruction, the basic training data comprises a first analysis data table, and the first analysis data table comprises texts and numbers; acquiring data features of the basic training data; forming a training data set of a target data generation model according to the first data acquisition instruction, the basic training data and the data features of the basic training data, and training the target data generation model through the training data set; a second data acquisition instruction and a first data feature combination are input into the trained target data generation model to form derived training data, the derived training data comprises a second analysis data table, and the second analysis data table comprises texts and numbers; and forming model training data of the target analysis model according to the basic training data and the derived training data. According to the embodiment of the invention, the capability of the artificial intelligence model for analysis can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more specifically, to an analytical model training data generation method and electronic device. Background Technology

[0002] Using artificial intelligence to automatically analyze data and thereby improve the operational efficiency of e-commerce platforms has become a trend in AI applications.

[0003] Before using artificial intelligence to analyze data, the AI ​​model must first be trained. However, due to the scarcity and limited quantity of certain types of data, AI models designed for analyzing these specific types of data are difficult to train sufficiently.

[0004] In related technologies, training data for artificial intelligence models that analyze specific types of data is usually constructed by perturbing a small amount of original training data based on random numbers (e.g., multiplying the random numbers with the original training data). However, the distribution of such training data is very random, lacks rationality and interpretability, resulting in insufficient capabilities of the trained artificial intelligence models.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this disclosure is to provide an analytical model training data generation method and electronic device to solve the problem of insufficient training data for artificial intelligence models for certain specific tasks.

[0007] According to a first aspect of the present disclosure, a method for generating training data for an analysis model is provided, comprising: acquiring basic training data of a target analysis model via a first data acquisition instruction, the basic training data including a first analysis data table, the first analysis data table including text and numbers; acquiring data features of the basic training data; forming a training dataset of a target data generation model based on the first data acquisition instruction, the basic training data, and the data features of the basic training data; training the target data generation model using the training dataset; inputting a second data acquisition instruction and a combination of the first data features into the trained target data generation model to form derived training data, the derived training data including a second analysis data table, the second analysis data table including text and numbers; and forming model training data of the target analysis model based on the basic training data and the derived training data.

[0008] According to a second aspect of the present disclosure, an electronic device is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to perform the method as described in any one of the preceding embodiments based on instructions stored in the memory.

[0009] According to a third aspect of this disclosure, a computer-readable storage medium is provided having a program stored thereon that, when executed by a processor, implements the analytical model training data generation method as described in any of the preceding claims.

[0010] According to a fourth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0011] This embodiment of the disclosure trains a target data generation model using the basic training data of the target analysis model. This enables the target data generation model to generate derived training data with a similar format and rich features to the basic training data, thereby greatly expanding the quantity and diversity of training data for the target analysis model. By using this training data to train the target analysis model, the analytical accuracy and generalization of the target analysis model can be greatly improved.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0014] Figure 1 This is a flowchart of the method for generating training data for the analysis model in an exemplary embodiment of this disclosure.

[0015] Figure 2 This is a sub-flowchart of step S3 in an exemplary embodiment of this disclosure.

[0016] Figure 3 This is a flowchart illustrating the self-iterative training of the target data model in an exemplary embodiment of this disclosure.

[0017] Figure 4 This is a flowchart illustrating the generation of derived training data in an exemplary embodiment of this disclosure.

[0018] Figure 5This is a block diagram of an analytical model training data generation apparatus according to an exemplary embodiment of the present disclosure.

[0019] Figure 6 This is a block diagram of an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0020] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0021] Furthermore, the accompanying drawings are merely illustrative of this disclosure, and the same reference numerals in the drawings denote the same or similar parts, thus repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0022] The exemplary embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart of the method for generating training data for the analysis model in an exemplary embodiment of this disclosure.

[0024] refer to Figure 1 The analysis model training data generation method 100 may include:

[0025] Step S1: Obtain the basic training data of the target analysis model through the first data acquisition instruction. The basic training data includes a first analysis data table, which includes text and numbers.

