A table data augmentation method and system based on a conditional inference diffusion model

Abstract

This invention provides a method and system for table data augmentation based on a conditional inference diffusion model, belonging to the field of table data processing technology. The method first acquires the original table data and preset constraints, and constructs a conditional inference diffusion model composed of a chain inference network and a conditional diffusion network. The chain inference network extracts implicit relationships between data based on the original table data, preset constraints, and a large language model, outputting intermediate table data containing these relationships. Subsequently, the conditional diffusion network uses this intermediate table data as constraints to perform data augmentation operations on the original table data, ultimately generating the target table data. This method achieves data augmentation through the conditional inference diffusion model composed of a chain inference network and a conditional diffusion network, effectively solving the problems of poor generation quality and weak data relationship preservation ability of traditional generation models in complex table scenarios, thereby achieving high-quality table data augmentation.

Description

Technical Field

[0001] This invention belongs to the field of tabular data processing technology, specifically relating to a tabular data enhancement method and system based on a conditional inference diffusion model. Background Technology

[0002] Tabular data, as an important carrier of structured information, has wide applications in many fields such as machine learning, data mining, and analysis. However, in practical applications, obtaining sufficient and high-quality tabular data often faces challenges, such as sample scarcity due to high data acquisition costs, or the inability to directly use the original data for privacy protection reasons. To address this, tabular data augmentation techniques aim to expand datasets, improve model robustness, or desensitize original information by generating synthetic data. Currently, methods based on generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have been attempted for tabular data synthesis. However, when dealing with tables containing complex inter-column relationships and heterogeneous features, it is often difficult to accurately maintain the internal structure and semantic relationships of the original data in the generated samples, resulting in limited generation quality and insufficient fidelity. In addition, existing methods often lack targeted guidance for specific constraints or implicit relationships, which limits the reliability and practicality of the generated results in scenarios where data is scarce or where strict adherence to predetermined business logic is required. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a table data augmentation method and system based on a conditional inference diffusion model to solve the aforementioned problems. This method completes data augmentation through a conditional inference diffusion model composed of a chain inference network and a conditional diffusion network, effectively improving the problems of insufficient generation quality and weak correlation maintenance of traditional generative models in complex table scenarios. It can achieve high-quality table data augmentation in scenarios with scarce data and sensitive privacy.

[0004] To address the aforementioned technical problems, this invention provides a table data augmentation method based on a conditional inference diffusion model, comprising the following steps: Retrieve raw table data and preset constraints; Construct a conditional reasoning diffusion model consisting of a chain reasoning network and a conditional diffusion network; The original table data and preset constraints are input into the conditional reasoning diffusion model, so that the chain reasoning network extracts implicit relationships based on the original table data, preset constraints, and preset large language model to obtain intermediate table data; the intermediate table data includes implicit relationship information. The intermediate table data and the original table data are input into the conditional inference diffusion model, so that the conditional diffusion network performs data augmentation on the original table data based on the intermediate table data to obtain the target table data.

[0005] In the above scheme, by constructing a conditional inference diffusion model composed of a chain-like inference network and a conditional diffusion network, data augmentation can be directly performed on the original table data and preset conditional constraints, thereby improving the targeting, data fidelity, and semantic consistency of the data augmentation process. The chain-like inference network, combined with a preset large language model, extracts implicit relationships based on the original table data and preset conditional constraints, generating intermediate table data including implicit relationship information. The conditional diffusion network uses the intermediate table data as a condition to augment the original table data to obtain the target table data, ensuring that the structure and correlation of the augmented data are highly consistent with the original table data. This effectively improves the problems of insufficient generation quality and weak correlation maintenance of traditional generative models in complex table scenarios, and can achieve high-quality table data expansion in data-scarce and privacy-sensitive scenarios.

[0006] It should be noted that the preset conditions and constraints include, but are not limited to, the user's query question, the target relationship to be revealed, or the target description of the generated task, which are used to guide the subsequent reasoning and generation direction.

[0007] Further, the original table data and preset constraints are input into the conditional inference diffusion model, so that the chain inference network extracts implicit relationships based on the original table data, preset constraints, and preset large language model to obtain intermediate table data, including: In the chained inference network, the following steps are performed: Initialize the current table state and current historical operation records based on the original table data and preset condition constraints; Based on the current table state, current historical operation record, preset condition constraints and preset large language model, iterative reasoning and evolution are performed until the preset termination condition is met, and then the final table state and complete historical operation record are output. Based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm, implicit relationship parsing is performed to obtain implicit relationship information; The intermediate table data is obtained based on the implicit relationship information and the final table state.

[0008] In the above scheme, a chain-based reasoning network is used to systematically process the original table data and preset constraints. First, the current table state and historical operation records are initialized, laying a standardized foundation for subsequent reasoning. Then, combined with a preset large language model, iterative reasoning evolution is carried out based on the real-time table state, historical operation records, and constraints until a preset termination condition is met, thus completely reconstructing the table generation logic. Subsequently, a structured relation extraction algorithm is used to deeply analyze the final state and operation records, accurately mining implicit relational information such as column dependencies, and forming intermediate table data that integrates logical relationships. This process achieves full-link automation from data initialization and reasoning evolution to relation parsing, significantly improving the accuracy and completeness of extracting implicit relations from complex tables. It effectively solves the problem of insufficient recognition by traditional methods in scenarios with sparse data and complex logic, enhances the model's ability to understand the deep logic of tables, and provides reliable support for the subsequent generation of high-quality table data.

[0009] It should be noted that the iterative reasoning evolution process of the chain-based reasoning network is preferably based on a predefined operation pool. The operation pool contains multiple atomic table operations, which are dynamically invoked by the large language model to gradually evolve the table and reveal implicit relationships.

[0010] Furthermore, the iterative reasoning and evolution based on the current table state, current historical operation records, preset condition constraints, and preset large language model, until a preset termination condition is met, and then outputting the final table state and complete historical operation records, includes: Perform iterative reasoning and evolution steps: Dynamic prompts are generated based on the current table status, current historical operation records, and preset condition constraints. The dynamic prompt information is input into a preset large language model, so that the preset large language model can perform inference and evolution based on the dynamic prompt information to obtain the current table operation instruction and the corresponding current operation parameters. Update the current table status based on the current table operation instruction; update the current historical operation record based on the current table operation instruction and the corresponding current operation parameters; Repeat the above iterative reasoning and evolution steps until the preset number of iterative reasoning and evolution steps are reached. Then, output the current table state and the current historical operation record as the final table state and the complete historical operation record.

