A memory-augmented LLM multi-agent system for automatic feature generation

Abstract

The application provides a memory-enhanced LLM multi-agent system for automatic feature generation, which comprises: a plurality of generation agents, each generation agent independently interacts with a large language model, and generates new features from original data based on different generation strategies; a multi-level memory module associated with each generation agent for storing operation records, performance feedback and abstract concepts in the feature generation process; a summary agent for summarizing the memory information of each generation agent, generating global concept memory, and guiding the next round of feature generation; an iterative optimization engine for dynamically adjusting the feature generation strategy of the generation agent based on the feedback of the memory module in multiple iterations. Through the cooperation of multiple agents and the unified multi-level memory mechanism, the semantic understanding ability of LLM is fully utilized, the space of feature generation is expanded, and the quality and diversity of the generated features are improved through iterative optimization.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically a memory-enhanced LLM multi-agent system for automatic feature generation. Background Technology

[0002] With the rapid development of Automated Machine Learning (AutoML) technology, the efficiency of data modeling has been significantly improved, attracting widespread attention from academia and industry. In particular, the emergence of automated feature generation technology has enabled the extraction of informative features from raw data without human intervention, becoming a key driving force for building accurate machine learning models with good generalization capabilities.

[0003] Traditional automated feature generation methods typically rely on a predefined set of operators, such as symbolic feature trees, exhaustive combinations, or feature maps. These methods generate new features by transforming the original features. However, these methods depend on a finite set of operators and do not consider the semantic information of the features, resulting in a limited feature space and making it difficult to meet the feature generation needs of complex tasks.

[0004] In recent years, LLM has demonstrated powerful capabilities in semantic understanding and generation, leading to the emergence of LLM-based feature generation methods. These methods generate features by extracting semantic information from task descriptions, showing initial potential to overcome the limitations of traditional methods. However, these methods typically rely on static, single generation strategies, which still limit the extensive exploration of the feature space. More importantly, they lack mechanisms for adapting generation strategies based on historical experience or task feedback, resulting in a disconnect between feature generation and learning objectives, and an inability to efficiently explore valuable feature transformations. This lack of feedback mechanisms often prevents efficient feature generation in complex tasks, limiting their performance.

[0005] To address these issues, this invention proposes a memory-enhanced LLM multi-agent system for automatic feature generation. Summary of the Invention

[0006] This invention addresses the problem that existing technologies rely on limited predefined operations, making it difficult to handle the generation of diverse features in complex tasks.

[0007] To address the aforementioned problems, this invention provides a memory-enhanced LLM multi-agent system for automatic feature generation, an automatic feature generation method, an electronic device, and a storage medium.

[0008] In a first aspect, the present invention provides a memory-enhanced LLM multi-agent system for automatic feature generation, the system comprising: Multiple generative agents, each interacting independently with a large language model, generate new features from the original data based on different generative strategies; A multi-level memory module, associated with each generating agent, is used to store operation records, performance feedback, and abstract concepts during the feature generation process; The summary agent is used to summarize the memory information of each generating agent, generate a global conceptual memory, and guide the next round of feature generation; An iterative optimization engine is used to dynamically adjust the feature generation strategy of the generated agent based on the feedback of the memory module in multiple iterations, thereby gradually optimizing the quality and diversity of the generated features.

[0009] Optionally, the multi-level memory module includes: Process memory is used to record the feature generation operations performed by each generative agent in each iteration, including basic columns, transformation type, generated feature name, transformation description and iteration round; Feedback memory is used to store the performance evaluation results of the generated features in downstream tasks, including feature name, evaluation metric, metric value and validity identifier; Conceptual memory is used to store high-level generative rules or patterns abstracted from procedural and feedback memory by large language models.

[0010] Optionally, the summarizing agent is further used for: It receives the conceptual memory and feedback memory of all generated agents and generates global conceptual memory through a large language model. The global conceptual memory is distributed to each generating agent for constructing prompts for the next round of feature generation.

[0011] Optionally, the generated agent includes, but is not limited to, one or more of the following types: Univariate feature intelligent agents, cross-combination intelligent agents, time feature intelligent agents, aggregate construction intelligent agents, local transformation intelligent agents, and local pattern intelligent agents.

[0012] Optionally, the generating agent performs the following operations in each iteration: Dynamic prompts are built based on data statistics, task metadata, local feedback memory, local conceptual memory, and global conceptual memory. Dynamic prompts are input into a large language model to generate transformation ideas and instantiate them into executable feature generation strategies. The generated new features are evaluated, and the evaluation results are stored in the feedback memory; Update process memory and concept memory.