[0026] Step S2: Obtain the data features of the basic training data;

[0027] Step S3: Based on the first data acquisition instruction, the basic training data, and the data features of the basic training data, a training dataset for the target data generation model is formed, and the target data generation model is trained using the training dataset;

[0028] Step S4: Input the second data acquisition instruction and the first data feature combination into the trained target data generation model to form derived training data. The derived training data includes a second analysis data table, which includes text and numbers.

[0029] Step S5: Based on the basic training data and the derived training data, form the model training data for the target analysis model.

[0030] The development process of a target analysis model (AI model) includes steps such as data collection and construction, model training, model testing, and iteration. In the data collection and construction section, for business analysis scenarios, data collection includes two channels: internal data collection and external data collection. Internal data is relatively direct, while external data collection includes, but is not limited to, open-source data, financial news reports, etc.

[0031] Business analytics scenarios require data retrieved from underlying data tables using SQL (e.g., actual GTV, traffic, year-on-year and month-on-month comparisons, first-, second-, and third-tier product category data for a specific business unit). This necessitates that the target analytics model be able to interpret complex data tables reasonably and accurately. However, for data collected through the two methods mentioned above, internal data is limited in quantity and accumulation; for example, data for new business activities may only be available from last year, or holiday data may only be available for a few years (data fluctuations or low richness are common). External data is mostly in the form of financial reports and news, fundamentally different from the underlying data tables targeted by the analysis model, making it difficult to match with the analytical scenario and often unavailable. Therefore, the scarcity and difficulty in acquiring large quantities of analytical data means that if such limited data is used as training data for the model, the model will have seen very limited data, resulting in little diversity in data variation and hindering the generalization ability of the target analytics model.

[0032] This embodiment of the disclosure trains a target data generation model using the basic training data of the target analysis model. This enables the target data generation model to generate derived training data with a similar format and rich features to the basic training data, thereby greatly expanding the quantity and diversity of training data for the target analysis model. By using this training data to train the target analysis model, the analytical accuracy and generalization of the target analysis model can be greatly improved.

[0033] For example, method 100 of this disclosure embodiment can be implemented based on a Large Language Model (LLM) to simulate a data retrieval tool (i.e., a tool for obtaining underlying data tables). Using the target data obtained by method 100 to generate a model, data retrieval requests can be sent (e.g., obtaining the actual GTV of all BUs from April 4th to April 6th, including year-on-year and month-on-month comparisons) just like using a real data retrieval tool, obtaining diverse underlying data tables. Furthermore, external factors can be introduced in a customized manner to achieve data diversity perturbation (beneficial to the generalization of model training). Method 100 starts from internal data, without the need for external data support, and can construct a data volume and data diversity far exceeding that of existing internal data.

[0034] The following section provides a detailed explanation of each step in the method 100 for generating training data for the analysis model.

[0035] In step S1, the basic training data of the target analysis model is obtained through the first data acquisition instruction. The basic training data includes a first analysis data table, which includes text and numbers.

[0036] Each step of the embodiments of this disclosure can be implemented by an artificial intelligence model, such as a large language model, for example, a modified version based on large language models such as ChatGPT and Deepseek.

[0037] In step S1, real and potentially sensitive internal data, namely the basic training data of the target analysis model, can be obtained first based on real data retrieval tools and various query commands. In this embodiment of the disclosure, the query command is referred to as the data acquisition command, and the query command input to the real database is referred to as the first data acquisition command.

[0038] Each piece of internal data retrieved from the real database according to the query command is recorded as a data pair, denoted as (q, table), where q is the query command, i.e., the first data retrieval command, and table is the first analysis data table returned by the real data retrieval tool (m rows, n columns, with the first column being the header), i.e., the basic training data. The amount of internal data obtained in this way is limited, possibly only a few hundred pieces, but each piece of internal data meets the model training requirements.