[0011] In the above scheme, prompts are dynamically constructed based on the current table state, historical operation records, and preset constraints. This ensures that the reasoning evolution of the large language model always aligns with the real-time context and constraints, effectively improving the accuracy and relevance of table operation instructions and corresponding operation parameters, and preventing instructions from deviating from the task objective. During iterative reasoning, the table state and historical operation records are updated step by step according to the operation instructions. Through multiple rounds of progressive evolution, the complex implicit logic and structural relationships in the original table are gradually sorted out and made explicit, ensuring the integrity and coherence of the reasoning chain. Using a preset number of iterations as the termination condition ensures that the reasoning process is fully executed while preventing infinite iteration from wasting computational resources and reducing efficiency. This approach makes the entire chain-like reasoning process more controllable and standardized, improves the interpretability and scenario adaptability of the reasoning logic, and ultimately outputs a structurally sound and logically complete final table state and complete historical operation records. This provides reliable support for subsequent conditional generation and data augmentation, effectively guaranteeing the high consistency and quality of table data in terms of structure, distribution, and semantics.

[0012] Further, the step of repeating the above iterative reasoning and evolution steps until a preset number of iterative reasoning and evolution steps are reached, and then outputting the current table state and the current historical operation record as the final table state and the complete historical operation record, includes: During the repeated iterative reasoning and evolution steps described above, the current table row number and the current key column structure are obtained based on the current table state. If the current table row count and the current key column structure meet the preset constraints, or reach the preset number of iterations and inferences, then the current table state and the current historical operation record will be output as the final table state and the complete historical operation record.

[0013] In the above scheme, during the iterative evolution of chained reasoning, the number of table rows and key column structure information are extracted in real time based on the current table state. This information serves as a dynamic criterion for terminating the iteration. When the relevant indicators meet the preset constraints, the iteration can be terminated early, allowing the reasoning process to be flexibly adjusted according to the actual table construction. This effectively reduces unnecessary iteration steps, avoids ineffective calculations and resource waste, and significantly improves the efficiency of iterative reasoning. Simultaneously, using a preset number of iterations as a hard safety net effectively prevents the reasoning process from looping infinitely due to abnormal situations, ensuring a stable and controllable iteration process. This dual termination mechanism enables the entire chained reasoning process to have adaptive judgment and dynamic adjustment capabilities, better adapting to table data processing scenarios with different structures and complexities. It ensures that the final output table state is standardized and reasonable, and that historical operation records are complete and clear, providing a stable and reliable data foundation for subsequent implicit relationship parsing and conditional table generation, effectively guaranteeing the overall effectiveness and quality of the table data augmentation task.

[0014] Further, the step of repeating the above iterative reasoning and evolution steps until a preset number of iterative reasoning and evolution steps are reached, and then outputting the current table state and the current historical operation record as the final table state and the complete historical operation record, includes: During the repeated iterative reasoning evolution steps, the preset large language model is subjected to end marker detection. If the preset large language model outputs a preset reasoning end marker, or reaches a preset number of iterative reasoning evolutions, the current table state and the current historical operation record are used as the final table state and the complete historical operation record and output.

[0015] In the above scheme, a pre-set large language model end-of-process marker detection mechanism is introduced during the iterative execution of the reasoning evolution steps, making the chained reasoning process more adaptive and dynamically controllable. The model output is monitored in real time during iteration. Once the pre-set reasoning end-of-process marker is detected, the reasoning process can be terminated early, allowing the model to autonomously determine its completion status based on its own reasoning progress, which better aligns with the real reasoning logic of generating tabular data. Simultaneously, a pre-set number of iterations is used as a hard termination condition, forming a dual termination strategy combining flexible judgment and hard constraints. This avoids redundant computation and resource consumption caused by excessive iteration, improving overall reasoning evolution efficiency, and effectively prevents the inability to terminate the reasoning due to abnormal output. This significantly improves the system's operational stability and generalization ability when processing tabular data of varying complexity, ensuring that the final output table state and historical operation records are logically complete, concise, and standardized, providing reliable support for subsequent relation parsing and data generation.

[0016] Furthermore, in the process of parsing implicit relationships based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm to obtain implicit relationship information, the preset structured relationship extraction algorithm specifically includes: Based on the final table state, statistical correlation indicators are calculated to obtain the statistical relationships between columns; Based on the complete historical operation records, semantic relationships between columns are obtained through parsing and inference. The implicit relationship information is obtained by performing a structured representation based on the statistical and semantic relationships between columns.

[0017] In the above scheme, by calculating statistical correlation indicators on the final table state, the statistical relationships between columns can be accurately extracted from the perspective of data distribution and numerical correlation, objectively reflecting the inherent patterns of the table data. Simultaneously, relying on complete historical operation records for layer-by-layer analysis and logical inference, the semantic relationships between columns are mined from the perspective of operational intent and field association, clearly revealing the business logic and dependencies between fields. Then, the statistical and semantic relationships are represented in a unified structure, merging to form comprehensive and accurate implicit relationship information. The implicit relationship information obtained in this way possesses both statistical reliability and semantic interpretability, effectively identifying complex implicit relationships that are difficult to capture using conventional methods. This significantly improves the understanding of the deep structure and internal logic of the table, providing a solid and reliable basis for subsequent conditional generation and logical constraints.

[0018] It should be noted that the calculation of the statistical correlation index includes, but is not limited to, calculating the Pearson correlation coefficient between columns. The analytical inference can be based on the logic of historical operation sequences to reveal the semantic dependencies and associations between columns.

[0019] Further, the step of inputting the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network uses the intermediate table data as a condition to perform data augmentation on the original table data to obtain the target table data, includes: In the conditional diffusion network, the following steps are performed: Encoding is performed based on the intermediate table data and the preset condition encoder to obtain a condition vector representation; Noise scheduling is performed based on the original table data to obtain the original noise data; Denoising sampling is performed based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network to obtain the initial synthesized table data; The target table data is obtained by post-processing the initial synthesized table data.

[0020] In the above scheme, the preset conditional encoder encodes the intermediate table data to obtain a conditional vector representation, ensuring that the generation process is always subject to the semantic constraints of implicit relationships; noise scheduling is performed on the original table data to obtain original noisy data, effectively preserving the inherent data features of the original table; a pre-trained conditional denoising sub-network combines the conditional vector representation with the original noisy data to complete denoising sampling, obtaining initial synthesized table data, which is then post-processed to obtain the target table data. This scheme significantly improves the fidelity, semantic consistency, and structural integrity of the generated data, effectively addressing real-world scenarios such as data sparsity, and enhancing its practicality and reliability in table data expansion and privacy protection.

[0021] It should be noted that the noise scheduling based on the original tabular data may specifically include applying different diffusion processes to the numerical features and categorical features in the table. For numerical features, a continuous-time diffusion process is used, for example, by adding Gaussian noise through a variance-explosive stochastic differential equation. For categorical features, after converting them into one-hot vectors, a categorical diffusion process is applied, using a mask scheduling function to control the probability of a feature being replaced by a masked state. The noise scheduling process involves feature-specific learnable noise scheduling parameters.

