Scientific research data credibility auditing method and device, electronic equipment and storage medium

Abstract

The application belongs to the technical field of data processing, and discloses a scientific research data credibility auditing method and device, an electronic device and a storage medium. The method comprises the following steps: obtaining to-be-audited scientific research data of each scientific research node, performing feature decomposition and orthogonal complementary projection processing on the to-be-audited scientific research data to obtain correction tensor data corresponding to the to-be-audited scientific research data, performing model construction based on the correction tensor data to obtain a scientific research data hypergraph model, performing coherent superposition processing on a Laplacian matrix corresponding to the scientific research data hypergraph model and the correction tensor data to obtain converged feature data of the to-be-audited scientific research data, and performing credit value calculation according to the converged feature data to obtain credibility auditing results of the to-be-audited scientific research data corresponding to each scientific research node. Through the above method, the convergence accuracy of the scientific research data is improved.

Description

Technical Field

[0001] This application belongs to the technical field of data processing, and specifically relates to a method, apparatus, electronic device and storage medium for auditing the credibility of scientific research data. Background Technology

[0002] In the process of distributed scientific research data aggregation, multi-source heterogeneity and spatiotemporal sparsity are common challenges. Existing techniques typically employ low-rank matrix completion or tensor decomposition frameworks to recover missing values ​​and reconstruct signals. These methods perform well under ideal conditions where the data distribution is uniform and the missing data patterns are random. However, in real-world large-scale observation networks, observational data are often accompanied by non-random missing data patterns and multi-source interference, making it difficult for traditional low-rank assumptions to accurately characterize the true structure of the data. For example, when data missing data is not random but related to specific observation nodes or time periods, low-rank approximation methods may fail to effectively capture the inherent correlations in the data, thus affecting the accuracy of data recovery.

[0003] Furthermore, in a distributed environment, the data quality of each observation node varies, potentially leading to systematic biases or anomalous contributions. Existing solutions generally lack dynamic evaluation mechanisms for the reliability of data sources, making it difficult to effectively identify low-quality or anomalous data. Specifically, during multi-source data fusion, endogenous bias information from different observation nodes can easily infiltrate the true signal, affecting the accuracy of the convergence results. Simultaneously, traditional low-rank approximation methods, when processing data with high missing rates, tend to filter out signal components with small amplitudes but consistent structure as noise, resulting in the loss of weak but important scientific signals. More importantly, in distributed observation scenarios, the lack of effective inter-node data quality assessment and anomaly detection methods makes it difficult to guarantee the reliability of the convergence results, thereby impacting the reliability of scientific research decisions.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, electronic device and storage medium for auditing the credibility of scientific research data, so as to solve the problem that the existing scientific research data aggregation process is unable to guarantee the credibility of the aggregation results due to the lack of a dynamic evaluation mechanism for the credibility of data sources, which affects scientific research decision-making. It can dynamically evaluate the credibility of data at each scientific research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of scientific research data aggregation.

[0006] Firstly, this application provides a method for auditing the credibility of scientific research data, including: Obtain the research data to be audited at each research stage; The research data to be audited is subjected to feature decomposition and orthogonal projection processing to obtain the corrected tensor data corresponding to the research data to be audited. Based on the aforementioned calibration tensor data, a model is constructed to obtain a scientific research data hypergraph model; The Laplacian matrix corresponding to the hypergraph model of the scientific research data is coherently superimposed with the correction tensor data to obtain the aggregated feature data of the scientific research data to be audited. The credibility audit results of the research data to be audited are obtained by calculating the reputation value based on the aggregated feature data for each research node.

[0007] The research data credibility auditing method provided in this application can achieve credibility auditing of research data. It corrects data bias through feature decomposition and orthogonal complementary projection, constructs a hypergraph model to capture the inherent correlation of data, and combines the Laplacian matrix to perform high-quality feature aggregation. Ultimately, it achieves accurate auditing of the credibility of research data, solving the problem that the existing research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, which makes it difficult to ensure the credibility of the aggregation results and thus affects research decisions. It can dynamically evaluate the data credibility of each research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of research data aggregation.

[0008] Optionally, the research data to be audited is subjected to feature decomposition and orthogonal projection processing to obtain the corrected tensor data corresponding to the research data to be audited, including: The research data to be audited is preprocessed to obtain a time-series observation data set; Based on the endogenous bias information of each research node, tensors are constructed on the time series observation data set corresponding to the research data to be audited to obtain the bias tensor data corresponding to the research data to be audited. Based on the deviation tensor data, eigenvalue decomposition and orthogonal projection processing are performed to obtain the correction tensor data corresponding to the research data to be audited.

[0009] Optionally, based on the deviation tensor data, eigenvalue decomposition and orthogonal complementary projection processing are performed to obtain the corrected tensor data corresponding to the research data to be audited, including: The deviation tensor data is subjected to eigenvalue decomposition to obtain the eigenvalue decomposition matrix; The eigenvalue matrix is ​​subjected to orthogonal complementary projection to obtain the corrected tensor data corresponding to the research data to be audited.

[0010] The research data credibility auditing method provided in this application can perform credibility auditing on research data. By constructing a bias tensor and performing eigenvalue decomposition and orthogonal complementary projection, it effectively removes systematic biases from the data, thereby obtaining purer and more accurate correction tensor data, and further improving the accuracy of data correction.

[0011] Optionally, a model is constructed based on the corrected tensor data to obtain a research data hypergraph model, including: An initial hypergraph model is constructed using the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association relationships between the sites and the sampling batches as edges. Based on the auditable scientific research data from different sites belonging to the same sampling batch in the correction tensor data, covariance is calculated to obtain the empirical covariance matrix between the sites belonging to the same sampling batch. Based on the maximum and minimum values ​​in the empirical covariance matrix, the covariance consistency is calculated to obtain the covariance consistency corresponding to the empirical covariance matrix. The covariance consistency is determined as the hyperedge weight of each edge in the initial hypergraph model to obtain the scientific research data hypergraph model.

[0012] Optionally, the Laplacian matrix corresponding to the hypergraph model of the scientific research data is coherently superimposed with the correction tensor data to obtain the convergent feature data of the scientific research data to be audited, including: Determine the Laplacian matrix corresponding to the hypergraph model of the scientific research data; Extract a predetermined number of target non-zero eigenvalues ​​from the Laplacian matrix; The target non-zero eigenvalues ​​are coherently superimposed with the correction tensor data to obtain the aggregated feature data of the research data to be audited.

[0013] The research data credibility auditing method provided in this application can perform credibility auditing on research data. By using the target non-zero eigenvalues ​​of the Laplacian matrix, it can effectively capture the global structural information of the hypergraph model and the inherent manifold structure of the data. By coherently superimposing it with the calibration tensor data, it can highlight the consistency patterns and important features in the data, thereby obtaining more representative and robust convergent feature data.

[0014] Optionally, the target non-zero eigenvalues ​​are coherently superimposed with the correction tensor data to obtain the aggregated feature data of the research data to be audited, including: Based on the aforementioned sites, construct the target non-zero feature matrix corresponding to the target non-zero feature values; The corrected tensor data is projected onto the target non-zero feature matrix to obtain the manifold trend matrix; Based on a preset threshold for preventing the removal of zero, the target non-zero feature matrix is ​​subjected to weak signal amplification processing to obtain the diagonal weighted matrix corresponding to the target non-zero feature matrix. The convergence processing is performed based on the target non-zero feature matrix, the manifold trend matrix, and the diagonal weighted matrix to obtain the converged feature data of the scientific research data to be audited.