[0013] Optionally, the evaluation metrics for the feature generation strategy include, but are not limited to, the correlation between the feature and the target variable, the improvement in model performance, and the diversity or stability of the feature.

[0014] Optionally, the iterative optimization engine is further used for: After each iteration, the top-ranked features are selected from the feature set generated by each generating agent. The selected features are merged into the global dataset for use in the next round of feature generation and model training.

[0015] Optionally, the generating agents operate in parallel, each generating features independently, and their generation strategies are gradually differentiated based on feedback from the memory module during the iteration process to avoid feature homogenization.

[0016] Secondly, embodiments of the present invention provide an automatic feature generation method based on the system described in the first aspect, comprising the following steps: S1. Initialize multiple generating agents and their corresponding multi-level memory modules; S2. In each iteration, each generative agent generates new features based on dynamic prompts and evaluates its performance. S3. Update the process memory, feedback memory, and conceptual memory of each agent; S4. Summarize the conceptual memories of each agent and generate a global conceptual memory. S5. Adjust the feature generation strategy for the next round based on global conceptual memory and feedback memory; S6. Filter and aggregate the best features generated in this round, and update the global dataset.

[0017] Optionally, the construction of the dynamic prompt further includes: Retrieve valid features generated in the previous round from the feedback memory; Extracting effective transformation patterns from conceptual memory; Retrieve shared knowledge across agents from global conceptual memory; By combining statistical data and task metadata, structured generation prompts are generated.

[0018] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the second aspect.

[0019] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the second aspect.

[0020] The beneficial effects of the memory-enhanced LLM multi-agent system for automatic feature generation of the present invention are as follows: The system of this invention effectively expands the feature generation space through multi-agent collaboration and memory enhancement mechanisms, avoiding the limitations of operators and generation strategies in traditional methods. The system can dynamically optimize the generation strategy based on historical experience and task feedback, improving the quality and diversity of features and ensuring a high degree of alignment between generated features and learning objectives. By summarizing the guidance of agents, this system improves feature discovery efficiency significantly, outperforming traditional methods, and has broad application prospects. Attached Figure Description

[0021] Figure 1 This is a block diagram illustrating the principle of a memory-enhanced LLM multi-agent system for automatic feature generation in an embodiment of the present invention. Figure 2 This is a flowchart of the automatic feature generation method in an embodiment of the present invention; Figure 3 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0022] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0023] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0024] First, a brief description of the present invention: This invention aims to address the challenges of automated feature generation, particularly how to generate diverse and high-quality features while maintaining efficiency through innovative generation strategies to support the optimization of machine learning models. Traditional methods rely on predefined operators and fixed generation strategies, which cannot adaptively optimize based on task requirements and complex relationships between features, thus limiting the diversity and quality of feature generation.

[0025] This invention proposes a multi-agent system framework for large-scale language models based on memory enhancement for automated feature generation. The core idea of ​​this scheme is to expand the feature generation space and achieve adaptive optimization during the feature generation process through collaboration among multiple agents and the integration of a multi-level memory mechanism. Specific technical solutions include: Multi-agent system design: By decomposing the feature generation process into multiple independent agents, each agent is responsible for different feature generation tasks and has a clearly defined role and responsibility. Each agent dynamically constructs prompts and generates new features based on the current feature set and feedback information through multi-turn dialogue with the LLM. This design effectively avoids the feature homogenization problem in traditional single-agent methods and expands the dimensionality and semantic space of feature generation.

[0026] Memory enhancement mechanism: To address the problem of insufficient adaptability in feedback-driven systems, this invention designs a multi-level memory module, comprising the following parts: Process memory: Records past feature generation operations to avoid repeatedly exploring the same features; Feedback memory: Stores performance feedback of generated features, directly linking the generated features to the utility of subsequent tasks; Conceptual memory: Extracting high-level patterns and abstractions from historical feedback to help intelligent agents reason and adapt in the long term.

[0027] Adaptive Generation Strategy: Based on a multi-level memory module, the agent can continuously optimize the generation strategy according to historical experience and task feedback, thereby improving the quality and diversity of feature generation. By introducing a summarizing agent, the system can integrate the feedback from various agents to form a global conceptual memory to guide the subsequent feature generation process.

[0028] This invention enables an agent to dynamically adjust its generation strategy and discover higher-value, task-relevant features by implementing this closed-loop generation-evaluation-adaptation system.

[0029] like Figure 1 As shown in the embodiment of the present invention, a memory-enhanced LLM multi-agent system for automatic feature generation is provided. The system includes: Multiple generative agents, each interacting independently with a large language model, generate new features from the original data based on different generative strategies; In this embodiment, the generated intelligent agent includes, but is not limited to, one or more of the following types: Univariate feature intelligent agents, cross-combination intelligent agents, time feature intelligent agents, aggregate construction intelligent agents, local transformation intelligent agents, and local pattern intelligent agents.

