A multi-agent based data augmentation method and device
By generating new samples through multi-agent data augmentation, the problems of inaccurate predictions and insufficient understanding of obscure objects in deep learning models under limited sample conditions are solved. This enables the model to efficiently recognize and accurately predict obscure objects, avoiding the formation of information cocoons.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU HIKVISION DIGITAL TECHNOLOGY CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-28
AI Technical Summary
Deep learning models are inaccurate in automatically predicting with a small number of samples, lack sufficient knowledge of obscure objects, and are inefficient to adjust manually.
A multi-agent-based data augmentation method is adopted. By planning the agent to generate a zero-shot thought chain, calling the feature-derived agent and the sample-generating agent to generate new samples, and synthesizing new samples through few-shot learning and the SMOTE algorithm, the problems of data imbalance and missing feature representation are solved.
It enhances the deep learning model's ability to recognize less common objects, avoids the machine illusion caused by information cocoons, and improves the model's prediction accuracy and efficiency.
Smart Images
Figure CN121434793B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a data augmentation method and apparatus based on multi-agent systems. Background Technology
[0002] In high-traffic network scenarios, the business data obtained by deep learning models from the network is often quite complex. The number of samples of popular objects in the network is often greater than that of unpopular objects, and a large amount of hidden information is contained in network traffic interactions. Human representation of sample data is not only highly dependent on expert experience but also time-consuming and inefficient. In this situation, if the deep learning model processing business data is consistently recommended with sample data of popular objects, an information cocoon will form, causing the deep learning model to experience illusions.
[0003] Therefore, how to properly improve the deep learning model's ability to recognize less popular objects has become an urgent problem to be solved. Summary of the Invention
[0004] This application provides a multi-agent-based data augmentation method to address the problems of inaccurate automatic predictions by deep learning models in the case of a few samples, insufficient understanding of obscure objects, and low efficiency of manually adjusting deep learning models.
[0005] In a first aspect, embodiments of this application provide a data augmentation method based on multiple agents, including:
[0006] The task description and data description for planning intelligent agents to acquire sample data include data distribution statistics and data feature details;
[0007] The planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data. The zero-shot thought chain includes the agent type and execution strategy called at each execution level.
[0008] The planning agent, based on the execution strategy of the zero-shot thought chain, determines the execution level of the agent that calls the enhancement strategy and generates new samples.
[0009] The planning agent evaluates the enhancement effect of the new sample on the downstream deep learning model. If the enhancement effect does not meet the set conditions, the execution hierarchy order of the zero-shot thought chain and the data augmentation steps in the execution hierarchy are adjusted, and the process returns to the step of determining the execution hierarchy of the agent that calls the enhancement strategy based on the zero-shot thought chain.
[0010] Optionally, the planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data, including:
[0011] Obtain the distribution characteristics of positive and negative samples in the sample data as the data distribution statistics of the sample data, and generate the execution strategy in the zero-shot thought chain based on the data distribution statistics.
[0012] Obtain the data column names, meanings, and data attributes of the sample data input by the user, and generate the feature-derived execution strategy in the zero-shot thought chain based on the data column names, meanings, and attributes.
[0013] Optionally, the enhanced strategy agent includes: a feature-derived agent and a sample-generating agent.
[0014] The determination of the execution level of the agent that invokes the enhanced policy includes:
[0015] The feature-derived agent is invoked to perform a derived feature code generation task in conjunction with the task description;
[0016] The feature-derived agent is invoked in conjunction with the task description to perform a few-short learning task using a small number of samples.
[0017] The sample generation agent is invoked to synthesize new samples from the derived features in the feature derivation execution results.
[0018] Optionally, invoking the feature-derived agent in conjunction with the task description to perform a derived feature code generation task includes:
[0019] The feature-derived agent, based on the data distribution statistics and data feature details in the data description, identifies imbalanced classifications and / or missing sample data with insufficient feature representation in the classification rules of the sample data.
[0020] Based on imbalanced classification in the classification rules of the sample data, and / or missing sample data with insufficient feature representation, a corresponding feature derivation algorithm is determined, wherein the derived feature code includes at least one of the following algorithms:
[0021] Multinomial feature derivation algorithm, cross-combination feature derivation algorithm, grouped statistical feature derivation algorithm, time series feature derivation algorithm, discretization feature derivation algorithm, and neuro-linguistic programming (NLP) feature derivation algorithm;
[0022] The feature-derived agent combines the feature-derived algorithm with the task description to generate corresponding derived feature codes.
