Knowledge graph generation method and system for risk control of scientific and technological projects

Abstract

The present invention provides a knowledge graph generation method and system for risk control of scientific and technological projects. By acquiring a multi-source heterogeneous data set of a target scientific and technological project, the structured indicator data is converted into a standard vector sequence, the unstructured text data is converted into a semantic vector sequence, and the time series behavior data is converted into a behavior pattern vector sequence through a heterogeneous data fusion mechanism. The three are then input into a risk quantification assessment model to generate a risk entity feature matrix and a risk association strength matrix. According to the entity feature vectors in the risk entity feature matrix, the node distribution topology of the knowledge graph is determined. According to the association strength values ​​in the risk association strength matrix, the entity relationship topology of the knowledge graph is determined, a dynamic knowledge graph of the target scientific and technological project is generated, and potential risk propagation paths in the dynamic knowledge graph are identified. The present invention can make the risk identification results have both the dynamic characteristics of real-time updating and the traceability of multi-dimensional risk characteristics.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs, and in particular to a method and system for generating a knowledge graph for risk control of scientific and technological projects. Background Art

[0002] With the deepening of digital transformation, risk control technologies for scientific and technological projects have gradually become a core means of ensuring scientific research safety and investment returns. Current mainstream risk control methods for scientific and technological projects typically employ structured data analysis models, extracting project indicators to construct risk assessment systems, or rely on expert experience to manually annotate unstructured documents for risks and generate risk analysis reports. However, traditional methods struggle to address the dynamic risk identification needs of complex scientific and technological projects, leading to lags in risk assessment and insufficient decision-making support capabilities. Summary of the Invention

[0003] The present invention provides a method and system for generating a knowledge graph for risk control of scientific and technological projects.

[0004] In the first aspect, an embodiment of the present invention provides a method for generating a knowledge graph for risk control of scientific and technological projects, the method comprising: obtaining a multi-source heterogeneous data set of a target scientific and technological project, the multi-source heterogeneous data set comprising structured indicator data, unstructured text data and time series behavior data; converting the structured indicator data into a standard vector sequence, converting the unstructured text data into a semantic vector sequence, and converting the time series behavior data into a behavior pattern vector sequence through a heterogeneous data fusion mechanism; inputting the standard vector sequence, semantic vector sequence and behavior pattern vector sequence into a risk quantification assessment model to generate a risk entity feature matrix and a risk association strength matrix; determining the node distribution topology of the knowledge graph based on the entity feature vectors in the risk entity feature matrix, and determining the entity relationship topology of the knowledge graph based on the association strength values ​​in the risk association strength matrix; generating a dynamic knowledge graph of the target scientific and technological project based on the node distribution topology and the entity relationship topology, and identifying potential risk propagation paths in the dynamic knowledge graph.

[0005] In a second aspect, an embodiment of the present invention provides a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the knowledge graph generation method for risk control of scientific and technological projects as described above.

[0006] The knowledge graph generation method based on scientific and technological project risk control provided by the present invention converts structured indicators, unstructured text and time series behavior data into a computable standardized vector sequence through the collaborative fusion mechanism of multi-source heterogeneous data, effectively eliminating the information island effect brought about by single data source analysis in traditional methods, and significantly improving the integrity of risk characteristics and the complementarity of data representation; using the risk quantification assessment model to jointly model the multimodal vector sequence, and synchronously generating the risk entity feature matrix and the risk association strength matrix through the cross-attention mechanism and two-way interactive calculation, realizing the collaborative optimization of entity attributes and association relationships, and overcoming the feature matching error caused by step-by-step modeling; based on dynamic knowledge The knowledge graph generation mechanism aligns and jointly encodes the node distribution topology and entity relationship topology in time and space, and constructs a dynamic graph structure with physical interpretability through density gradient-driven node cluster optimization and adaptive threshold segmentation edge connection pruning, so that the risk transmission path analysis can accurately capture the propagation law of potential risks in time and space dimensions; through the design of the present invention, the deep coupling of knowledge graph generation and risk transmission analysis is achieved, and the dynamic edge weight distribution and path integrity verification mechanism are used to make the risk identification results have the dynamic characteristics of real-time updates and maintain the traceability of multi-dimensional risk characteristics, comprehensively improving the intelligence level and decision-making support efficiency of scientific and technological project risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flow chart of a method for generating a knowledge graph for risk control of scientific and technological projects provided by an embodiment of the present invention;

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0010] See also Figure 1 , Figure 1 A flowchart of a knowledge graph generation method for scientific and technological project risk control provided in an embodiment of the present invention. The knowledge graph generation method for scientific and technological project risk control can be executed by a computer system and may include the following steps: Step S100: Obtain a multi-source heterogeneous data set for a target scientific and technological project, where the multi-source heterogeneous data set includes structured indicator data, unstructured text data, and time series behavior data.

[0011] In an embodiment of the present invention, the target scientific and technological project is a specific scientific and technological project that requires risk control and the construction of a knowledge graph. A multi-source heterogeneous data set indicates that the data comes from different sources and has different structures and formats. Structured indicator data is data with a clear structure and format, which can be expressed in a tabular form, such as the project's budget data, progress data, etc. These data can be organized and stored according to preset rules to facilitate statistics and analysis. Unstructured text data is text information without a fixed structure, such as project documents, reports, meeting minutes, etc. These text data may contain various descriptions and analyses of project risks, but due to their non-uniform format, special processing is required to extract useful information. Time series behavior data is a record of behavior related to time, such as the operating behavior of project team members, the operating status of the system, etc. These data are arranged in chronological order and reflect the behavioral characteristics of the project at different time points.

[0012] Step S200: Through the heterogeneous data fusion mechanism, the structured indicator data is converted into a standard vector sequence, the unstructured text data is converted into a semantic vector sequence, and the time series behavior data is converted into a behavior pattern vector sequence.

[0013] The heterogeneous data fusion mechanism integrates and converts data of varying structures and formats, aiming to transform multi-source heterogeneous data into a unified vector representation for subsequent analysis and processing. A standard vector sequence is a vector representation obtained by converting structured indicator data. It features unified dimensions and metrics, facilitating comparison and analysis in subsequent calculations. A semantic vector sequence is a vector obtained by processing unstructured text data. It reflects the semantic information of the text and helps identify risk factors embedded within it. A behavioral pattern vector sequence is derived by converting time-series behavioral data, reflecting the behavioral pattern characteristics of the project.

[0014] As an implementation method, step S200, through a heterogeneous data fusion mechanism, converts the structured indicator data into a standard vector sequence, converts the unstructured text data into a semantic vector sequence, and converts the time series behavior data into a behavior pattern vector sequence. Specifically, the following steps may be included: Step S210: Perform abnormal state detection on the structured indicator data, generate a data correction mask matrix based on the detection results, use the data correction mask matrix to interpolate and reconstruct the original indicator data, and output the corrected structured indicator data.

[0015] Anomaly detection is the process of identifying and marking outliers in structured indicator data. Outliers can be caused by data entry errors, system failures, and other factors. These outliers can affect subsequent data analysis and processing, necessitating detection and correction. The data correction mask matrix is ​​a binary matrix used to identify anomalous and normal data points in the original indicator data. Interpolation and reconstruction is the process of filling in and correcting anomalous data points based on information from normal data points.

[0016] When performing abnormal state detection, a method based on a statistical model can be used. Specifically, first obtain the historical indicator data of similar scientific and technological projects, calculate the mean vector and covariance matrix of each indicator dimension, and generate a multidimensional normal distribution model based on the mean vector and covariance matrix. The historical indicator data of similar scientific and technological projects can be obtained from the project database. By performing statistical analysis on these data, the mean and covariance of each indicator dimension can be obtained. The multidimensional normal distribution model is an existing probability distribution model that can describe the distribution of data in multiple dimensions. Then input the structured indicator data of the current project into the multidimensional normal distribution model to calculate the probability density value of each data point. When the probability density value is lower than the preset quantile of the distribution model, it means that the data point may be an outlier, and an abnormal marker sequence is generated. The preset quantile can be set according to the actual situation. For example, it can be set to 0.05, that is, when the probability density value of a data point is lower than 0.05, the data point is considered to be an outlier.

[0017] Based on the location information of outliers in the anomaly marker sequence, a binary mask matrix is ​​generated. Positions corresponding to outliers in the binary mask matrix are assigned a value of zero, while positions corresponding to normal points are assigned a value of one. This allows the binary mask matrix to clearly identify outliers in the original indicator data. To eliminate isolated outlier markers, a morphological closing operation is performed on the binary mask matrix. This morphological closing operation connects adjacent outliers to form connected regions. This closing operation prevents occasional outliers from being misclassified as true outliers. After generating the connected region correction mask matrix, the outlier data points are located in the original indicator data based on the anomaly regions marked by this matrix. The indicator values ​​of the adjacent data points within the time window before and after the outlier data point are extracted. The size of the time window can be adjusted based on the actual situation; for example, it can be set to three data points before and after the outlier data point.

[0018] The indicator values ​​of adjacent data points are fed into a sliding average filter to generate an interpolated filling sequence. The sliding average filter smooths the input data and generates new values ​​by calculating the average of adjacent data points. The zero positions of the mask matrix are corrected based on the connected regions. The interpolated filling sequence replaces the original outlier data points, and the replaced structured indicator data is output as the corrected structured indicator data. This effectively corrects outliers in the structured indicator data, improving data quality and reliability.

[0019] As an implementation method, step S210 performs abnormal state detection on the structured indicator data, generates a data correction mask matrix based on the detection results, uses the data correction mask matrix to interpolate and reconstruct the original indicator data, and outputs the corrected structured indicator data. Specifically, the following steps may be included: Step S211: Obtain historical indicator data of similar scientific and technological projects, calculate the mean vector and covariance matrix of each indicator dimension, and generate a multidimensional normal distribution model based on the mean vector and covariance matrix.

[0020] The historical indicator data of similar scientific and technological projects refer to the indicator data recorded in the past period of time by other scientific and technological projects with similar nature, scale and business fields as the current target scientific and technological project. The mean vector is a vector composed of the average values ​​of all data points in each indicator dimension. For a data set with n indicator dimensions, the mean vector can be expressed as an n-dimensional vector, in which each element is the average value of all data points in the corresponding indicator dimension. The method to calculate the mean vector is to add all the data points in each indicator dimension and then divide it by the total number of data points. The covariance matrix is ​​an n×n matrix used to describe the correlation between the various indicator dimensions. The elements of the covariance matrix are Represents the covariance between the i-th indicator dimension and the j-th indicator dimension. The calculation formula of covariance is: ;in, and Respectively represent the values ​​of the k-th data point in the i-th and j-th indicator dimensions, and are the means of the i-th and j-th indicator dimensions respectively.

[0021] The multidimensional normal distribution model can be used to describe the joint distribution between multiple variables. In an embodiment of the present invention, the multidimensional normal distribution model is generated using the calculated mean vector and covariance matrix. The probability density function of the multidimensional normal distribution can be expressed as: ; Where x is an n-dimensional vector, representing the value of the data point; is the mean vector; is the covariance matrix; is the determinant of the covariance matrix; is the inverse matrix of the covariance matrix. Through this multidimensional normal distribution model, we can calculate the probability density of the structured indicator data of the current project and determine whether the data point is abnormal.

[0022] Step S212: Input the structured indicator data of the current project into the multidimensional normal distribution model, calculate the probability density value of each data point, and generate an abnormal mark sequence when the probability density value is lower than the preset quantile of the distribution model.

[0023] The structured indicator data of the current project refers to the indicator data with a clear structure and format collected from the target scientific and technological project, such as project progress data and cost data. This data is input into the multidimensional normal distribution model generated previously, and the probability density value of each data point is calculated based on the probability density function of the multidimensional normal distribution. The probability density value represents the probability of a data point appearing in the multidimensional normal distribution. For an n-dimensional data point x, its probability density value can be calculated by substituting x into the probability density function f(x) of the multidimensional normal distribution. The preset quantile of the distribution model is a pre-set threshold used to determine whether a data point is abnormal. The quantile divides the probability distribution into a number of equal numerical points.