[0015] Optionally, a reputation score is calculated based on the aggregated feature data to obtain the credibility audit results of the research data to be audited corresponding to each research node, including: Based on the converged feature data and its corresponding truncated mean, alignment is calculated to obtain global consensus feature data; Based on the correction tensor data and the median of the correction tensor data corresponding to each scientific research node, deviation anomaly calculation is performed to obtain the deviation anomaly degree. The on-time reporting rate and service availability rate are obtained by statistically analyzing the historical upload records of the aggregated feature data of each research node. Based on the global consensus feature data, the deviation anomaly degree, the on-time reporting rate, and the service availability rate, combined with the preset calculation weight and the preset forgetting factor, the reputation value is calculated to obtain the reputation value corresponding to each research node. Based on the tiered threshold range in which the reputation value falls, the credibility audit result of the research data to be audited corresponding to each research node is determined.

[0016] Secondly, this application provides a scientific research data credibility auditing device, comprising: The acquisition module is used to acquire the research data to be audited at each research node; The projection module is used to perform feature decomposition and orthogonal projection processing on the research data to be audited to obtain the corrected tensor data corresponding to the research data to be audited. The construction module is used to construct a model based on the calibration tensor data to obtain a scientific research data hypergraph model; The overlay module is used to coherently overlay the Laplacian matrix corresponding to the hypergraph model of the scientific research data with the correction tensor data to obtain the convergent feature data of the scientific research data to be audited. The calculation module is used to calculate the reputation value based on the aggregated feature data to obtain the credibility audit result of the research data to be audited corresponding to each research node.

[0017] This scientific research data credibility auditing device corrects data deviations through feature decomposition and orthogonal complementary projection, constructs a hypergraph model to capture the inherent correlations of data, and combines the Laplacian matrix to perform high-quality feature aggregation. Ultimately, it achieves accurate auditing of the credibility of scientific research data, solving the problem that the existing scientific research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, which makes it difficult to ensure the credibility of the aggregation results and thus affects scientific research decisions. It can dynamically evaluate the data credibility of each scientific research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of scientific research data aggregation.

[0018] Thirdly, this application provides an electronic device, including a processor and a memory, wherein the memory stores a computer program executable by the processor, and when the processor executes the computer program, it performs the steps in the scientific research data credibility auditing method described above.

[0019] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the scientific research data credibility auditing method described above.

[0020] Beneficial Effects: The research data credibility auditing method, device, electronic equipment, and storage medium provided in this application correct data bias through feature decomposition and orthogonal complementary projection, construct a hypergraph model to capture the intrinsic correlation of data, and combine it with the Laplacian matrix for high-quality feature aggregation, ultimately achieving accurate auditing of research data credibility. This solves the problem that the existing research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, making it difficult to guarantee the credibility of aggregation results and thus affecting research decisions. It focuses on the endogenous bias removal and dynamic credibility evaluation of distributed research node data. Through bias tensor modeling, orthogonal complementary projection correction, and reputation value calculation, it achieves data quality auditing, can dynamically evaluate the data credibility of each research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of research data aggregation. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the scientific research data credibility auditing method provided in this application embodiment.

[0022] Figure 2 A schematic diagram of the structure of the scientific research data credibility auditing device provided in the embodiments of this application.

[0023] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0024] Labeling Explanation: 1. Acquisition Module; 2. Projection Module; 3. Construction Module; 4. Overlay Module; 5. Calculation Module; 301. Processor; 302. Memory; 303. Communication Bus. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Please refer to Figure 1 , Figure 1 This application provides a method for auditing the credibility of scientific research data, as described in some embodiments. The method includes the following steps: Step S101: Obtain the research data to be audited for each research node; Step S102: Perform feature decomposition and orthogonal projection processing on the research data to be audited to obtain the corrected tensor data corresponding to the research data to be audited. Step S103: Based on the calibration tensor data, a model is constructed to obtain the scientific research data hypergraph model; Step S104: The Laplacian matrix corresponding to the hypergraph model of scientific research data is coherently superimposed with the correction tensor data to obtain the aggregated feature data of the scientific research data to be audited. Step S105: Calculate the reputation value based on the aggregated feature data to obtain the credibility audit results of the research data to be audited for each research node.

[0028] This research data credibility auditing method corrects data bias through feature decomposition and orthogonal complementary projection, constructs a hypergraph model to capture the inherent correlation of data, and combines it with the Laplacian matrix for high-quality feature aggregation. Ultimately, it achieves accurate auditing of research data credibility, thus solving the problem that the existing research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, making it difficult to guarantee the credibility of aggregation results and thus affecting research decisions. It can dynamically evaluate the data credibility of each research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of research data aggregation.

[0029] Specifically, in step S101, the research data to be audited from each research node is acquired. Here, a research node refers to a data acquisition, processing, or storage unit in a distributed research network, such as a sensor node, laboratory server, or personal research terminal. The research data to be audited refers to the raw or preliminarily processed research data that requires credibility assessment; its form can be time series, images, text, etc.

[0030] Specifically, in step S102, the research data to be audited is subjected to feature decomposition and orthogonal projection processing to obtain the corrected tensor data corresponding to the research data to be audited, which specifically includes: Preprocessing of the research data to be audited yields a time-series observation data set. Based on the endogenous bias information of each research node, tensors are constructed on the time series observation data set corresponding to the research data to be audited, and the bias tensor data corresponding to the research data to be audited are obtained. Based on the bias tensor data, eigenvalue decomposition and orthogonal projection processing are performed to obtain the corrected tensor data corresponding to the scientific research data to be audited.

[0031] In step S102, the research data to be audited is preprocessed to eliminate inconsistencies, noise, and redundant information, making it suitable for subsequent processing and obtaining a time-series observation dataset corresponding to the research data to be audited. This time-series observation dataset is cleaned, aligned, and normalized, exhibiting higher consistency and usability, laying the foundation for subsequent tensor construction and eigenvalue decomposition.

[0032] Preprocessing includes spatiotemporal alignment and normalization. Spatiotemporal alignment addresses the mismatch in time or space between data from different research nodes or sampling batches. For example, methods such as timestamp calibration, interpolation, or resampling can ensure that all data points are compared and analyzed on a unified spatiotemporal reference, eliminating data bias caused by differences in sampling frequency, recording time, or geographical location. Normalization transforms data with different dimensions or numerical ranges to a uniform scale. Methods such as min-max normalization and Z-score normalization can eliminate the impact of dimensional differences on feature decomposition and projection processing, preventing certain features from dominating the calculation due to their large values.

[0033] Endogenous bias information refers to the inherent, systematic errors or deviations of each research node during data acquisition, transmission, or processing. Specifically, it includes three source dimensions: (1) Equipment dimension, systematic errors exist in hardware such as different models of sequencers / mass spectrometers; (2) Batch dimension, conditional drift exists in different experimental batches of the same equipment; (3) Site dimension (a site refers to the physical observation location corresponding to a research node, including the geographical location (latitude and longitude information) and environmental information (such as the upstream and downstream location of a river)), there are protocol differences between research nodes in different geographical locations (wherein, a site can contain multiple research nodes, and a research node can contain multiple devices). This bias information usually has a certain degree of stability and identifiability, and can be obtained through historical data analysis, expert experience, or specialized calibration experiments, and can be represented in the form of vectors, matrices, or tensors.

[0034] When acquiring research data to be audited, each data point (research data to be audited) is accompanied by information from the three source dimensions mentioned above: equipment model, batch number, and site identifier. For datasets with the same equipment model, batch number, and site identifier, the difference between each data point and its corresponding dataset mean (or gold standard data, i.e., a special sample or data point that has been measured by authoritative methods, has precise numerical knowledge, and can serve as an absolute reference benchmark) is calculated. This yields endogenous bias information, which is then combined with the preprocessed time-series observation dataset and the endogenous bias information of each research node to form a high-dimensional bias tensor dataset. ,in, For bias tensor data, This represents the number of equipment model categories. This refers to the batch number. The number of the site identifier. For dimensional space, This represents the dimensional space of equipment model, batch number, and site identifier. This bias tensor data aims to comprehensively capture and quantify the endogenous bias patterns of each research node at different time points and across different observation dimensions.

