Robust network reconstruction search method and device based on large language model
By combining a large language model with sparse constraints and structural prior constraints, this network reconstruction method solves the problems of poor adaptability and reliance on expert experience in existing technologies, achieving efficient and accurate reconstruction of complex networks and providing interpretable results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-16
Smart Images
Figure CN122221418A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of complex network reconstruction technology, specifically to a robust network reconstruction search method and apparatus based on a large language model. Background Technology
[0002] Complex networks serve as the fundamental framework for describing numerous real-world systems, such as neural networks in the brain, social networks, and gene regulatory networks. However, the true connectivity relationships within these systems are often not directly observable; typically, only multivariate time-series data generated at their nodes can be obtained. The core task of network reconstruction is to infer the underlying connectivity topology from these observed time series data. This is crucial for revealing the system's intrinsic mechanisms and enabling prediction and intervention control. Traditional network reconstruction methods heavily rely on manually pre-designed models and algorithms, often encountering technical problems such as adaptation difficulties and limited inference accuracy when facing unknown or highly complex network dynamics.
[0003] In existing technologies, the following main approaches exist to address the aforementioned problems: One type is regression methods based on fixed models, which assume that network dynamics follow specific parameterized models such as linear dynamical systems, and then directly estimate the adjacency matrix using techniques such as sparse regression. This method performs well when the model assumptions are accurate. Another type is model-free association measurement methods, such as determining the existence of edges by calculating statistical or information-theoretic metrics like transitive entropy or convergent cross-mapping, without explicitly relying on a specific dynamic model. Furthermore, recent explorations in the field of automatic algorithm design have also emerged, such as using large language models combined with evolutionary computation to automatically generate heuristic algorithms, but these works mainly focus on combinatorial optimization problems.
[0004] However, existing technologies still have significant shortcomings. First, methods based on pre-defined models have poor adaptability and struggle to handle unknown or time-varying dynamic characteristics in real-world systems. Second, most methods focus only on minimizing data fitting errors, lacking effective regularization or structural prior constraints to ensure the rationality and uniqueness of solutions. Third, the selection and parameter tuning of existing methods heavily rely on expert experience, lacking an end-to-end automated framework capable of automatically searching and optimizing reconstruction strategies based on data characteristics. Finally, some known structural information that may exist in practice has not yet been effectively integrated to guide the overall reconstruction process. Summary of the Invention
[0005] To address the aforementioned problems in the prior art, this invention provides a robust network reconstruction search method and apparatus based on a large language model.
[0006] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a robust network reconstruction search method based on a large language model, comprising: S101. Obtain multivariate time series data of the social scene to be reconstructed, prior data of network structure, and population of initial dynamic functions; S102. Using multivariate time series data, prior network structure data, and the population of initial dynamic functions, an inference network optimization model is established from the perspectives of data fitting, sparsity constraints, and structural prior constraints. The current optimal predicted network structure corresponding to each initial dynamic function is obtained using the inference network optimization model. The inference network optimization model uses the minimum sum of the corresponding data fitting terms, sparsity constraints, and structural prior constraints as the optimization condition. S103. Use the current optimal prediction network structure to search for the current optimal dynamic function corresponding to the social scene to be reconstructed; S104. Input the current optimal dynamic function into the large language model, and perform a reverse-driven network dynamic function search based on the preset prompt words to obtain the current optimal dynamic function population. Use the current optimal dynamic function population as the initial dynamic function population. S105. Repeat steps S102-S104 until the preset number of iterations is reached. The current optimal predicted network structure in S102 when the preset number of iterations is reached is taken as the final reconstructed social network, and the current optimal dynamic function in S103 is taken as the final dynamic function.