[0024] When the calculated probability density value of a data point is lower than the preset quantile, the data point is marked as an outlier. By traversing all data points, an outlier label sequence is generated. The outlier label sequence has the same length as the structured indicator data, where each element corresponds to a data point. If the data point is an outlier, it is marked as 1, otherwise it is marked as 0.

[0025] Step S213: Generate a binary mask matrix based on the position information of the abnormal points in the abnormal mark sequence, where the positions corresponding to the abnormal points in the binary mask matrix take a value of zero, and the positions corresponding to the normal points take a value of one.

[0026] The anomaly labeling sequence records whether each data point in the structured indicator data is an anomaly. Based on the location of the anomaly points in this sequence, a binary mask matrix can be generated. The binary mask matrix has the same dimensions as the structured indicator data and is used to mark the abnormal and normal states of the data points using 0s and 1s.

[0027] Assume that the structured indicator data is an m×n matrix, where m represents the number of data points and n represents the number of indicator dimensions. The anomaly marker sequence is a one-dimensional sequence of length m. For each element in the anomaly marker sequence, if its value is 1, it indicates that the data point is an anomaly, and the corresponding row and column position in the binary mask matrix is ​​zero; if its value is 0, it indicates that the data point is a normal point, and the corresponding row and column position in the binary mask matrix is ​​one.

[0028] Step S214: performing a morphological closing operation on the binary mask matrix to eliminate isolated abnormal markers and generate a connected region correction mask matrix.

[0029] In this embodiment of the present invention, the binary mask matrix is ​​considered a binary image, where 0 represents an outlier (black pixel) and 1 represents a normal pixel (white pixel). The morphological closing process includes two basic operations: dilation and erosion. The dilation operation expands the white regions (normal pixels) in the binary mask matrix outward, connecting adjacent white regions; the erosion operation contracts the white regions inward, removing small, isolated white regions. By performing the dilation operation followed by the erosion operation, isolated outlier markers in the binary mask matrix can be eliminated and adjacent outlier points can be connected into connected regions.

[0030] Specifically, the dilation operation can be performed by defining a structural element (such as a 3×3 square), sliding the structural element on the binary mask matrix, and setting the pixel value at a position in the matrix to 1 when the structural element completely overlaps with the position in the matrix. The erosion operation sets the pixel value at that position to 1 only when the structural element is completely contained in the white area.

[0031] Through morphological closing operations, the original isolated abnormal markers are eliminated and a connected region correction mask matrix is ​​generated. This matrix can more accurately reflect the true abnormal regions in the structured indicator data, providing more reliable information for subsequent interpolation and reconstruction.

[0032] Step S215: Correct the abnormal area marked by the mask matrix based on the connected area, locate the abnormal data point in the original indicator data, and extract the indicator values ​​of the adjacent data points in the time window before and after the abnormal data point.

[0033] The connected component correction mask matrix identifies the locations of anomalous regions in structured indicator data. Using this matrix, anomalous data points can be accurately located within the raw indicator data. Raw indicator data refers to structured indicator data collected before any processing. The time window is a predefined range used to determine the number of adjacent data points before and after an anomalous data point. For example, a time window can be set to three data points before and after each. For each anomalous data point, the indicator values ​​of the adjacent data points within the time window before and after it are extracted.

[0034] Assume that the original indicator data is a time series containing indicator values ​​at multiple time points. For an outlier at time point i, based on the time window settings, extract the indicator values ​​of the adjacent data points from time points i-3 to i-1, and from time points i+1 to i+3. By extracting the indicator values ​​of the adjacent data points within the time window before and after the outlier, the information from these normal data points can be used to interpolate and reconstruct the outlier, thereby correcting the outlier data.

[0035] Step S216: Input the index values ​​of the adjacent data points into the sliding average filter to generate an interpolation filling sequence, correct the zero value position of the mask matrix according to the connected area, and replace the original abnormal data points with the interpolation filling sequence.

[0036] The sliding average filter is used to smooth time series data. The basic principle is to calculate the average value of the data within a window of fixed length and use this average value as the new value at the center of the window. The indicator values ​​of the adjacent data points extracted previously are input into the sliding average filter. When calculating the interpolation filling sequence, starting from the first element of the sequence, the average value of the data within each window is calculated in sequence. The zero-value positions of the connected component correction mask matrix correspond to abnormal data points in the original indicator data. Based on these zero-value positions, the original abnormal data points are replaced with values ​​from the interpolation filling sequence. In this way, the abnormal data points are corrected using information from adjacent normal data points, improving the quality of the structured indicator data.

[0037] Step S217: outputting the replaced structured index data as the corrected structured index data.

[0038] After replacing outliers in the original indicator data, the replaced structured indicator data is output as the revised structured indicator data. The revised structured indicator data removes outliers and more accurately reflects the actual situation of the target technology project, providing a reliable data foundation for subsequent analysis and processing.

[0039] Step S220: input the corrected structured indicator data into the indicator encoder, map the indicator values ​​of different dimensions into a unified metric space through a feature dimension alignment algorithm, and generate a standard vector sequence.

[0040] The indicator encoder is used to encode structured indicator data, converting the corrected structured indicator data into a vector representation that is easier to process and analyze. The feature dimension alignment algorithm is a method used to address the dimensional inconsistency of indicator values. It maps indicator values ​​of different dimensions into a unified metric space, making them comparable. Indicator values ​​of different dimensions refer to indicators in structured indicator data that use different measurement units and scales. The feature dimension alignment algorithm can be implemented in various ways, such as normalization. Normalization converts the value of each indicator to a value with a mean of 0 and a standard deviation of 1. Specifically, for each indicator dimension, the mean and standard deviation are first calculated, and then each data point is normalized. After feature dimension alignment, the converted values ​​of each data point in each indicator dimension are combined to form a vector. In this way, the corrected structured indicator data is converted into a sequence of standard vectors, where each vector corresponds to a data point and the dimensionality of the vector equals the number of indicator dimensions.

[0041] Step S230: Perform semantic segmentation on the unstructured text data to identify text paragraphs containing risk factor descriptions, input the text paragraphs into a pre-trained semantic encoder to generate primary semantic vectors, perform spatial transformation on the primary semantic vectors through a context-aware network, and output a semantic vector sequence.

[0042] Semantic segmentation divides unstructured text data into multiple meaningful segments based on semantics and structure. Unstructured text data, such as project documents and reports, consists of large amounts of text without a fixed format or structure. Semantic segmentation can break this text data into smaller, more manageable units.

[0043] Semantic segmentation can be performed using sentence segmentation and paragraph division algorithms in natural language processing. For example, based on punctuation and grammatical rules, the text can be segmented into sentences, and then related sentences can be grouped into paragraphs based on the semantic relevance between the sentences.

[0044] Identifying text paragraphs containing risk factor descriptions involves filtering out those paragraphs containing information related to project risks from the segmented text paragraphs. This can be achieved through methods such as keyword matching and semantic analysis. For example, pre-define risk-related keywords such as "risk," "crisis," and "uncertainty," and then search for the occurrence of these keywords in each text paragraph. If a paragraph contains these keywords and their contextual semantics are relevant to project risk, the paragraph is considered to contain a risk factor description.

[0045] A pretrained semantic encoder is a model that has been pretrained on large-scale text data and can convert input text paragraphs into vector representations. For example, the BERT (Bidirectional Encoder Representations from Transformers) model is a pretrained semantic encoder. A text paragraph containing risk factor descriptions is fed into the pretrained semantic encoder, which extracts features and transforms the text to generate a primary semantic vector. This primary semantic vector is a vector representation of the text paragraph that reflects the text's semantic information.

[0046] A context-aware network is a neural network model that takes into account text context. It can perform spatial transformations on primary semantic vectors to further explore semantic information and contextual relationships within the text. For example, a common LSTM (Long Short-Term Memory) network or Gated Recurrent Unit (GRU) network can be used as a context-aware network. When a primary semantic vector is input into the context-aware network, the network adjusts and transforms it based on the text context and outputs a sequence of semantic vectors.

[0047] Step S240: Perform event window division on the time series behavior data, extract the type distribution characteristics and time interval characteristics of the behavior operations in the window, perform orthogonal projection transformation on the type distribution characteristics and time interval characteristics, and generate a behavior pattern vector sequence.

[0048] Event windowing divides time series behavior data into multiple windows of fixed lengths according to chronological order. Time series behavior data is a record of time-related behaviors, such as the operational behavior of project team members and the operating status of the system. Event windowing can be used to segment continuous time series data into multiple discrete windows, facilitating analysis of the behavior data within each window.

[0049] The time interval feature is the time interval information between adjacent behavioral operations within an event window. Statistics such as the mean and standard deviation of the time interval between adjacent behavioral operations can be calculated as the time interval feature.

[0050] In a non-essential, optional embodiment, orthogonal projection transformation is a method for fusing and transforming type distribution features and time interval features. It projects these two different types of features into a new vector space, generating a sequence of behavioral pattern vectors. The basis vector set for the orthogonal projection transformation is determined by the singular value decomposition results of the covariance matrix of the standard vector sequence and the semantic vector sequence.

[0051] Specifically, we first calculate the covariance matrix of the standard vector sequence and the semantic vector sequence, then perform singular value decomposition on the covariance matrix to obtain a set of basis vectors. The feature vectors composed of type distribution features and time interval features are projected onto these basis vectors to obtain a sequence of behavior pattern vectors.

[0052] Step S300: Input the standard vector sequence, semantic vector sequence and behavior pattern vector sequence into the risk quantification assessment model to generate a risk entity feature matrix and a risk association strength matrix.

[0053] The risk quantification assessment model is used to quantitatively analyze and evaluate the risks of scientific and technological projects. It comprehensively considers the information contained in standard vector sequences, semantic vector sequences, and behavioral pattern vector sequences to mine risk entities and risk associations within a project.

[0054] The standard vector sequence is a vector sequence derived from encoding and aligning the modified structured indicator data. It reflects the structured indicator characteristics of the project. The semantic vector sequence is a vector sequence derived from processing unstructured text data. It contains the risk semantic information in the text. The behavioral pattern vector sequence is a vector sequence derived from processing and transforming time-series behavioral data. It reflects the behavioral pattern characteristics of the project.

[0055] After inputting these three vector sequences into the risk quantitative assessment model, the model analyzes and processes them to generate a risk entity feature matrix and a risk correlation strength matrix. The risk entity feature matrix is ​​a matrix in which each row corresponds to a risk entity, each column corresponds to a feature dimension, and the matrix elements represent the value of the risk entity on that feature dimension. A risk entity can be a task, resource, or phase in a project. The risk correlation strength matrix is ​​a matrix used to represent the strength of correlations between different risk entities, with the matrix elements representing the degree of association between two risk entities.

[0056] For example, suppose the risk quantification assessment model is based on a neural network and consists of an input layer, a hidden layer, and an output layer. The input layer uses a concatenation of standard vector sequences, semantic vector sequences, and behavioral pattern vector sequences. After feature extraction and transformation in the hidden layer, the output layer outputs a risk entity feature matrix and a risk association strength matrix, respectively.

[0057] As an implementation method, step S300 inputs the standard vector sequence, semantic vector sequence and behavior pattern vector sequence into the risk quantification assessment model to generate a risk entity feature matrix and a risk association strength matrix, which may specifically include the following steps: Step S310: Perform cross-attention calculation on the standard vector sequence and the semantic vector sequence to generate a cross-attention weight matrix of indicators and semantics, where each element of the cross-attention weight matrix reflects the association strength between the indicator vector and the semantic vector.