[0035] Specifically, in step S102, eigenvalue decomposition and orthogonal projection processing are performed based on the deviation tensor data to obtain the correction tensor data corresponding to the research data to be audited, including: Eigenvalue decomposition is performed on the bias tensor data to obtain the eigenvalue matrix; The eigenvalue matrix is ​​subjected to orthogonal complementary projection to obtain the corrected tensor data corresponding to the research data to be audited.

[0036] In step S102, eigenvalue decomposition is performed on the deviation tensor data to decompose the complex deviation pattern into a series of independent feature components, thereby revealing the main structure and potential patterns of the deviation data and obtaining the eigenvalue decomposition matrix. The eigenvalue decomposition matrix includes the core tensor, equipment model factor matrix, batch number factor matrix, and site identifier factor matrix. The specific calculation process of the eigenvalue decomposition is as follows: ; in, For the first Equipment model factor matrix for each deviation pollution item; For the first Batch number factor matrix of each deviation contamination item; For the first A matrix of site identification factors for each deviation pollution item; This represents the total number of deviation contamination items, i.e., the preset number of block items, whose value is determined based on physical priors (such as the number of physical component types, including equipment baseline items, batch drift items, and fast-changing signal items). This is the symbol for the outer product operator.

[0037] Orthogonal projection of the eigenvalue matrix can separate components related to endogenous bias from the original data, yielding the corrected tensor data corresponding to the research data to be audited. Orthogonal projection projects the data onto a subspace orthogonal to the bias feature space, effectively removing or suppressing the influence of endogenous bias, resulting in a purer and more accurate corrected tensor data that more realistically reflects the intrinsic characteristics of the research data.

[0038] The calculation process for orthogonal projection is as follows: ; in, To correct tensor data; It is the identity matrix; Equipment model factor matrix for all deviation contamination items Batch number factor matrix of all deviation contamination items Site identification factor matrix for all deviation pollution items The constructed subspace basis can be obtained by applying the equipment model factor matrix to all deviation contamination terms. Batch number factor matrix of all deviation contamination items Site identification factor matrix for all deviation pollution items After splicing, the spliced ​​matrix is ​​orthogonalized (i.e., QR decomposition or singular value decomposition (SVD)).

[0039] Specifically, in step S103, a model is constructed based on the calibration tensor data to obtain a research data hypergraph model, including: An initial hypergraph model is constructed by taking the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association between sites and sampling batches as edges. Based on the auditable scientific research data from different sites belonging to the same sampling batch (same batch number) in the calibration tensor data, covariance is calculated to obtain the empirical covariance matrix between sites belonging to the same sampling batch. Based on the maximum and minimum values ​​in the empirical covariance matrix, the covariance consistency is calculated to obtain the covariance consistency corresponding to the empirical covariance matrix. By defining the covariance consistency as the hyperedge weight of each edge in the initial hypergraph model, the research data hypergraph model is obtained.

[0040] In step S103, an initial hypergraph model is constructed using the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association relationships between sites and sampling batches as edges. This establishes a basic topological structure that reflects the fundamental relationships of research data in the spatial (site) and temporal (sampling batch) dimensions. Nodes can specifically refer to the physical location (site) of data collection or the time window (sampling batch) of data collection, while edges represent the preset logical or physical connections between these sites and sampling batches. Each node's edges connect all sites within the same sampling batch, or all sampling batches within the same site.

[0041] Furthermore, based on the auditable research data from different sites belonging to the same sampling batch in the calibration tensor data, covariance is calculated to obtain the empirical covariance matrix among the sites within the same sampling batch. This matrix can quantify the synchronicity or correlation of the changing trends of research data provided by different sites under the same data collection batch. By calculating the covariance, the consistency or potential differences in the data behavior of each site can be revealed.

[0042] The specific process of calculating covariance is as follows: ; in, Let be the empirical covariance matrix, which is a A square matrix, where m is the number of stations; For time windows; This refers to the auditable research data (sorted by time series) from m sites acquired within a time window t.

[0043] Based on this, covariance consistency is calculated using the maximum and minimum values ​​in the empirical covariance matrix to obtain the covariance consistency corresponding to the empirical covariance matrix. Covariance consistency is a normalization or quantification process that transforms complex covariance relationships into a unified and comparable indicator to measure the level of consistency between data from different sites.

[0044] The calculation process for covariance consistency is as follows: ; in, For covariance consistency; This is the ratio of the maximum to the minimum value in the empirical covariance matrix; To prevent small amounts from being removed, settings can be configured according to actual needs.

[0045] Finally, the covariance consistency was determined as the hyperedge weights of each edge in the initial hypergraph model, resulting in the research data hypergraph model. The hyperedge weights reflect the strength or reliability of the associations between nodes. By assigning weights, the hypergraph model can more precisely express the complex dependencies between research data, with higher weights typically indicating stronger associations or higher data consistency.

[0046] Specifically, in step S104, the Laplacian matrix corresponding to the hypergraph model of the research data is coherently superimposed with the correction tensor data to obtain the aggregated feature data of the research data to be audited, including: Determine the Laplacian matrix corresponding to the hypergraph model of scientific research data; Extract a predetermined number of target non-zero eigenvalues ​​from the Laplacian matrix; The non-zero eigenvalues ​​of the target data are coherently superimposed with the correction tensor data to obtain high-quality converged features of the research data to be audited.

[0047] In step S104, the Laplacian matrix of the constructed research data hypergraph model is calculated. In graph theory, the Laplacian matrix is ​​the difference between the adjacency matrix and the degree matrix of a graph, reflecting its connectivity and structural information. For the hypergraph model, the Laplacian matrix is ​​constructed by considering the characteristics of hyperedges, such as through hyperedge association matrices and hyperedge weight matrices, to capture the complex relationships between research nodes, sites, and sampling batches.

[0048] After obtaining the Laplacian matrix, its eigenvalues ​​are decomposed, and several eigenvalues ​​with the smallest non-zero values ​​are selected to obtain the target non-zero eigenvalues ​​(i.e., the minimum non-zero eigenvalues). These target non-zero eigenvalues ​​are usually associated with the connectivity, clustering structure, and low-frequency information of the graph, effectively capturing the intrinsic structure and latent patterns in the hypergraph model while filtering out high-frequency eigenvalues ​​related to noise or irrelevant information. The preset number of eigenvalues ​​can be adjusted according to the actual application scenario and data characteristics, for example, by using empirical values, cross-validation, or information criteria.

[0049] Optionally, in step S104, the target non-zero eigenvalues ​​and the correction tensor data are coherently superimposed to obtain the aggregated feature data of the research data to be audited, specifically including: Based on the site, construct the target non-zero feature matrix corresponding to the target non-zero feature values; The corrected tensor data is projected onto the target non-zero feature matrix to obtain the manifold trend matrix; Based on a preset threshold for preventing the elimination of zero, the non-zero feature matrix of the target is subjected to weak signal amplification processing to obtain the diagonal weighted matrix corresponding to the non-zero feature matrix of the target. The aggregated feature data of the research data to be audited is obtained by performing aggregation processing based on the target non-zero feature matrix, manifold trend matrix and diagonal weighted matrix.

[0050] Specifically, in step S104, a predetermined number of target non-zero eigenvalues ​​extracted from the Laplacian matrix are organized and arranged according to their respective sites to form a matrix that can characterize the intrinsic structure and correlation of the data from each site. This target non-zero eigenvalue matrix can be regarded as a basis in a low-dimensional space, used to capture the main trends and intrinsic relationships of scientific research data under the hypergraph structure.

[0051] By using linear or nonlinear mapping, high-dimensional calibration tensor data is reduced to a low-dimensional manifold space defined by the target non-zero eigenvalue matrix, resulting in a manifold trend matrix. This projection process filters out noise and redundant information from the data, extracting the essential trends and structures of the data on the manifold, thus yielding a purer and more representative manifold trend matrix.