[0039] In an exemplary embodiment, the basic training data may be, for example, internal operational data or internal data of an enterprise, and may be presented in the form of an analysis data table. The analysis data table may include multiple rows and columns, with the first row or column serving as the header. It contains multiple data cells, some of which may be empty; however, at least one row or column of the multiple data cells must be non-empty. In an exemplary embodiment, the analysis data table includes both text and numbers.

[0040] To provide as much basic training data as possible, multiple basic training data sets can be acquired through multiple first data acquisition instructions, where each first data acquisition instruction corresponds to one or more basic training data sets. When acquiring multiple basic training data sets through multiple first data acquisition instructions, the correspondence between the first data acquisition instructions and the basic training data sets is recorded to facilitate subsequent processing.

[0041] In step S2, the data features of the basic training data are obtained.

[0042] In an exemplary embodiment, data features can be extracted from the basic training data using a large model. When the basic training data is a business analysis data table, the data features of the basic training data may be data distribution characteristics such as "the GTV contribution of the dining-in business is the largest, and the contribution of the Maoyan business is the smallest." A single basic training data set (business analysis data table) can have multiple data features. For example, from a time perspective, the data feature "the actual GTV paid during holidays increased by XX% compared to the average weekday, with significant fluctuations in payment order volume" can be extracted; from a merchant perspective, the data feature "the top XX% of online merchants contribute over XX% of the actual GTV, exhibiting a clear Matthew effect" can be extracted; and from the user perspective, the data feature "although high-transaction-frequency users (those with XX or more transactions per month) only account for XX% of DAU, their transaction frequency contributes XX%, having a crucial impact on overall business stickiness" can be extracted.

[0043] When the model executing method 100 is a custom model, the custom model can be trained in advance to enable it to extract data features from the basic training data; or, method 100 can be executed directly through a pre-trained large model to achieve step S2.

[0044] When multiple basic training data are obtained in step S1, multiple data features are extracted for each basic training data (first data analysis table) in step S2, and the correspondence between each data feature and the basic training data (first data analysis table) is recorded.

[0045] In step S3, a training dataset for the target data generation model is formed based on the first data acquisition instruction, the basic training data, and the data features of the basic training data, and the target data generation model is trained using the training dataset.

[0046] In this embodiment, a target data generation model is configured to mimic a real data retrieval tool (i.e., a tool that returns basic training data in response to data retrieval commands). This enables the target data generation model to respond to data retrieval commands and return derived training data with features similar to but richer in characteristics than the basic training data. In an exemplary embodiment, the derived training data includes a second analysis data table, which comprises text and numbers. The second analysis data table has a similar format and features to the first analysis data table, meeting the requirements for training data of the target analysis model.

[0047] Therefore, in this embodiment of the disclosure, a training dataset for the target data generation model is formed based on multiple first data acquisition instructions of real data, multiple basic training data (real data) corresponding to each first data acquisition instruction, and multiple data features corresponding to each basic training data, and the target data generation model is trained using the training dataset.

[0048] Figure 2 This is a sub-flowchart of step S3 in an exemplary embodiment of this disclosure.

[0049] refer to Figure 2 In an exemplary embodiment, step S3 may include:

[0050] Step S31: Obtain multiple data features corresponding to the i-th basic training data;

[0051] Step S32: Arrange and combine multiple data features to form Mi feature combinations, 1≤i≤N, where N is the number of basic training data corresponding to the first data acquisition instruction, and Mi≥1;

[0052] Step S33: Based on the first data acquisition instruction, the i-th basic training data corresponding to the first data acquisition instruction, and the j-th feature combination corresponding to the i-th basic training data, a training data group corresponding to the first data acquisition instruction is formed, where 1≤j≤Mi;

[0053] Step S34: The training dataset for the target data generation model is formed based on the training data groups corresponding to the multiple first data acquisition instructions.