[0007] Secondly, the present invention provides a robust network reconstruction search device based on a large language model, which includes: an acquisition unit, a model building unit, a search unit, a reverse driving unit, and a loop unit. The acquisition unit is used to acquire multivariate time-series data of the social scene to be reconstructed, prior data of the network structure, and the population of initial dynamic functions; The model building unit is used to establish an inference network optimization model from the perspectives of data fitting, sparse constraints, and structural prior constraints using multivariate time series data, prior network structure data, and a population of initial dynamic functions. The inference network optimization model is used to obtain the current optimal predicted network structure for each initial dynamic function. The optimization condition of the inference network optimization model is to minimize the sum of the corresponding data fitting terms, sparse constraints, and structural prior constraints. The search unit is used to search for the current optimal dynamic function corresponding to the social scene to be reconstructed by utilizing the current optimal prediction network structure. The reverse-drive unit is used to input the current optimal dynamic function into the large language model and perform a reverse-drive network dynamic function search based on preset prompt words to obtain the current optimal dynamic function population, and use the current optimal dynamic function population as the initial dynamic function population; The loop unit is used to repeatedly execute the steps of the model building unit, the search unit, and the back-driving unit until a preset number of iterations is reached. The current optimal prediction network structure in the model building unit at the preset number of iterations is used as the final reconstructed social network, and the current optimal dynamic function in the search unit is used as the final dynamic function.
[0008] Thirdly, the present invention provides a robust network reconstruction search device based on a large language model, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the robust network reconstruction search device based on a large language model is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the robust network reconstruction search method based on a large language model as described in the first aspect above.
[0009] This invention provides a robust network reconstruction search method and apparatus based on a large language model. The robust network reconstruction search method based on a large language model includes: S101, acquiring multivariate time-series data of the social scenario to be reconstructed, prior network structure data, and an initial dynamics function population; S102, using the multivariate time-series data, prior network structure data, and the initial dynamics function population, establishing an inference network optimization model from the perspectives of data fitting, sparsity constraints, and structural prior constraints, and using the inference network optimization model to obtain the current optimal predicted network structure corresponding to each initial dynamics function; the inference network optimization model uses the minimum sum of the corresponding data fitting terms, sparsity constraints, and structural prior constraints as the optimization condition. S103. Utilize the current optimal predictive network structure to search for the current optimal dynamic function corresponding to the social scenario to be reconstructed; S104. Input the current optimal dynamic function into the large language model and perform a reverse-driven network dynamic function search based on preset prompts to obtain a population of current optimal dynamic functions, which is then used as the initial dynamic function population; S105. Repeat steps S102-S104 until a preset number of iterations is reached, and use the current optimal predictive network structure in S102 when the preset number of iterations is reached as the final reconstructed social network, and use the current optimal dynamic function in S103 as the final dynamic function. In this invention, firstly, by initializing the dynamic function population and using an iterative optimization framework, combined with the reverse-driven search of the large language model, the network dynamic function is dynamically adjusted and explored, solving the problem of poor adaptability and difficulty in dealing with unknown or time-varying dynamic characteristics in real systems based on preset model methods. Then, data fitting terms, sparse constraint terms, and structural prior constraint terms are explicitly introduced into the inference network optimization model. These multi-faceted optimization conditions ensure the rationality and uniqueness of the solution, addressing the problem that most methods only focus on data fitting errors and lack effective regularization or structural prior constraints. Next, an end-to-end automated loop process automatically searches for and optimizes reconstruction strategies, reducing manual intervention and addressing the issue that existing methods heavily rely on expert experience and lack automated frameworks for parameter selection and tuning. Finally, prior network structure data is integrated into the optimization model, effectively utilizing some known structural information to guide the reconstruction process. This ultimately achieves efficient and accurate reconstruction of social network dynamics functions and social network structures, thereby comprehensively improving the method's practicality and robustness.
[0010] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0011] Figure 1 A flowchart illustrating a robust network reconstruction search method based on a large language model, provided in an embodiment of the present invention; Figure 2A schematic diagram of a robust network reconstruction search device based on a large language model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a robust network reconstruction search device based on a large language model, provided as an embodiment of the present invention. Detailed Implementation
[0012] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0013] To achieve efficient and accurate reconstruction of social network dynamics functions and social network structures, this invention provides a robust network reconstruction search method based on a large language model. Figure 1 This is a flowchart illustrating a robust network reconstruction search method based on a large language model, provided as an embodiment of the present invention. Figure 1 As shown, it includes: S101. Obtain multivariate time series data of the social scene to be reconstructed, prior data of network structure, and population of initial dynamic functions.