[0052] The calculation process of the manifold trend matrix is ​​as follows: ; in, This is the manifold trend matrix; For the target non-zero characteristic matrix, Let be a k×c matrix, where k is a preset quantity and c is the number of different physical quantities or different feature dimensions collected simultaneously at the same observation time and the same station. The preset quantity k is selected based on the eigenvalue spectral gap (eigengap) of the Laplace matrix, which is calculated as the difference between adjacent eigenvalues ​​in the non-zero feature matrix of the target. , (in, Let p+1 be the (p+1)th non-zero eigenvalue of the target in the non-zero eigenvalue matrix. (where p is the target non-zero eigenvalue in the target non-zero eigenvalue matrix), and select the one that makes p a target non-zero eigenvalue. The value of p when the maximum value is obtained is taken as the value of k. This selection method can automatically determine the most important manifold structure dimension in the hypergraph model and avoid subjective bias caused by manual setting.

[0053] Based on a preset threshold for preventing division by zero, the target non-zero feature matrix undergoes weak signal amplification processing to obtain a diagonal weighted matrix corresponding to the target non-zero feature matrix. This enhances important feature signals that might be overlooked in the original data due to their small amplitude. The construction of the diagonal weighted matrix allows each eigenvalue or its corresponding component to receive different weights based on its importance or signal strength. The threshold for preventing division by zero can be set according to actual needs to avoid division by zero in subsequent calculations, while ensuring that even weak signals receive appropriate weighting and amplification, thereby preventing the loss of important information.

[0054] The specific calculation process for the diagonal weighted matrix is ​​as follows: ; ; in, It is a diagonal weighted matrix; It is a diagonal matrix function; This is a function for amplifying and processing weak signals. The k-th non-zero eigenvalue of the target; The preset threshold for preventing zero-value loss can be set according to actual needs.

[0055] Furthermore, convergence processing is performed based on the target non-zero feature matrix, manifold trend matrix, and diagonal weighted matrix to obtain converged feature data of the research data to be audited. This convergence processing can employ various mathematical operations, such as matrix multiplication, to effectively integrate the information after manifold projection and weak signal amplification processing, ultimately generating converged feature data that comprehensively, accurately, and robustly reflects the essential characteristics of the research data to be audited.

[0056] The specific calculation process for the aggregated feature data is as follows: ; in, To aggregate feature data.

[0057] Specifically, in step S105, a reputation value is calculated based on the aggregated feature data to obtain the credibility audit results of the research data to be audited for each research node, including: Alignment is calculated based on the converged feature data and its corresponding truncated mean to obtain global consensus feature data; Based on the calibration tensor data and the median of the calibration tensor data corresponding to each scientific research node, deviation anomaly calculation is performed to obtain the deviation anomaly degree. The on-time reporting rate and service availability rate are obtained by statistically analyzing the historical upload records of the aggregated characteristic data of each scientific research node. Based on global consensus feature data, deviation anomaly degree, on-time reporting rate and service availability rate, combined with preset calculation weights and preset forgetting factors, reputation value is calculated to obtain the reputation value corresponding to each scientific research node. Based on the tiered threshold range of the reputation value, the credibility audit results of the research data to be audited corresponding to each research node are determined.

[0058] In step S105, by comparing the difference between the aggregated feature data and the truncated mean, the deviation of each research node's data from the overall consensus is quantified, resulting in global consensus feature data. The truncated mean can be understood as the mean obtained by removing a portion of extreme values ​​(e.g., a preset percentage of the top and bottom rankings, which can be set according to actual needs) when calculating the average. This reduces the impact of outliers on the global consensus feature data, making it more representative. The global consensus feature data reflects the overall trend or standard of all research node data after removing the influence of outliers.

[0059] The specific formula for calculating the global consensus feature data is as follows: ; in, This is global consensus feature data; This represents the aggregated feature data of the i-th research node; The global truncated mean is obtained by processing the aggregated feature data of all research nodes in the same round of credibility audit after coordinate-level truncation mean processing (after sorting the aggregated feature data of all research nodes in the same round of credibility audit, removing the data of the top and bottom preset percentages, and then calculating the arithmetic mean of the remaining data to obtain the truncated mean). This refers to the aggregated feature data of the j-th research node in the same round of credibility audit.

[0060] Based on the median of the calibration tensor data and the corresponding calibration tensor data for each research node, deviation anomaly calculation is performed to assess the degree of deviation of each research node's data from its own median, thus obtaining the deviation anomaly score. The median, as a robust statistic, effectively resists the interference of outliers; therefore, calculating the deviation anomaly score based on the median can more accurately measure the anomalous behavior of individual research node data. A higher deviation anomaly score indicates a greater deviation of the node's data from the performance of other nodes, and its reliability may be lower.

[0061] The specific formula for calculating the deviation anomaly degree is as follows: ; in, The degree of deviation anomaly; It is a sigmoid function; is the Frobenius norm notation, representing the square root of the sum of squares of all elements; The median norm of the corrected tensor data for all research nodes.

[0062] Based on the historical upload records of aggregated characteristic data from each research node, statistical analysis is performed to obtain the on-time reporting rate and service availability rate. The on-time reporting rate and service availability rate are important indicators for measuring the reliability of research node behavior. The on-time reporting rate refers to the proportion of research nodes that submit data within the stipulated time, reflecting the node's fulfillment capability. The service availability rate refers to the stability and continuity of the data services provided by research nodes, reflecting the node's reliability. These indicators are obtained by statistically analyzing the historical upload records of aggregated characteristic data from each research node, enabling an assessment of the research node's reputation from a behavioral perspective.

[0063] Furthermore, by combining preset calculation weights and a preset forgetting factor, the aforementioned multi-dimensional indicators (global consensus feature data, deviation anomaly, on-time reporting rate, and service availability) are comprehensively considered to calculate a reputation score. The preset calculation weights are used to adjust the importance of different indicators in the reputation score calculation; for example, data quality indicators can be assigned higher weights based on actual needs. The preset forgetting factor is used to attenuate the influence of historical data, ensuring that the reputation score more timely reflects the latest performance of research nodes and avoids outdated information unduly affecting the current reputation assessment. This factor can be set according to actual needs. Through this weighting and attenuation mechanism, a dynamic and comprehensive reputation score for research nodes can be obtained.

[0064] The specific formula for calculating the credit score is as follows: ; ; in, Let be the reputation value of the i-th research node in the n-th round of credibility audit; Let be the reputation value of the i-th research node in the (n-1)th round of credibility audit; The preset forgetting factor; The behavior score of the i-th research node in the n-th round of credibility audit; This represents the global consensus feature data for the i-th research node; This represents the deviation anomaly at the i-th research node; Let be the on-time reporting rate of the i-th research node; The service availability of the i-th research node; , , , The preset calculation weight corresponds to the i-th research node. , Let be the weight for calculating the global consensus feature data of the i-th research node. Let be the weight for calculating the deviation anomaly of the i-th research node. Let be the weight used to calculate the on-time reporting rate of the i-th research node. The weight for calculating the service availability of the i-th research node; when n=1, At this time It is also assigned a value of 0, starting from the second round (i.e., n=2). Restore to the default value.

[0065] In addition, to prevent long-term offline nodes from bringing back expired reputation values, a reputation value decay mechanism is introduced: if a research node fails to report aggregated feature data for multiple consecutive rounds (e.g., 5 rounds), its reputation value will be reduced. Multiply by the attenuation factor when re-launching. ( It can be set according to actual needs, such as 0.9), that is... ,in, The new (upon relaunch) reputation value. This is the old (before relaunching) reputation value. This mechanism ensures that the reputation value reflects the node's recent performance, avoiding a "one-and-done" accumulation of reputation.