[0022] Furthermore, the preset condition encoder includes an embedding layer and an encoding network layer; the encoding based on the intermediate table data and the preset condition encoder to obtain a condition vector representation includes: Based on the intermediate table data, obtain cell content, column identifiers, and type information; The cell content, column identifier, and type information are input into a preset condition encoder so that the embedding layer performs feature mapping based on the cell content, column identifier, and type information to obtain continuous feature embedding. The continuous feature embeddings are input into the encoding network layer, so that the encoder network layer performs fusion modeling based on the continuous feature embeddings to obtain the encoded feature sequence; Pooling is performed based on the encoded feature sequence to obtain the conditional vector representation.

[0023] In the above scheme, the preset conditional encoder uses intermediate table data as the processing object, accurately extracting cell content, column identifiers, and type information as encoding input. This comprehensively captures the core features of the table data from multiple dimensions, including semantics, structure, and data type, avoiding the omission of crucial information. The embedding layer performs unified feature mapping on the aforementioned multi-dimensional information, transforming discrete and heterogeneous table information into continuous feature embeddings, achieving standardized expression of heterogeneous features. Based on this, the encoding network layer performs deep fusion modeling on the continuous feature embeddings, fully exploring the correlations between features to form a high-dimensional encoded feature sequence. Subsequently, pooling operations are used to aggregate and compress the feature sequence, ultimately generating a compact, concise conditional vector representation rich in complete semantic and structural information. This scheme effectively improves the expression accuracy and information completeness of conditional information, enabling the subsequent generation process to strictly follow the implicit logical relationships extracted through chain reasoning. It significantly enhances the consistency of the generated data in terms of column structure standardization, statistical distribution rationality, and semantic level, greatly improving the matching degree with the original table data, and providing solid support for achieving high-quality, high-fidelity conditional table data generation.

[0024] It should be noted that the encoding network layer preferably employs an encoder based on the Transformer architecture. The conditional vector representation is obtained by average pooling the encoded feature sequence.

[0025] Further, the denoising sampling based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network to obtain the initial synthesized table data includes: Based on the original noise data and a preset mixed-type random sampler, noise sampling is performed to obtain the current noise data; Iterative conditional denoising is performed based on the current noise data, wherein the following steps are executed in each iteration of conditional denoising: The current noise data and conditional vector representation are input into the conditional denoising subnetwork so that the conditional denoising subnetwork performs denoising prediction based on the current noise data and conditional vector representation to obtain the current denoising update amount. Update the current noise data based on the current denoising update amount; Repeat the above steps until the preset iteration condition for denoising is reached, and then use the current noise data as the initial synthesized table data.

[0026] In the above scheme, noise sampling is first performed based on the original noise data and a preset mixed-type random sampler to obtain the current noise data. Then, iterative conditional denoising is carried out. Each time, the current noise data and the conditional vector representation are input into the pre-trained conditional denoising sub-network to obtain and update the current noise data based on the current denoising update amount, until the preset number of iterations of conditional denoising is reached to obtain the initial synthesized table data. The continuous guidance of the conditional vector ensures that each denoising step fits the implicit relationship and structural constraints. Multiple iterations effectively fuse the conditional semantics with the original data distribution, restore the correlation and type features between columns, and generate a synthesized table with reasonable structure and consistent semantics. This significantly improves the stability and controllability of the generation process, enhances the fidelity and usability of the synthesized data in complex relationship and multi-type mixed scenarios, and provides reliable technical support for high-quality table data enhancement.

[0027] It should be noted that the conditional denoising sub-network is built on the Transformer architecture, which injects the conditional vector representation into the denoising process through a cross-attention mechanism. The denoising prediction outputs noise prediction values ​​and original data class probability distributions for numerical features and classification features, respectively. The preset hybrid-type random sampler is used to correct accumulated decoding errors during the sampling process, for example, by resampling the mask state to correct the generation bias of classification features.

[0028] It should be further explained that the tabular data augmentation method based on the conditional inference diffusion model provided by this invention may also include an ensemble optimization step. This step uses a feedback loop mechanism to feed the initial synthetic tabular data or target tabular data generated by the conditional diffusion network back to the chained inference network as new input, thereby initiating a new round of implicit relation extraction and table evolution. The parameters of the conditional inference diffusion model are adjusted using an optimization algorithm until the convergence condition is met. The optimization algorithm combines end-to-end training and reinforcement learning. For example, by defining a reward function based on the accuracy of implicit relation extraction and the correlation error of the generated data columns, a proximal policy optimization algorithm is used to adjust the operation selection strategy in the chained inference network, thereby achieving collaborative optimization and performance improvement of the entire framework.

[0029] This invention also provides a tabular data augmentation system based on a conditional inference diffusion model, comprising: The data acquisition module is used to acquire raw table data and preset condition constraints; The model building module is used to build a conditional reasoning diffusion model consisting of a chain reasoning network and a conditional diffusion network. The implicit relation extraction module is used to input the original table data and preset conditional constraints into the conditional reasoning diffusion model, so that the chain reasoning network can extract implicit relations based on the original table data, preset conditional constraints and preset large language model to obtain intermediate table data, wherein the intermediate table data includes implicit relation information; The conditional diffusion module is used to input the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network performs data augmentation on the original table data based on the intermediate table data to obtain the target table data.

[0030] The above solution achieves end-to-end intelligent processing from data input to augmented output by constructing a multi-module collaborative table data augmentation system based on a conditional inference diffusion model. The system relies on a data acquisition module to accurately acquire the original table data and preset constraints. A model construction module builds a conditional inference diffusion model containing a chain inference network and a conditional diffusion network, ensuring a clear and logically rigorous overall architecture. The implicit relation extraction module uses a preset large language model to drive the chain inference network, efficiently extracting deep implicit relations from the table and generating intermediate table data containing relational information, providing accurate semantic guidance for subsequent augmentation. The conditional diffusion module uses the intermediate table data as constraints to drive the conditional diffusion network to perform targeted augmentation on the original table data, ultimately obtaining standardized and reliable target table data. This modular architecture significantly improves the targeting, structural fidelity, and semantic consistency of table data augmentation, effectively solving the problem of unstable generation results in traditional methods. It also possesses good scalability, adapting to complex application scenarios such as data scarcity and privacy sensitivity, providing efficient and stable data support for various intelligent table data processing tasks. Attached Figure Description

[0031] Figure 1 A schematic diagram of a table data augmentation method based on a conditional inference diffusion model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a tabular data augmentation system architecture based on a conditional inference diffusion model, provided as an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1 This embodiment provides a table data augmentation method based on a conditional inference diffusion model, including the following steps: Step S1: Obtain the original table data and preset condition constraints; Step S2: Construct a conditional reasoning diffusion model consisting of a chain-based reasoning network and a conditional diffusion network; Step S3: Input the original table data and preset conditional constraints into the conditional reasoning diffusion model, so that the chain reasoning network can extract implicit relations based on the original table data, preset conditional constraints and preset large language model to obtain intermediate table data; the intermediate table data includes implicit relation information; Step S4: Input the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network uses the intermediate table data as a condition to perform data augmentation on the original table data to obtain the target table data.