[0030] To address the issue that a single generator cannot deeply explore new features during the generation phase, this system overcomes this limitation by splitting the feature generation task into multiple agents with specific responsibilities, thereby naturally increasing the diversity and adaptability of generated features.

[0031] Specifically, the system includes Each agent generates an intelligent agent, and each agent generates an intelligent agent. These agents are responsible for generating features based on different and specialized strategies. They operate in parallel, each utilizing its unique approach to explore the interactions, transformations, and combinations between features. This multi-agent design allows the system to flexibly adapt to different feature types and modeling needs, applying heterogeneous strategies through multi-round iterative feature generation. The entire process can be represented as: ; in, Indicates the first The aggregate set of features generated by all agents in the round, and Indicates the first Each agent generates a subset of features in the same round. Each agent applies its specified policy from the original dataset. A new set of features is generated independently. The combined outputs of all agents ultimately form a richer and more comprehensive feature space, which is used for subsequent model training.

[0032] The feature generation process in this system is systematically divided into several well-defined roles, each agent focusing on a specific aspect of feature construction. These roles aim to ensure comprehensive exploration across different feature dimensions and generate a large number of high-value features applicable to complex tasks. They also enable each agent to specialize within its domain and develop a deeper understanding of its functional scope.

[0033] Each agent in this system All will be on the original dataset Apply different transformation strategies Generate a feature subset that matches its specified target. Formalistically, intelligent agents The generated features are defined as follows: .

[0034] This embodiment instantiates this design through a set of specific agent policies, where each transformation policy This corresponds to a predefined strategy category, as described below: Univariate feature agent: for the original dataset We apply univariate transformations to individual features to explore basic yet informative feature transformations.

[0035] Cross-combination agents: by combining the original dataset By combining multiple inputs, new features are constructed, capturing higher-order feature interactions.

[0036] Temporal Feature Agent: When the dataset contains time series data, it focuses on extracting temporal patterns. These features are crucial for understanding the relationship between past and current observations in the task.

[0037] Aggregate to build intelligent agents: generate statistical summary features of variable groups, such as mean, maximum value, or count.

[0038] Local transformation agent: applies region-specific transformations, such as discretization or piecewise functions, to capture local information patterns.

[0039] Local pattern agents: use clustering or local interaction modeling techniques to identify latent patterns within a subset of features.

[0040] A multi-level memory module, associated with each generating agent, is used to store operation records, performance feedback, and abstract concepts during the feature generation process; In this embodiment, the multi-level memory module includes: Process memory is used to record the feature generation operations performed by each generative agent in each iteration, including basic columns, transformation type, generated feature name, transformation description and iteration round; Feedback memory is used to store the performance evaluation results of the generated features in downstream tasks, including feature name, evaluation metric, metric value and validity identifier; Conceptual memory is used to store high-level generative rules or patterns abstracted from procedural and feedback memory by large language models.

[0041] It should be noted that the feature generation process in this system is not static, nor is it limited to a single policy. Instead, it is an iterative, memory-enhanced process, in which each agent maintains and updates its memory to optimize the feature generation policy over time. Specifically, each agent... It is equipped with a dedicated memory module that tracks past generation operations, evaluation feedback, and abstract conceptual patterns. This module consists of three parts: process memory, feedback memory, and conceptual memory.

[0042] Procedural memory: Procedural memory aims to record the exact operations performed by each agent during feature generation. In the... In the round of iteration, if the agent Generated If a new feature is defined, then process memory is defined as: ; in, Represents the basic column, It is a transformation type. It is the name of the generated feature. It is a transformation description. This is the current iteration round. This component ensures traceability, promotes reproducibility, and prevents redundant exploration during iteration.

[0043] Feedback memory: The feedback memory stores the performance evaluation of the features generated by each agent. After each iteration, the system records whether the features improved the performance of downstream tasks through predefined metrics. For each agent... If in the first Generated in round iteration The feature, feedback memory is defined as: ; in, It is the feature name. These are evaluation indicators. It is an indicator value. Indicates whether the feature is valid. It refers to the number of iterations. Feedback memory guides the agent's future decisions by recording which features positively contributed to the model's performance.

[0044] Conceptual memory: Conceptual memory captures high-level patterns extracted from the agent's own feature generation history. After each iteration, LLM abstracts generalized rules from the agent's procedural and feedback memories: ; These local conceptual rules enable each agent to optimize its feature generation strategy in subsequent iterations.