[0058] The standard vector sequence is a vector sequence converted from structured indicator data, while the semantic vector sequence is a vector sequence converted from unstructured text data. The cross-attention weight matrix is ​​a two-dimensional matrix with the number of rows equal to the number of vectors in the standard vector sequence and the number of columns equal to the number of vectors in the semantic vector sequence. Each element in the matrix represents the strength of the association between an indicator vector and a semantic vector.

[0059] The cross-attention calculation process can be implemented through the following steps: First, define the query (Query), key (Key) and value (Value) for the standard vector sequence and the semantic vector sequence respectively. For example, the standard vector sequence can be used as the query and the semantic vector sequence as the key and value. Then, calculate the similarity between the query vector and the key vector. The similarity calculation method can use the dot product operation. Then, the calculated similarity value is normalized, such as the softmax function. The role of the softmax function is to convert the similarity value into a probability distribution so that the sum of all elements is 1. For each query vector q i , calculate the softmax value of its similarity with all key vectors: ; Where n is the number of vectors in the semantic vector sequence. attn ij It is the element in the i-th row and j-th column of the cross attention weight matrix, which reflects the indicator vector q i With the semantic vector k j The strength of the association between them.

[0060] Step S320: Perform weighted aggregation on the semantic vector sequence based on the cross-attention weight matrix to generate a semantically enhanced indicator feature sequence, and multiply the semantically enhanced indicator feature sequence with the standard vector sequence element by element to generate a cross-fusion feature tensor.

[0061] Weighted aggregation of the semantic vector sequence based on the cross-attention weight matrix involves performing a weighted summation of each vector in the semantic vector sequence according to the weight values ​​in the cross-attention weight matrix. Each element in the cross-attention weight matrix represents the strength of the association between the indicator vector and the semantic vector. Through weighted aggregation, the information in the semantic vector sequence can be integrated into the indicator vector, generating a semantically enhanced indicator feature sequence.

[0062] The calculated semantic enhancement indicator feature sequence is element-wise multiplied with the standard vector sequence to generate a cross-fusion feature tensor. Element-wise multiplication involves multiplying each vector in the semantic enhancement indicator feature sequence with each element of the vector at the corresponding position in the standard vector sequence.

[0063] Step S330: Expand the cross-fused feature tensor into a two-dimensional matrix along the feature dimension, and perform time alignment operation with the behavior pattern vector sequence, adjust the row and column distribution of the two-dimensional matrix according to the timestamp information of the behavior pattern vector sequence, and generate a time-aligned feature matrix.

[0064] Expanding the cross-fused feature tensor along the feature dimension into a two-dimensional matrix converts the three-dimensional cross-fused feature tensor into a two-dimensional matrix. For example, a cross-fused feature tensor has three dimensions: sample dimension, time dimension, and feature dimension. By expanding along the feature dimension, the feature information in the tensor is flattened to form a two-dimensional matrix.

[0065] The time alignment operation matches and aligns the expanded two-dimensional matrix with the behavioral pattern vector sequence in terms of time. The behavioral pattern vector sequence contains timestamp information. Based on this timestamp information, the row and column distribution of the two-dimensional matrix is ​​adjusted to ensure that the data in the two-dimensional matrix and the data in the behavioral pattern vector sequence are aligned in time. After the time alignment operation, a time-aligned feature matrix is ​​generated. The data in this matrix is ​​temporally consistent with the behavioral pattern vector sequence, facilitating further analysis and processing.

[0066] Step S340: Perform an outer product operation on the time-aligned feature matrix and the behavior pattern vector sequence to generate an initial interaction matrix, and perform channel attention weighting on the initial interaction matrix to calculate the contribution of each channel dimension to the risk association and generate a channel weight vector.

[0067] The outer product operation is used to calculate the interaction information between two matrices. In an embodiment of the present invention, the outer product operation is performed on the time-aligned feature matrix and the behavior pattern vector sequence to obtain an initial interaction matrix. The initial interaction matrix reflects the interaction relationship between the time-aligned feature matrix and the behavior pattern vector sequence. Channel attention weighting is a method for calculating the contribution of each channel dimension to risk association. The channel dimension is the third dimension of the initial interaction matrix. Through channel attention weighting, a weight can be assigned to each channel dimension to reflect the importance of the channel dimension to the risk association.

[0068] Specifically, a global average pooling operation is first performed on each channel dimension of the initial interaction matrix to obtain the average eigenvalue of each channel. These average eigenvalues ​​are then input into a fully connected layer and processed by an activation function (such as the sigmoid function) to obtain a channel weight vector. The length of the channel weight vector is equal to the number of channel dimensions in the initial interaction matrix, and each element represents the weight of the corresponding channel dimension.

[0069] In one embodiment, step S340, the time-aligned feature matrix is ​​outer-producted with the behavior pattern vector sequence to generate an initial interaction matrix, which can specifically include the following steps: Step S341: According to the row and column distribution of the time-aligned feature matrix, the time-aligned feature matrix is ​​divided into multiple feature sub-blocks, each feature sub-block corresponds to a behavior pattern vector in the same time window in the behavior pattern vector sequence.

[0070] The time-aligned feature matrix is ​​obtained after the time alignment operation. Its row and column distribution reflects data information at different time points and across different feature dimensions. Based on the row and column distribution of the time-aligned feature matrix, it is divided into multiple feature sub-blocks. Each feature sub-block is a subset of the time-aligned feature matrix and corresponds to the behavior pattern vectors in the same time window in the behavior pattern vector sequence.

[0071] For example, assume the time-aligned feature matrix has dimensions m × n, and the behavior pattern vector sequence is divided into k windows based on the time window, with each window length being w. The time-aligned feature matrix can then be partitioned based on the time window length to obtain k feature sub-blocks, each with dimensions w × n. Each feature sub-block corresponds to a behavior pattern vector within a time window in the behavior pattern vector sequence.

[0072] Step S342: perform outer product operations on each feature sub-block and the behavior pattern vector of the corresponding time window to generate a sub-block interaction matrix.

[0073] Perform outer product operations on each feature sub-block with the behavior pattern vector of the corresponding time window. The outer product operation process is similar to the outer product operation described above. For each element in the feature sub-block and each element in the behavior pattern vector of the corresponding time window, their product is calculated to obtain the sub-block interaction matrix.

[0074] Step S343: performing an edge filling operation on the sub-block interaction matrix to keep its dimension consistent with the time-aligned feature matrix, thereby generating a filled sub-block interaction matrix.

[0075] The edge padding operation adds extra elements to the edges of the sub-block interaction matrix to keep its dimensions consistent with the time-aligned feature matrix. Since the sub-block interaction matrix is ​​obtained by performing the outer product operation on the feature sub-blocks and the behavior pattern vector, its dimensions may be different from the time-aligned feature matrix. Through the edge padding operation, the dimensions of the sub-block interaction matrix can be expanded to the same as the time-aligned feature matrix, facilitating subsequent splicing operations. Possible edge padding methods include zero padding, which adds zero elements to the edges of the sub-block interaction matrix. For example, assuming the dimensions of the sub-block interaction matrix are w×n×p and the dimensions of the time-aligned feature matrix are m×n, zero elements can be added above, below, left, and right to make the dimensions of the sub-block interaction matrix m×n×p.

[0076] Step S344: concatenate the filled sub-block interaction matrices in the order of the time windows to generate an interaction tensor of the complete time dimension.

[0077] After the edge filling operation, the sub-block interaction matrices are concatenated in the order of the time windows. Since each sub-block interaction matrix corresponds to a time window, concatenating them in time order can obtain an interaction tensor with a complete time dimension.

[0078] Step S345: Perform channel dimension compression on the interaction tensor, map the multi-dimensional channel into a single correlation strength channel through a summation operation, and generate an initial interaction matrix.

[0079] Channel dimension compression compresses the multiple channel dimensions of the interaction tensor into a single channel dimension. By summing the multiple channel elements at each position of the interaction tensor, the multi-dimensional channels are mapped into a single correlation strength channel to obtain the initial interaction matrix.

[0080] Step S346: Perform a residual connection on the initial interaction matrix and the time-aligned feature matrix, retain the original time-aligned feature information, and generate an enhanced initial interaction matrix.

[0081] Residual connections are used to address vanishing gradients and model degradation. In this embodiment of the present invention, a residual connection is performed between the initial interaction matrix and the time-aligned feature matrix. This involves adding the corresponding elements of these two matrixes to produce an enhanced initial interaction matrix. This enhanced initial interaction matrix incorporates the interaction information from the initial interaction matrix while retaining the information from the original time-aligned feature matrix. This prevents the loss of important feature information during the calculation process and improves model performance.

[0082] Step S350: Perform a Hadamard product operation on the channel weight vector and the initial interaction matrix to generate a weighted interaction matrix, and input the weighted interaction matrix into the bidirectional recurrent neural network. The forward propagation path captures the historical dependency features, and the backward propagation path captures the potential correlation features to generate a bidirectional fusion feature matrix.

[0083] The Hadamard product multiplies the elements of two matrices at corresponding positions. In this embodiment of the present invention, a Hadamard product is performed on the channel weight vector and the initial interaction matrix to obtain a weighted interaction matrix. The channel weight vector reflects the contribution of each channel dimension to the risk association. The Hadamard product can be used to assign a weight to each element in the initial interaction matrix, highlighting important channel information.

[0084] A bidirectional recurrent neural network (BRNN) is a neural network model that simultaneously considers both past and future information about sequence data. It consists of a forward propagation path and a backward propagation path. The forward propagation path starts at the beginning of the sequence and processes each element in turn, capturing historical dependencies. The backward propagation path starts at the end of the sequence and processes each element in reverse order, capturing potential correlations. A weighted interaction matrix is ​​input into the bidirectional recurrent neural network, which then processes the sequence data in the weighted interaction matrix. During forward propagation, the network predicts the current element based on information from previous elements. During backward propagation, the network predicts the current element based on information from subsequent elements. Finally, the outputs of the forward and backward propagation paths are combined to generate a bidirectional fused feature matrix.

[0085] Step S360: Perform a strided convolution operation on the bidirectional fusion feature matrix to extract the multi-scale spatiotemporal correlation pattern, and add the weighted interaction matrix to the strided convolution output through a skip connection to generate a correlation strength matrix.

[0086] Strided convolutions skip elements during the convolution process, expanding the receptive field of the convolution kernel and extracting multi-scale spatiotemporal correlation patterns. The bidirectional fusion feature matrix contains both historical dependency features and latent correlation features. Strided convolutions can further mine spatiotemporal correlation information at different scales within the matrix.

[0087] The process of the strided convolution operation is as follows: define a convolution kernel, and the size and step size of the convolution kernel are pre-set. Then the convolution kernel is slid on the bidirectional fusion feature matrix, and the step size of each slide is determined by the step size parameter. At each sliding position, the sum of the product of the convolution kernel and the elements of the corresponding area of ​​the matrix is ​​calculated to obtain the convolution output. Jump connection is a method of connecting the outputs of different layers. In an embodiment of the present invention, the weighted interaction matrix is ​​added to the strided convolution output through a jump connection. In this way, the original information in the weighted interaction matrix can be fused with the multi-scale spatiotemporal correlation information extracted by the strided convolution to generate a correlation strength matrix.

[0088] Step S370: Input the association strength matrix into the feature distillation network, extract the local association pattern through the separable convolution layer, compress the redundant features through the pooling layer, and output the risk association strength matrix.

[0089] The feature distillation network is a network model used to extract and compress features. It can further process the correlation strength matrix, extract features from it, and compress redundant information.

[0090] A separable convolution layer is a convolution layer that decomposes the traditional convolution operation into two steps: depthwise separable convolution and pointwise convolution. Depthwise separable convolution performs a convolution operation on each input channel separately to extract local correlation patterns; pointwise convolution performs a linear combination of the outputs of the depthwise separable convolution to generate the final output. Through the separable convolution layer, local correlation patterns in the correlation strength matrix can be effectively extracted while reducing the amount of calculation. The pooling layer can reduce the number of features and remove redundant information. Pooling operations include, for example, maximum pooling and average pooling. In an embodiment of the present invention, the output of the separable convolution layer is processed by the pooling layer to compress redundant features.

