Financial risk control system based on big data and artificial intelligence

Abstract

The application discloses a financial risk control system based on big data and artificial intelligence and belongs to the technical field of financial data processing. The system comprises a data acquisition module for acquiring heterogeneous financial data; a modal alignment module for performing semantic mapping on a financial statement text sequence and a transaction record time sequence to generate a multi-modal embedding vector; a graph attention module for updating a supply chain association graph node representation by using an attention mechanism in a graph attention network to obtain a risk enhanced node representation; a residual fusion module for splicing the risk enhanced node representation and a multi-modal embedding vector residual to obtain a fusion risk feature; a gated recurrent module for inputting the fusion risk feature into a gated recurrent unit according to a time sequence and outputting a risk gating value; and a risk warning module for triggering a warning and storing a risk anchor point vector when the risk gating value exceeds a dynamic threshold.

Description

Technical Field

[0001] This invention relates to the field of financial data processing technology, specifically to a financial risk control system based on big data and artificial intelligence. Background Technology

[0002] The financial risk control field has long faced the challenge of heterogeneous data fusion and processing. Existing technologies mainly rely on rule engines or traditional statistical models for single-modal data, making it difficult to effectively integrate cross-modal information such as financial statement text, transaction flow time series, and supply chain relationship graphs. These systems typically execute text analysis, time series detection, and graph mining as independent tasks, leading to the neglect of implicit correlations between different data sources and the attenuation or loss of risk signals in fragmented processing flows. When facing complex financial risks with strong concealment and cross-entity transmission, isolated data processing methods cannot capture the collaborative evolution patterns of risks in multimodal data, easily resulting in delayed warnings and false alarms. Existing methods have significant shortcomings in handling time series dependencies. Conventional risk control models often use aggregate statistics with fixed time windows or simple moving average calculations, lacking the ability to fine-tune the nonlinear evolution of risk states over time. There is often a time lag between localized abnormal fluctuations in transaction flows and negative semantic clues in financial statement texts. However, existing systems cannot dynamically weigh the contribution of multimodal information at different time steps, causing models to either be overly sensitive to short-term disturbances or slow to react to gradually accumulating risk trends. Meanwhile, risk transmission paths in supply chain mapping are often simplified to static topology analysis, ignoring the iterative update requirements of node states in the temporal context, making it difficult to identify cross-enterprise cascading default risks in advance.

[0003] To address the aforementioned issues, this invention focuses on how to adaptively enhance the representation of supply chain graph nodes based on aligning cross-modal semantics and timestamps, and how to robustly assess risk states during temporal evolution through a dynamic threshold mechanism. The former involves precise semantic mapping between financial statement text units and transaction numerical units in the time dimension, and using this mapping to drive a graph attention network to capture the intensity of risk propagation between nodes. The latter addresses overcoming the limitations of fixed thresholds in adapting to changes in the risk environment, enabling the early warning mechanism to autonomously adjust its judgment boundaries based on the distribution characteristics of historical risk threshold values, thereby balancing the sensitivity and robustness of early warnings without relying on manual parameter tuning. Summary of the Invention

[0004] This invention aims to provide a financial risk control system based on big data and artificial intelligence, which solves the problems of the difficulty in effectively integrating cross-modal financial data and the difficulty in robustly judging the temporal evolution of risks, and achieves accurate capture and adaptive early warning of complex financial risks.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a financial risk control system based on big data and artificial intelligence. This system achieves accurate perception and dynamic early warning of financial risks through deep fusion and time-series evolution modeling of multi-source heterogeneous financial data.

[0006] In this invention, the data acquisition module acquires heterogeneous financial data, including financial statement text sequences, transaction record time series sequences, and supply chain relationship graphs, ensuring the comprehensiveness and multidimensionality of the information upon which risk analysis is based. The modal alignment module performs modal alignment operations on the heterogeneous financial data, semantically mapping each text unit in the financial statement text sequence to numerical units with the same timestamp in the transaction record time series, generating a multimodal embedding vector. This mapping mechanism enables precise matching of textual semantic information and transaction numerical information at specific time points, eliminating the semantic gap between modalities. The generated multimodal embedding vector can more compactly express the synergistic relationship between a company's financial description and capital behavior at the same point in time.