[0054] exist Figure 2 In the illustrated embodiment, multiple feature combinations can be formed for each basic training data set. Some feature combinations may contain one data feature, while others may contain multiple data features. The maximum number of feature combinations corresponding to each basic training data set can be 2. n -1, where n is the number of features corresponding to the basic training data. For example, if 10 data features are extracted from a basic training data, then the maximum number of feature combinations formed by the permutations and combinations of the data features can be 1023.

[0055] In the process of forming the training data set, it is not necessary to exhaustively list all feature combinations. For the i-th basic training data, Mi feature combinations can be randomly selected from all its feature combinations, or, according to a preset permutation and combination logic, Mi feature combinations can be generated based on the ni data features of the i-th basic training data, where 1≤Mi≤2. ni -1, where ni is the number of data features of the i-th basic training data.

[0056] Different basic training data can have different numbers of features; for different basic training data, regardless of whether the number of features they correspond to is the same or different, the number of feature combinations formed by them can be the same or different, that is, Mi is not necessarily equal to M(i+1).

[0057] Next, based on the feature combination corresponding to each basic training data, a training data group is formed according to the correspondence of (first data acquisition instruction, basic training data, feature combination). Thus, multiple training data groups can be formed for one basic training data. For multiple basic training data acquired by multiple first data acquisition instructions, a large number of training data groups can be formed. When the basic training data is limited, the training dataset formed by the basic training data can be greatly expanded, thereby improving the model training efficiency.

[0058] During the training of the target data generation model, the correspondence between the input and output data to be learned by the target data generation model is set as follows: Input data (first data acquisition instruction, feature combination) - Output data (basic training data). The basic training data is used to evaluate the data output by the target data generation model based on the input data (first data acquisition instruction, feature combination), and the parameters of the target data generation model are iteratively adjusted until convergence is achieved when the target data generation model is trained on a training dataset in the form of (first data acquisition instruction, basic training data, feature combination).

[0059] Specifically, during the iterative adjustment process, optimization algorithms such as gradient descent can be used. When there is a deviation between the output data of the target data generation model and the basic training data, a loss function (such as mean squared error or cross-entropy loss) is calculated, and the model parameters are adjusted according to the gradient direction of the loss value, so that the model can output results that are closer to the real basic training data when dealing with the same or similar input data (first data acquisition instructions, feature combination).

[0060] Furthermore, to avoid overfitting, regularization methods such as L1 or L2 regularization can be introduced during training to constrain model parameters and enhance the model's generalization ability. Simultaneously, training, validation, and test sets can be created. The validation set can be used to evaluate the model during training, dynamically adjusting hyperparameters such as the learning rate and number of iterations to ensure the model maintains good performance even on unseen data.

[0061] Once the target data generation model converges, it can quickly generate derived training data that is similar to and richer in features than the real base training data, upon receiving data acquisition instructions and specific feature combinations. This derived data can not only further expand training resources but also assist in data exploration in real-world business scenarios, such as simulating business data performance under different feature combinations in business analysis.

[0062] In step S4, the trained target data generation model is input with a second data acquisition instruction and a first data feature combination to form derived training data. The derived training data includes a second analysis data table, which includes text and numbers.

[0063] In an exemplary embodiment, a second data acquisition instruction and a first data feature combination can be randomly generated by executing method 100 using a large model.

[0064] The second data acquisition instruction can be the same as or different from the first data acquisition instruction. The second data acquisition instruction is used to limit the derived data generated by the target data generation model to conform to certain format characteristics (e.g., outputting a data table with m rows and n columns, the table header including p attributes, each attribute being...).

[0065] The first data feature combination includes Q data features, where Q ≥ 0. That is, the first data feature combination can be empty. When the first data feature combination is empty, it means that there are no feature requirements for the generated derived training data, and a data table can be generated randomly.

[0066] Randomly generating a second data acquisition instruction combined with the first data feature can create richer and more abundant derived training data, suitable for rapidly expanding the amount of training data. However, the generated derived training data may differ significantly from the basic training data, which may prevent the target analysis model from being trained effectively in certain scenarios.