[0023] Optionally, based on imbalanced classification and missing sample data with insufficient feature representation in the classification rules of the sample data, a corresponding feature derivation algorithm is determined, including:
[0024] When the missing sample data is classified as user features, a derived feature code containing a multinomial feature derivation algorithm is generated based on the numerical values and / or characters of the existing user features.
[0025] When the missing sample data is classified as user behavior, a derived feature code containing a cross-combination feature derivation algorithm is generated based on the numerical and / or vector representation of existing user behavior.
[0026] When the missing sample data is classified as an item attribute, a derived feature code containing a grouped statistical feature derivation algorithm is generated based on the numerical value and / or character of the existing item attribute.
[0027] When the missing sample data is classified as user interaction data, a derived feature code containing NLP feature derivation algorithm is generated based on the numerical or character values of the existing user interaction data.
[0028] Optionally, the sample generation agent is invoked to synthesize new samples from the derived features in the feature derivation execution result, including:
[0029] The derived features generated by the feature-derived agent are transformed into categorical features and numerical features. The categorical features are used as keys and the numerical features are used as values to obtain key-value pairs.
[0030] Based on the task description, obtain the corresponding scenario domain knowledge, and combine the key-value pairs to obtain structured data;
[0031] New samples are synthesized based on the structured data.
[0032] Optionally, the planning agent evaluates the enhancement effect of the new samples on the downstream deep learning model, including:
[0033] When the task description characterizes the downstream deep learning model for performing a classification task, the planning agent verifies the enhancement effect of the downstream deep learning model through the accuracy of the classification task.
[0034] When the task description characterizes the downstream deep learning model for performing a regression task, the planning agent verifies the enhancement effect of the downstream deep learning model through the mean square error of the regression task results.
[0035] Optionally, the planning agent, based on the execution strategy of the zero-shot thought chain, determines the execution level of the agent that calls the enhancement strategy and generates new samples, including:
[0036] The feature-derived agent determines one or more missing items based on the data feature details, and performs one or more derivative feature code generation tasks for the one or more missing items to obtain at least one derivative feature code generation layer.
[0037] The feature-derived agent determines one or more less popular data points based on the data distribution statistics, and performs a few-short learning task on the less popular data points as a few-short learning layer.
[0038] The sample generation agent synthesizes new samples using the SMOTE sampling algorithm based on the derived feature code and / or the execution result of the few-short learning task, and treats the process of synthesizing a new sample as a sample synthesis layer.
[0039] The planning agent sorts and combines at least one layer of derived feature code generation layer, at least one layer of few-short learning layer, and at least one layer of sample synthesis layer to obtain the execution level of the zero-shot thought chain;
[0040] New samples are generated according to the execution hierarchy of the zero-sample thinking chain.
[0041] Optionally, the method further includes:
[0042] The sample data includes network traffic data, and the data description includes: the network traffic data includes popular data and unpopular data, the data distribution statistics include the recommendation frequency of popular objects and the recommendation frequency of unpopular objects, and the data feature details include: parameter features of popular objects and parameter features of unpopular objects.
[0043] Secondly, embodiments of this application provide a data augmentation device based on multiple agents, comprising:
[0044] A planning agent is used to acquire task descriptions and data descriptions of sample data, wherein the data descriptions include data distribution statistics and data feature details;
[0045] The planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data. The zero-shot thought chain includes the agent type and execution strategy called at each execution level.
[0046] The planning agent determines the execution level for calling the feature derivation agent or the sample generation agent based on the execution strategy of the zero-shot thought chain. The feature derivation agent combines the task description to execute the derived feature code generation task and / or the few-short learning task to obtain the feature derivation execution result. The sample generation agent synthesizes the derived features in the feature derivation execution result into a new sample.
[0047] The planning agent evaluates the enhancement effect of the new samples on the downstream deep learning model.
[0048] If the enhancement effect does not meet the set conditions, a new zero-sample thought chain is generated, and the step of determining the execution level of the feature-derived agent or the sample-generating agent based on the zero-sample thought chain is returned.