[0034] In this embodiment, by constructing a conditional inference diffusion model composed of a chain-based inference network and a conditional diffusion network, data augmentation can be directly performed on the original table data and preset conditional constraints, thereby improving the targeting, data fidelity, and semantic consistency of the data augmentation process. The chain-based inference network, combined with a preset large language model, extracts implicit relationships based on the original table data and preset conditional constraints, generating intermediate table data including implicit relationship information. The conditional diffusion network uses the intermediate table data as a condition to augment the original table data to obtain the target table data, ensuring that the structure and association of the augmented data are highly consistent with the original table data. This effectively improves the problems of insufficient generation quality and weak association preservation in complex table scenarios of traditional generation models, and can achieve high-quality table data expansion in data-scarce and privacy-sensitive scenarios.

[0035] In one embodiment, a method for augmenting tabular data based on a conditional inference diffusion model is provided. First, original tabular data and preset conditional constraints are obtained. The original table may contain a mixture of numerical and categorical features, and the conditional constraints guide subsequent inference and generation processes. Next, a conditional inference diffusion model composed of a chain-based inference network and a conditional diffusion network is constructed, and the original tabular data and conditional constraints are input into this model. The chain-based inference network, based on a preset large language model, iteratively generates operation chains and transforms the original table, gradually evolving into intermediate table sequences and extracting implicit relationship information from them. The intermediate table data carries these structured implicit relationships. Subsequently, the conditional diffusion network uses the intermediate table data as conditions to augment and generate the original tabular data. The conditional diffusion network employs feature-specific learnable noise scheduling to jointly diffuse and denoise numerical features and categorical features, ensuring that the generated synthetic data conforms to the conditional constraints.

[0036] It should be noted that, in this embodiment, the intermediate table serves as a condition vector, which is injected into the denoising network through a cross-attention mechanism, thereby making the generation process explicitly guided by implicit relationships. Specifically, in its internal Transformer decoder layer, the denoising network treats the representation of noisy data as a query and the representation of the condition vector as a key and value, so that each step of the denoising process can "pay attention" to the table structure information extracted by the chain inference network.

[0037] It should be noted that this embodiment also supports a feedback iteration mechanism, which re-inputs the generated synthetic data into the chain inference network to further evolve the table and optimize the understanding of implicit relationships. Through multiple iterations, errors are reduced and accuracy is improved. The final output is enhanced table data and its implicit relationship analysis results, including a synthetic sample set, inter-column correlation indicators, and semantic interpretations, achieving high-quality, faithful, and semantically consistent table data expansion.

[0038] Further, the original table data and preset constraints are input into the conditional inference diffusion model, so that the chain inference network extracts implicit relationships based on the original table data, preset constraints, and preset large language model to obtain intermediate table data, including: In the chained inference network, the following steps are performed: Initialize the current table state and current historical operation records based on the original table data and preset condition constraints; Based on the current table state, current historical operation record, preset condition constraints and preset large language model, iterative reasoning and evolution are performed until the preset termination condition is met, and then the final table state and complete historical operation record are output. Based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm, implicit relationship parsing is performed to obtain implicit relationship information; The intermediate table data is obtained based on the implicit relationship information and the final table state.

[0039] In this embodiment, the iterative reasoning evolution mechanism of the chain-based reasoning network enables automated and accurate extraction of complex implicit relationships in tables. First, the current table state and current historical operation records are initialized based on the original table data and preset constraints. Then, relying on a preset large language model, the iterative reasoning evolution is completed to a preset termination condition, combining the current table state, current historical operation records, and preset constraints. Finally, implicit relationship parsing is performed using a preset structured relationship extraction algorithm, outputting intermediate table data rich in implicit relationship information. This improves the accuracy and reliability of implicit relationship extraction, adapts to complex scenarios such as inter-column dependencies, and effectively overcomes the limitations of traditional methods when data is sparse.

[0040] In one embodiment, the chained inference network performs implicit relation extraction based on a preset large language model. The process includes: first, defining an operation pool containing multiple atomic operations, the operation pool including f_add_column() for adding a new column to a table, f_select_row() for filtering relevant row subsets based on conditions, f_select_column() for selecting and querying relevant column subsets, f_group_by() for grouping by column and performing aggregation operations, and f_sort_by() for sorting rows by column value. Based on the original table data and preset constraints, initialize the current table state and the current historical operation record; The current table status, historical operation records, and conditional constraints are input into the preset large language model. The next operation and its parameters are sampled through the context learning mechanism. For example, for the query "which country has the most top three cyclists", the preset large language model can sample the f_add_column() operation to extract the country column.

[0041] The selected operation is then executed programmatically to update the table, forming an intermediate table and recording the operation history. The operation history is a sequence of operations and their parameters recorded in the execution order, and its structure is represented as [(op_1,args_1),(op_2,args_2),…,(op_n,args_n)], where op_i represents the selected operation and args_i represents the corresponding parameter. The sampling and execution steps described above are repeated iteratively until a preset termination condition is met, such as the table being simplified to a single row containing the core relationship or the model output ending marker, to obtain the final table state and complete operation history; finally, based on the final table state and operation history, implicit relationship information is obtained by parsing through a structured relationship extraction algorithm, and intermediate table data is generated accordingly.

[0042] In this embodiment, the operation history records all steps in the reasoning process, making the evolution of the table state traceable, thereby supporting the structured parsing of implicit relationships and guiding the subsequent conditional diffusion process. Through the above iterative reasoning evolution mechanism, the chain-based reasoning network can achieve automated and accurate extraction of complex implicit relationships in the table, improving its adaptability and reliability in scenarios with sparse data or complex inter-column dependencies.

[0043] Furthermore, the iterative reasoning and evolution based on the current table state, current historical operation records, preset condition constraints, and preset large language model, until a preset termination condition is met, and then outputting the final table state and complete historical operation records, includes: Perform iterative reasoning and evolution steps: Dynamic prompts are generated based on the current table status, current historical operation records, and preset condition constraints. The dynamic prompt information is input into a preset large language model, so that the preset large language model can perform inference and evolution based on the dynamic prompt information to obtain the current table operation instruction and the corresponding current operation parameters. Update the current table status based on the current table operation instruction; update the current historical operation record based on the current table operation instruction and the corresponding current operation parameters; Repeat the above iterative reasoning and evolution steps until the preset number of iterative reasoning and evolution steps are reached. Then, output the current table state and the current historical operation record as the final table state and the complete historical operation record.