[0091] The correlation strength matrix is ​​fed into the feature distillation network. It first passes through a separable convolutional layer to extract local correlation patterns, then through a pooling layer to compress redundant features. Finally, the output is the risk correlation strength matrix. This further processed and compressed matrix more concisely represents the correlation strength between different risk entities.

[0092] Step S380: Input the cross-fused feature tensor into the feature encoder, expand the feature receptive field through the dilated convolution layer, and filter the output results of the dilated convolution layer based on the feature selection weights generated by the risk association strength matrix to generate a risk entity feature matrix.

[0093] The cross-fused feature tensor is input into the feature encoder for further processing and feature extraction. A dilated convolution layer is a convolution layer that inserts holes between the elements of the convolution kernel, thereby expanding the receptive field of the convolution kernel. This allows for the extraction of a wider range of feature information without increasing the size of the convolution kernel. For example, a traditional convolution kernel is 3×3 in size. A dilated convolution layer can insert holes between the convolution kernel elements, expanding the receptive field to 5×5 or larger. Feature selection weights generated based on the risk association strength matrix assign a weight to each feature in the cross-fused feature tensor based on the element values ​​in the risk association strength matrix. These weights reflect the importance of each feature to the risk entity. The feature selection weights can be used to filter the output of the dilated convolution layer, retaining important features and removing unimportant ones.

[0094] Specifically, feature selection weights are calculated based on the risk association strength matrix. Normalization can be used to convert the element values ​​in the risk association strength matrix into feature selection weights. The output of the dilated convolutional layer is then element-wise multiplied by the feature selection weights to obtain filtered features. Finally, these filtered features are combined into the risk entity feature matrix.

[0095] Step S400: Determine the node distribution topology of the knowledge graph based on the entity feature vectors in the risk entity feature matrix, and determine the entity relationship topology of the knowledge graph based on the association strength values ​​in the risk association strength matrix.

[0096] In an embodiment of the present invention, the node distribution topology of the knowledge graph is determined according to the entity feature vectors in the risk entity feature matrix, and the entity relationship topology of the knowledge graph is determined according to the association strength value in the risk association strength matrix.

[0097] The entity feature vector in the risk entity feature matrix is ​​a characteristic description of each risk entity, which contains various attributes and characteristic information of the risk entity. By processing and analyzing these entity feature vectors, the distribution of nodes in the knowledge graph can be determined. The node distribution topology reflects the position and layout of nodes in the knowledge graph, which can help us intuitively understand the relative relationships between risk entities. The association strength values ​​in the risk association strength matrix represent the degree of association between different risk entities. Based on these association strength values, the relationship between entities in the knowledge graph, including the type and strength of the relationship, can be determined. The entity relationship topology reflects the connection method and relationship strength between entities in the knowledge graph, which can help us analyze the spread and impact of risks between different entities.

[0098] As an implementation method, in step S400, the node distribution topology of the knowledge graph is determined based on the entity feature vectors in the risk entity feature matrix, which can specifically include the following steps: Step S410: perform nonlinear dimensionality reduction processing on each entity feature vector in the risk entity feature matrix to generate a low-dimensional embedding vector, and map the low-dimensional embedding vector to a topological space coordinate system.

[0099] The goal of nonlinear dimensionality reduction is to reduce the dimensionality of data while preserving its key features, facilitating subsequent visualization and analysis. In an embodiment of the present invention, nonlinear dimensionality reduction is performed on each entity feature vector in the risk entity feature matrix. Examples of nonlinear dimensionality reduction methods include t-distributed stochastic neighbor embedding (t-SNE) and autoencoders. Taking t-SNE as an example, it constructs probability distributions between data points in both high-dimensional and low-dimensional spaces, then finds the optimal embedding of the high-dimensional data in the low-dimensional space by minimizing the difference between these two distributions. Specifically, in the high-dimensional space, a Gaussian distribution is used to measure the similarity between data points; in the low-dimensional space, a t-distribution is used to measure similarity. Through continuous iterative optimization, similar data points in the high-dimensional space remain similar in the low-dimensional space. The resulting low-dimensional embedding vector is then mapped to a topological coordinate system. The topological coordinate system can be a two-dimensional plane coordinate system or a three-dimensional spatial coordinate system.

[0100] Step S420: Based on the embedded vector distribution density in the topological space coordinate system, the spatial grid units are divided, and the density gradient value of the entity feature vector in each spatial grid unit is calculated.

[0101] After mapping the low-dimensional embedding vectors to a topological space coordinate system, the space is partitioned based on the distribution density of these embedding vectors. Spatial grid cells divide the topological space into several small regions, each of which is a spatial grid cell. The partitioning method can be selected based on the actual situation, for example, a regular rectangular grid or a hexagonal grid can be used.

[0102] Calculate the density of entity eigenvectors within each spatial grid cell. Density can be defined as the ratio of the number of entity eigenvectors within a grid cell to the volume of the grid cell. For example, in a two-dimensional plane, where the grid cell is a rectangle, the density is the number of entity eigenvectors within the rectangle divided by the area of ​​the rectangle.

[0103] The density gradient is a physical quantity that describes the rate of change of density. For each spatial grid cell, its density gradient is calculated. The density gradient can be approximated using the finite difference method. Taking a two-dimensional plane as an example, for a grid cell (i, j), its density gradient in the x-direction can be calculated by dividing the density difference between that cell and its adjacent cells in the x-direction by the distance between them; the same applies to the y-direction. This yields a density gradient vector for each spatial grid cell, which reflects the changing trend of the density around that cell.

[0104] Step S430: performing a clustering operation on the spatial grid units according to the density gradient value to generate an initial node cluster set, where each node cluster in the initial node cluster set corresponds to a local high-density area in the topological space.

[0105] Clustering is the process of dividing data points into groups or clusters based on similarity. In this embodiment of the present invention, spatial grid cells are clustered based on their density gradient values. Density-based clustering algorithms such as DBSCAN (Density-Based Spatial Clustering of Applications with Noise) can be used.

[0106] The DBSCAN algorithm divides clusters based on the density of data points. It defines two important parameters: the neighborhood radius and the minimum number of points. For a spatial grid cell, if the number of grid cells in its neighborhood is greater than or equal to the minimum number of points, the cell is considered a core point. If a cell is not a core point but is within the neighborhood of a core point, it is considered a boundary point; otherwise, it is considered a noise point.

[0107] Starting from a core point, the core points and boundary points in its neighborhood are continuously expanded to form a cluster. This process is repeated until all core points are classified into a cluster, ultimately obtaining an initial set of node clusters. Each node cluster corresponds to a local high-density region in the topological space, meaning that the entity feature vectors within this region have high similarity and may represent a similar class of risky entities.

[0108] Step S440: extracting the boundary overlap region between adjacent node clusters, calculating the feature similarity of entity feature vectors within the boundary overlap region, merging the node clusters whose feature similarity exceeds the inter-cluster merging threshold, and generating an optimized node cluster distribution.

[0109] The boundary overlap area between adjacent node clusters is the part where two adjacent node clusters overlap in the topological space. The entity feature vector in this overlapping area is extracted.

[0110] Calculate the feature similarity of entity feature vectors within the boundary overlap region. Feature similarity can be calculated using a variety of methods, such as cosine similarity, Euclidean distance, etc.

[0111] A cluster merging threshold is set. When the average feature similarity of entity feature vectors in the overlapping region of the boundary between two adjacent node clusters exceeds this threshold, the two node clusters are considered to have high similarity and are merged. The fusion process can be to merge all entity feature vectors of the two node clusters into a new node cluster.

[0112] In this way, the initial node cluster set is optimized, some unnecessary segmentations are removed, and the optimized node cluster distribution is obtained, making the node distribution of the knowledge graph more reasonable.

[0113] Step S450: Based on the optimized node cluster distribution, the spatial coordinates of the center point of each node cluster are calculated, and the spatial coordinates are weighted interpolated with the risk level score in the risk entity feature matrix to generate a node distribution topology.

[0114] For each optimized node cluster, the spatial coordinates of its center point are calculated. The calculation method of the center point can be selected according to the specific situation, such as calculating the average value of all entity feature vectors in the node cluster in the topological space.

[0115] The risk entity feature matrix contains the risk level score for each risk entity. A weighted interpolation is performed between the spatial coordinates of the node cluster center and the risk level score in the risk entity feature matrix. The purpose of weighted interpolation is to comprehensively consider the spatial location and risk level information of the node, so that the node distribution topology can more comprehensively reflect the risk situation. For example, a linear weighted interpolation method can be used, where the weight is determined based on the distance from the entity feature vector to the center point. The closer the distance to the center point, the greater the weight. A comprehensive risk level value is obtained through weighted summation, which is combined with the spatial coordinates of the center point to generate the node distribution topology.

[0116] Step S460: Perform reverse projection matching on the node cluster center points in the node distribution topology and the entity feature vectors in the risk entity feature matrix, establish a many-to-one mapping relationship from entity to node, and generate the final node distribution topology structure.

[0117] Backprojection matching is the process of matching the node cluster centers in the node distribution topology with the entity feature vectors in the risk entity feature matrix. Since the entity feature vectors were processed through dimensionality reduction and clustering in the previous steps to obtain the node cluster centers, the original entity feature vectors now need to be remapped to these nodes.

[0118] Matching can be achieved by calculating the similarity between the entity feature vector and the node cluster center. For example, using Euclidean distance to measure similarity, the entity feature vector is assigned to the node corresponding to the node cluster center closest to it. This establishes a many-to-one mapping relationship between entities and nodes, meaning that multiple entity feature vectors can be mapped to the same node. This mapping relationship generates a final node distribution topology that clearly demonstrates the correspondence between risk entities and knowledge graph nodes, providing a foundation for subsequent knowledge graph construction and risk analysis.

[0119] As an implementation method, in step S400, the entity relationship topology of the knowledge graph is determined based on the association strength value in the risk association strength matrix, which may specifically include the following steps: Step S470: perform symmetry processing on the risk association strength matrix, set the diagonal elements of the matrix to zero and take the maximum value of the upper and lower triangular matrices to generate an undirected association strength matrix.

[0120] The risk association strength matrix is ​​a square matrix whose elements represent the strength of associations between different risk entities. In practical applications, risk associations are often undirected; that is, the strength of association between entity A and entity B should be the same as the strength of association between entity B and entity A. Therefore, the risk association strength matrix needs to be symmetric.

[0121] First, set the diagonal elements of the matrix to zero. These elements represent the strength of an entity's association with itself. In a knowledge graph, these associations are usually meaningless, so they are set to zero. Then, take the maximum value of the upper and lower triangular matrices. This process results in a symmetric matrix, the undirected association strength matrix.

[0122] Step S480: input the undirected association strength matrix into the graph structure generation network, extract the distribution quantiles of the association strength values ​​through the adaptive threshold segmentation algorithm, and generate a dynamic association threshold interval.

[0123] The undirected association strength matrix is ​​input into a graph structure to generate a network, which analyzes and processes the association strength values ​​in the matrix. An adaptive threshold segmentation algorithm is an algorithm that automatically determines a threshold based on the data distribution. In an embodiment of the present invention, this algorithm is used to extract the distribution quantiles of the association strength values. Distribution quantiles are numerical points that divide the data distribution into several equal parts. For example, the median is the 50th percentile.

[0124] The quantile regression method can be used to determine the distribution quantile. Quantile regression is a regression method for estimating conditional quantiles. It can calculate the values ​​corresponding to different quantiles based on the distribution of data.

[0125] Generate a dynamic correlation threshold interval based on the calculated distribution quantiles. For example, the 25th and 75th percentiles can be used as the upper and lower limits of the threshold interval. This dynamic correlation threshold interval can be adjusted based on the actual distribution of correlation strength values, more accurately reflecting the characteristics of the data.