[0049] In this embodiment, the planning agent, based on the task description and data description of the sample data, obtains an initial zero-shot thought chain to enhance the weak parts of the sample data. The feature-derived agent and the sample-generating agent are invoked to execute the execution level in the initial zero-shot thought chain, thereby matching a suitable data augmentation execution strategy for each deep learning model. The data imbalance problem of unpopular objects is solved by few-shot learning, and the problem of missing feature representation is solved by derivative feature code generation. New samples that are more suitable for downstream deep learning models are obtained, enabling downstream deep learning models to improve their cognition of unpopular objects and avoid the machine illusion caused by the information cocoon resulting from the over-concentration of answers. Attached Figure Description
[0050] Figure 1 This application provides a schematic diagram of a data augmentation method based on multiple agents.
[0051] Figure 2 This is a schematic diagram of a zero-sample thought process provided for an embodiment of this application;
[0052] Figure 3 This is a schematic flowchart of a multi-agent data augmentation method provided in an embodiment of this application;
[0053] Figure 4 This is a schematic diagram of the internal modules of the feature-derived intelligent agent provided in the embodiments of this application. Detailed Implementation
[0054] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.
[0055] In the field of artificial intelligence, the imbalance in data distribution and the lack of feature representation are the biggest obstacles to the accurate prediction of deep learning models.
[0056] Imbalanced data distribution refers to a situation where there is a significant difference in the sample size between different categories, or where a few samples are more important and represent key business data. In such cases, deep learning models need to learn from these few samples. However, due to the scarcity of this data, deep learning models often fail to achieve the predictive performance required by the business after learning from these few samples.
[0057] Feature representation deficiency refers to insufficient feature description of a sample. Deep learning models cannot learn from samples based on simple feature descriptions, and therefore cannot complete prediction tasks or support business operations.
[0058] The data requirements of deep learning models may change due to changes in task strategies. If we rely on expert experience to integrate strategies and define data to solve the problems of imbalanced data distribution and missing feature representations, it will not only be unable to cope with new scenarios, but the learning efficiency of deep learning models will also be relatively low due to manual integration of strategies and definition of data.
[0059] like Figure 1 As shown, this application provides a data augmentation method based on multiple agents, including:
[0060] S101: The task description and data description of the planning agent for acquiring sample data, wherein the data description includes data distribution statistics and data feature details.
[0061] S102: The planning agent determines the execution level of the agent that calls the enhancement strategy based on the execution strategy of the zero-shot thought chain, and generates new samples.
[0062] Among them, the enhanced policy agent can include feature-derived agents and sample-generated agents.
[0063] An intelligent agent can be a software or hardware entity with autonomy, responsiveness, initiative, and interactivity. It can perceive the environment through sensors and exert influence through actuators. In the embodiments of this application, the intelligent agent may include a policy agent composed of a large language model. A multi-agent system can be a networked system containing multiple interacting intelligent agents.
[0064] S103: The planning agent evaluates the enhancement effect of the new sample on the downstream deep learning model. If the enhancement effect does not meet the set conditions, the execution hierarchy order of the zero-shot thought chain and the data augmentation steps in the execution hierarchy are adjusted, and the process returns to S102.
[0065] If the enhancement effect meets the set conditions, it is determined that the current data enhancement of a small number of samples meets the needs of the downstream deep learning model.
[0066] For example, the features and labels of new and historical samples can be validated, requiring that the features and / or labels of new and historical samples remain consistent. In this way, after extending the thought chain in different ways for multiple downstream deep learning models, the generated new samples can be used by multiple downstream deep learning models as long as their features and / or labels remain consistent with those of historical samples.
[0067] For example, the determination of the enhancement effect can be performed by the planning agent based on an evaluation and feedback mechanism. For instance, the enhancement effect can be evaluated based on the accuracy of downstream tasks and / or the minimum mean square error of regression tasks, thereby quantifying the reasoning performance of the enhancement policy agent during task execution. The planning agent can dynamically adjust the weights for evaluating the enhancement effect according to the type of downstream tasks and environmental changes, enabling the enhancement policy agent to flexibly adapt to different scenarios.
[0068] For example, adjusting the execution hierarchy order of the zero-shot thought chain and the data augmentation steps in the execution hierarchy can be achieved through the ε-greedy algorithm, which enables the augmentation policy agent to gradually increase the proportion of environmental information utilized during the training process, thereby improving the stability and efficiency of the policy.
[0069] Feature-derived agents in augmented policy agents can perform policy optimization using algorithms such as Actor-Critic (policy gradient-estimation), A3C (Asynchronous Advantage Actor-Critic), and DDPG (Deep Deterministic Policy Gradient). If the Actor-Critic algorithm is used, the policy gradient (Actor) and value function (Critic) can be combined to improve learning efficiency, and it is suitable for both continuous and discrete action spaces.