[0044] In this embodiment, dynamic prompts are constructed based on the current table state, current historical operation records, and preset conditional constraints. This ensures that each step of reasoning is highly consistent with the current context and preset conditional constraints, significantly improving the relevance and accuracy of the generated table operation instructions and parameters. Through iterative reasoning evolution until a preset number of iterations is reached, the complex implicit relationships in the original table can be gradually extracted and made explicit. This ensures the integrity of the reasoning process and avoids computational redundancy caused by infinite loops. This embodiment enhances the adaptability, interpretability, and stability of chained reasoning, providing a structurally clear and semantically explicit final table state and complete historical operation records for the subsequent conditional diffusion stage, thereby supporting high-quality, strongly consistent table data generation and enhancement tasks.

[0045] In one embodiment, the chain-based reasoning network iteratively evolves based on the current table state, historical operation records, and preset conditional constraints. The process specifically includes: in each iteration, firstly, constructing dynamic prompt information, which includes the currently evolved table, user query or conditional constraints, executed operation history, task description and operation definition, and several high-quality few-sample instances; inputting the dynamic prompt information into a preset large language model; the large language model, based on a context learning mechanism, analyzes the current table content, query intent, and executed steps, and performs analogical reasoning with the provided examples, thereby generating the next operation instruction and its corresponding parameters in an autoregressive manner; subsequently, the current table state is updated programmatically according to the generated operation instruction and parameters, and the operation and parameters are appended to the historical operation record, forming an updated operation history sequence; repeating the above steps of constructing prompts, model reasoning, state updating, and record appending until a preset number of iterations is reached or other termination conditions are met (such as the table being simplified to a single row containing core relationships or the model output ending marker), finally outputting the final table state and complete operation history.

[0046] In this embodiment, the construction of the dynamic prompts integrates the current state, task definition, and example guidance, enabling the large language model to fully combine context and prior knowledge in each step of reasoning, achieving dynamic planning and adaptive sampling of the operation chain. During the generation process, a greedy search can be used to directly select the output with the highest probability, or a more complex sampling strategy can be used to balance diversity and accuracy. Through the above iterative mechanism, the chain-like reasoning network can gradually extract and reveal the complex implicit relationships in the table, ensuring the integrity of reasoning while avoiding invalid loops, thereby providing structurally clear and semantically explicit conditional inputs for the subsequent conditional diffusion stage.

[0047] Further, the step of repeating the above iterative reasoning and evolution steps until a preset number of iterative reasoning and evolution steps are reached, and then outputting the current table state and the current historical operation record as the final table state and the complete historical operation record, includes: During the repeated iterative reasoning and evolution steps described above, the current table row number and the current key column structure are obtained based on the current table state. If the current table row count and the current key column structure meet the preset constraints, or reach the preset number of iterations and inferences, then the current table state and the current historical operation record will be output as the final table state and the complete historical operation record.

[0048] In this embodiment, during the iterative reasoning evolution steps, the current table row number and key column structure are obtained in real time based on the current table state. These are used as the criteria for determining iteration termination. Evolution can be terminated early when preset constraints are met, effectively avoiding invalid iterations and reducing computational overhead. Simultaneously, a preset number of iterative reasoning evolutions serves as a safety net to prevent reasoning from getting stuck in a loop, ensuring the integrity and usability of the final output table state and complete historical operation records. This embodiment enables the iterative process of chained reasoning to have adaptive judgment capabilities, adapting to table data with different characteristics, improving the response efficiency and stability of the reasoning process, and thus laying a reliable foundation for high-quality table data enhancement.

[0049] Further, the step of repeating the above iterative reasoning and evolution steps until a preset number of iterative reasoning and evolution steps are reached, and then outputting the current table state and the current historical operation record as the final table state and the complete historical operation record, includes: During the repeated iterative reasoning evolution steps, the preset large language model is subjected to end marker detection. If the preset large language model outputs a preset reasoning end marker, or reaches a preset number of iterative reasoning evolutions, the current table state and the current historical operation record are used as the final table state and the complete historical operation record and output.

[0050] In this embodiment, by introducing an end-of-inference marker detection mechanism based on a preset large language model during the iterative reasoning evolution process, the adaptability and controllability of the chained reasoning process are significantly improved. By performing end-of-inference marker detection on the preset large language model, the evolution terminates when a preset reasoning end-of-inference marker is detected, allowing the reasoning process to autonomously determine its completion status, aligning with actual reasoning logic. Combined with the hard constraint of the preset number of iterative reasoning evolutions, a multi-layered termination strategy is formed, effectively avoiding redundant reasoning, improving evolution efficiency, enhancing stability and generalization ability for tables of varying complexity, and ensuring that the final output table state and complete historical operation records are intact and concise.

[0051] Furthermore, in the process of parsing implicit relationships based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm to obtain implicit relationship information, the preset structured relationship extraction algorithm specifically includes: Based on the final table state, statistical correlation indicators are calculated to obtain the statistical relationships between columns; Based on the complete historical operation records, semantic relationships between columns are obtained through parsing and inference. The implicit relationship information is obtained by performing a structured representation based on the statistical and semantic relationships between columns.

[0052] In this embodiment, statistical correlation indicators are calculated based on the final table state, capturing statistical relationships between columns from the data distribution level, such as calculating the Pearson correlation coefficient between columns. Simultaneously, analytical inference is performed using complete historical operation records to mine semantic relationships between columns from the operational logic level. The two are then fused through structured representation to obtain implicit relationship information. This ensures that the output implicit relationship information possesses both statistical reliability and semantic interpretability, effectively identifying complex implicit relationships in the table and enhancing the understanding of the table's deep structure.

[0053] Further, the step of inputting the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network uses the intermediate table data as a condition to perform data augmentation on the original table data to obtain the target table data, includes: In the conditional diffusion network, the following steps are performed: Encoding is performed based on the intermediate table data and the preset condition encoder to obtain a condition vector representation; Noise scheduling is performed based on the original table data to obtain the original noise data; Denoising sampling is performed based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network to obtain the initial synthesized table data; The target table data is obtained by post-processing the initial synthesized table data.

[0054] In this embodiment, the preset conditional encoder encodes the intermediate table data to obtain a conditional vector representation, ensuring that the generation process is always semantically constrained by implicit relationships. Noise scheduling is performed on the original table data to obtain original noisy data, effectively preserving the inherent data characteristics of the original table. A pre-trained conditional denoising sub-network combines the conditional vector representation with the original noisy data to complete denoising sampling, obtaining initial synthesized table data, which is then post-processed to obtain the target table data. This embodiment significantly improves the fidelity, semantic consistency, and structural integrity of the generated data, effectively addressing real-world scenarios such as data sparsity, and enhancing its practicality and reliability in table data expansion and privacy protection.

[0055] In one embodiment, the conditional diffusion network augments the original table data using the intermediate table data as a condition, specifically including: for the numerical features in the original table, forward noise is added using a variance explosion stochastic differential equation, in the form of... .