[0126] Step S490: Divide the correlation strength values ​​into a strong correlation interval, a weak correlation interval, and a noise interval according to the dynamic correlation threshold interval, and retain the matrix elements in the strong correlation interval to generate an initial edge connection set.

[0127] Based on the dynamic correlation threshold interval, the correlation strength values ​​in the undirected correlation strength matrix are divided into three intervals: strong correlation interval, weak correlation interval, and noise interval. The strong correlation interval, defined as the interval where the correlation strength value exceeds the upper limit of the dynamic correlation threshold interval, is relatively strong, indicating a close connection between the two risk entities. The weak correlation interval, defined as the interval where the correlation strength value falls within the dynamic correlation threshold interval, is relatively weak. The noise interval, defined as the interval where the correlation strength value falls below the lower limit of the dynamic correlation threshold interval, indicates that the association within this interval is likely due to noise or random factors and is therefore considered to be of no practical significance.

[0128] The matrix elements within the strongly associated interval are retained, and the entity pairs corresponding to these elements are used as the initial edge connection set. The initial edge connection set represents the strong association relationship between entities in the knowledge graph and provides a foundation for the subsequent graph structure construction.

[0129] Step S4100: Perform path connectivity verification on each edge in the initial edge connection set. When there are multiple association paths between two entity nodes, calculate the cumulative association strength of each path and retain the path with the maximum value.

[0130] Path connectivity verification checks whether the edges in the initial edge connection set can form a valid path connecting two entity nodes. For each edge in the initial edge connection set, starting from one entity node, a search is performed along the edge to see if it can reach the next entity node. When multiple association paths exist between two entity nodes, the cumulative association strength of each path is calculated. The cumulative association strength is the product of the association strengths of all edges on the path (or other appropriate accumulation method). The path with the highest cumulative association strength is retained. This removes some redundant paths, making the edge connections in the knowledge graph more concise and effective. Through path connectivity verification and path selection, the association relationships between entities in the knowledge graph are ensured to be the most representative and strongest.

[0131] Step S4110: Generate a sparse incidence matrix based on the retained edge connection set, perform a subgraph partitioning operation on the sparse incidence matrix, and detect fully connected subgraphs and bridging edge structures.

[0132] Generate a sparse incidence matrix based on the set of retained edge connections. A sparse incidence matrix is ​​a matrix of the same dimensions as the undirected incidence strength matrix, but with most of its elements zero and only the elements corresponding to the retained edge connections non-zero.

[0133] Subgraph partitioning is the process of dividing a graph represented by a sparse incidence matrix into multiple subgraphs. Graph theory algorithms, such as the Kosaraju algorithm or the Tarjan algorithm, can be used for subgraph partitioning. A fully connected subgraph is one in which any two nodes in the subgraph are connected by an edge. Bridging edges, edges connecting different subgraphs, play a key role in graph connectivity. By analyzing the results of subgraph partitioning, the presence of fully connected subgraphs and bridging edge structures can be detected.

[0134] For example, a graph consisting of multiple nodes and edges may be partitioned into several independent subgraphs, some of which are fully connected, and the edges connecting these subgraphs are bridge edges. Examining the structure of fully connected subgraphs and bridge edges helps us understand the overall structure and connectivity of the knowledge graph.

[0135] Step S4120: compress the fully connected subgraph into super nodes, calculate the bridge edge association strength between super nodes, and generate a hierarchical entity relationship topology.

[0136] Compress the detected fully connected subgraphs into supernodes. A supernode is an abstract representation that combines multiple nodes into a single node, representing all nodes in the fully connected subgraph. By compressing the fully connected subgraph into supernodes, the structure of the knowledge graph can be simplified and the number of nodes can be reduced. Calculate the strength of the bridging edge associations between supernodes. The bridging edge association strength can be calculated based on the association strengths between nodes in the fully connected subgraphs connected by the bridging edge. For example, the average of the association strengths between all pairs of nodes in the two fully connected subgraphs connected by the bridging edge can be taken as the bridging edge association strength between supernodes.

[0137] Based on the compressed supernodes and the calculated bridging edge strengths, a hierarchical entity relationship topology is generated. This topology displays entity relationships at different levels in the knowledge graph, reflecting the macro-level connections between entities at the supernode level, providing a clearer structure for subsequent risk analysis and propagation path identification.

[0138] Step S4130: Slice the hierarchical entity relationship topology in time windows according to the temporal attributes in the risk entity feature matrix to generate a dynamic entity relationship topology structure with temporal dependencies.

[0139] The risk entity feature matrix may contain time series attributes, such as the time when the risk occurs, the duration of the risk, etc. Based on these time series attributes, the hierarchical entity relationship topology is sliced ​​into time windows.

[0140] Time window slicing divides the hierarchical entity relationship topology into multiple time windows in chronological order, with each time window corresponding to a time period. Within each time window, the relationships between entities are analyzed to generate the entity relationship topology within that time window.

[0141] For example, assuming that the time series attribute in the risk entity feature matrix is ​​the time when the risk occurs, the time range is divided into multiple time periods. For each time period, the entities within the time period and the association relationships between them are extracted to generate the corresponding entity relationship topology.

[0142] By slicing time windows, a dynamic entity relationship topology with temporal dependencies is generated. This structure can reflect the changes in entity relationships in the knowledge graph over time, helping to analyze the spread and evolution of risks at different time points.

[0143] Step S4140: spatially align the dynamic entity relationship topology structure with the node distribution topology, and generate a weighted edge connection topology based on the mapping relationship from entities to nodes.

[0144] Spatial alignment is the process of spatially matching and aligning the dynamic entity relationship topology and node distribution topology. Since the dynamic entity relationship topology describes the relationships between entities, while the node distribution topology describes the distribution of nodes, spatial alignment can combine the two.

[0145] Based on the mapping relationship between entities and nodes, the edge connections between entities in the dynamic entity relationship topology are mapped to the nodes in the node distribution topology. At the same time, these edge connections are assigned weights, which can be determined based on the strength of the association between the entities.

[0146] In this way, a weighted edge connection topology is generated, which integrates the node distribution and entity relationship information, and the edge connections have weights, which can more accurately reflect the association strength and structure between entities in the knowledge graph.

[0147] Step S500: Generate a dynamic knowledge graph of the target scientific and technological project based on the node distribution topology and entity relationship topology, and identify potential risk propagation paths in the dynamic knowledge graph.

[0148] A dynamic knowledge graph is a graph structure that can reflect how knowledge changes over time. It combines information on node distribution topology and entity relationship topology, and shows how the distribution of risk entities in target scientific and technological projects and the relationships between them change over time.

[0149] The process of generating a dynamic knowledge graph based on node distribution topology and entity relationship topology involves integrating the nodes in the node distribution topology and the edge connections in the entity relationship topology into a single graph structure. The node distribution topology determines the location and attributes of the nodes, while the entity relationship topology determines the connections and weights of the edges. Furthermore, given the dynamic nature of the entity relationship topology, a time dimension is added to the graph structure, allowing the knowledge graph to reflect the situation at different points in time.

[0150] Identifying potential risk propagation paths in dynamic knowledge graphs is the process of analyzing how risks propagate from one node to other nodes in the knowledge graph.

[0151] As an implementation method, step S500 generates a dynamic knowledge graph of the target scientific and technological project based on the node distribution topology and the entity relationship topology, which can specifically include the following steps: Step S510: extract the cosine similarity of the feature vectors of each entity in the node distribution topology, generate a topological similarity matrix, and channel-concatenate the topological similarity matrix with the risk association intensity matrix to form a multimodal association tensor.

[0152] The generated topological similarity matrix and the risk association strength matrix are channel-concatenated. Channel concatenation involves concatenating the two matrices along the channel dimension to form a multimodal association tensor. Assuming the dimensions of the topological similarity matrix are n×n and the dimensions of the risk association strength matrix are also n×n, the dimensions of the multimodal association tensor formed after channel concatenation are n×n×2, where the third dimension represents the channel. Through channel concatenation, two different types of information, topological similarity and risk association strength, are integrated into a single tensor, providing richer features for subsequent graph convolutional network processing.

[0153] Step S520: Input the multimodal correlation tensor into the graph convolutional network, aggregate the risk features of adjacent nodes through the neighborhood message passing mechanism, and generate a node enhanced feature matrix.

[0154] A graph convolutional network (GCN) is a neural network model specifically designed for processing graph-structured data. When a multimodal correlation tensor is input into a GCN, the network extracts and updates features from the nodes and edges in the graph.

[0155] The neighborhood message passing mechanism is one of the core mechanisms of graph convolutional networks. It transmits and aggregates feature information from adjacent nodes to the current node through edge connections between nodes. Specifically, for each node in the graph, it collects the features of its neighboring nodes and performs a weighted sum of these features based on the weights of the edges. The summed result is then merged and transformed with the features of the current node to obtain the updated node features.

[0156] In this embodiment of the present invention, a graph convolutional network aggregates risk features of adjacent nodes through a neighborhood message passing mechanism. Through multiple iterations of message passing and feature updates, the network learns the complex relationships between nodes and the risk propagation patterns. Ultimately, a node-enhanced feature matrix is ​​generated, where each element represents the enhanced node feature.

[0157] Step S530: Calculate the directional derivative of the associated edge in the entity relationship topology based on the eigenvector of each node in the node enhancement feature matrix to generate an edge weight gradient matrix.

[0158] The directional derivative is the rate of change of a vector along a certain direction. In an embodiment of the present invention, the directional derivative of the associated edge in the entity relationship topology is calculated based on the eigenvector of each node in the node enhancement feature matrix. For example, the finite difference method can be used to approximate the directional derivative. All associated edges in the entity relationship topology are calculated to obtain an edge weight gradient matrix. Each element of the edge weight gradient matrix represents the directional derivative of the corresponding associated edge, which reflects the change of the node feature in the edge direction and provides a basis for subsequent edge weight adjustment.

[0159] Step S540: perform element-wise multiplication of the edge weight gradient matrix and the risk association strength matrix to generate a dynamic edge weight distribution matrix, perform quantile normalization on the dynamic edge weight distribution matrix, and map it to the probability distribution space.

[0160] Element-wise multiplication multiplies the corresponding elements of two matrices. Element-wise multiplication of the edge weight gradient matrix and the risk correlation strength matrix yields the dynamic edge weight distribution matrix. This matrix combines edge directional derivative information with risk correlation strength information, more comprehensively reflecting the importance and dynamic changes of edges.

[0161] Quantile normalization is a method for mapping the elements of the dynamic edge weight distribution matrix to a probability distribution space. First, the quantiles of the elements in the dynamic edge weight distribution matrix are calculated, such as the median, 25th percentile, and 75th percentile. Then, the elements in the matrix are normalized based on these quantiles, so that the values ​​of the elements are between 0 and 1 and conform to the characteristics of the probability distribution.

[0162] Through quantile normalization, the elements in the dynamic edge weight distribution matrix are converted into probability values, which facilitates subsequent edge pruning and risk analysis.

[0163] Step S550: Based on the normalized dynamic edge weight distribution matrix, adaptively prune the edge connections in the entity relationship topology, and retain the associated edges whose gradient accumulation exceeds the average similarity between nodes.

[0164] Adaptive pruning is the process of screening and removing edge connections in an entity relationship topology based on their importance. Based on the normalized dynamic edge weight distribution matrix, the cumulative gradient of each edge is calculated. The cumulative gradient can be defined as the cumulative value of the dynamic edge weight within a preset time window. The average inter-node similarity is the average cosine similarity between all pairs of nodes in the node distribution topology. The cumulative gradient of each edge is compared with the average inter-node similarity. Edges with gradient cumulative values ​​exceeding the average inter-node similarity are retained, while other edges are removed. Adaptive pruning removes unnecessary edge connections, making the knowledge graph more concise and clear while retaining the most important risk-related information.

[0165] Step S560: perform spatiotemporal alignment of the node enhanced feature matrix and the pruned entity relationship topology, and adjust the temporal dependency of the node features according to the timestamp of the risk association strength matrix.