[0070] Data augmentation can be a process of performing deterministic or stochastic transformations on the original training data in a supervised / unsupervised manner to generate new samples that are semantically equivalent but have different feature space distributions. It can be used to improve the model's generalization ability, alleviate data sparsity, and suppress model overfitting.
[0071] For example, in S101, the planning agent can obtain the distribution characteristics of positive and negative samples in the sample data as the data distribution statistics of the sample data, and generate the execution strategy in the zero-shot thought chain based on the data distribution statistics.
[0072] The system obtains the data column names, meanings, and attributes of the sample data input by the user. Based on these data column names, meanings, and attributes, it generates the feature derivation execution strategy in the zero-shot thought chain. This feature derivation execution strategy, generated by combining data column names, meanings, and attributes, is derived according to the actual needs of the data characteristics, better matching user requirements. Data derived in this way, when subsequently used to enhance deep learning models, will be more targeted and can avoid the deep learning model from becoming misled due to overly broad data attributes or meanings.
[0073] See Figure 2 As shown, each execution level in the zero-shot thought chain can include an augmentation policy agent and a corresponding initial data augmentation scheme. The initial data augmentation scheme can include multiple execution levels, such as data augmentation steps T1, T2, ..., TN.
[0074] The aforementioned determination of the execution level of the agent that invokes the enhanced policy may include, but is not limited to:
[0075] First execution level: Invoke the feature-derived agent to execute the derived feature code generation task in conjunction with the task description;
[0076] Second execution level: Invoke the feature-derived agent to perform a few-short learning task based on the task description;
[0077] Third execution level: Invoke the sample generation agent to synthesize new samples from the derived features in the feature derivation execution results.
[0078] The first execution level can be implemented, but is not limited to, in the following ways:
[0079] (1) The feature-derived agent identifies missing sample data with insufficient feature representation based on the data feature details;
[0080] (2) Based on the classification of the missing sample data, a derived feature code containing at least one of the following feature derivation algorithms is obtained:
[0081] Multinomial feature derivation algorithm, cross-combination feature derivation algorithm, grouped statistical feature derivation algorithm, time series feature derivation algorithm, discretization feature derivation algorithm, NLP (Neuro-Linguistic Programming) feature derivation algorithm.
[0082] For example, when the missing sample data is classified as user features, a derived feature code containing a multinomial feature derivation algorithm is generated based on the numerical values and / or characters of the existing user features.
[0083] When the missing sample data is classified as user behavior, a derived feature code containing a cross-combination feature derivation algorithm is generated based on the numerical and / or vector representation of existing user behavior.
[0084] When the missing sample data is classified as an item attribute, a derived feature code containing a grouped statistical feature derivation algorithm is generated based on the numerical value and / or character of the existing item attribute.
[0085] When the missing sample data is classified as user interaction data, a derived feature code containing NLP feature derivation algorithm is generated based on the numerical or character values of the existing user interaction data.
[0086] Optionally, the enhancement policy agent first determines whether the feature description is sufficient based on the original feature distribution. If the feature description is insufficient, it will call the feature derivation agent to infer the derivation code of the derived features. The derivation code of the derived features is applied to the original data to generate the derived features.
[0087] For example, a feature-derived agent can perform the following feature-derived tasks:
[0088] (1) Assume that the sample data includes user age, user hobbies, user historical shopping records and user historical click rate for a certain product.
[0089] (2) Feature-derived task description: Please summarize the classification rules based on the above examples.
[0090] (3) The enhanced strategy agent detected that the original data lacked sufficient feature description, and called the feature-derived agent to summarize the classification rules based on the task description and few-shot samples as follows:
[0091] The categories that users would purchase are:
[0092] User age <50
[0093] User purchase history = 0
[0094] Historical click-through rate > 0.5
[0095] The categories that users will not purchase are:
[0096] User age > 50
[0097] User purchase history = 1
[0098] Historical click-through rate < 0.5
[0099] The feature-derived agent generates feature derivation based on classification rules, that is, it directly generates a series of data to determine whether the user meets the above conditions.
[0100] (4) After completing the feature derivation task, the enhancement strategy agent detects that the distribution ratio of positive and negative samples in the data is unbalanced, with negative samples far less than positive samples. It then calls the large model sample synthesis tool to synthesize negative samples until a better ratio of positive and negative samples is obtained, so that the downstream deep learning model has a better evaluation index.