[0056] It should be noted that, among them This represents the original, clean tabular data sample taken from the real data distribution, which is the starting point of the diffusion process; Indicates at time step The data sample contaminated by noise describes the intermediate state of the data as it evolves from its original state to a noisy state; For continuous-time variables, t=0 represents the original data state, and t=1 represents the pure noise state. The noise vector sampled from the standard Gaussian distribution is used to randomly perturb the numerical features. Its mean is 0, and its covariance matrix is ​​the identity matrix I, representing isotropic Gaussian noise. A feature-specific learnable noise scheduling function is used to control the time step. The noise intensity added to the time-varying data, and its specific functional form, are optimized and adjusted according to the data characteristics during model training.

[0057] For the categorical features in the original table, they are converted into one-hot vectors and masked classes are appended. Then, a classification diffusion process is applied. .

[0058] It should be noted that, among them This represents a class distribution whose class probabilities are represented by a vector. The definition describes the probability of each discrete class being sampled given parameters; For the mask vector, Regarding time The strictly decreasing mask scheduling function describes the probability of each discrete class being sampled given parameters; Indicates from raw data To noise data The conditional probability distribution of the forward diffusion process (noise-adding process).

[0059] The intermediate table output by the chained inference network is used as condition c and input into the denoising network. .

[0060] It should be noted that the denoising network mentioned above... To enable neural networks to handle mixed data types, where This represents the complete set of trainable parameters constituting the neural network, including the weight matrix and bias vector, which are determined by optimizing the loss function on the training data; the denoising network The goal is: for the numerical feature portion, the prediction is applied to... The noise ε; for the classification feature part, predict the original data. The category probability distribution.

[0061] The denoising network fuses the semantic information of the intermediate table through a conditional injection mechanism. The specific process is as follows: First, the encoder processes the intermediate table... Encoded as a fixed-dimensional conditional vector The network input includes noisy data. The current time step encoded as a time embedding vector and condition vector Conditional injection is achieved through a cross-attention mechanism, that is... The representation of the query, the condition vector The representation is used as key and value, so that each step of the denoising process can be aware of conditional information; the network output layer is designed to process numerical and categorical features simultaneously, predicting noise ε for the numerical part and predicting the original data for the categorical part. The category probability distribution is used to generate a synthetic table data sample with the same structure as the original table, enhanced content, and conforming to implicit relationships.

[0062] Furthermore, the preset condition encoder includes an embedding layer and an encoding network layer; the encoding based on the intermediate table data and the preset condition encoder to obtain a condition vector representation includes: Based on the intermediate table data, obtain cell content, column identifiers, and type information; The cell content, column identifier, and type information are input into a preset condition encoder so that the embedding layer performs feature mapping based on the cell content, column identifier, and type information to obtain continuous feature embedding. The continuous feature embeddings are input into the encoding network layer, so that the encoder network layer performs fusion modeling based on the continuous feature embeddings to obtain the encoded feature sequence; Pooling is performed based on the encoded feature sequence to obtain the conditional vector representation.

[0063] In this embodiment, the preset conditional encoder extracts cell content, column identifiers, and type information from the intermediate table data as encoding input, comprehensively capturing the semantic, structural, and type features of the table. The embedding layer performs feature mapping on the above information to obtain continuous feature embeddings, and the encoding network layer performs fusion modeling based on the continuous feature embeddings to obtain an encoded feature sequence. Finally, a compact and information-rich conditional vector representation is generated through pooling operations. This embodiment significantly improves the expression quality and completeness of conditional information, ensuring that the generation process is fully controlled by the implicit relationships extracted by chain reasoning, enhancing the consistency of the generated data with the original table in terms of column structure, statistical distribution, and semantic consistency, and laying a reliable foundation for the generation of high-quality, high-fidelity conditional table data.

[0064] In one embodiment, the preset condition encoder is implemented based on the Transformer Encoder architecture and is used to encode the intermediate table output by the chained inference network into a conditional vector representation usable by the diffusion model. The specific steps include: first, obtaining the intermediate table. .

[0065] It should be noted that, among them This represents the intermediate table representing the output of the chained inference network in the k-th iteration; These represent row index and column index respectively. For the position located at the Line number The content of the cells in the column.

[0066] Embedding functions in tables Cell content, column structure, and type information are mapped to continuous vector representations to form an initial embedding sequence; this embedding sequence is then input into a... The encoding network layer consists of Transformer Encoder layers.

[0067] It should be noted that, among them The total number of encoding network layers is a positive integer hyperparameter used to control the depth and capacity of the model in modeling tabular data.

[0068] By fusing multi-layer self-attention and feedforward networks to model the encoded feature sequence, we obtain the encoded feature sequence. .

[0069] It should be noted that N is the number of encoded tokens, corresponding to the length of the table embedding sequence; The hidden layer dimension is the size of the vector space to which each token is mapped. Represents the encoded feature sequence Let be an N-row, d-column matrix over the real number field.

[0070] Finally, for Perform average pooling along the token dimension to obtain the final conditional vector representation. This serves as the external condition input for the subsequent conditional diffusion denoising network.

[0071] It should be noted that c is a fixed-dimensional conditional vector representation obtained after pooling, which integrates all the key information from the intermediate table; This represents the pooling function, which is an average pooling operation along the length of the sequence.

[0072] In this embodiment, the table embedding function By integrating representation learning of cell content, column identifiers, and feature types, it can fully preserve the structured and semantic information of the table. The encoding network models the global dependencies of the table through a multi-layer Transformer, while the pooling operation generates compact and representative condition vectors, ensuring that the diffusion generation process is subject to explicit constraints of the implicit relationships extracted by chain reasoning.

[0073] The injection of the conditional vector into the denoising network is achieved through a cross-attention mechanism. Let the noise table embedding at the current time step be represented as... The conditional coding output is In the l In the Transformer Decoder layer, cross-attention is calculated as follows: It should be noted that Q is the query matrix, which is embedded from the noise table of the current layer. After weight matrix Obtained by linear transformation, i.e. K is the key matrix, output by conditional encoding. After weight matrix Obtained by linear transformation, i.e. V is the value matrix, output by conditional coding. After weight matrix Obtained by linear transformation, i.e. ; This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to an excessively large dot product result; This is a normalized exponential function used to convert attention scores into a probability distribution; superscript This represents the matrix transpose operation.

[0074] The cross-attention mechanism enables the denoising network to dynamically focus on the table structure information encoded by the condition vector at each generation step, thereby ensuring that the synthesized data remains consistent with the chain inference results in terms of inter-column correlation and semantic relationships.

[0075] Further, the denoising sampling based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network to obtain the initial synthesized table data includes: Based on the original noise data and a preset mixed-type random sampler, noise sampling is performed to obtain the current noise data; Iterative conditional denoising is performed based on the current noise data, wherein the following steps are executed in each iteration of conditional denoising: The current noise data and conditional vector representation are input into the conditional denoising subnetwork so that the conditional denoising subnetwork performs denoising prediction based on the current noise data and conditional vector representation to obtain the current denoising update amount. Update the current noise data based on the current denoising update amount; Repeat the above steps until the preset iteration condition for denoising is reached, and then use the current noise data as the initial synthesized table data.