[0007] The graph attention module inputs the multimodal embedding vector into the first layer of the graph attention network, calculates the attention coefficient between any two adjacent nodes in the supply chain relationship graph, and updates the initial representation of each node based on the attention coefficient to obtain the risk-enhanced node representation. This process enables the system to adaptively explore the differences in the intensity of risk transmission between different enterprises in the supply chain network, giving higher attention to nodes with high correlation, thereby effectively injecting the risk propagation information contained in the graph structure into the node representation. The residual fusion module performs residual concatenation of the risk-enhanced node representation and the multimodal embedding vector in the second layer of the graph attention network to obtain the fused risk features. Through residual concatenation, the integrity of the original multimodal information is preserved while incorporating the network relationship features enhanced by graph attention, so that the fused features have the dual risk discrimination capabilities of both the internal attributes of enterprises and external relationships.

[0008] The gating loop module inputs the fused risk features into the gating loop unit, which iteratively updates the hidden state according to time steps and outputs a risk gating value at each time step. The gating loop unit continuously models the temporal evolution of the fused risk features, and its output risk gating value quantifies the systemic risk level at the current time step in real time. The risk warning module triggers a risk warning when the risk gating value exceeds a dynamic threshold and stores the hidden state at the trigger moment as a risk anchor vector. By setting a dynamic threshold, the system can adaptively adjust the warning sensitivity based on historical risk fluctuation levels, reducing the false alarm rate; while the stored risk anchor vector provides a valuable state snapshot for risk attribution and subsequent trend inference.

[0009] As a technical solution of this invention, in the modal alignment operation, the system extracts the semantic feature vector of each text unit in the financial statement text sequence, and simultaneously extracts the distribution feature vector of each numerical unit in the transaction record time sequence; calculates the cosine similarity matrix between the semantic feature vector and the distribution feature vector, and selects the column index corresponding to the maximum value of each row from this matrix, and determines the timestamp represented by the column index as the matching timestamp of the text unit; the semantic feature vector of the text unit and the distribution feature vector of the numerical unit corresponding to the matching timestamp are concatenated to obtain the multimodal embedding vector. An automatic matching strategy based on maximizing similarity is adopted to achieve cross-modal time alignment of text semantics and numerical distribution, effectively overcoming the inefficiency and inconsistency problems of traditional manual alignment methods. Preferably, when the maximum value of a certain row of the cosine similarity matrix is ​​lower than a preset similarity threshold, the multimodal embedding vector corresponding to that text unit is set to a zero vector, avoiding interference from low-confidence matching to subsequent models.

[0010] As a technical solution of this invention, in graph attention calculation, the system initializes each node in the supply chain association graph as a vector fragment corresponding to that node in the multimodal embedding vector; for any first node and second node connected by an edge, the current vectors of the two are concatenated and input into a single-layer feedforward neural network to output the original attention value; a normalized exponential function is applied to the original attention value to obtain the normalized attention coefficient; the normalized attention coefficients of all adjacent nodes of the first node are multiplied by the current vector of the corresponding adjacent node and summed, and then input into a nonlinear activation function to output a new representation of the first node; the new representations of all nodes are stacked to obtain the risk-enhanced node representation. The single-layer feedforward neural network preferably uses the LeakyReLU activation function to perform a nonlinear transformation on the original attention value to enhance the model's ability to identify subtle risk differences.

[0011] As a technical solution of this invention, during the residual fusion process, the system adds the risk enhancement node representation and the multimodal embedding vector element-wise according to their corresponding node positions to obtain a first residual intermediate result. The first residual intermediate result is then input into a first linear transformation layer to obtain a first linear transformation vector, and the multimodal embedding vector is input into a second linear transformation layer to obtain a second linear transformation vector. The two linear transformation vectors are concatenated along the feature dimension, and then a normalized concatenated vector is output as the fused risk feature via a batch normalization layer. This design utilizes dual-line linear transformation and feature concatenation to selectively enhance and normalize features from different sources, ensuring the stability of the numerical distribution of the fused risk feature and accelerating the convergence of subsequent models.