[0067] In some embodiments, a preset input database including multiple data acquisition instructions can be pre-formed. Then, a data acquisition instruction is randomly selected from the data acquisition instructions in the preset input database as a second data acquisition instruction, so as to improve the efficiency of input data acquisition instructions for the target data generation model and the efficiency of generating derived training data while enriching the diversity of derived training data.

[0068] Correspondingly, a data feature library can be formed based on the data features of the basic training data corresponding to multiple first data acquisition instructions. The data feature library includes multiple feature combinations, and each feature combination includes Q data features, where Q≥0.

[0069] Then, select the first data feature combination from the data feature library, select the second data acquisition instruction from the preset data input library, and input it into the target data generation model to obtain derived training data.

[0070] Therefore, by selecting the second data acquisition instruction and the first data feature combination from the preset data input library and the data feature library formed by the data features of the basic training data, the generated derived training data can be closer to the form or features of the basic training data, with higher data quality and more accurate achievement of the training objective of the target analysis model.

[0071] In some embodiments, the first data acquisition instruction used in step S1 may also be randomly selected from a preset data input library. The second data acquisition instruction may be the same as or different from the first data acquisition instruction.

[0072] Example:

[0073] The second data acquisition command is, for example: to acquire the xx and yy indicators grouped by BU dimension from January 1, 2024 to January 1, 2025, and to calculate the year-on-year and month-on-month comparisons of xx and yy.

[0074] The first data feature combination is, for example: the xx indicator of the catering business contributes the most.

[0075] The derived training data output by the target data generation model based on the combination of the second data acquisition instruction and the first data feature is, for example, the following data table:

[0076] BU XX Indicator YY Index …(omitted) Meal 13,342,333.00 16,000 To Comprehensive 10,998,222.00 17,500

[0077] Because the amount of basic training data is limited, even if the training dataset corresponding to the basic training data is expanded by constructing the above training data set and trained until the target data generation model converges, the performance of the target data generation model trained using the training dataset formed by the basic training data may still be worse than expected.

[0078] Therefore, in this embodiment of the disclosure, the target data generation model is also configured to generate its own training data to perform iterative training on itself, thereby further improving the ability to generate target data.

[0079] Figure 3 This is a flowchart illustrating the self-iterative training of the target data model in an exemplary embodiment of this disclosure.

[0080] refer to Figure 3 The process of iteratively training the target data generation model can include:

[0081] Event 301: For the derived training data output by the target data generation model 200, determine the first data that does not conform to the first data feature combination or whose format does not conform to the second data acquisition instruction.

[0082] Event 302: Determine whether the proportion of the first data in the derived training data is higher than the preset value.

[0083] Event 303: When the proportion of the first data in the derived training data is higher than the preset value, the first data is removed from the derived training data, and the second data whose format conforms to the second data acquisition instruction and conforms to the first data feature combination is retained.

[0084] Event 304: Update the training dataset 201 of the target data generation model 200 based on the second data, the second data acquisition instruction corresponding to the second data, and the combination of the first data features.

[0085] Event 305, train the target data generation model 200 again using the training dataset 201 of the target data generation model 200.

[0086] Event 306: Input the second data acquisition instruction and the first data feature combination into the trained target data generation model to form derived training data again.

[0087] Repeat the above steps until the proportion of the first data in the derived training data is less than or equal to the preset value, then end the iterative training and fix the target data to generate the parameters of model 200.

[0088] The first data point can be briefly described as unqualified data, and the second data point can be briefly described as qualified data.

[0089] In an exemplary embodiment, the large language model of execution method 100 can be used to determine the first data in the derived data that does not conform to the first data feature combination or whose format does not conform to the second data acquisition instruction.