[0066] Finally, the calculated reputation score is compared with a preset grading threshold range to determine the credibility audit result. For example, the reputation score can be divided into multiple levels such as "high credibility," "medium credibility," and "low credibility," with each level corresponding to a reputation score range. For instance, the reputation score range can be divided into... , and When the reputation value of a research node falls within a certain range, the credibility audit result of the corresponding research data to be audited can be determined. For example, when the reputation value falls within the range... When the credibility of the corresponding research data to be audited is determined, the aggregated feature data corresponding to the research data to be audited is directly incorporated into the global tensor; when the credibility value falls within the interval At this time, staff at the corresponding research node need to witness and sign for confirmation before the aggregated feature data corresponding to the research data to be audited can be directly incorporated into the global tensor; when the reputation value falls within the interval If a research node is deemed unreliable, its data for that round will not be adopted. Furthermore, if the reputation score remains below 0.3 for multiple consecutive rounds, the corresponding research node will be blacklisted and reported to the management center. This tiered mechanism makes the audit results more operational and interpretable.

[0067] As shown above, this scientific research data credibility auditing method acquires the research data to be audited from each research node, performs feature decomposition and orthogonal complementary projection processing on the research data to be audited to obtain the corresponding correction tensor data, constructs a model based on the correction tensor data to obtain a research data hypergraph model, and coherently superimposes the Laplacian matrix corresponding to the research data hypergraph model with the correction tensor data to obtain the aggregated feature data of the research data to be audited. Based on the aggregated feature data, a reputation value is calculated to obtain the credibility audit results of the research data to be audited for each research node. Thus, by correcting data bias through feature decomposition and orthogonal complementary projection, constructing a hypergraph model to capture the inherent correlation of data, and combining it with the Laplacian matrix for high-quality feature aggregation, the method ultimately achieves accurate auditing of the credibility of research data. This solves the problem that the existing research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, making it difficult to guarantee the credibility of the aggregation results and thus affecting research decisions. It can dynamically evaluate the data credibility of each research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of research data aggregation.

[0068] refer to Figure 2 This application provides a scientific research data credibility auditing device for auditing the credibility of scientific research data, including: Module 1 is used to acquire the research data to be audited at each research node; Projection module 2 is used to perform feature decomposition and orthogonal projection processing on the research data to be audited, so as to obtain the corrected tensor data corresponding to the research data to be audited. Module 3 is used to build a model based on the calibration tensor data to obtain a hypergraph model of scientific research data. Overlay module 4 is used to coherently overlay the Laplacian matrix corresponding to the hypergraph model of scientific research data with the correction tensor data to obtain the aggregated feature data of the scientific research data to be audited. Calculation module 5 is used to calculate the reputation value based on the aggregated feature data, and obtain the credibility audit results of the research data to be audited corresponding to each research node.

[0069] This scientific research data credibility auditing device corrects data deviations through feature decomposition and orthogonal complementary projection, constructs a hypergraph model to capture the inherent correlations of data, and combines it with the Laplacian matrix to perform high-quality feature aggregation. Ultimately, it achieves accurate auditing of the credibility of scientific research data, thereby solving the problem that the existing scientific research data aggregation process is unable to guarantee the credibility of the aggregation results due to the lack of a dynamic evaluation mechanism for the credibility of data sources, which affects scientific research decisions. It can dynamically evaluate the data credibility of each scientific research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of scientific research data aggregation.

[0070] Specifically, during execution, module 1 acquires the research data to be audited from each research node. Here, a research node refers to a data acquisition, processing, or storage unit in a distributed research network, such as a sensor node, laboratory server, or personal research terminal. The research data to be audited refers to the raw or preliminarily processed research data that requires credibility assessment; its form can be time series, images, text, etc.

[0071] Specifically, when projection module 2 performs feature decomposition and orthogonal projection processing on the research data to be audited to obtain the corrected tensor data corresponding to the research data to be audited, it executes: Preprocessing of the research data to be audited yields a time-series observation data set. Based on the endogenous bias information of each research node, tensors are constructed on the time series observation data set corresponding to the research data to be audited, and the bias tensor data corresponding to the research data to be audited are obtained. Based on the bias tensor data, eigenvalue decomposition and orthogonal projection processing are performed to obtain the corrected tensor data corresponding to the scientific research data to be audited.

[0072] During execution, projection module 2 preprocesses the research data to be audited to eliminate inconsistencies, noise, and redundant information, ensuring it meets the requirements for subsequent processing and yielding a time-series observation dataset corresponding to the research data to be audited. This time-series observation dataset is cleaned, aligned, and normalized, exhibiting higher consistency and usability, laying the foundation for subsequent tensor construction and eigenvalue decomposition.

[0073] Preprocessing includes spatiotemporal alignment and normalization. Spatiotemporal alignment addresses the mismatch in time or space between data from different research nodes or sampling batches. For example, methods such as timestamp calibration, interpolation, or resampling can ensure that all data points are compared and analyzed on a unified spatiotemporal reference, eliminating data bias caused by differences in sampling frequency, recording time, or geographical location. Normalization transforms data with different dimensions or numerical ranges to a uniform scale. Methods such as min-max normalization and Z-score normalization can eliminate the impact of dimensional differences on feature decomposition and projection processing, preventing certain features from dominating the calculation due to their large values.

[0074] Endogenous bias information refers to the inherent, systematic errors or deviations of each research node during data acquisition, transmission, or processing. Specifically, it includes three source dimensions: (1) Equipment dimension, systematic errors exist in hardware such as different models of sequencers / mass spectrometers; (2) Batch dimension, conditional drift exists in different experimental batches of the same equipment; (3) Site dimension (a site refers to the physical observation location corresponding to a research node, including the geographical location (latitude and longitude information) and environmental information (such as the upstream and downstream location of a river)), there are protocol differences between research nodes in different geographical locations (wherein, a site can contain multiple research nodes, and a research node can contain multiple devices). This bias information usually has a certain degree of stability and identifiability, and can be obtained through historical data analysis, expert experience, or specialized calibration experiments, and can be represented in the form of vectors, matrices, or tensors.

[0075] When acquiring research data to be audited, each data point (research data to be audited) is accompanied by information from the three source dimensions mentioned above: equipment model, batch number, and site identifier. For datasets with the same equipment model, batch number, and site identifier, the difference between each data point and its corresponding dataset mean (or gold standard data, i.e., a special sample or data point that has been measured by authoritative methods, has precise numerical knowledge, and can serve as an absolute reference benchmark) is calculated. This yields endogenous bias information, which is then combined with the preprocessed time-series observation dataset and the endogenous bias information of each research node to form a high-dimensional bias tensor dataset. ,in, For bias tensor data, This represents the number of equipment model categories. This refers to the batch number. The number of the site identifier. For dimensional space, This represents the dimensional space of equipment model, batch number, and site identifier. This bias tensor data aims to comprehensively capture and quantify the endogenous bias patterns of each research node at different time points and across different observation dimensions.

[0076] Specifically, when projection module 2 performs eigenvalue decomposition and orthogonal projection processing based on the bias tensor data to obtain the corrected tensor data corresponding to the research data to be audited, it executes: Eigenvalue decomposition is performed on the bias tensor data to obtain the eigenvalue matrix; The eigenvalue matrix is ​​subjected to orthogonal complementary projection to obtain the corrected tensor data corresponding to the research data to be audited.

[0077] During execution, projection module 2 performs feature decomposition on the deviation tensor data to break down complex deviation patterns into a series of independent feature components. This reveals the main structure and underlying patterns of the deviation data, resulting in a feature decomposition matrix. The feature decomposition matrix includes a core tensor, an equipment model factor matrix, a batch number factor matrix, and a site identifier factor matrix. The specific calculation process for feature decomposition is as follows: ; in, For the first Equipment model factor matrix for each deviation pollution item; For the first Batch number factor matrix of each deviation contamination item; For the first A matrix of site identification factors for each deviation pollution item; This represents the total number of deviation contamination items, i.e., the preset number of block items, whose value is determined based on physical priors (such as the number of physical component types, including equipment baseline items, batch drift items, and fast-changing signal items). This is the symbol for the outer product operator.