[0014] Optionally, the multivariate time series data is a time series set consisting of the behavior or state data of multiple user nodes in the social scenario to be reconstructed in a continuous time dimension; the prior data of the network structure is a set of all known nodes and all known edges in the social scenario to be reconstructed.
[0015] Optionally, multivariate time series data can be represented as follows: ; in, Representing multivariate time series data, This indicates the number of user nodes in the social scenario to be reconstructed. Indicates the length of the time step. Indicates the first Each user node in time The state or behavioral characteristics; The prior data of the network structure are represented as follows: ; in, This represents prior data about the network structure. Represents the set of all known nodes. This represents the set of all known edges.
[0016] Optionally, the initialization population of dynamic functions includes at least: linearly coupled functions, nonlinear polynomial functions, exponential functions, sigmoid functions, and combined functions based on symbolic regression.
[0017] In this embodiment, the acquisition of multivariate time series data can follow the following process.
[0018] First, data collection and extraction: core raw data recording user interactions and states are obtained from server logs, user behavior databases, or open API interfaces on the platform backend. This includes explicit interactive behaviors such as posting, forwarding, commenting, and liking, as well as implicit state data such as online status, active periods, and timestamps of specific information dissemination events, providing rich raw materials for subsequent analysis.
[0019] Next, time normalization and discretization are performed: to transform the massive, continuously occurring events into a computable time series, the time axis needs to be divided into fixed-length analysis windows (e.g., 1 hour, 1 day). Within each time window, the behavior of each user (i.e., network node) is aggregated and statistically analyzed, such as calculating the number of posts, total interactions, or average activity level. This step transforms the unstructured streaming data into a well-organized, time-stepped multidimensional data table, which is the foundation for constructing the time series.
[0020] Finally, feature engineering for analysis is completed: based on the aforementioned discretized data, time-series feature vectors with clear semantic and network dynamic meanings are constructed for each node in the network. These features typically include, but are not limited to: a "post count sequence" reflecting the frequency of individual expression, a "retweet count sequence" reflecting the degree of participation in information diffusion, a "topic participation intensity sequence" measuring the degree of involvement in community topics, and an "information dissemination range sequence" assessing the node's influence. Thus, each node is characterized by a set of interrelated yet distinct multivariate time series, which together constitute the core input data describing the dynamic evolution of the entire network, laying a solid data foundation for subsequent network structure inference and dynamic function identification.
[0021] The prior data for the network structure consists of known information, such as known friend relationships, known follow relationships, and known connections that appear frequently in the historical interaction records.
[0022] Furthermore, the network to be reconstructed in this embodiment is not limited to social network scenarios, but can also be dynamic systems such as information dissemination networks and financial linkage networks. The implementation steps are similar to this method, and will not be repeated in this embodiment.
[0023] S102. Using multivariate time series data, prior data of network structure, and population of initial dynamic functions, establish an inference network optimization model from the perspectives of data fitting, sparsity constraints, and prior structural constraints, and use the inference network optimization model to obtain the current optimal predicted network structure corresponding to each initial dynamic function.
[0024] The optimization condition for the inference network optimization model is to minimize the sum of the corresponding data fitting terms, sparse constraint terms, and structural prior constraint terms.
[0025] Alternatively, the inference network optimization model can be expressed as: ; in, This represents the current optimal prediction network structure. This represents the structural variables of the social scenario to be reconstructed. Representing multivariate time series data, This represents any dynamic function in the population of initialization dynamic functions. This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the structure mask matrix corresponding to the prior data of the network structure. This represents prior data about the network structure. This represents the Hadamard product. Describing the L1 norm, This represents the L2 norm.
[0026] The solution for the current optimal prediction network structure can be obtained by using L1 regularization with proximal gradients, or by using algorithms suitable for sparse optimization such as Alternating Directional Multipliers (ADMM) or coordinate descent.
[0027] S103. Use the current optimal prediction network structure to search for the current optimal dynamic function corresponding to the social scene to be reconstructed.