[0090] By setting a target data generation model and using self-generated, compliant derived training data for iterative training, the deficiency of insufficient basic training data can be compensated for, further improving the quality and accuracy of the model-generated data. In this process, the reasonable setting of preset values ​​is particularly crucial. If the preset value is too high, the model may have difficulty converging, leading to low training efficiency; if the preset value is too low, the model may stop training before sufficient optimization, failing to achieve optimal performance. Therefore, in practical applications, a suitable preset value can be determined through multiple rounds of experimentation and tuning, based on the data scale, business scenario characteristics, and model complexity.

[0091] Monitoring model performance changes is crucial during iterative training. Besides focusing on the proportion of data that doesn't meet requirements, various evaluation metrics can be introduced, such as the similarity between the generated derived training data and the real data (i.e., the base training data) (e.g., cosine similarity, Euclidean distance), and the diversity of the generated data. Visualization tools can be used to display the trends of these metrics in real time, allowing for the timely detection of anomalies during model training, such as overfitting and underfitting, and enabling targeted adjustments to the training strategy.

[0092] Once the target data generation model completes its self-iterative training and fixes its parameters, the derived training data it generates in response to subsequent inputs—the second data acquisition command and the combination of the first data features—not only meets diverse format requirements but also satisfies the requirement of having features similar to and rich in the basic training data. Therefore, it is sufficient to replace existing data acquisition tools, providing rich derived training data for the target analysis model that meets various feature requirements of the internal data.

[0093] In an exemplary embodiment, in order to enhance the feature consistency between the derived training data output by the target data generation model and the basic training data, the first data acquisition instruction and the second data acquisition instruction may both include a first label. The first label is used to restrict the target data generation model from generating derived training data based on the training dataset of the target data generation model.

[0094] In an exemplary embodiment, the first label can be a punctuation mark. For example, the first data acquisition instruction can be set to "[Task A] Acquire xx data", and the second data acquisition instruction can be set to "[Task A] Acquire yy data". This restricts the target data generation model from generating derived training data with similar features to the basic training data of the training dataset labeled "[Task A]" in response to the second data acquisition instruction corresponding to "[Task A]".

[0095] It should be noted that the second data acquisition instruction refers to the data acquisition instruction for generating model inputs using target data trained to convergence using basic training data. It can be various data acquisition instructions, rather than just one or a certain type of data acquisition instruction.

[0096] That is, the first data acquisition instruction corresponds to the basic training data, and the second data acquisition instruction corresponds to the derived training data.

[0097] In step S5, model training data for the target analysis model is formed based on the basic training data and the derived training data.

[0098] After being filtered (e.g., removing the first data mentioned above), the derived training data can be directly added to the training data of the target analysis model or to the training dataset of the target data generation model to iteratively train the target data generation model. That is, the basic training data and the derived training data can be used as training data for both the target analysis model and the target data generation model.

[0099] When the target data generation model performs self-iterative training, the derived training data of the training dataset 201 added in each iteration can be used as the model training data of the target analysis model to expand the model training data of the target analysis model.

[0100] In an exemplary embodiment, there is a requirement for the quantity of training data for the target analysis model. In this case, the model can be generated based on the target data after the self-iteration ends, and the number of derived training data generated can be set to meet the quantity requirement.

[0101] Figure 4 This is a flowchart illustrating the generation of derived training data in an exemplary embodiment of this disclosure.

[0102] refer to Figure 4 Generating derived training data can include:

[0103] Event 401: Determine the target number of derived training data. This target number can be obtained by subtracting the amount of basic training data already obtained from the required amount of training data for the target analysis model.

[0104] Event 402: Input the second data acquisition instruction and the first data feature combination into the trained target data generation model 200, and update the total number of derived training data based on the resulting derived training data. A set of second data acquisition instructions and the first data feature combination can correspond to one or more derived training data. The number of derived training data corresponding to a set of second data acquisition instructions and the first data feature combination can be set by the second data acquisition instruction or determined by the target data generation model itself.

[0105] Event 403: Determine if the total number of derived training data is less than the target number.