[0166] Spatiotemporal alignment matches and aligns the node-enhanced feature matrix and the pruned entity relationship topology in space and time. Since the node-enhanced feature matrix reflects the feature information of the nodes, and the pruned entity relationship topology reflects the connection between nodes, spatiotemporal alignment can combine the two.

[0167] The temporal dependencies of node features are adjusted based on the timestamps of the risk association strength matrix. The timestamps in the risk association strength matrix record the risk associations at different points in time. Based on these timestamps, the node features in the node enhancement feature matrix are adjusted so that the node features can reflect the risk status at different points in time.

[0168] For example, for a certain node, at different time points, its feature value may change according to the change in the strength of risk association. By adjusting the temporal dependency of node features, the dynamic knowledge graph can more accurately reflect the evolution of risk over time.

[0169] Step S570: The node features and associated edges after spatiotemporal alignment are jointly encoded through a graph structure encoder to generate a graph embedding representation of the dynamic knowledge graph, where the node embedding vector contains the risk propagation potential coefficient and the edge embedding vector contains the risk attenuation factor.

[0170] The graph structure encoder is a model used to process graph data and convert it into a low-dimensional vector representation. The spatiotemporally aligned node features contain the risk status information of the nodes at different time points, and the associated edges reflect the connection relationship between nodes and the path of risk propagation.

[0171] Taking the graph neural network (GNN) as an example, as a graph structure encoder, it can encode nodes and edges through a message passing mechanism. This message passing mechanism is the core of GNNs. Its basic idea is that nodes update their feature representations by exchanging information with neighboring nodes. Specifically, for each node, it collects feature information from its neighboring nodes and performs a weighted sum of this information based on edge weights. The summed result is then combined with the node's own features and transformed to obtain the updated node features.

[0172] During the encoding process, the spatiotemporally aligned node features and associated edge information are first input into the GNN's input layer. The input layer converts this information into a format suitable for model processing. Next, in the hidden layer, the GNN continuously absorbs information from neighboring nodes through multiple iterations of message passing, thereby learning the node's global characteristics within the graph structure. Each message passing can be viewed as an aggregation and propagation of information, allowing the node's features to gradually incorporate more graph structural information. For example, for node i, its update formula during the kth message passing can be expressed as: ;in, is the feature representation of node i after the kth iteration, N(i) is the set of adjacent nodes of node i, is the normalized coefficient of edge (i, j), and is a learnable weight matrix, is the activation function.

[0173] For associated edges, they can also be encoded during the message passing process. The information of the edge can affect the transmission weight of the information between nodes. For example, the risk attenuation factor of the edge can be used as part of the weight to adjust the strength of the information transmission between nodes.

[0174] Finally, in the output layer, GNN maps the learned feature representations of nodes and edges to a low-dimensional space to generate a graph embedding representation of the dynamic knowledge graph.

[0175] Node embeddings contain a risk propagation potential coefficient. In risk analysis scenarios, each node has a different potential influence in the risk propagation process. The risk propagation potential coefficient reflects the node's ability to serve as a risk source or an intermediary node for risk propagation. During the encoding process, the graph encoder encodes this potential influence into the node embedding by learning the node's characteristics and its position in the graph. For example, if a node is connected to multiple high-risk nodes and has high resources or influence, its risk propagation potential coefficient will be high. Edge embeddings contain a risk attenuation factor. Risk propagation along edges is affected by various factors and attenuates. The risk attenuation factor describes the degree to which risk is attenuated as it propagates along an edge connection. The graph encoder encodes the risk attenuation factor into the edge embedding by learning edge characteristics (such as edge type and weight) and its role in the graph. For example, an edge connecting two distant or weakly connected nodes may have a large risk attenuation factor, indicating that risk attenuates rapidly as it propagates along this edge.

[0176] Through the joint encoding of the graph structure encoder, the generated graph embedding representation can comprehensively reflect the characteristics of nodes and edges in the dynamic knowledge graph and the relationships between them, providing an effective feature representation for subsequent risk propagation path identification and risk analysis.

[0177] As an implementation method, in step S500, identifying potential risk propagation paths in the dynamic knowledge graph may specifically include the following steps: Step S580: extracting the risk propagation potential coefficient from the node attributes of the dynamic knowledge graph, and extracting the risk attenuation factor from the edge attributes.

[0178] The node attributes of a dynamic knowledge graph contain various node information. The risk propagation potential coefficient is a key parameter in the node embedding vector generated by the graph structure encoder in the previous step. By accessing the node attributes, the risk propagation potential coefficient of each node is extracted.

[0179] Edge attributes also contain information about the edge. The risk attenuation factor is a parameter in the edge embedding vector. The risk attenuation factor of each edge is extracted from the edge attributes.

[0180] Step S590: Nodes whose risk propagation potential coefficient is greater than the average risk value of the same type in the node attributes are selected as candidate risk source nodes, and the reverse propagation path priority is generated by arranging them in reverse order according to the risk attenuation factors of the edge attributes.

[0181] Calculate the average value of similar risks within the node attributes. Similar risks are the risks faced by nodes with similar risk characteristics. For each node, compare its risk propagation potential coefficient with the average value of similar risks. Nodes with a risk propagation potential coefficient greater than the average value are considered candidate risk source nodes.

[0182] Edges are sorted in reverse order based on their risk attenuation factors. A smaller risk attenuation factor indicates less risk attenuation when propagating along that edge, which is more conducive to risk propagation. Therefore, edges with smaller risk attenuation factors are prioritized to generate reverse propagation path priorities.

[0183] Step S5100: Perform breadth-first traversal on the candidate risk source nodes along the priority of the reverse propagation path, and record the node sequence passed during the traversal and the risk attenuation factor of the corresponding edge.

[0184] In the embodiment of the present invention, a breadth-first traversal is performed on the candidate risk source nodes along the back propagation path priority. Starting from each candidate risk source node, the next node to be visited is selected according to the back propagation path priority. During the traversal process, the node sequence passed and the risk attenuation factor of the corresponding edge are recorded. For example, starting from candidate risk source node A, according to the back propagation path priority, node B is visited first, and the node sequence [A, B] and the risk attenuation factor d of the edge (A, B) are recorded. AB ; Then continue to visit the next node from node B, and so on, until all reachable nodes are traversed.

[0185] Step S5110: Perform conduction path integrity verification on each node sequence. When the risk propagation potential energy coefficients of adjacent nodes in the sequence satisfy the exponential decay law, calculate the cumulative conduction probability of the conduction path.

[0186] The integrity verification of the transmission path is the process of checking whether the node sequence constitutes a complete risk transmission path. For each node sequence, check whether the risk transmission potential coefficient of the adjacent nodes in the sequence satisfies the exponential decay law. The exponential decay law can be expressed as , where p i and p i+1 is the risk propagation potential coefficient of adjacent nodes, is the attenuation coefficient. When the risk propagation potential energy coefficient of adjacent nodes satisfies the exponential decay law, the cumulative transmission probability of the transmission path is calculated. The cumulative transmission probability can be calculated based on the risk attenuation factor of each edge on the path.

[0187] Step S5120: Perform a convolution operation on the cumulative transmission probability and the risk propagation potential coefficient of the node sequence to generate a comprehensive risk assessment value of the path.

[0188] The comprehensive risk assessment value of a path comprehensively considers the risk propagation potential of the node and the transmission probability of the path, and more comprehensively evaluates the possibility of risk propagation and the degree of impact on the path.

[0189] Step S5130: Based on the comparison of the comprehensive risk assessment value of the path with the benchmark value of similar historical risk cases, screen out abnormal conduction paths whose assessment values ​​deviate from the standard deviation range of the benchmark value.

[0190] Similar historical risk cases are risk cases from historical projects with similar characteristics and risk profiles to the current target technology project. For these similar historical risk cases, the baseline value and standard deviation of their path comprehensive risk assessment values ​​are calculated.

[0191] The comprehensive risk assessment value of each conduction path is compared with the benchmark value of similar historical risk cases. If the assessment value deviates from the standard deviation range of the benchmark value, the conduction path is considered to be an abnormal conduction path.

[0192] Step S5140: Perform a bidirectional risk diffusion simulation on the abnormal conduction path to capture the risk amplification nodes in the forward propagation path and the key inhibition nodes in the reverse blocking path.

[0193] Bidirectional risk diffusion simulation simultaneously simulates both forward risk diffusion and reverse risk blocking. For an abnormal transmission path, a forward risk diffusion simulation is performed, starting from the risk source node and simulating the risk's propagation along the path. The authors observe which nodes amplify the risk; these nodes are known as risk amplification nodes. Simultaneously, a reverse risk blocking simulation is performed, starting from the path's end node and simulating the risk blocking process in reverse. The authors identify which nodes effectively suppress the risk's propagation; these nodes are known as key suppression nodes.

[0194] Step S5150: Generate a multidimensional risk transmission map including path weights and intervention priorities based on the spatial distribution density of risk amplification nodes and key inhibition nodes.

[0195] Calculate the spatial distribution density of risk amplification nodes and key inhibition nodes. Spatial distribution density can be defined as the ratio of the number of risk amplification nodes or key inhibition nodes to the area or volume of a predefined spatial range.

[0196] Based on the spatial distribution density of risk amplification nodes and key inhibition nodes, a path weight is assigned to each abnormal conduction path. The path weight can be determined by the size of the spatial distribution density. The greater the spatial distribution density, the greater the path weight.

[0197] At the same time, the intervention priority is determined based on the path weight and the severity of the risk. The larger the path weight and the more serious the risk, the higher the intervention priority.

[0198] By integrating pathway weights and intervention priority information into a single graph, a multidimensional risk transmission map with pathway weights and intervention priorities is generated. This map can intuitively demonstrate the importance of abnormal transmission pathways and the order of intervention, providing strong support for risk control and management.

[0199] It should be noted that, in the various calculation links involved above, if it involves the fusion calculation of variables of different dimensions, or the fusion calculation of variables of different dimensions, those skilled in the art can perform pre-processing such as dimension elimination and dimension alignment according to conventional techniques in this field (such as normalization, linear interpolation, truncation, etc.), and then perform subsequent corresponding processing. Since this belongs to the conventional technical means in this field, it is not described in detail in the embodiments of the present invention.

[0200] As an implementation method, the method provided by an embodiment of the present invention also includes a dynamic knowledge graph update mechanism, which specifically includes the following steps: Step S600: Real-time monitoring of new heterogeneous data of the target scientific and technological project, and triggering a graph update event when it is detected that the numerical change of the structured indicator data exceeds the set sensitivity or new risk keywords appear in the unstructured text data.

[0201] Real-time monitoring of newly added heterogeneous data from target technology projects involves continuously collecting and analyzing new structured indicator data, unstructured text data, and time-series behavioral data generated within the target technology projects. Numerical changes in structured indicator data may reflect actual project progress and changes in risk profiles, while the appearance of new risk keywords in unstructured text data may indicate the emergence of new risk factors. A pre-set sensitivity threshold is used to determine whether a numerical change in structured indicator data is significant. For example, if a project's budget indicator changes by more than a set percentage (e.g., 5%), the change is considered to have exceeded the set sensitivity. New risk keywords are keywords that appear in unstructured text data and have not previously been identified as risk-related. New risk keywords can be identified through keyword matching or natural language processing techniques. A graph update event is triggered when a numerical change in structured indicator data exceeds the set sensitivity or when a new risk keyword appears in unstructured text data. The triggering of a graph update event indicates that the dynamic knowledge graph needs to be updated to reflect the latest risk profile of the project.

[0202] Step S700: performing incremental feature extraction on the newly added heterogeneous data to generate incremental entity feature vectors and incremental association strength matrices.