[0028] Optionally, S103 includes: Based on the inference network optimization model, the initial dynamic function that minimizes the sum of preset fitting terms in the current optimal prediction network structure is taken as the current optimal dynamic function. The preset fitting terms include: data fitting terms and structural prior constraints.
[0029] In addition, some implementations can also use a consistency scoring method based on known partial structural priors to search for the current optimal dynamic function corresponding to the social scene to be reconstructed.
[0030] S104. Input the current optimal dynamic function into the large language model, and perform a reverse-driven network dynamic function search based on the preset prompt words to obtain the current optimal dynamic function population. Use the current optimal dynamic function population as the initial dynamic function population.
[0031] Optionally, the preset prompts include: social network dynamics algorithm expert, searching for the current optimal dynamics function population, and the current optimal dynamics function population being searched possesses differentiability and interpretability.
[0032] Optionally, the large language model is an LLM model.
[0033] In addition, the input, output and calling processes of the large language model are all controlled by computer programs. Its generation process does not depend on subjective judgment, but is embedded in the evolutionary computation process as a deterministic or semi-deterministic program generation module.
[0034] In addition to allowing LLM to freely combine basic operators to generate a population of dynamic functions, a "dynamic basis function library" (such as polynomials, trigonometric functions, common activation functions, etc.) can be pre-built, making the task of LLM to select and combine basis functions from the library. This can reduce the search space and improve efficiency.
[0035] Furthermore, the core "LLM search + evolution" paradigm can be implemented on different platforms. Besides ShinkaEvolve and EoH platforms, the process of this invention can also be implemented on automated algorithm design platforms such as LLM4AD. What is protected is the application of this paradigm to network reconstruction problems.
[0036] S105. Repeat steps S102-S104 until the preset number of iterations is reached. The current optimal predicted network structure in S102 when the preset number of iterations is reached is taken as the final reconstructed social network, and the current optimal dynamic function in S103 is taken as the final dynamic function.
[0037] It should be noted that the final reconstructed social network in this embodiment is typically represented as an NxN adjacency matrix or edge list, where element values represent connection strength or probability of existence. This forms the basis for all subsequent analyses. The final dynamic function derived from evolution is an interpretable mathematical expression that describes how a node's state is influenced by its neighbors (e.g., it is "white-box" knowledge about how the system operates).
[0038] Specifically, reconstructing social networks can ultimately pinpoint key nodes, identify the true "information hubs" or "opinion leaders" (highly connected or highly central nodes) within the network, and discover interest communities or factions based on actual interaction patterns rather than self-proclaimed ones through cluster analysis. Furthermore, by analyzing information dissemination paths, it becomes clear which connection paths information, behavior, or trends are most likely to spread along.
[0039] The final dynamics function can decipher the "mathematical laws of social interaction." As an interpretable mathematical formula, the final dynamics function reveals how a node's state (such as activity level and opinion) is influenced by its neighbors. Together, they constitute a complete toolbox for understanding and managing complex social systems.
[0040] This invention provides a robust network reconstruction search method based on a large language model. First, by initializing a population of dynamic functions and an iterative optimization framework, combined with a back-driven search using a large language model, the network dynamic functions are dynamically adjusted and explored, addressing the problem of poor adaptability and difficulty in handling unknown or time-varying dynamic characteristics in real-world systems based on preset models. Then, data fitting terms, sparse constraints, and structural prior constraints are explicitly introduced into the inference network optimization model to ensure the rationality and uniqueness of the solution through multi-faceted optimization conditions, solving the problem that most methods only focus on data fitting errors and lack effective regularization or structural prior constraints. Next, an end-to-end automated loop process automatically searches for and optimizes reconstruction strategies, reducing manual intervention and addressing the problem that existing methods heavily rely on expert experience and lack automated frameworks for parameter selection and tuning. Finally, prior network structure data is integrated into the optimization model, effectively utilizing some known structural information to guide the reconstruction process, ultimately achieving efficient and accurate reconstruction of social network dynamics and structure, thereby comprehensively improving the method's practicality and robustness.
[0041] In summary, the present invention has the following technical advantages: 1. Enhanced adaptability and discovery capabilities: Compared to fixed model methods, this invention can automatically discover and adapt to the unknown and complex network dynamics behind the data, breaking through the limitations of model assumptions and significantly enhancing generalization ability.