[0076] In this embodiment, noise sampling is first performed based on the original noise data and a preset mixed-type random sampler to obtain the current noise data. Then, iterative conditional denoising is carried out. Each time, the current noise data and the conditional vector representation are input into the pre-trained conditional denoising sub-network to obtain and update the current noise data based on the current denoising update amount, until the preset number of iterative conditional denoising iterations is reached to obtain the initial synthesized table data. The continuous guidance of the conditional vector ensures that each denoising step conforms to the implicit relationships and structural constraints. Multiple iterations effectively fuse the conditional semantics with the original data distribution, restore the correlation and type features between columns, and generate a synthesized table with reasonable structure and consistent semantics. This significantly improves the stability and controllability of the generation process, enhances the fidelity and usability of the synthesized data in complex relationships and multi-type mixed scenarios, and provides reliable technical support for high-quality table data enhancement.

[0077] In one embodiment, the process of denoising and sampling based on conditional vector representation, original noisy data, and a pre-trained conditional denoising sub-network to obtain initial synthetic table data specifically includes: first, inputting the intermediate table output by the chained inference network as condition c into a denoising neural network with parameter θ. The denoising neural network, based on the Transformer architecture, can handle mixed input types of numerical and categorical features. Its goal is to reduce the noise ε applied to numerical feature prediction and reduce the noise ε applied to the original data for categorical feature prediction. The distribution of categories.

[0078] It should be noted that the denoising network described in this embodiment is trained using a continuous-time evidence lower bound loss function. To optimize the reverse diffusion process.

[0079] It should be noted that, The mathematical expectation operator is used to approximate the expected value by averaging small batches of samples during training. This represents the square of the L2 norm, which is the sum of the squares of the elements of the vector, and is used to calculate the Euclidean distance between the predicted noise and the actual noise.

[0080] During the sampling phase, a hybrid-type random sampler is used for progressive denoising. Specifically, noise is sampled based on the original noise data and a preset hybrid-type random sampler to obtain the current noise data; then iterative conditional denoising is performed: in each iteration, the current noise data and the conditional vector representation cc are input into a pre-trained conditional denoising subnetwork. The conditional denoising subnetwork performs denoising prediction based on the current input and outputs the current denoising update amount; it updates the current noise data according to the update amount; it repeats this iterative step until the preset number of denoising iterations is reached, and finally outputs the updated noise data as the initial synthesized table data.

[0081] In this embodiment, the sampling process introduces a classifier-free guidance mechanism, and its sampling probability is determined by... Decision, among which The intensity hyperparameter is used to guide the trade-off between the diversity of generated samples and the degree of adherence to the conditional constraints.

[0082] Furthermore, the hybrid random sampler possesses automatic correction capabilities, such as resampling the mask state to correct decoding biases that may accumulate in the denoising process of classification features. Through this method, the conditional vector provides continuous guidance during the iterative denoising process, ensuring that each update aligns with the implicit relationships and structural constraints extracted by chained inference, thereby generating synthetic data that highly matches the original table in terms of column correlation, type features, and semantic consistency.

[0083] Please see Figure 2 This embodiment also provides a tabular data augmentation system based on a conditional inference diffusion model, including: The data acquisition module is used to acquire raw table data and preset condition constraints; The model building module is used to build a conditional reasoning diffusion model consisting of a chain reasoning network and a conditional diffusion network. The implicit relation extraction module is used to input the original table data and preset conditional constraints into the conditional reasoning diffusion model, so that the chain reasoning network can extract implicit relations based on the original table data, preset conditional constraints and preset large language model to obtain intermediate table data, wherein the intermediate table data includes implicit relation information; The conditional diffusion module is used to input the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network performs data augmentation on the original table data based on the intermediate table data to obtain the target table data.

[0084] In this embodiment, a multi-module collaborative conditional reasoning diffusion model table data augmentation system is constructed, achieving high-quality table data augmentation guided by semantic understanding. The data acquisition module obtains the original table data and preset constraints. The model construction module builds a conditional reasoning diffusion model composed of a chain-based reasoning network and a conditional diffusion network. The implicit relation extraction module drives the chain-based reasoning network to extract implicit relations, obtaining intermediate table data containing implicit relation information. The conditional diffusion module uses this intermediate table data as a condition to augment the original table data through the conditional diffusion network, obtaining the target table data. This achieves end-to-end automated processing, improving the targeting, fidelity, and semantic consistency of data augmentation, effectively addressing the generation quality issues of traditional generative models, adapting to data-scarce and privacy-sensitive scenarios, and possessing good scalability due to its modular architecture, providing efficient and reliable data support for various table data processing tasks.

[0085] In one embodiment, a tabular data augmentation system based on a conditional inference diffusion model is provided, which further includes an integrated optimization module. The integrated optimization module works during the operation or training phase of the tabular data augmentation system based on the conditional inference diffusion model, and is used to collaboratively optimize the overall performance of the chain inference network and the conditional diffusion network through a mechanism that combines end-to-end training and reinforcement learning.

[0086] The workflow of this integrated optimization module is aligned with that of the system, specifically including: first, initializing system parameters, including the diffusion model parameters θ and the prompt word templates of the large language model; then, during the training phase, optimizing the conditional diffusion network using the evidence lower bound loss, and simultaneously calculating the reward function for the chain inference process, which is composed of the accuracy of implicit relation extraction and the column correlation error of the generated data; based on this, dynamically adjusting the probability strategy of operation selection in chain inference through reinforcement learning algorithms to maximize cumulative rewards, thereby incentivizing operations that can improve the quality of conditional generation.

[0087] It should be noted that the reinforcement learning process described in this embodiment is formalized as a Markov decision process. At step t, the system state is... From the current intermediate table Operation history The query or conditional constraint Q are jointly represented, that is... ,in The network is a state-encoding network; the action space A corresponds to the operation pool defined by the chained reasoning network; the reward function R extracts the accuracy reward from the implicit relation. Correlation error penalty Weighted combination, i.e. ,in , , The Pearson correlation coefficient is used. This is the set of key column pairs.

[0088] The tabular data augmentation system based on a conditional inference diffusion model described in this embodiment also includes a feedback loop mechanism managed by the integrated optimization module: in each iteration, the synthetic data generated by the conditional diffusion network is fed back to the chain inference network as new input, updating the intermediate table and retraining the diffusion model accordingly. The triggering conditions for this feedback loop include the inference evaluation index falling below a preset threshold, or the column correlation error between the generated data and the conditional table exceeding a threshold; its termination conditions include the correlation error change being less than a set threshold in multiple consecutive iterations, or reaching a preset maximum number of iterations. Finally, the stability of the optimization process is ensured by monitoring indicators such as KL (relative entropy) divergence and Pearson correlation coefficient. This module achieves deep synergy and iterative self-reinforcement between chain inference and conditional diffusion within the system through the above methods.