[0203] Incremental feature extraction involves extracting useful feature information from newly added heterogeneous data. For newly added structured indicator data, preprocessing and feature extraction are required, such as removing outliers and normalizing the data. Feature vectors are then extracted. For newly added unstructured text data, semantic analysis and feature extraction are required, such as identifying entities and relationships and extracting semantic features from the text. For newly added time-series behavioral data, behavioral patterns and temporal characteristics need to be analyzed to extract corresponding feature vectors.

[0204] Generating incremental entity feature vectors involves integrating feature vectors extracted from newly added heterogeneous data into feature vectors representing newly added entities. The incremental association strength matrix reflects the strength of associations between newly added entities and existing entities, as well as between newly added entities themselves.

[0205] As an implementation method, step S700 performs incremental feature extraction on the newly added heterogeneous data to generate incremental entity feature vectors and incremental association strength matrices, which can specifically include the following steps: Step S710: performs sliding mean filtering on the newly added structured indicator data to eliminate sudden noise interference.

[0206] Through sliding mean filtering, the newly added structured indicator data can be made smoother, reducing the impact of noise on subsequent feature extraction and analysis.

[0207] Step S720: Perform named entity recognition on the newly added unstructured text data to extract new risk entities that are not registered in the existing knowledge graph.

[0208] Step S730: Perform operation pattern cluster analysis on the newly added time series behavior data to identify new pattern categories whose distance from the cluster center of the existing behavior pattern vector sequence exceeds a set radius.

[0209] Operation pattern cluster analysis is the process of classifying and clustering the operation patterns in newly added time series behavior data. Clustering algorithms, such as K-Means or DBSCAN, can be used to cluster newly added time series behavior data.

[0210] First, the cluster centers of the existing behavior pattern vector sequence are calculated. Then, for each operation pattern vector in the newly added time-series behavior data, its distance from the existing cluster center is calculated. A radius threshold is set. When the distance between an operation pattern vector and the existing cluster center exceeds this radius, the operation pattern is considered to belong to a new pattern category.

[0211] Step S740: Map the new risk entity and the new pattern category to the feature space to generate an incremental entity feature vector.

[0212] Mapping new risk entities and new pattern categories into the feature space involves converting them into vector representations. For new risk entities, features can be extracted based on their attributes and related information and converted into vectors. For new pattern categories, feature vectors can be generated based on their operational characteristics, such as frequency and time interval. The feature vectors for the new risk entities and new pattern categories are combined to generate an incremental entity feature vector.

[0213] Step S750: Calculate the association strength between the new risk entity and the existing entity, and the association between the new pattern category and the existing behavior pattern, and generate an incremental association strength matrix.

[0214] Calculating the strength of association between new risk entities and existing entities can be done using a variety of methods, such as feature-based or rule-based methods. For example, the cosine similarity between the feature vectors of the new risk entity and the existing entity can be calculated, and the similarity value can be used as the strength of association. Calculating the strength of association between new pattern categories and existing behavior patterns can be achieved by comparing their operational features. For example, the similarity between the new pattern category and existing behavior patterns based on features such as operation frequency and operation time interval can be calculated, and the similarity value can be used as the degree of association.

[0215] The association strengths between new risk entities and existing entities, as well as the associations between new pattern categories and existing behavioral patterns, are combined into a matrix to generate an incremental association strength matrix.

[0216] Step S800: perform similarity matching between the incremental entity feature vector and the existing node distribution topology. When the similarity is lower than the matching threshold, a new node is created; otherwise, the feature vector of the existing node is updated.

[0217] The incremental entity feature vector is matched with the node feature vector in the existing node distribution topology by similarity. Similarity matching can use methods such as cosine similarity and Euclidean distance.

[0218] A matching threshold is set. When the similarity between the incremental entity feature vector and the existing node feature vector is lower than the threshold, the incremental entity is considered to represent a new risk entity and a new node is created to represent the entity. The feature vector of the new node is the incremental entity feature vector.

[0219] When the similarity is higher than the matching threshold, it means that the incremental entity has a high similarity with the existing node, and the feature vector of the existing node is updated. A weighted average method can be used to perform a weighted average of the incremental entity feature vector and the existing node feature vector to obtain the updated node feature vector.

[0220] Step S900: superimpose the incremental association strength matrix and the existing entity relationship topology to generate an updated edge weight distribution.

[0221] Matrix overlay adds the elements of the incremental association strength matrix to the corresponding elements of the association strength matrix in the existing entity relationship topology. The association strength matrix in the existing entity relationship topology represents the association strength between existing entities, while the incremental association strength matrix represents the association strength between new entities and existing entities, as well as between new entities themselves.

[0222] By overlaying the matrix, the newly added association strength information is integrated into the existing entity relationship topology to generate an updated edge weight distribution. The updated edge weight distribution more comprehensively reflects the association relationships between entities in the knowledge graph.

[0223] Step S1000: Based on the updated node distribution and edge weight distribution, re-execute the risk transmission path analysis and update the potential risk transmission path set.

[0224] Based on the updated node and edge weight distributions, the risk transmission path analysis process described in the previous steps is re-executed. This includes identifying candidate risk source nodes, generating backpropagation path priorities, performing a breadth-first traversal, calculating cumulative transmission probabilities, and generating a comprehensive risk assessment for each path. This re-analysis yields a new set of potential risk transmission paths. Updating this set of potential risk transmission paths enables the dynamic knowledge graph to more accurately reflect the latest risk status of the target technology project, providing more timely and effective information for risk control and management.

[0225] As an implementation method, the method provided by an embodiment of the present invention also includes a process for generating a risk mitigation strategy, which specifically includes the following steps: Step S1100: After identifying a potential risk propagation path, extract a set of influencing factors of key nodes in the path.

[0226] After identifying potential risk propagation paths, identify key nodes along the path. Key nodes are nodes that play an important role in the risk propagation process, such as risk amplification nodes or key inhibition nodes. Extract a set of influencing factors for key nodes. Influencing factors are various factors associated with key nodes that influence their role in risk propagation. For example, for a project task node, influencing factors might include task completion time, resource requirements, technical difficulty, and so on.

[0227] By analyzing the attributes and related information of key nodes, a set of influencing factors is extracted, which will serve as input for subsequent risk mitigation strategy optimization.

[0228] Step S1200: Construct a risk mitigation strategy optimization model. The risk mitigation strategy optimization model takes minimizing the risk level score of key nodes as the objective function and takes project resource constraints as the boundary conditions.

[0229] The risk mitigation strategy optimization model is a mathematical model used to find the optimal risk mitigation strategy. The goal of this model is to minimize the risk level score of key nodes, that is, to reduce the risk level of key nodes by rationally allocating resources and taking appropriate measures.

[0230] The objective function can be expressed as the sum of the risk level scores for key nodes. Project resource constraints refer to the limitations on available project resources when implementing risk mitigation strategies. Resource constraints can include limitations on human, material, and financial resources. For example, the total project budget cannot exceed a preset amount, or available human resources are limited.

[0231] Project resource constraints are incorporated into the risk mitigation strategy optimization model as boundary conditions to ensure that the optimized risk mitigation strategy is within the scope allowed by project resources.

[0232] Step S1300: Solve the risk mitigation strategy optimization model using a gradient descent algorithm to obtain the optimal resource allocation solution for each key node.

[0233] The optimal resource allocation scheme for each key node is obtained by using a gradient descent algorithm. This scheme can minimize the risk level score of key nodes while meeting the project resource constraints.

[0234] Step S1400: Convert the optimal resource allocation plan into a natural language description and associate it with the corresponding node in the knowledge graph.

[0235] The optimal resource allocation plan is converted into a natural language description to facilitate understanding and implementation by project managers. The optimal resource allocation plan, described in natural language, is associated with the corresponding node in the knowledge graph. In the knowledge graph, each key node has corresponding attributes and relationships, and the optimal resource allocation plan is associated with the node as an attribute. This allows for a direct understanding of the risk mitigation strategy for each key node when viewing the knowledge graph.

[0236] Step S1500: recalculate the comprehensive risk value of the potential risk propagation path according to the change in node risk level after resource allocation.

[0237] After allocating resources to key nodes according to the optimal resource allocation plan, the risk level of these nodes will change. This is because resource allocation may improve the node's condition and reduce its risk level. The comprehensive risk value of the potential risk propagation path is recalculated. The calculation method for the comprehensive risk value is the same as described in the previous step. This method considers the risk propagation potential coefficient of each node on the path and the risk attenuation factor of the edges, calculates the cumulative transmission probability, and convolves this cumulative transmission probability with the risk propagation potential coefficient of the node sequence.

[0238] By recalculating the comprehensive risk value of potential risk transmission paths, the effectiveness of risk mitigation strategies can be evaluated, providing a basis for further risk control and management.

[0239] As an implementation method, the risk mitigation strategy optimization model includes: defining a decision variable matrix, wherein each element represents the type and quantity of resources allocated to a node.

[0240] The decision variable matrix is ​​a crucial component of the risk mitigation strategy optimization model. It represents the resources allocated to each key node. Each row of the matrix corresponds to a key node, and each column to a resource type. Each element in the matrix represents the quantity of that resource type allocated to the node.

[0241] Construct a resource benefit function and calculate the attenuation coefficient of each resource type on the node risk level score.

[0242] The resource benefit function describes the relationship between resource input and the reduction in a node's risk rating. Different resource types have varying degrees of impact on a node's risk rating. By constructing a resource benefit function, we can calculate the attenuation coefficient of each resource type on the node's risk rating.

[0243] By analyzing and modeling historical data, the specific form and parameters of the resource benefit function can be determined, thereby calculating the attenuation coefficient of each resource type on the node risk level score.

[0244] Establish a resource constraint inequality group to ensure that the total resource consumption does not exceed the project budget and complies with resource allocation rules.

[0245] Resource constraint inequalities are designed to ensure that total resource consumption during resource allocation does not exceed the project budget and adheres to resource allocation rules. The project budget represents the total resources available for the project, while resource allocation rules are the constraints on resource allocation, such as the allocation ratio of certain resources cannot exceed a preset range.

[0246] A risk conduction inhibition factor is introduced, which is inversely proportional to the edge conduction probability in the knowledge graph.

[0247] The risk transmission suppression factor is a parameter used to measure the degree to which risk is suppressed on edge connections. It is inversely proportional to the edge transmission probability in the knowledge graph. That is, the larger the risk transmission suppression factor, the smaller the edge transmission probability, and the lower the possibility of risk propagation on that edge.

[0248] Introducing risk conduction inhibitors can more accurately describe the spread of risks in the knowledge graph, and considering risk conduction inhibitors in the risk mitigation strategy optimization model can better control the spread of risks.

[0249] By adjusting the value of the risk transmission inhibition factor, the edge transmission probability can be changed, thereby affecting the risk transmission path and the comprehensive risk value.

[0250] The constraints are integrated into the objective function through the Lagrange multiplier method to form an unconstrained optimization problem.

[0251] In the risk mitigation strategy optimization model, the objective function is to minimize the risk level score of key nodes, while also subject to a set of resource constraint inequalities. By introducing Lagrange multipliers, these resource constraint inequalities are converted into equality constraints and integrated into the objective function, forming an unconstrained optimization problem. Specifically, for each resource constraint inequality, a Lagrange multiplier is introduced, multiplied by the inequality constraint, and added to the objective function.

[0252] By solving this unconstrained optimization problem, we can obtain the optimal solution that satisfies resource constraints. We use an adaptive learning rate optimization algorithm to iteratively update the decision variable matrix until the objective function converges to a local optimal solution.

[0253] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 may be connected via a bus or other means. The processor 101 (also known as the Central Processing Unit (CPU)) is the computing and control core of the computer system, capable of parsing various instructions within the computer system and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, a mobile communication interface, etc.), which can be used to send and receive data under the control of the processor 101. The communication interface 102 may also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system for storing programs and data. It is understood that the memory 103 herein may include both the built-in memory of the computer system and, of course, the extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system, but this is not limited to this in the present invention.

[0254] In one embodiment, the processor 101 executes the knowledge graph generation method for scientific and technological project risk control provided in the above embodiment of the present invention by running the computer program in the memory 103.