[0106] Event 404: When the total number of derived training data is less than the target number, adjust the second data acquisition instruction and / or the first data feature combination.

[0107] In some embodiments, if the target data generation model can generate one derived training data each time, or if the derived training data generated each time is not exhaustive, the same set of second data acquisition instructions and first data feature combinations can be output to the target data generation model multiple times in event 404, that is, the second data acquisition instructions and first data feature combinations are not adjusted.

[0108] Event 405: Input the adjusted second data acquisition instruction and the first data feature combination into the training target data generation model to form a second number of derived training data.

[0109] Repeat the above process until the number of derived training data reaches the target number, then stop inputting and save the current basic training data and derived training data as the model training data for the target analysis model.

[0110] The method provided in this disclosure can generate derived training data that meets quantity requirements and has a high degree of data consistency with real data. It can generate diverse data distributions as needed, providing sufficient training data for the target analysis model, thereby greatly optimizing the training effect of the target analysis model and improving its capabilities.

[0111] Corresponding to the above method embodiments, this disclosure also provides an analytical model training data generation apparatus, which can be used to execute the above method embodiments. This analytical model training data generation apparatus is, for example, an apparatus corresponding to a large language model.

[0112] Figure 5 This is a block diagram of an analytical model training data generation apparatus according to an exemplary embodiment of the present disclosure.

[0113] refer to Figure 5 The analysis model training data generation device 500 may include:

[0114] The basic training data acquisition module 51 is configured to acquire basic training data of the target analysis model through a first data acquisition instruction. The basic training data includes a first analysis data table, which includes text and numbers.

[0115] The data feature extraction module 52 is configured to obtain the data features of the basic training data;

[0116] The data generation model training module 53 is configured to form a training dataset for the target data generation model based on the first data acquisition instruction, the basic training data, and the data features of the basic training data, and to train the target data generation model using the training dataset.

[0117] The derived training data acquisition module 54 is configured to input a second data acquisition instruction and a first data feature combination into the trained target data generation model to form derived training data. The derived training data includes a second analysis data table, which includes text and numbers.

[0118] The training data aggregation module 55 is configured to generate model training data for the target analysis model based on the basic training data and the derived training data.

[0119] Since the functions of the device 500 have been described in detail in their respective method embodiments, they will not be repeated here.

[0120] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0121] In an exemplary embodiment of this disclosure, an electronic device capable of implementing the above-described method is also provided, the electronic device being used to perform at least one step of method 100 as described above.

[0122] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: entirely hardware implementations, entirely software implementations (including firmware, microcode, etc.), or implementations combining hardware and software aspects, collectively referred to herein as “circuits,” “modules,” or “systems.”

[0123] The following reference Figure 6 To describe an electronic device 600 according to this embodiment of the present invention. Figure 6 The electronic device 600 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0124] like Figure 6As shown, the electronic device 600 is manifested in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, and a bus 630 connecting different system components (including storage unit 620 and processing unit 610).

[0125] The storage unit stores program code that can be executed by the processing unit 610, causing the processing unit 610 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform the method shown in the embodiments of this disclosure.

[0126] Storage unit 620 may include readable media in the form of volatile storage units, such as random access memory (RAM) 6201 and / or cache 6202, and may further include read-only memory (ROM) 6203.

[0127] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0128] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0129] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. As shown, network adapter 660 communicates with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0130] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0131] In exemplary embodiments of this disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible embodiments, various aspects of the invention may also be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps of the various exemplary embodiments of the invention described in the "Exemplary Methods" section of this specification.

[0132] The program product for implementing the above-described method according to embodiments of the present invention may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this document, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0133] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0134] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0135] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0136] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0137] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0138] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and concept of this disclosure are indicated by the claims.