[0042] 2. Fundamentally alleviate the ill-conditioned problem: By explicitly incorporating sparsity and partial structural priors into the joint optimization framework, a strong constraint is provided for the inverse problem of ill-conditioned problems, making the solution more stable and closer to the real structure, overcoming the defect of non-unique solutions in traditional methods.
[0043] 3. Achieving high-level automation: Shifting the main responsibility for algorithm design from human experts to AI systems. The framework of this invention can automatically complete the entire process from "function search" to "parameter optimization" and then to "model selection," significantly reducing the technical threshold for network reconstruction applications.
[0044] 4. Higher data utilization efficiency: By effectively utilizing some known prior structural information, this invention can still achieve more accurate network reconstruction than methods without prior knowledge, even when the amount of data is limited or the noise is high.
[0045] 5. Provides interpretable intermediate results: The final dynamics function is a readable code function that not only reconstructs the network but also provides interpretable insights into the underlying dynamics of the system, which is not available in pure black-box deep learning reconstruction methods.
[0046] The method provided in this embodiment of the invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc., and this embodiment of the invention does not limit the application to such devices.
[0047] Based on the same inventive concept, embodiments of the present invention also provide a robust network reconstruction search device based on a large language model. Figure 2 A schematic diagram of a robust network reconstruction search device based on a large language model provided in an embodiment of the present invention is shown below. Figure 2 As shown, it includes: an acquisition unit 201, a model building unit 202, a search unit 203, a reverse driving unit 204, and a loop unit 205; Acquisition unit 201 is used to acquire multivariate time series data of the social scene to be reconstructed, prior data of network structure, and population of initial dynamic functions; The model building unit 202 is used to establish an inference network optimization model from the perspectives of data fitting, sparsity constraints, and structural prior constraints using multivariate time series data, prior network structure data, and initial dynamic function population. The inference network optimization model is used to obtain the current optimal predicted network structure corresponding to each initial dynamic function. The inference network optimization model uses the minimum sum of the corresponding data fitting terms, sparsity constraints, and structural prior constraints as the optimization condition. Search unit 203 is used to search for the current optimal dynamic function corresponding to the social scene to be reconstructed by utilizing the current optimal prediction network structure. The reverse driving unit 204 is used to input the current optimal dynamic function into the large language model and perform reverse driving network dynamic function search based on preset prompt words to obtain the current optimal dynamic function population, and use the current optimal dynamic function population as the initial dynamic function population; The loop unit 205 is used to repeatedly execute the steps of the model building unit 202, the search unit 203 and the back-driving unit 204 until a preset number of iterations is reached. The current optimal prediction network structure in the model building unit 202 when the preset number of iterations is reached is used as the final reconstructed social network, and the current optimal dynamic function in the search unit 203 is used as the final dynamic function.
[0048] Figure 3This invention provides a schematic diagram of a robust network reconstruction search device based on a large language model, comprising a processor 310, a storage medium 320, and a bus 330. The storage medium 320 stores machine-readable instructions executable by the processor 310. When the robust network reconstruction search device based on a large language model is running, the processor 310 communicates with the storage medium 320 via the bus 330, and the processor 310 executes the machine-readable instructions to perform the steps of the above-described method embodiment. Specific implementation methods and technical effects are similar and will not be repeated here.
[0049] The storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the storage medium may also be at least one storage device located remotely from the aforementioned processor.
[0050] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0051] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings and the disclosure, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In this description, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.