Claims

1. A knowledge graph generation method for risk control of scientific and technological projects, characterized by: The method comprises: Acquire a multi-source heterogeneous data set of a target scientific and technological project, wherein the multi-source heterogeneous data set includes structured indicator data, unstructured text data, and time series behavior data; Obtain historical indicator data of similar scientific and technological projects, calculate the mean vector and covariance matrix of each indicator dimension, and generate a multidimensional normal distribution model based on the mean vector and covariance matrix; Input the structured indicator data of the current project into the multidimensional normal distribution model, calculate the probability density value of each data point, and generate an abnormal marker sequence when the probability density value is lower than the preset quantile of the distribution model; Generate a binary mask matrix based on the position information of the abnormal points in the abnormal mark sequence, wherein the positions corresponding to the abnormal points in the binary mask matrix take a value of zero, and the positions corresponding to the normal points take a value of one; Performing a morphological closing operation on the binary mask matrix to eliminate isolated abnormal markers and generate a connected region correction mask matrix; Correcting the abnormal area marked by the mask matrix based on the connected area, locating the abnormal data point in the original indicator data, and extracting the indicator values ​​of the adjacent data points in the time window before and after the abnormal data point; Inputting the index values ​​of the adjacent data points into a sliding average filter to generate an interpolation filling sequence, correcting the zero value position of the mask matrix according to the connected region, and replacing the original abnormal data points with the interpolation filling sequence; Outputting the replaced structured indicator data as the corrected structured indicator data; Inputting the modified structured indicator data into an indicator encoder, mapping indicator values ​​of different dimensions into a unified metric space through a feature dimension alignment algorithm, and generating a standard vector sequence; Performing semantic segmentation on the unstructured text data to identify text paragraphs containing risk factor descriptions, inputting the text paragraphs into a pre-trained semantic encoder to generate primary semantic vectors, performing spatial transformation on the primary semantic vectors through a context-aware network, and outputting a sequence of semantic vectors; Performing event window division on the time series behavior data, extracting type distribution features and time interval features of the behavior operations within the window, performing orthogonal projection transformation on the type distribution features and time interval features, and generating a behavior pattern vector sequence; Inputting the standard vector sequence, semantic vector sequence, and behavior pattern vector sequence into a risk quantitative assessment model to generate a risk entity feature matrix and a risk association strength matrix; Determine the node distribution topology of the knowledge graph based on the entity feature vectors in the risk entity feature matrix, and determine the entity relationship topology of the knowledge graph based on the association strength values ​​in the risk association strength matrix; Based on the node distribution topology and entity relationship topology, a dynamic knowledge graph of the target scientific and technological project is generated, and potential risk propagation paths in the dynamic knowledge graph are identified.

2. The method according to claim 1, characterized in that The step of inputting the standard vector sequence, semantic vector sequence, and behavior pattern vector sequence into a risk quantification assessment model to generate a risk entity feature matrix and a risk association strength matrix includes: Performing cross-attention calculation on the standard vector sequence and the semantic vector sequence to generate a cross-attention weight matrix of indicators and semantics, wherein each element of the cross-attention weight matrix reflects the strength of association between the indicator vector and the semantic vector; Performing weighted aggregation on the semantic vector sequence based on the cross attention weight matrix to generate a semantically enhanced indicator feature sequence, and performing element-by-element multiplication of the semantically enhanced indicator feature sequence with the standard vector sequence to generate a cross-fusion feature tensor; Expanding the cross-fused feature tensor into a two-dimensional matrix along the feature dimension, performing a time alignment operation with the behavior pattern vector sequence, adjusting the row and column distribution of the two-dimensional matrix according to the timestamp information of the behavior pattern vector sequence, and generating a time-aligned feature matrix; Performing an outer product operation on the time-aligned feature matrix and the behavior pattern vector sequence to generate an initial interaction matrix, performing channel attention weighting on the initial interaction matrix, calculating the contribution of each channel dimension to the risk association, and generating a channel weight vector; Performing a Hadamard product operation on the channel weight vector and the initial interaction matrix to generate a weighted interaction matrix, and inputting the weighted interaction matrix into a bidirectional recurrent neural network to capture historical dependency features through a forward propagation path and potential correlation features through a backward propagation path to generate a bidirectional fusion feature matrix; Performing a strided convolution operation on the bidirectional fusion feature matrix to extract multi-scale spatiotemporal correlation patterns, and adding the weighted interaction matrix to the strided convolution output through skip connections to generate a correlation strength matrix; The association strength matrix is ​​input into the feature distillation network, the local association pattern is extracted through the separable convolution layer, and the redundant features are compressed through the pooling layer to output the risk association strength matrix; The cross-fused feature tensor is input into the feature encoder, the feature receptive field is expanded through the dilated convolution layer, and the output result of the dilated convolution layer is filtered based on the feature selection weight generated by the risk association strength matrix to generate a risk entity feature matrix.

3. The method according to claim 2, characterized in that The step of performing an outer product operation on the time-aligned feature matrix and the behavior pattern vector sequence to generate an initial interaction matrix includes: Dividing the time-aligned feature matrix into a plurality of feature sub-blocks according to the row and column distribution of the time-aligned feature matrix, each feature sub-block corresponding to a behavior pattern vector in the same time window in the behavior pattern vector sequence; Perform outer product operations on each feature sub-block and the behavior pattern vector of the corresponding time window to generate a sub-block interaction matrix; Performing an edge filling operation on the sub-block interaction matrix to keep its dimension consistent with the time-aligned feature matrix, thereby generating a filled sub-block interaction matrix; The filled sub-block interaction matrices are concatenated in the order of time windows to generate the interaction tensor of the complete time dimension; Performing channel dimension compression on the interaction tensor, mapping the multi-dimensional channels into a single correlation intensity channel through a summation operation, and generating an initial interaction matrix; The initial interaction matrix is ​​residually connected with the time-aligned feature matrix to retain the original time-aligned feature information and generate an enhanced initial interaction matrix.

4. The method according to claim 1, wherein Determining the node distribution topology of the knowledge graph based on the entity feature vectors in the risk entity feature matrix includes: Performing nonlinear dimensionality reduction processing on each entity feature vector in the risk entity feature matrix to generate a low-dimensional embedding vector, and mapping the low-dimensional embedding vector to a topological space coordinate system; Dividing the spatial grid cells based on the embedded vector distribution density in the topological space coordinate system, and calculating the density gradient value of the entity feature vector in each of the spatial grid cells; performing a clustering operation on the spatial grid units according to the density gradient value to generate an initial node cluster set, wherein each node cluster in the initial node cluster set corresponds to a local high-density area in the topological space; Extracting the boundary overlap region between adjacent node clusters, calculating the feature similarity of entity feature vectors within the boundary overlap region, and merging the node clusters whose feature similarity exceeds the inter-cluster merging threshold to generate an optimized node cluster distribution; Calculating the spatial coordinates of the center point of each node cluster based on the optimized node cluster distribution, performing weighted interpolation between the spatial coordinates and the risk level score in the risk entity feature matrix to generate a node distribution topology; The node cluster center points in the node distribution topology are back-projected and matched with the entity feature vectors in the risk entity feature matrix to establish a many-to-one mapping relationship from entity to node, thereby generating a final node distribution topology structure.

5. The method according to claim 4, characterized in that Determining the entity relationship topology of the knowledge graph according to the association strength values ​​in the risk association strength matrix includes: Performing symmetry processing on the risk correlation strength matrix, setting the diagonal elements of the matrix to zero and taking the maximum value of the upper and lower triangular matrices to generate an undirected correlation strength matrix; Inputting the undirected association strength matrix into a graph structure generation network, extracting the distribution quantiles of the association strength values ​​through an adaptive threshold segmentation algorithm, and generating a dynamic association threshold interval; Dividing the association strength value into a strong association interval, a weak association interval, and a noise interval according to the dynamic association threshold interval, and retaining the matrix elements in the strong association interval to generate an initial edge connection set; Performing path connectivity verification on each edge in the initial edge connection set, and when there are multiple association paths between two entity nodes, calculating the cumulative association strength of each path and retaining the path with the maximum value; generating a sparse incidence matrix according to the retained edge connection set, performing a subgraph partitioning operation on the sparse incidence matrix, and detecting a fully connected subgraph and a bridging edge structure; Compressing the fully connected subgraph into super nodes, calculating the bridge edge association strength between the super nodes, and generating a hierarchical entity relationship topology; Slicing the hierarchical entity relationship topology in a time window according to the time series attributes in the risk entity feature matrix to generate a dynamic entity relationship topology structure with time series dependencies; The dynamic entity relationship topology structure is spatially aligned with the node distribution topology, and a weighted edge connection topology is generated according to the entity-to-node mapping relationship.

6. The method according to claim 1, characterized in that Generating a dynamic knowledge graph of a target scientific and technological project based on the node distribution topology and the entity relationship topology includes: Extracting the cosine similarity of the feature vectors of each entity in the node distribution topology to generate a topology similarity matrix, and performing channel concatenation on the topology similarity matrix and the risk association strength matrix to form a multimodal association tensor; Inputting the multimodal correlation tensor into a graph convolutional network, aggregating the risk features of adjacent nodes through a neighborhood message passing mechanism, and generating a node enhanced feature matrix; According to the eigenvector of each node in the node enhancement feature matrix, the directional derivative of the associated edge in the entity relationship topology is calculated to generate the edge weight gradient matrix; Performing element-wise multiplication of the edge weight gradient matrix and the risk association strength matrix to generate a dynamic edge weight distribution matrix, and performing quantile normalization processing on the dynamic edge weight distribution matrix to map it to a probability distribution space; Based on the normalized dynamic edge weight distribution matrix, the edge connections in the entity relationship topology are adaptively pruned, and the associated edges whose gradient accumulation exceeds the average similarity between nodes are retained; The node enhancement feature matrix is ​​spatially and temporally aligned with the pruned entity relationship topology, and the temporal dependency of the node features is adjusted according to the timestamp of the risk association strength matrix; The spatiotemporally aligned node features and associated edges are jointly encoded through a graph structure encoder to generate a graph embedding representation of the dynamic knowledge graph, where the node embedding vector contains the risk propagation potential coefficient and the edge embedding vector contains the risk attenuation factor.

7. The method according to claim 6, characterized in that The identification of potential risk propagation paths in the dynamic knowledge graph includes: Extract the risk propagation potential coefficient from the node attributes of the dynamic knowledge graph and the risk attenuation factor from the edge attributes; Nodes whose risk propagation potential coefficient is greater than the average value of similar risks in node attributes are selected as candidate risk source nodes, and the reverse propagation path priority is generated by sorting them in reverse order according to the risk attenuation factors of edge attributes; Perform breadth-first traversal on the candidate risk source nodes along the priority of the back-propagation path, and record the node sequence passed through during the traversal and the risk attenuation factor of the corresponding edge; Perform conduction path integrity verification on each node sequence. When the risk propagation potential energy coefficient of adjacent nodes in the sequence satisfies the exponential decay law, calculate the cumulative conduction probability of the conduction path. The cumulative transmission probability is convolved with the risk propagation potential coefficient of the node sequence to generate a comprehensive risk assessment value of the path; By comparing the comprehensive risk assessment value of the path with the benchmark value of similar historical risk cases, abnormal conduction paths whose assessment values ​​deviate from the standard deviation range of the benchmark value are screened out; Perform bidirectional risk diffusion simulation on abnormal conduction paths to capture risk amplification nodes in the forward propagation path and key inhibition nodes in the reverse blocking path; Based on the spatial distribution density of risk amplification nodes and key inhibition nodes, a multidimensional risk transmission map including path weights and intervention priorities is generated.

8. A computer system, characterized in that: include: a memory storing a computer program; A processor, used to load the computer program to implement the knowledge graph generation method for scientific and technological project risk control as described in any one of claims 1 to 7.

Images