[0012] As a technical solution of this invention, in the gated loop processing, the system divides the fused risk features into multiple time slices according to time steps; at the t-th time step, it obtains a subset of the fused risk features of the previous hidden state and the current time slice, concatenates them, and inputs them into the reset gate, outputting a reset gate vector; the reset gate vector is multiplied element-wise with the previous hidden state to obtain the reset hidden state, and concatenates it with the current time slice, inputting it into the candidate hidden state calculation unit to obtain the candidate hidden state; simultaneously, the previous hidden state is concatenated with the current time slice and input into the update gate, outputting an update gate vector; linear interpolation is performed on the previous hidden state and the candidate hidden state according to the update gate vector to obtain the current hidden state; finally, the current hidden state is input into the risk gating output layer to obtain the risk gating value. The design of the gating mechanism enables the system to selectively forget historical redundant information and absorb current valuable risk signals, achieving smooth recursion of risk states. Preferably, the risk gating output layer uses the Sigmoid activation function to map the current hidden state to a standard interval, so that the output risk gating value has an intuitive probabilistic interpretation.

[0013] As a technical solution of this invention, in the dynamic threshold early warning mechanism, the system initializes a sliding window queue with a fixed window size to store the most recent preset number of historical risk gate values; calculates the arithmetic mean of all historical risk gate values ​​in the queue and multiplies it by a preset relaxation coefficient to obtain the dynamic threshold; compares the current risk gate value with the dynamic threshold in real time, and generates a risk trigger flag if the threshold is exceeded; in response to the flag, extracts and stores the current hidden state as a risk anchor vector, associates it with the flag, stores it in the risk event log, and clears the sliding window queue. The dynamic threshold can fluctuate smoothly with the recent risk level, avoiding the problems of delayed reporting or oversensitivity caused by fixed thresholds; the window queue clearing mechanism ensures the independence of a single risk event trigger and prevents continuous oscillations. Preferably, the preset relaxation coefficient can be adaptively adjusted according to the frequency of historical risk events, further improving the threshold's adaptability to environmental changes.

[0014] As a technical solution of this invention, after generating the multimodal embedding vector, the system can further input it into a temporal convolutional network, using dilated convolutional kernels to perform convolution operations on the time dimension to obtain a multi-scale temporal feature map. This feature map is then input into a channel attention module, where it undergoes global average pooling, global max pooling, and shared multilayer perceptron processing sequentially to obtain channel attention weights. These weights are then multiplied channel-by-channel with the feature map to obtain weighted temporal features. The residual connection between these weighted temporal features and the multimodal embedding vector generates an enhanced multimodal embedding vector, which replaces the original vector input into the graph attention network. This approach captures long-term trends and short-term mutations in trading behavior across multiple time scales and automatically enhances risk-sensitive feature channels through channel attention, enabling the subsequent graph attention module to model based on more discriminative features.

[0015] As a technical solution of this invention, after obtaining the risk-enhanced node representation, the system can input it into a graph pooling layer, calculate the score of each node using trainable projection vectors, select the top K nodes with the highest scores in descending order as retained nodes, search for the first-order neighbors of the retained nodes in the original graph, and establish a connection if the neighbor is also a retained node to generate a pooled subgraph structure; the risk-enhanced node representations of all retained nodes in the subgraph are weighted and summed according to their scores to obtain a graph-level risk summary vector, which is then concatenated with the fused risk features and input into the gating recurrent unit. This solution, through a learnable node selection mechanism, focuses on the most core substructure of risk-bearing capacity in the supply chain, and uses graph-level summary vectors to assist the gating recurrent unit in performing risk discrimination at a high level, thereby enhancing the system's control over the overall risk situation.

[0016] As a technical solution of this invention, after a risk warning is triggered, the system inputs the risk anchor vector into the back attention module to calculate the back attention weights of the historical hidden states and the risk anchor vector at each historical time step before the triggering time of the gated loop unit. The historical hidden states are then weighted and summed to obtain a contribution feature vector. This contribution feature vector is input into the explanation generation module, where discretization and risk explanation vocabulary mapping are performed sequentially to obtain risk explanation phrases. Finally, these phrases are concatenated in chronological order to generate the risk warning explanation text and stored accordingly. This process achieves automatic backtracking from the risk anchor vector to the readable text explanation, clearly revealing the key temporal events and influencing factors that led to the risk triggering, providing a solid basis for risk control decision-makers' understanding and handling.