[0089] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for augmenting tabular data based on a conditional reasoning diffusion model, characterized in that, Includes the following steps: Retrieve raw table data and preset constraints; Construct a conditional reasoning diffusion model consisting of a chain reasoning network and a conditional diffusion network; The original table data and preset constraints are input into the conditional reasoning diffusion model, so that the chain reasoning network extracts implicit relationships based on the original table data, preset constraints, and preset large language model to obtain intermediate table data; the intermediate table data includes implicit relationship information. The intermediate table data and the original table data are input into the conditional inference diffusion model, so that the conditional diffusion network performs data augmentation on the original table data based on the intermediate table data to obtain the target table data.

2. The table data augmentation method based on a conditional inference diffusion model according to claim 1, characterized in that, The original table data and preset constraints are input into the conditional inference diffusion model, so that the chain inference network extracts implicit relationships based on the original table data, preset constraints, and preset large language model to obtain intermediate table data, including: In the chained inference network, the following steps are performed: Initialize the current table state and current historical operation records based on the original table data and preset condition constraints; Based on the current table state, current historical operation record, preset condition constraints and preset large language model, iterative reasoning and evolution are performed until the preset termination condition is met, and then the final table state and complete historical operation record are output. Based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm, implicit relationship parsing is performed to obtain implicit relationship information; The intermediate table data is obtained based on the implicit relationship information and the final table state.

3. The tabular data augmentation method based on a conditional inference diffusion model according to claim 2, characterized in that, The iterative reasoning and evolution based on the current table state, current historical operation records, preset condition constraints, and preset large language model continues until a preset termination condition is met, at which point the final table state and complete historical operation records are output, including: Perform iterative reasoning and evolution steps: Dynamic prompts are generated based on the current table status, current historical operation records, and preset condition constraints. The dynamic prompt information is input into a preset large language model, so that the preset large language model can perform inference and evolution based on the dynamic prompt information to obtain the current table operation instruction and the corresponding current operation parameters. Update the current table status based on the current table operation instruction; update the current historical operation record based on the current table operation instruction and the corresponding current operation parameters; Repeat the above iterative reasoning and evolution steps until the preset number of iterative reasoning and evolution steps are reached. Then, output the current table state and the current historical operation record as the final table state and the complete historical operation record.

4. The tabular data augmentation method based on a conditional inference diffusion model according to claim 3, characterized in that, The iterative reasoning and evolution steps described above are repeated until a preset number of iterative reasoning and evolution steps are reached. Then, the current table state and the current historical operation record are output as the final table state and complete historical operation record, including: During the repeated iterative reasoning and evolution steps described above, the current table row number and the current key column structure are obtained based on the current table state. If the current table row count and the current key column structure meet the preset constraints, or reach the preset number of iterations and inferences, then the current table state and the current historical operation record will be output as the final table state and the complete historical operation record.

5. A method for table data augmentation based on a conditional inference diffusion model according to claim 3, characterized in that, The iterative reasoning and evolution steps described above are repeated until a preset number of iterative reasoning and evolution steps are reached. Then, the current table state and the current historical operation record are output as the final table state and complete historical operation record, including: During the repeated iterative reasoning evolution steps, the preset large language model is subjected to end marker detection. If the preset large language model outputs a preset reasoning end marker, or reaches a preset number of iterative reasoning evolutions, the current table state and the current historical operation record are used as the final table state and the complete historical operation record and output.

6. The table data augmentation method based on a conditional inference diffusion model according to claim 2, characterized in that, In the process of parsing implicit relationships based on the final table state, complete historical operation records, and a preset structured relationship extraction algorithm to obtain implicit relationship information, the preset structured relationship extraction algorithm specifically includes: Based on the final table state, statistical correlation indicators are calculated to obtain the statistical relationships between columns; Based on the complete historical operation records, semantic relationships between columns are obtained through parsing and inference. The implicit relationship information is obtained by performing a structured representation based on the statistical and semantic relationships between columns.

7. The table data augmentation method based on a conditional inference diffusion model according to claim 1, characterized in that, The step of inputting the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network uses the intermediate table data as a condition to perform data augmentation on the original table data to obtain the target table data, includes: In the conditional diffusion network, the following steps are performed: Encoding is performed based on the intermediate table data and the preset condition encoder to obtain a condition vector representation; Noise scheduling is performed based on the original table data to obtain the original noise data; Denoising sampling is performed based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network to obtain the initial synthesized table data; The target table data is obtained by post-processing the initial synthesized table data.

8. A method for enhancing tabular data based on a conditional inference diffusion model according to claim 7, characterized in that, The preset condition encoder includes an embedding layer and a coding network layer; The process of encoding based on the intermediate table data and a preset condition encoder to obtain a condition vector representation includes: Based on the intermediate table data, obtain cell content, column identifiers, and type information; The cell content, column identifier, and type information are input into a preset condition encoder so that the embedding layer performs feature mapping based on the cell content, column identifier, and type information to obtain continuous feature embedding. The continuous feature embeddings are input into the encoding network layer, so that the encoder network layer performs fusion modeling based on the continuous feature embeddings to obtain the encoded feature sequence; Pooling is performed based on the encoded feature sequence to obtain the conditional vector representation.

9. A method for table data augmentation based on a conditional inference diffusion model according to claim 7, characterized in that, The denoising sampling based on the conditional vector representation, the original noise data, and the pre-trained conditional denoising sub-network yields the initial synthesized table data, including: Based on the original noise data and a preset mixed-type random sampler, noise sampling is performed to obtain the current noise data; Iterative conditional denoising is performed based on the current noise data, wherein the following steps are executed in each iteration of conditional denoising: The current noise data and conditional vector representation are input into the conditional denoising subnetwork so that the conditional denoising subnetwork performs denoising prediction based on the current noise data and conditional vector representation to obtain the current denoising update amount. Update the current noise data based on the current denoising update amount; Repeat the above steps until the preset iteration condition for denoising is reached, and then use the current noise data as the initial synthesized table data.

10. A tabular data augmentation system based on a conditional reasoning diffusion model, characterized in that, include: The data acquisition module is used to acquire raw table data and preset condition constraints; The model building module is used to build a conditional reasoning diffusion model consisting of a chain reasoning network and a conditional diffusion network. The implicit relation extraction module is used to input the original table data and preset conditional constraints into the conditional reasoning diffusion model, so that the chain reasoning network can extract implicit relations based on the original table data, preset conditional constraints and preset large language model to obtain intermediate table data, wherein the intermediate table data includes implicit relation information; The conditional diffusion module is used to input the intermediate table data and the original table data into the conditional inference diffusion model, so that the conditional diffusion network performs data augmentation on the original table data based on the intermediate table data to obtain the target table data.

Images