[0101] (5) Repeat the above process to repeatedly verify the representational properties of the features until a better training effect of the downstream deep learning model is obtained.
[0102] See Figure 3 As shown, the raw information input by the user may include task description, sample distribution, and feature details.
[0103] For example, the raw information entered by the user can be as follows:
[0104] You are an expert who, based on the task description, feature list, and data distribution statistics, needs to customize specific data augmentation strategies to improve the accuracy of the product recommendation algorithm.
[0105] The task description can be as follows:
[0106] Will this user purchase the corresponding product? Yes or no?
[0107] The feature details can be as follows:
[0108] 1. User characteristics
[0109] Age: Column name: Age, Type: Numeric
[0110] Location: Column name: Location, Type: Character
[0111] 2. User behavior history
[0112] Historical Clicks: Column name: Click, Type: Numeric
[0113] History browsing: Column name: Browse, Type: Numeric
[0114] Purchase History: Column Name: Purchase, Type: Numeric
[0115] 3. Item attributes
[0116] Price: Column name: Price, Type: Numeric
[0117] Category: Column Name Type: Numeric
[0118] Rating: Column name: rate; Type: Numeric
[0119] 4. Interaction characteristics - User interaction data with items (such as click-through rate, dwell time, rating).
[0120] Click-through rate: Column name: click_rate, Type: numeric
[0121] Dwell Time: Column name: Dwell Time, Type: Numeric
[0122] Sample distribution:
[0123] Sample size: 1,200,000
[0124] Sample types: Two types of samples, one is positive samples of historical purchases, with a label value of 1; the other is negative samples of historical clicks without purchases, with a label value of 0.
[0125] Positive and negative sample ratio: 0.2 (positive samples only account for 20% of negative samples)
[0126] Based on the above description, please generate a data augmentation strategy.
[0127] Since the data augmentation step T1-TN may have defects, causing the augmented model to appear disorienting, in this embodiment of the application, a downstream deep learning model can be used to correct the thinking of the data augmentation step and adjust the thinking chain CoT.
[0128] The process of mind correction can be carried out based on the performance evaluation of the downstream deep learning model. Different tasks can correspond to different verification methods for the enhancement effect.
[0129] For example, when the task description characterizes the downstream deep learning model for performing a classification task, the planning agent verifies the enhancement effect of the downstream deep learning model through the accuracy of the classification task.
[0130] When the task description characterizes the downstream deep learning model for performing a regression task, the planning agent verifies the enhancement effect of the downstream deep learning model through the mean square error of the regression task results.
[0131] For example, raw data can be input into a planning agent, which can formulate a data augmentation strategy based on the data distribution. Based on the data augmentation strategy, the raw data can be augmented. The augmented data can then be used to train the model corresponding to the downstream task. The algorithm performance (mean squared error, accuracy) of the downstream task model on the test set can be fed back to the agent for strategy iterative optimization.
[0132] For example, planning agents can leverage experience replay mechanisms to enhance policy intelligence by learning from historical experience, improving data utilization, and reducing relevance interference during training.
[0133] Adjusting one's thought process can be achieved, but is not limited to, through the following methods:
[0134] For example, when verifying that the enhancement effect of the downstream deep learning model is an illusion, the execution level of the feature-derived agent performing the task of generating derived feature codes can be adjusted. Or, when verifying that the accuracy of the downstream deep learning model is declining, the execution level of the feature-derived agent performing the task of learning a few short samples can be increased.
[0135] Suppose the thought chain consists of T1-Tn-1 data augmentation steps, and these steps have been optimized and stored. The execution strategy is then fixed within these T1-Tn-1 steps during subsequent optimization processes. When adjusting the thought chain, a new step Tn can be added, and this step can be optimized based on the augmentation effect on the downstream deep learning model. After optimizing Tn, a new step Tn+1 is added. Due to data consistency requirements, the input, output, and execution strategy type of Tn+1 must correspond to those of Tn. Therefore, during subsequent optimization of the thought chain, once Tn is optimized, the input and execution strategy of Tn+1 will also change accordingly.
[0136] For example, Tn is used to call the feature-derived agent to perform a derived feature code generation task in conjunction with the task description. After optimization, Tn becomes: performing a few-short learning task in conjunction with the task description. At this time, the output of Tn is no longer the derived feature code, but the features in the learned sample data. Therefore, the input of Tn+1 will also be transformed from the derived feature code into the features in the sample data, and Tn+1 will change accordingly, thereby meeting the requirements of downstream deep learning models for data consistency.