[0017] As a technical solution of this invention, after storing risk anchor vectors, the system acquires multiple risk anchor vectors stored from historical risk warnings, calculates the average Euclidean distance between each anchor vector as a clustering feature value, sorts them in ascending order, selects the anchor vector with the smallest clustering feature value as a reference anchor, calculates the difference vector between the remaining anchor vectors and the reference anchor, and inputs it into the encoding part of the autoencoder to obtain a potential evolution vector; inputs this potential evolution vector into the trend prediction layer, outputting a predicted risk gating sequence within the future time window; finally, based on whether there are any threshold exceedances in the predicted sequence, it determines whether to generate an early warning signal. This solution is based on the evolution trajectory of historical risk state points in the potential space, realizing a forward-looking prediction of future risk trends, enabling the system to extend from passive response to active defense.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This approach extracts the semantic feature vector of each text unit in the financial statement text sequence and the distribution feature vector of each numerical unit in the transaction record time sequence using a modal alignment module. A cosine similarity matrix between the two feature vectors is calculated, and the column index corresponding to the maximum value in each row is selected as the matching timestamp of the text unit. The semantic feature vector and the distribution feature vector corresponding to the matching timestamp are concatenated to form a multimodal embedding vector. This scheme enables unstructured risk semantic descriptions in the text to be accurately anchored to the numerical features of transactions at the same timestamp, avoiding the loss of correlation signals caused by information isolation between modalities and enhancing the model's ability to capture subtle risk clues such as discrepancies between accounts and actual transactions. A graph attention network is used to calculate the attention coefficients of adjacent nodes in the supply chain correlation graph based on the multimodal embedding vector and update the node representations, generating risk-enhanced node representations. These representations are then concatenated with the original multimodal embedding vectors using residual concatenation. The concatenated result is fed into a gate control loop unit to iteratively update the hidden state step by step and output the risk gate value. This design enables the intensity of risk transmission between nodes in the graph to be dynamically enhanced under attention weighting. At the same time, the residual structure preserves the original multimodal information from being submerged by the deep network, while the gated recurrent unit further continuously models the nonlinear evolution of the risk state in the time dimension, capturing the complete trajectory of risk accumulation from quantitative to qualitative change.

[0019] The risk warning module stores historical risk threshold values ​​by maintaining a sliding window queue with a fixed window size. It calculates the arithmetic mean of all threshold values ​​in the queue and multiplies it by a relaxation coefficient to obtain a dynamic threshold. The current risk threshold is compared with the dynamic threshold in real time to determine whether a warning is triggered. This mechanism allows the judgment boundary to adaptively shift with the overall fluctuations of the risk environment. During low-risk periods, the threshold is tightened to reduce false alarms, while during high-risk periods, the threshold is appropriately relaxed to avoid missed alarms, thus solving the problem of insufficient adaptability of fixed thresholds under different market conditions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of a financial risk control system based on big data and artificial intelligence. Figure 2 This is a flowchart of the heterogeneous financial data modality alignment and multimodal embedding vector generation process; Figure 3 This is a flowchart of risk enhancement node representation and feature fusion based on graph attention networks; Figure 4This is a flowchart of a process that integrates risk characteristics into a gating loop unit and outputs a risk gating value. Figure 5 It is a flowchart of risk warning, interpretation generation, and trend prediction based on dynamic thresholds; Figure 6 This is a flowchart of risk summary extraction based on temporal convolution and graph pooling; Figure 7 It is a distribution density map of the maximum cosine similarity between the financial statement text and the transaction record numerical distribution characteristics; Figure 8 It is the attention coefficient sorting curve of adjacent nodes in the supply chain relationship graph; Figure 9 It is a time series curve of the risk threshold value; Figure 10 It shows the curves of risk threshold and dynamic threshold changes, as well as a schematic diagram of risk warning triggers. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1 This invention provides a financial risk control system based on big data and artificial intelligence, including a data acquisition module that acquires heterogeneous financial data such as financial statement text sequences, transaction record time-series sequences, and supply chain correlation graphs; a modal alignment module that performs modal alignment operations on the heterogeneous financial data, semantically mapping each text unit in the financial statement text sequence to a numerical unit with the same timestamp in the transaction record time-series sequence to generate a multimodal embedding vector; and a graph attention module that inputs the multimodal embedding vector into the first layer of a graph attention network, calculates the attention coefficient between any two adjacent nodes in the supply chain correlation graph, and then... The attention coefficient updates the initial representation of each node in the supply chain association graph to obtain a risk-enhanced node representation; the residual fusion module performs residual concatenation of the risk-enhanced node representation and the multimodal embedding vector in the second layer of the graph attention network to obtain a fused risk feature; the gating loop module inputs the fused risk feature into the gating loop unit, which iteratively updates the hidden state according to time steps and outputs a risk gating value at each time step; the risk warning module triggers a risk warning when the risk gating value exceeds a dynamic threshold and stores the hidden state at the trigger time as a risk anchor vector.