[0045] Global Conceptual Memory: To facilitate knowledge sharing among agents, the conceptual memories of all agents are aggregated into a global conceptual summary by summarizing the agents' collective knowledge. This collective knowledge guides subsequent iterations, reduces redundancy, and optimizes feature generation strategies.

[0046] The summary agent is used to summarize the memory information of each generating agent, generate a global conceptual memory, and guide the next round of feature generation; In summary, intelligent agents can be further used for: It receives the conceptual memory and feedback memory of all generated agents and generates global conceptual memory through a large language model. The global conceptual memory is distributed to each generating agent for constructing prompts for the next round of feature generation.

[0047] After each iteration, the summarizing agent synthesizes a global conceptual memory from the local conceptual memories and feedback memories of all agents. This summary aggregates the effective transformation strategies identified by each agent, thereby supporting the adjustment of the generation strategy in the next iteration.

[0048] An iterative optimization engine is used to dynamically adjust the feature generation strategy of the generated agent based on the feedback of the memory module in multiple iterations, thereby gradually optimizing the quality and diversity of the generated features.

[0049] The generated agent performs the following operations in each iteration: Dynamic prompts are built based on data statistics, task metadata, local feedback memory, local conceptual memory, and global conceptual memory. Dynamic prompts are input into a large language model to generate transformation ideas and instantiate them into executable feature generation strategies. The generated new features are evaluated, and the evaluation results are stored in the feedback memory; Update process memory and concept memory.

[0050] Evaluation metrics for feature generation strategies include, but are not limited to, the correlation between features and target variables, the extent to which features improve model performance, and the diversity or stability of features.

[0051] The iterative optimization engine is further used for: After each iteration, the top-ranked features are selected from the feature set generated by each generating agent. The selected features are merged into the global dataset for use in the next round of feature generation and model training.

[0052] The generating agents operate in parallel, each generating features independently, and their generation strategies gradually differentiate based on feedback from the memory module during the iteration process to avoid feature homogenization.

[0053] Based on the above, the iterative feature generation mechanism of a multi-agent system is described in detail. The agents utilize local and global memory to progressively optimize the generation strategy, refining the feature set across multiple iterations. This process is as follows: Figure 1 As shown. In each iteration In the middle, each intelligent agent Features are autonomously generated through a series of structured and consistent steps: building prompts, feature generation and evaluation, and memory updates (see the definition of memory above).

[0054] Hint Construction: Each agent constructs a dynamic cue that integrates: data statistics and task-specific metadata. ; Effective features retrieved from the previous round of feedback memory From the previous round of local concept memory and global memory Insights extracted from [the text].

[0055] Feature generation and evaluation: Guided by the construction prompts, the agent first generates a high-level transformation idea through LLM, and then instantiates it into an executable policy. This strategy uses predefined evaluation metrics for assessment and stores performance feedback in the agent's feedback memory for future optimization.

[0056] Memory update: The agent updates its procedural and conceptual memories to reflect effective transformations and meaningful patterns discovered during the evaluation process, thereby providing evidence-based guidance for feature generation in future iterations.

[0057] In this invention, the iterative optimization process systematically and gradually optimizes the global feature set through accumulated feedback and memory. Each iteration... At the end, the best features generated by all agents are carefully selected and integrated into the dataset. Specifically, each agent... Its local feature set application Only the highest-ranking and most informative features based on the evaluation indicators are retained.

[0058] These selected features are aggregated to expand the global dataset, enabling the system to gradually accumulate high-quality transformations and adapt based on previous results, thereby achieving incremental and composite improvements in model performance.

[0059] In summary, the system of this invention effectively expands the feature generation space through multi-agent collaboration and memory enhancement mechanisms, avoiding the limitations of operators and generation strategies in traditional methods. The system can dynamically optimize the generation strategy based on historical experience and task feedback, improving the quality and diversity of features and ensuring a high degree of alignment between generated features and learning objectives. By summarizing the guidance of agents, this system improves feature discovery efficiency significantly, outperforming traditional methods, and has broad application prospects.

[0060] This invention also provides an automatic feature generation method based on the above system, such as... Figure 2 As shown, Figure 2 This is a flowchart of the automatic feature generation method in this invention, which includes the following steps: S1. Initialize multiple generating agents and their corresponding multi-level memory modules; S2. In each iteration, each generative agent generates new features based on dynamic prompts and evaluates its performance. S3. Update the process memory, feedback memory, and conceptual memory of each agent; S4. Summarize the conceptual memories of each agent and generate a global conceptual memory. S5. Adjust the feature generation strategy for the next round based on global conceptual memory and feedback memory; S6. Filter and aggregate the best features generated in this round, and update the global dataset.