[0078] Orthogonal projection of the eigenvalue matrix can separate components related to endogenous bias from the original data, yielding the corrected tensor data corresponding to the research data to be audited. Orthogonal projection projects the data onto a subspace orthogonal to the bias feature space, effectively removing or suppressing the influence of endogenous bias, resulting in a purer and more accurate corrected tensor data that more realistically reflects the intrinsic characteristics of the research data.

[0079] The calculation process for orthogonal projection is as follows: ; in, To correct tensor data; It is the identity matrix; Equipment model factor matrix for all deviation contamination items Batch number factor matrix of all deviation contamination items Site identification factor matrix for all deviation pollution items The constructed subspace basis can be obtained by applying the equipment model factor matrix to all deviation contamination terms. Batch number factor matrix of all deviation contamination items Site identification factor matrix for all deviation pollution items After splicing, the spliced ​​matrix is ​​orthogonalized (i.e., QR decomposition or singular value decomposition (SVD)).

[0080] Specifically, when building the model based on the calibration tensor data to obtain the research data hypergraph model, module 3 executes the following: An initial hypergraph model is constructed by taking the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association between sites and sampling batches as edges. Based on the auditable scientific research data from different sites belonging to the same sampling batch (same batch number) in the calibration tensor data, covariance is calculated to obtain the empirical covariance matrix between sites belonging to the same sampling batch. Based on the maximum and minimum values ​​in the empirical covariance matrix, the covariance consistency is calculated to obtain the covariance consistency corresponding to the empirical covariance matrix. By defining the covariance consistency as the hyperedge weight of each edge in the initial hypergraph model, the research data hypergraph model is obtained.

[0081] When module 3 is executed, it constructs an initial hypergraph model using the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association relationships between sites and sampling batches as edges. This establishes a basic topological structure that reflects the fundamental relationships between research data in the spatial (site) and temporal (sampling batch) dimensions. Nodes can specifically refer to the physical location (site) of data collection or the time window (sampling batch) of data collection, while edges represent the pre-defined logical or physical connections between these sites and sampling batches. Each node's edges connect all sites within the same sampling batch, or all sampling batches within the same site.

[0082] Furthermore, based on the auditable research data from different sites belonging to the same sampling batch in the calibration tensor data, covariance is calculated to obtain the empirical covariance matrix among the sites within the same sampling batch. This matrix can quantify the synchronicity or correlation of the changing trends of research data provided by different sites under the same data collection batch. By calculating the covariance, the consistency or potential differences in the data behavior of each site can be revealed.

[0083] The specific process of calculating covariance is as follows: ; in, Let be the empirical covariance matrix, which is a A square matrix, where m is the number of stations; For time windows; This refers to the auditable research data (sorted by time series) from m sites acquired within a time window t.

[0084] Based on this, covariance consistency is calculated using the maximum and minimum values ​​in the empirical covariance matrix to obtain the covariance consistency corresponding to the empirical covariance matrix. Covariance consistency is a normalization or quantification process that transforms complex covariance relationships into a unified and comparable indicator to measure the level of consistency between data from different sites.

[0085] The calculation process for covariance consistency is as follows: ; in, For covariance consistency; This is the ratio of the maximum to the minimum value in the empirical covariance matrix; To prevent small amounts from being removed, settings can be configured according to actual needs.

[0086] Finally, the covariance consistency was determined as the hyperedge weights of each edge in the initial hypergraph model, resulting in the research data hypergraph model. The hyperedge weights reflect the strength or reliability of the associations between nodes. By assigning weights, the hypergraph model can more precisely express the complex dependencies between research data, with higher weights typically indicating stronger associations or higher data consistency.

[0087] Specifically, when the overlay module 4 coherently overlays the Laplacian matrix corresponding to the hypergraph model of the research data with the correction tensor data to obtain the aggregated feature data of the research data to be audited, it performs the following: Determine the Laplacian matrix corresponding to the hypergraph model of scientific research data; Extract a predetermined number of target non-zero eigenvalues ​​from the Laplacian matrix; The non-zero eigenvalues ​​of the target data are coherently superimposed with the correction tensor data to obtain high-quality converged features of the research data to be audited.

[0088] During execution, overlay module 4 calculates the corresponding Laplacian matrix based on the constructed hypergraph model of scientific research data. In graph theory, the Laplacian matrix is ​​the difference between the adjacency matrix and the degree matrix of a graph, reflecting its connectivity and structural information. For the hypergraph model, the construction of its Laplacian matrix takes into account the characteristics of hyperedges, such as calculating it through the hyperedge association matrix and hyperedge weight matrix, to capture the complex relationships between research nodes, sites, and sampling batches.

[0089] After obtaining the Laplacian matrix, its eigenvalues ​​are decomposed, and several eigenvalues ​​with the smallest non-zero values ​​are selected to obtain the target non-zero eigenvalues ​​(i.e., the minimum non-zero eigenvalues). These target non-zero eigenvalues ​​are usually associated with the connectivity, clustering structure, and low-frequency information of the graph, effectively capturing the intrinsic structure and latent patterns in the hypergraph model while filtering out high-frequency eigenvalues ​​related to noise or irrelevant information. The preset number of eigenvalues ​​can be adjusted according to the actual application scenario and data characteristics, for example, by using empirical values, cross-validation, or information criteria.

[0090] Optionally, when the overlay module 4 performs coherent overlay processing on the target non-zero eigenvalues ​​and the correction tensor data to obtain the aggregated feature data of the research data to be audited, it specifically performs the following: Based on the site, construct the target non-zero feature matrix corresponding to the target non-zero feature values; The corrected tensor data is projected onto the target non-zero feature matrix to obtain the manifold trend matrix; Based on a preset threshold for preventing the elimination of zero, the non-zero feature matrix of the target is subjected to weak signal amplification processing to obtain the diagonal weighted matrix corresponding to the non-zero feature matrix of the target. The aggregated feature data of the research data to be audited is obtained by performing aggregation processing based on the target non-zero feature matrix, manifold trend matrix and diagonal weighted matrix.

[0091] Specifically, during execution, the overlay module 4 extracts a preset number of target non-zero eigenvalues ​​from the Laplacian matrix, organizes and arranges them according to their respective sites, forming a matrix that can characterize the intrinsic structure and correlation of the data from each site. This target non-zero eigenvalue matrix can be regarded as a basis in a low-dimensional space, used to capture the main trends and intrinsic relationships of scientific research data under the hypergraph structure.

[0092] By using linear or nonlinear mapping, high-dimensional calibration tensor data is reduced to a low-dimensional manifold space defined by the target non-zero eigenvalue matrix, resulting in a manifold trend matrix. This projection process filters out noise and redundant information from the data, extracting the essential trends and structures of the data on the manifold, thus yielding a purer and more representative manifold trend matrix.

[0093] The calculation process of the manifold trend matrix is ​​as follows: ; in, This is the manifold trend matrix; For the target non-zero characteristic matrix, Let be a k×c matrix, where k is a preset quantity and c is the number of different physical quantities or different feature dimensions collected simultaneously at the same observation time and the same station. The preset quantity k is selected based on the eigenvalue spectral gap (eigengap) of the Laplace matrix, which is calculated as the difference between adjacent eigenvalues ​​in the non-zero feature matrix of the target. , (in, Let p+1 be the (p+1)th non-zero eigenvalue of the target in the non-zero eigenvalue matrix. (where p is the target non-zero eigenvalue in the target non-zero eigenvalue matrix), and select the one that makes p a target non-zero eigenvalue. The value of p when the maximum value is obtained is taken as the value of k. This selection method can automatically determine the most important manifold structure dimension in the hypergraph model and avoid subjective bias caused by manual setting.

[0094] Based on a preset threshold for preventing division by zero, the target non-zero feature matrix undergoes weak signal amplification processing to obtain a diagonal weighted matrix corresponding to the target non-zero feature matrix. This enhances important feature signals that might be overlooked in the original data due to their small amplitude. The construction of the diagonal weighted matrix allows each eigenvalue or its corresponding component to receive different weights based on its importance or signal strength. The threshold for preventing division by zero can be set according to actual needs to avoid division by zero in subsequent calculations, while ensuring that even weak signals receive appropriate weighting and amplification, thereby preventing the loss of important information.