[0137] See Figure 4 As shown, when the user inputs raw information including task description and feature details, the thought process chain includes the following execution levels:
[0138] Execution Level 1 (Based on few-shot analysis of key feature derivation methods): The feature derivation agent performs few-short learning tasks based on key features;
[0139] Execution Level 2 (Generating Feature Derived Code): The feature derivation agent generates feature derived code;
[0140] Execution Level 3 (Execution Code Derivation Features): The feature derivation agent executes the feature derivation code to obtain derived features;
[0141] Execution Level 4 (Derived Feature Concatenation): The sample generating agent synthesizes the derived features generated in Execution Level 3 to obtain new samples.
[0142] For example, the feature derivation code mentioned above may include: polynomial feature derivation algorithm, cross-combination feature derivation algorithm, grouped statistical feature derivation algorithm, time series feature derivation algorithm, discretization feature derivation algorithm, NLP feature derivation algorithm, etc., without limitation.
[0143] Assume the sample data includes user age, user preferences, user shopping history, and user click-through rate for a particular product.
[0144] Task Description: Please summarize the classification rules based on the above examples.
[0145] Feature-derived agents can include classification rules based on task descriptions and sample summaries, such as:
[0146] The categories that users would purchase are:
[0147] User age <50
[0148] User purchase history = 0
[0149] Historical click-through rate > 0.5
[0150] The categories that users will not purchase are:
[0151] User age > 50
[0152] User purchase history = 1
[0153] Historical click-through rate < 0.5
[0154] The feature-derived agent can generate feature-derived codes based on classification rules, including:
[0155] When the missing sample data is classified as user features, a derived feature code containing a multinomial feature derivation algorithm is generated based on the numerical values and / or characters of the existing user features.
[0156] When the missing sample data is classified as user behavior, a derived feature code containing a cross-combination feature derivation algorithm is generated based on the numerical and / or vector representation of existing user behavior.
[0157] When the missing sample data is classified as an item attribute, a derived feature code containing a grouped statistical feature derivation algorithm is generated based on the numerical value and / or character of the existing item attribute.
[0158] When the missing sample data is classified as user interaction data, a derived feature code containing NLP feature derivation algorithm is generated based on the numerical or character values of the existing user interaction data.
[0159] The sample generation agent transforms the derived features generated by the feature derivation agent into categorical features and numerical features, using the categorical features as keys and the numerical features as values to obtain key-value pairs;
[0160] Based on the task description, obtain the corresponding scenario domain knowledge, and combine the key-value pairs to obtain structured data;
[0161] New samples are synthesized based on the structured data.
[0162] For example, the feature-derived agent determines one or more missing items based on the data feature details, and performs one or more derivative feature code generation tasks for the one or more missing items to obtain at least one derivative feature code generation layer;
[0163] The feature-derived agent identifies one or more less popular data points based on the data distribution statistics, and performs a few-short learning task on the less popular data points as a few-short learning layer;
[0164] The sample generation agent synthesizes new samples using the SMOTE sampling algorithm based on the derived feature code and / or the execution result of the few-short learning task, and the process of synthesizing a new sample once is regarded as a sample synthesis layer.
[0165] The planning agent sorts and combines at least one layer of derived feature code generation layer, at least one layer of few-short learning layer, and at least one layer of sample synthesis layer to obtain the execution level of the zero-shot thought chain;
[0166] New samples are generated according to the execution hierarchy of the zero-sample thinking chain.
[0167] Assuming the sample data includes network traffic data, the data description includes: the network traffic data includes popular data and unpopular data; the data distribution statistics include the recommendation frequency of popular objects and the recommendation frequency of unpopular objects; and the data feature details include: parameter features of popular objects and parameter features of unpopular objects.
[0168] Let's assume the less popular object is the hammer, and the more popular object is the screwdriver. The popular object's characteristics include the type of screwdriver: flathead, Phillips head, pentagonal, hexagonal, etc. The less popular object's characteristics include: wooden hammer, iron hammer, steel hammer, etc.
[0169] After identifying the less popular object, the hammer, the feature-derived agent determines the less popular data of the less popular object, including the shape and size of the hammer, based on the statistical distribution of data. Then, it performs one or more few-short learning tasks on the shape and size of the hammer.
[0170] The samples in the embodiments of this application can be product data.