Claims

1. A method for generating training data for an analytical model, characterized in that, include: The basic training data of the target analysis model is obtained through the first data acquisition instruction. The basic training data includes a first analysis data table, which includes text and numbers. Obtain the data features of the basic training data; A training dataset for the target data generation model is formed based on the first data acquisition instruction, the basic training data, and the data features of the basic training data, and the target data generation model is trained using the training dataset. The trained target data generation model is input with a second data acquisition instruction and a first data feature combination to form derived training data, which includes a second analysis data table, which includes text and numbers. The model training data for forming the target analysis model is based on the basic training data and the derived training data.

2. The method for generating training data for the analytical model as described in claim 1, characterized in that, The training dataset for forming the target data generation model based on the first data acquisition instruction, the basic training data, and the data features of the basic training data includes: Obtain multiple data features corresponding to the i-th basic training data; The multiple data features are arranged and combined to form Mi feature combinations, 1≤i≤N, where N is the number of basic training data corresponding to the first data acquisition instruction, and Mi≥1; A training data group corresponding to the first data acquisition instruction is formed by combining the first data acquisition instruction, the i-th basic training data corresponding to the first data acquisition instruction, and the j-th feature corresponding to the i-th basic training data, where 1≤j≤Mi; The training dataset for the target data generation model is formed by the training data groups corresponding to multiple first data acquisition instructions.

3. The method for generating training data for the analytical model as described in claim 1, characterized in that, Also includes: In the derived training data, first data that does not conform to the first data feature combination or whose format does not conform to the second data acquisition instruction is determined; When the proportion of the first data in the derived training data is higher than a preset value, the first data is removed from the derived training data, and the second data whose format conforms to the second data acquisition instruction and conforms to the combination of the first data features is retained. The training dataset of the target data generation model is updated according to the second data and the second data acquisition instruction and the first data feature combination corresponding to the second data, and the target data generation model is trained again using the training dataset of the target data generation model; The second data acquisition instruction and the first data feature combination are input into the trained target data generation model to form derived training data again; Repeat the above steps until the proportion of the first data in the derived training data is less than or equal to the preset value.

4. The method for generating training data for the analytical model as described in claim 1, characterized in that, The second data acquisition instruction and the first data feature combination are input into the trained target data generation model to form derived training data, including: Determine the target number of the derived training data; The trained target data generation model is input with a second data acquisition instruction and a first data feature combination, and the total number of derived training data is updated according to the generated derived training data. When the total number of derived training data is less than the target number, the second data acquisition instruction and / or the first data feature combination are adjusted, and the second data acquisition instruction and the first data feature combination are input into the trained target data generation model to update the total number of derived training data again based on the generated derived training data. Repeat the above process until the total number of derived training data reaches the target number.

5. The method for generating training data for the analytical model as described in any one of claims 1 to 4, characterized in that, The input to the trained target data generation model, consisting of a second data acquisition instruction and a first data feature combination, includes: The second data acquisition instruction and the first data feature combination are randomly generated, wherein the first data feature combination includes Q data features, where Q ≥ 0.

6. The method for generating training data for the analytical model as described in any one of claims 1 to 4, characterized in that, The input to the trained target data generation model, consisting of a second data acquisition instruction and a first data feature combination, includes: A data feature library is formed based on the data features of the basic training data corresponding to multiple first data acquisition instructions. The data feature library includes multiple feature combinations, and each feature combination includes Q data features, where Q ≥ 0. Select a data acquisition instruction from the data acquisition instructions in the preset input database as the second data acquisition instruction, and select the first data feature combination from the data feature library. The preset input database includes multiple data acquisition instructions, and the multiple data acquisition instructions include the first data acquisition instruction.

7. The method for generating training data for the analytical model as described in claim 1, characterized in that, Both the first data acquisition instruction and the second data acquisition instruction include a first label, which is used to restrict the target data generation model from generating the derived training data based on the training dataset of the target data generation model.

8. An electronic device, characterized in that, include: Memory; as well as A processor coupled to the memory, the processor being configured to perform the method as described in any one of claims 1-7 based on instructions stored in the memory.

9. A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

Images