[0095] The specific calculation process for the diagonal weighted matrix is ​​as follows: ; ; in, It is a diagonal weighted matrix; It is a diagonal matrix function; This is a function for amplifying weak signals. The k-th non-zero eigenvalue of the target; The preset threshold for preventing zero-value loss can be set according to actual needs.

[0096] Furthermore, convergence processing is performed based on the target non-zero feature matrix, manifold trend matrix, and diagonal weighted matrix to obtain converged feature data of the research data to be audited. This convergence processing can employ various mathematical operations, such as matrix multiplication, to effectively integrate the information after manifold projection and weak signal amplification processing, ultimately generating converged feature data that comprehensively, accurately, and robustly reflects the essential characteristics of the research data to be audited.

[0097] The specific calculation process for the aggregated feature data is as follows: ; in, To aggregate feature data.

[0098] Specifically, when calculation module 5 calculates the reputation value based on the aggregated feature data and obtains the credibility audit results of the research data to be audited for each research node, it executes: Alignment is calculated based on the converged feature data and its corresponding truncated mean to obtain global consensus feature data; Based on the calibration tensor data and the median of the calibration tensor data corresponding to each scientific research node, deviation anomaly calculation is performed to obtain the deviation anomaly degree. The on-time reporting rate and service availability rate are obtained by statistically analyzing the historical upload records of the aggregated characteristic data of each scientific research node. Based on global consensus feature data, deviation anomaly degree, on-time reporting rate and service availability rate, combined with preset calculation weights and preset forgetting factors, reputation value is calculated to obtain the reputation value corresponding to each scientific research node. Based on the tiered threshold range of the reputation value, the credibility audit results of the research data to be audited corresponding to each research node are determined.

[0099] During execution, calculation module 5 quantifies the deviation of each research node's data from the overall consensus by comparing the difference between the aggregated feature data and the truncated mean, thus obtaining the global consensus feature data. The truncated mean can be understood as the mean obtained after removing a portion of extreme values ​​(such as a preset percentage of the top and bottom rankings, which can be set according to actual needs) when calculating the average. This reduces the impact of outliers on the global consensus feature data, making it more representative. The global consensus feature data reflects the overall trend or standard of all research node data after removing the influence of outliers.

[0100] The specific formula for calculating the global consensus feature data is as follows: ; in, This is global consensus feature data; This represents the aggregated feature data of the i-th research node; The global truncated mean is obtained by processing the aggregated feature data of all research nodes in the same round of credibility audit after coordinate-level truncated mean processing (after sorting the aggregated feature data of all research nodes in the same round of credibility audit, removing the data of the top and bottom preset percentages, and then calculating the arithmetic mean of the remaining data to obtain the truncated mean). This refers to the aggregated feature data of the j-th research node in the same round of credibility audit.

[0101] Based on the median of the calibration tensor data and the corresponding calibration tensor data for each research node, deviation anomaly calculation is performed to assess the degree of deviation of each research node's data from its own median, thus obtaining the deviation anomaly score. The median, as a robust statistic, effectively resists the interference of outliers; therefore, calculating the deviation anomaly score based on the median can more accurately measure the anomalous behavior of individual research node data. A higher deviation anomaly score indicates a greater deviation of the node's data from the performance of other nodes, and its reliability may be lower.

[0102] The specific formula for calculating the deviation anomaly degree is as follows: ; in, The degree of deviation anomaly; It is a sigmoid function; is the Frobenius norm notation, representing the square root of the sum of squares of all elements; The median norm of the correction tensor data corresponding to all research nodes.

[0103] Based on the historical upload records of aggregated characteristic data from each research node, statistical analysis is performed to obtain the on-time reporting rate and service availability rate. The on-time reporting rate and service availability rate are important indicators for measuring the reliability of research node behavior. The on-time reporting rate refers to the proportion of research nodes that submit data within the stipulated time, reflecting the node's fulfillment capability. The service availability rate refers to the stability and continuity of the data services provided by research nodes, reflecting the node's reliability. These indicators are obtained by statistically analyzing the historical upload records of aggregated characteristic data from each research node, enabling an assessment of the research node's reputation from a behavioral perspective.

[0104] Furthermore, by combining preset calculation weights and a preset forgetting factor, the aforementioned multi-dimensional indicators (global consensus feature data, deviation anomaly, on-time reporting rate, and service availability) are comprehensively considered to calculate a reputation score. The preset calculation weights are used to adjust the importance of different indicators in the reputation score calculation; for example, data quality indicators can be assigned higher weights based on actual needs. The preset forgetting factor is used to attenuate the influence of historical data, ensuring that the reputation score more timely reflects the latest performance of research nodes and avoids outdated information unduly affecting the current reputation assessment. This factor can be set according to actual needs. Through this weighting and attenuation mechanism, a dynamic and comprehensive reputation score for research nodes can be obtained.

[0105] The specific formula for calculating the credit score is as follows: ; ; in, Let be the reputation value of the i-th research node in the n-th round of credibility audit; Let be the reputation value of the i-th research node in the (n-1)th round of credibility audit; The preset forgetting factor; The behavior score of the i-th research node in the n-th round of credibility audit; This represents the global consensus feature data for the i-th research node; This represents the deviation anomaly at the i-th research node; Let be the on-time reporting rate of the i-th research node; The service availability of the i-th research node; , , , The preset calculation weight corresponds to the i-th research node. , Let be the weight for calculating the global consensus feature data of the i-th research node. Let be the weight for calculating the deviation anomaly of the i-th research node. Let be the weight used to calculate the on-time reporting rate of the i-th research node. The weight for calculating the service availability of the i-th research node; when n=1, At this time It is also assigned a value of 0, starting from the second round (i.e., n=2). Restore to the default value.

[0106] In addition, to prevent long-term offline nodes from bringing back expired reputation values, a reputation value decay mechanism is introduced: if a research node fails to report aggregated feature data for multiple consecutive rounds (e.g., 5 rounds), its reputation value will be reduced. Multiply by the attenuation factor when re-launching. ( It can be set according to actual needs, such as 0.9), that is... ,in, The new (upon relaunch) reputation value. This is the old (before relaunching) reputation value. This mechanism ensures that the reputation value reflects the node's recent performance, avoiding a "one-and-done" accumulation of reputation.

[0107] Finally, the calculated reputation score is compared with a preset grading threshold range to determine the credibility audit result. For example, the reputation score can be divided into multiple levels such as "high credibility," "medium credibility," and "low credibility," with each level corresponding to a reputation score range. For instance, the reputation score range can be divided into... , and When the reputation value of a research node falls within a certain range, the credibility audit result of the corresponding research data to be audited can be determined. For example, when the reputation value falls within the range... When the credibility of the corresponding research data to be audited is determined, the aggregated feature data corresponding to the research data to be audited is directly incorporated into the global tensor; when the credibility value falls within the interval At this time, staff at the corresponding research node need to witness and sign for confirmation before the aggregated feature data corresponding to the research data to be audited can be directly incorporated into the global tensor; when the reputation value falls within the interval If a research node is deemed unreliable, its data for that round will not be adopted. Furthermore, if the reputation score remains below 0.3 for multiple consecutive rounds, the corresponding research node will be blacklisted and reported to the management center. This tiered mechanism makes the audit results more operational and interpretable.