[0171] In this embodiment, the planning agent, based on the task description and data description of the sample data, obtains an initial zero-shot thought chain to enhance the weak parts of the sample data. The feature-derived agent and the sample-generating agent are invoked to execute the execution level in the initial zero-shot thought chain, thereby matching a suitable data augmentation execution strategy for each deep learning model. The data imbalance problem of unpopular objects is solved by few-shot learning, and the problem of missing feature representation is solved by derivative feature code generation. New samples that are more suitable for downstream deep learning models are obtained, enabling downstream deep learning models to improve their cognition of unpopular objects and avoid the machine illusion caused by the information cocoon resulting from the over-concentration of answers.
[0172] Based on the same inventive concept, this application provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described data augmentation method.
[0173] The memory can be random access memory, read-only memory, non-volatile, programmable ROM, erasable PROM, electrically erasable memory, flash memory, optical memory, and registers, etc. The processor can be a general-purpose processor, which is a processor that performs specific steps and / or operations by reading and executing computer programs stored in the memory. The general-purpose processor may use the memory during the execution of these steps and / or operations. The general-purpose processor can be a central processing unit, ASIC, and FPGA, etc. In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor.
[0174] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a solid-state drive (SSD), etc.
[0175] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0176] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above descriptions are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.
Claims
1. A data augmentation method based on multi-agent systems, characterized in that, The method includes: The plan includes a task description and data description for the intelligent agent to acquire sample data. The data description includes data distribution statistics and data feature details. The sample data includes network traffic data. The data description includes: the network traffic data includes popular data and unpopular data; the data distribution statistics include the recommendation frequency of popular objects and the recommendation frequency of unpopular objects; and the data feature details include: parameter features of popular objects and parameter features of unpopular objects. The planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data. The zero-shot thought chain includes the agent type and execution strategy called at each execution level. The planning agent, based on the execution strategy of the zero-shot thought chain, determines the execution level of the agent that calls the enhancement strategy and generates new samples. The planning agent evaluates the enhancement effect of the new sample on the downstream deep learning model. If the enhancement effect does not meet the set conditions, the execution order of the zero-shot thought chain and the data augmentation steps in the execution layer are adjusted, and the step of determining the execution layer of the agent that calls the enhancement strategy based on the zero-shot thought chain is returned. The enhanced strategy agent includes: a feature-derived agent and a sample-generated agent. The determination of the execution level of the agent that invokes the enhanced policy includes: The feature-derived agent is invoked to perform a derived feature code generation task in conjunction with the task description; The feature-derived agent is invoked to perform a few-short learning task based on the task description. The sample generation agent is invoked to synthesize new samples from the derived features in the feature derivation execution result: The feature-derived agent, based on the data distribution statistics and data feature details in the data description, identifies imbalanced classifications and / or missing sample data with insufficient feature representation in the classification rules of the sample data. Based on imbalanced classification in the classification rules of the sample data, and / or missing sample data with insufficient feature representation, a corresponding feature derivation algorithm is determined, wherein the derived feature code includes at least one of the following algorithms: Multinomial feature derivation algorithm, cross-combination feature derivation algorithm, grouped statistical feature derivation algorithm, time series feature derivation algorithm, discretization feature derivation algorithm, and neuro-linguistic programming (NLP) feature derivation algorithm; The feature-derived agent combines the feature-derived algorithm with the task description to generate corresponding derived feature codes.
2. The method as described in claim 1, characterized in that, The planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data, including: Obtain the distribution characteristics of positive and negative samples in the sample data as the data distribution statistics of the sample data, and generate the execution strategy in the zero-shot thought chain based on the data distribution statistics. Obtain the data column names, meanings, and data attributes of the sample data input by the user, and generate the feature-derived execution strategy in the zero-shot thought chain based on the data column names, meanings, and attributes.
3. The method as described in claim 1, characterized in that, Based on the imbalanced classification and missing sample data with insufficient feature representation in the classification rules of the sample data, the corresponding feature derivation algorithm is determined, including: When the missing sample data is classified as user features, a derived feature code containing a multinomial feature derivation algorithm is generated based on the numerical values and / or characters of the existing user features. When the missing sample data is classified as user behavior, a derived feature code containing a cross-combination feature derivation algorithm is generated based on the numerical and / or vector representation of existing user behavior. When the missing sample data is classified as an item attribute, a derived feature code containing a grouped statistical feature derivation algorithm is generated based on the numerical value and / or character of the existing item attribute. When the missing sample data is classified as user interaction data, a derived feature code containing NLP feature derivation algorithm is generated based on the numerical or character values of the existing user interaction data.