[0061] Furthermore, the construction of dynamic suggestions further includes: Retrieve valid features generated in the previous round from the feedback memory; Extracting effective transformation patterns from conceptual memory; Retrieve shared knowledge across agents from global conceptual memory; By combining statistical data and task metadata, structured generation prompts are generated.

[0062] This method is based on the memory-enhanced LLM multi-agent system for automatic feature generation described above. Its technical principles and the resulting technical effects are the same as those of this method, so they will not be repeated here.

[0063] like Figure 3 As shown in the figure, an electronic device provided by an embodiment of the present invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following method: S1. Initialize multiple generating agents and their corresponding multi-level memory modules; S2. In each iteration, each generative agent generates new features based on dynamic prompts and evaluates its performance. S3. Update the process memory, feedback memory, and conceptual memory of each agent; S4. Summarize the conceptual memories of each agent and generate a global conceptual memory. S5. Adjust the feature generation strategy for the next round based on global conceptual memory and feedback memory; S6. Filter and aggregate the best features generated in this round, and update the global dataset.

[0064] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0066] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A memory-enhanced LLM multi-agent system for automatic feature generation, characterized in that, The system includes: Multiple generative agents, each interacting independently with a large language model, generate new features from the original data based on different generative strategies; A multi-level memory module, associated with each generating agent, is used to store operation records, performance feedback, and abstract concepts during the feature generation process; The summary agent is used to summarize the memory information of each generating agent, generate a global conceptual memory, and guide the next round of feature generation; An iterative optimization engine is used to dynamically adjust the feature generation strategy of the generated agent based on the feedback of the memory module in multiple iterations, thereby gradually optimizing the quality and diversity of the generated features.

2. The system according to claim 1, characterized in that, The multi-level memory module includes: Process memory is used to record the feature generation operations performed by each generative agent in each iteration, including basic columns, transformation type, generated feature name, transformation description and iteration round; Feedback memory is used to store the performance evaluation results of the generated features in downstream tasks, including feature name, evaluation metric, metric value and validity identifier; Conceptual memory is used to store high-level generative rules or patterns abstracted from procedural and feedback memory by large language models.

3. The system according to claim 2, characterized in that, The summarizing agent is further used for: It receives the conceptual memory and feedback memory of all generated agents and generates global conceptual memory through a large language model. The global conceptual memory is distributed to each generating agent for constructing prompts for the next round of feature generation.

4. The system according to claim 3, characterized in that, The generated agent includes, but is not limited to, one or more of the following types: Univariate feature intelligent agents, cross-combination intelligent agents, time feature intelligent agents, aggregate construction intelligent agents, local transformation intelligent agents, and local pattern intelligent agents.

5. The system according to claim 1, characterized in that, The generated agent performs the following operations in each iteration: Dynamic prompts are built based on data statistics, task metadata, local feedback memory, local conceptual memory, and global conceptual memory. Dynamic prompts are input into a large language model to generate transformation ideas and instantiate them into executable feature generation strategies. The generated new features are evaluated, and the evaluation results are stored in the feedback memory; Update process memory and concept memory.

6. The system according to claim 5, characterized in that, The evaluation metrics for the feature generation strategy include, but are not limited to, the correlation between the feature and the target variable, the improvement in model performance, and the diversity or stability of the feature.

7. The system according to claim 1, characterized in that, The iterative optimization engine is further used for: After each iteration, the top-ranked features are selected from the feature set generated by each generating agent. The selected features are merged into the global dataset for use in the next round of feature generation and model training.

8. The system according to claim 1, characterized in that, The generating agents operate in parallel, each generating features independently, and their generation strategies are gradually differentiated based on feedback from the memory module during the iteration process to avoid feature homogenization.

9. An automatic feature generation method based on the system according to any one of claims 1 to 8, characterized in that, Includes the following steps: Initialize multiple generating agents and their corresponding multi-level memory modules; In each iteration, each generative agent generates new features based on dynamic prompts and evaluates its performance; Update the process memory, feedback memory, and conceptual memory of each agent; The summary agent aggregates the conceptual memories of each agent and generates a global conceptual memory. Based on global conceptual memory and feedback memory, adjust the feature generation strategy for the next round; Filter and aggregate the best features generated in this round, and update the global dataset.

10. The method according to claim 9, characterized in that, The construction of the dynamic prompts further includes: Retrieve valid features generated in the previous round from the feedback memory; Extracting effective transformation patterns from conceptual memory; Retrieve shared knowledge across agents from global conceptual memory; By combining statistical data and task metadata, structured generation prompts are generated.

Images