[0108] As shown above, this scientific research data credibility auditing device acquires the research data to be audited from various research nodes, performs feature decomposition and orthogonal complementary projection processing on the research data to be audited to obtain the corresponding correction tensor data, constructs a model based on the correction tensor data to obtain a research data hypergraph model, and coherently superimposes the Laplacian matrix corresponding to the research data hypergraph model with the correction tensor data to obtain the aggregated feature data of the research data to be audited. Based on the aggregated feature data, a reputation value is calculated to obtain the credibility audit results for the research data to be audited corresponding to each research node. Thus, by correcting data deviations through feature decomposition and orthogonal complementary projection, constructing a hypergraph model to capture the inherent correlation of data, and combining it with the Laplacian matrix for high-quality feature aggregation, it ultimately achieves accurate auditing of the credibility of research data. This solves the problem that the existing research data aggregation process lacks a dynamic evaluation mechanism for the credibility of data sources, making it difficult to guarantee the credibility of the aggregation results and thus affecting research decisions. It can dynamically evaluate the data credibility of each research node, effectively identify low-quality or abnormally contributed data, and improve the accuracy of research data aggregation.

[0109] Please refer to Figure 3 , Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device includes a processor 301 and a memory 302. The processor 301 and the memory 302 are interconnected and communicate with each other via a communication bus 303 and / or other connection mechanisms (not shown). The memory 302 stores a computer program executable by the processor 301. When the electronic device is running, the processor 301 executes the computer program to perform the research data credibility auditing method in any optional implementation of the above embodiments, to achieve the following functions: acquiring research data to be audited from each research node; performing feature decomposition and orthogonal projection processing on the research data to be audited to obtain the correction tensor data corresponding to the research data to be audited; constructing a model based on the correction tensor data to obtain a research data hypergraph model; coherently superimposing the Laplacian matrix corresponding to the research data hypergraph model with the correction tensor data to obtain converged feature data of the research data to be audited; calculating the reputation value based on the converged feature data to obtain the credibility auditing result of the research data to be audited corresponding to each research node.

[0110] This application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it executes the scientific research data credibility auditing method in any optional implementation of the above embodiments to achieve the following functions: acquiring the scientific research data to be audited from each scientific research node; performing feature decomposition and orthogonal projection processing on the scientific research data to be audited to obtain the correction tensor data corresponding to the scientific research data to be audited; constructing a model based on the correction tensor data to obtain a scientific research data hypergraph model; performing coherent superposition processing on the Laplacian matrix corresponding to the scientific research data hypergraph model and the correction tensor data to obtain the convergent feature data of the scientific research data to be audited; calculating the reputation value based on the convergent feature data to obtain the credibility audit results of the scientific research data to be audited for each scientific research node. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), Magnetic Storage, Flash Memory, Disk or Optical Disk.

[0111] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0112] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0113] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0114] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0115] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for auditing the credibility of scientific research data, characterized in that, include: Obtain the research data to be audited at each research stage; The research data to be audited is subjected to feature decomposition and orthogonal projection processing to obtain the corrected tensor data corresponding to the research data to be audited. Based on the aforementioned calibration tensor data, a model is constructed to obtain a scientific research data hypergraph model; The Laplacian matrix corresponding to the hypergraph model of the scientific research data is coherently superimposed with the correction tensor data to obtain the aggregated feature data of the scientific research data to be audited. The credibility audit results of the research data to be audited are obtained by calculating the reputation value based on the aggregated feature data for each research node.

2. The scientific research data credibility auditing method according to claim 1, characterized in that, The research data to be audited is subjected to feature decomposition and orthogonal projection processing to obtain the corrected tensor data corresponding to the research data to be audited, including: The research data to be audited is preprocessed to obtain a time-series observation data set; Based on the endogenous bias information of each research node, tensors are constructed on the time series observation data set corresponding to the research data to be audited to obtain the bias tensor data corresponding to the research data to be audited. Based on the deviation tensor data, eigenvalue decomposition and orthogonal projection processing are performed to obtain the correction tensor data corresponding to the research data to be audited.

3. The scientific research data credibility auditing method according to claim 2, characterized in that, Based on the aforementioned deviation tensor data, eigenvalue decomposition and orthogonal projection processing are performed to obtain the corrected tensor data corresponding to the research data to be audited, including: The deviation tensor data is subjected to eigenvalue decomposition to obtain the eigenvalue decomposition matrix; The eigenvalue matrix is ​​subjected to orthogonal complementary projection to obtain the corrected tensor data corresponding to the research data to be audited.

4. The scientific research data credibility auditing method according to claim 1, characterized in that, Based on the aforementioned calibration tensor data, a model is constructed to obtain a scientific research data hypergraph model, including: An initial hypergraph model is constructed using the sites to which each research node belongs and the sampling batches corresponding to the research data to be audited as nodes, and the preset association relationships between the sites and the sampling batches as edges. Based on the auditable scientific research data from different sites belonging to the same sampling batch in the correction tensor data, covariance is calculated to obtain the empirical covariance matrix between the sites belonging to the same sampling batch. Based on the maximum and minimum values ​​in the empirical covariance matrix, the covariance consistency is calculated to obtain the covariance consistency corresponding to the empirical covariance matrix. The covariance consistency is determined as the hyperedge weight of each edge in the initial hypergraph model to obtain the scientific research data hypergraph model.

5. The scientific research data credibility auditing method according to claim 4, characterized in that, The Laplacian matrix corresponding to the hypergraph model of the scientific research data is coherently superimposed with the correction tensor data to obtain the aggregated feature data of the scientific research data to be audited, including: Determine the Laplacian matrix corresponding to the hypergraph model of the scientific research data; Extract a predetermined number of target non-zero eigenvalues ​​from the Laplacian matrix; The target non-zero eigenvalues ​​are coherently superimposed with the correction tensor data to obtain the aggregated feature data of the research data to be audited.

6. The scientific research data credibility auditing method according to claim 5, characterized in that, The target non-zero eigenvalues ​​are coherently superimposed with the correction tensor data to obtain the aggregated feature data of the research data to be audited, including: Based on the aforementioned sites, construct the target non-zero feature matrix corresponding to the target non-zero feature values; The corrected tensor data is projected onto the target non-zero feature matrix to obtain the manifold trend matrix; Based on a preset threshold for preventing the removal of zero, the target non-zero feature matrix is ​​subjected to weak signal amplification processing to obtain the diagonal weighted matrix corresponding to the target non-zero feature matrix. The convergence processing is performed based on the target non-zero feature matrix, the manifold trend matrix, and the diagonal weighted matrix to obtain the converged feature data of the scientific research data to be audited.

7. The scientific research data credibility auditing method according to claim 1, characterized in that, The credibility score is calculated based on the aggregated feature data to obtain the credibility audit results of the research data to be audited for each research node, including: Based on the converged feature data and its corresponding truncated mean, alignment is calculated to obtain global consensus feature data; Based on the correction tensor data and the median of the correction tensor data corresponding to each scientific research node, deviation anomaly calculation is performed to obtain the deviation anomaly degree. The on-time reporting rate and service availability rate are obtained by statistically analyzing the historical upload records of the aggregated feature data of each research node. Based on the global consensus feature data, the deviation anomaly degree, the on-time reporting rate, and the service availability rate, combined with the preset calculation weight and the preset forgetting factor, the reputation value is calculated to obtain the reputation value corresponding to each research node. Based on the tiered threshold range in which the reputation value falls, the credibility audit result of the research data to be audited corresponding to each research node is determined.

8. A scientific research data credibility auditing device, characterized in that, include: The acquisition module is used to acquire the research data to be audited at each research node; The projection module is used to perform feature decomposition and orthogonal projection processing on the research data to be audited to obtain the corrected tensor data corresponding to the research data to be audited. The construction module is used to construct a model based on the calibration tensor data to obtain a scientific research data hypergraph model; The overlay module is used to coherently overlay the Laplacian matrix corresponding to the hypergraph model of the scientific research data with the correction tensor data to obtain the convergent feature data of the scientific research data to be audited. The calculation module is used to calculate the reputation value based on the aggregated feature data to obtain the credibility audit result of the research data to be audited corresponding to each research node.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, and when the processor executes the computer program, it performs the steps in the scientific research data credibility auditing method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps in the scientific research data credibility auditing method as described in any one of claims 1-7.

Images