4. The method as described in claim 2, characterized in that, The sample generation agent is invoked to synthesize new samples from the derived features in the feature derivation execution result, including: The derived features generated by the feature-derived agent are transformed into categorical features and numerical features. The categorical features are used as keys and the numerical features are used as values to obtain key-value pairs. Based on the task description, obtain the corresponding scenario domain knowledge, and combine the key-value pairs to obtain structured data; New samples are synthesized based on the structured data.
5. The method as described in claim 1, characterized in that, The planning agent evaluates the enhancement effect of the new samples on the downstream deep learning model, including: When the task description characterizes the downstream deep learning model for performing a classification task, the planning agent verifies the enhancement effect of the downstream deep learning model through the accuracy of the classification task. When the task description characterizes the downstream deep learning model for performing a regression task, the planning agent verifies the enhancement effect of the downstream deep learning model through the mean square error of the regression task results.
6. The method as described in claim 1, characterized in that, The planning agent, based on the execution strategy of the zero-shot thought chain, determines the execution level of the agent that calls the enhancement strategy, and generates new samples, including: The feature-derived agent determines one or more missing items based on the data feature details, and performs one or more derivative feature code generation tasks for the one or more missing items to obtain at least one derivative feature code generation layer. The feature-derived agent determines one or more less popular data points based on the data distribution statistics, and performs a few-short learning task on the less popular data points as a few-short learning layer. The sample generation agent synthesizes new samples using the SMOTE sampling algorithm based on the derived feature code and / or the execution result of the few-short learning task, and treats the process of synthesizing a new sample as a sample synthesis layer. The planning agent sorts and combines at least one layer of derived feature code generation layer, at least one layer of few-short learning layer, and at least one layer of sample synthesis layer to obtain the execution level of the zero-shot thought chain; New samples are generated according to the execution hierarchy of the zero-sample thinking chain.
7. A data augmentation device based on multi-agent systems, characterized in that, include: A planning agent is used to acquire task descriptions and data descriptions of sample data. The data descriptions include data distribution statistics and data feature details. The sample data includes network traffic data. The data descriptions include: the network traffic data includes popular data and unpopular data; the data distribution statistics include the recommendation frequency of popular objects and the recommendation frequency of unpopular objects; and the data feature details include: parameter features of popular objects and parameter features of unpopular objects. The planning agent generates an initial zero-shot thought chain based on the data distribution statistics and data feature details of the sample data. The zero-shot thought chain includes the agent type and execution strategy called at each execution level. The planning agent, based on the execution strategy of the zero-shot thought chain, determines the execution level of the agent that calls the enhancement strategy and generates new samples. The planning agent evaluates the enhancement effect of the new sample on the downstream deep learning model. If the enhancement effect does not meet the set conditions, the execution order of the zero-shot thought chain and the data augmentation steps in the execution layer are adjusted, and the step of determining the execution layer of the agent that calls the enhancement strategy based on the zero-shot thought chain is returned. The enhanced strategy agent includes: a feature-derived agent and a sample-generated agent. The determination of the execution level of the agent that invokes the enhanced policy includes: The feature-derived agent is invoked to perform a derived feature code generation task in conjunction with the task description; The feature-derived agent is invoked to perform a few-short learning task based on the task description. The sample generation agent is invoked to synthesize new samples from the derived features in the feature derivation execution result: The feature-derived agent, based on the data distribution statistics and data feature details in the data description, identifies imbalanced classifications and / or missing sample data with insufficient feature representation in the classification rules of the sample data. Based on imbalanced classification in the classification rules of the sample data, and / or missing sample data with insufficient feature representation, a corresponding feature derivation algorithm is determined, wherein the derived feature code includes at least one of the following algorithms: Multinomial feature derivation algorithm, cross-combination feature derivation algorithm, grouped statistical feature derivation algorithm, time series feature derivation algorithm, discretization feature derivation algorithm, and neuro-linguistic programming (NLP) feature derivation algorithm; The feature-derived agent combines the feature-derived algorithm with the task description to generate corresponding derived feature codes.
Citation Information
Patent Citations
E-commerce personalized recommendation data system fused with deep learning
CN120146966A
Data enhancement method and system based on multi-agent self-evolution and hybrid evaluation
CN121211014A