[0052] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A robust network reconstruction search method based on a large language model, characterized in that, include: S101. Obtain multivariate time series data of the social scene to be reconstructed, prior data of network structure, and population of initial dynamic functions; S102. Using the multivariate time series data, the prior data of the network structure, and the population of the initial dynamic functions, an inference network optimization model is established from the perspectives of data fitting, sparsity constraints, and structural prior constraints. The current optimal predicted network structure corresponding to each of the initial dynamic functions is obtained using the inference network optimization model. The inference network optimization model uses the minimum sum of the corresponding data fitting terms, sparsity constraints, and structural prior constraints as the optimization condition. S103. Use the current optimal prediction network structure to search for the current optimal dynamic function corresponding to the social scene to be reconstructed; S104. Input the current optimal dynamic function into the large language model, and perform a reverse-driven network dynamic function search based on preset prompt words to obtain the current optimal dynamic function population, and use the current optimal dynamic function population as the initial dynamic function population; S105. Repeat steps S102-S104 until a preset number of iterations is reached, and use the current optimal prediction network structure in S102 when the preset number of iterations is reached as the final reconstructed social network, and use the current optimal dynamic function in S103 as the final dynamic function.
2. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, The multivariate time series data is a time series set consisting of the behavior or state data of multiple user nodes in the social scenario to be reconstructed in a continuous time dimension; the network structure prior data is a set of all known nodes and all known edges in the social scenario to be reconstructed.
3. The robust network reconstruction search method based on a large language model according to claim 2, characterized in that, The multivariate time series data is represented as follows: ； in, This refers to the multivariate time series data. This indicates the number of user nodes in the social scenario to be reconstructed. Indicates the length of the time step. Indicates the first Each user node in time The state or behavioral characteristics; The prior data of the network structure are represented as follows: ； in, This represents the prior data of the network structure. Represents the set of all known nodes. This represents the set of all known edges.
4. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, The initial dynamic function population includes at least: linear coupling functions, nonlinear polynomial functions, exponential functions, sigmoid functions, and combined functions based on symbolic regression.
5. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, The inference network optimization model is expressed as follows: ； in, This represents the current optimal prediction network structure. Represents the structural variables of the social scene to be reconstructed. This refers to the multivariate time series data. This represents any one of the dynamic functions in the population of initialization dynamic functions. This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the structure mask matrix corresponding to the prior data of the network structure. This represents the prior data of the network structure. This represents the Hadamard product. Describing the L1 norm, This represents the L2 norm.
6. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, S103 includes: Based on the inference network optimization model, the initial dynamic function that minimizes the sum of preset fitting terms in the current optimal prediction network structure is taken as the current optimal dynamic function; The preset fitting terms include: the data fitting terms and the structural prior constraints.
7. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, The preset prompts include: social network dynamics algorithm expert, searching for the current optimal dynamics function population, and the current optimal dynamics function population being searched is differentiable and interpretable.
8. The robust network reconstruction search method based on a large language model according to claim 1, characterized in that, The large language model is an LLM model.
9. A robust network reconstruction search device based on a large language model, characterized in that, The robust network reconstruction search device based on a large language model includes: an acquisition unit, a model building unit, a search unit, a reverse driving unit, and a loop unit; The acquisition unit is used to acquire multivariate time series data of the social scene to be reconstructed, prior data of the network structure, and a population of initial dynamic functions; The model building unit is used to establish an inference network optimization model from the perspectives of data fitting, sparse constraints, and structural prior constraints using the multivariate time series data, the prior data of the network structure, and the population of initialization dynamic functions. The inference network optimization model is then used to obtain the current optimal predicted network structure corresponding to each of the initialization dynamic functions. The inference network optimization model uses the minimum sum of the corresponding data fitting terms, sparse constraint terms, and structural prior constraint terms as the optimization condition. The search unit is used to search for the current optimal dynamic function corresponding to the social scene to be reconstructed using the current optimal prediction network structure. The reverse driving unit is used to input the current optimal dynamic function into the large language model, and perform reverse driving network dynamic function search based on preset prompt words to obtain the current optimal dynamic function population, and use the current optimal dynamic function population as the initial dynamic function population; The loop unit is used to repeatedly execute the steps of the model building unit, the search unit, and the back-driving unit until a preset number of iterations is reached. The current optimal prediction network structure in the model building unit when the preset number of iterations is reached is used as the final reconstructed social network, and the current optimal dynamic function in the search unit is used as the final dynamic function.
10. A robust network reconstruction search device based on a large language model, characterized in that, include: The device includes a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the robust network reconstruction search device based on a large language model is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the robust network reconstruction search method based on a large language model as described in any one of claims 1-8.