[0024] Example 1 In specific implementation, please refer to Figure 2 The process of performing modal alignment on heterogeneous financial data, semantically mapping each text unit in the financial statement text sequence to a numerical unit with the same timestamp in the transaction record time sequence, and generating a multimodal embedding vector includes the following steps.

[0025] Extract the semantic feature vector of each text unit in the financial statement text sequence. The financial statement text sequence consists of multiple text units arranged in chronological order, each text unit corresponding to a paragraph or sentence in a financial report. Each text unit is processed by a pre-trained financial domain language model, which contains a 12-layer Transformer encoder, with a hidden dimension of 768 in each layer. The pre-trained financial domain language model was trained on a large-scale financial corpus using a masked language model task. For any text unit, the pre-trained financial domain language model takes the 768-dimensional vector corresponding to the [CLS] tag output by the last layer as the semantic feature vector of that text unit. The semantic feature vectors of all text units constitute the semantic feature vector set.

[0026] Extract the distribution feature vector for each numerical unit in the transaction record time series. The transaction record time series consists of numerical units arranged by timestamps. Each numerical unit contains multi-dimensional transaction indicators, such as transaction amount, number of transactions, and average transaction amount. For each numerical unit, obtain the local numerical unit set within the range of 5 timestamps before and after the timestamp of the numerical unit. Calculate the mean, standard deviation, skewness, and kurtosis of the local numerical unit set in each dimension. Concatenate the calculated mean, standard deviation, skewness, and kurtosis in dimensional order to form the distribution feature vector. If there are insufficient samples in a certain dimension in the local numerical unit set, fill them with the global mean.

[0027] Calculate the cosine similarity matrix between the semantic feature vector and the distribution feature vector. Suppose the financial statement text sequence contains M text units, and the transaction record time sequence contains N timestamps, each timestamp corresponding to a numerical unit. The semantic feature vector of the i-th text unit is denoted as... The distribution feature vector corresponding to the j-th timestamp is denoted as The element in the i-th row and j-th column of the cosine similarity matrix C Calculate using the following formula:

[0028] in, Let represent the semantic feature vector of the i-th text unit. This represents the distribution feature vector of the numerical unit corresponding to the j-th timestamp. This represents the dot product operation of vectors. Representing vectors L2 norm, Representing vectors The L2 norm of the cosine similarity matrix C. The row indices of the cosine similarity matrix C correspond to the order of text units in the financial statement text sequence, and the column indices correspond to the timestamp order in the transaction record time sequence.

[0029] In the cosine similarity matrix C, select the column index corresponding to the maximum value in each row, and determine the timestamp represented by that column index as the matching timestamp of the text unit. For the i-th row, compare all elements in the row vector and determine the column index of the maximum element. Then the first one in the time sequence of transaction records The timestamp is used as the matching timestamp for the i-th text unit.

[0030] The semantic feature vector of a text unit is concatenated with the distribution feature vector of the numerical unit corresponding to the matching timestamp to obtain a multimodal embedding vector. For the i-th text unit, its semantic feature vector... Distribution feature vector corresponding to the matching timestamp The vectors are concatenated to form a multimodal embedding vector, the dimension of which is the sum of the semantic feature vector dimension and the distribution feature vector dimension.

[0031] When the maximum value of a row in the cosine similarity matrix C is lower than a preset similarity threshold, the multimodal embedding vector corresponding to that text unit is set to a zero vector. The preset similarity threshold is set to 0.55. This threshold is determined based on the Pareto front between the modal alignment accuracy calculated on the validation dataset and different thresholds, taking the maximum threshold when the accuracy decreases by no more than 1%. If the maximum cosine similarity value of the i-th row is less than 0.55, it indicates that the semantic feature vector of the i-th text unit does not match the distribution feature vectors corresponding to all timestamps sufficiently. In this case, the multimodal embedding vector corresponding to the i-th text unit is directly set to a zero vector, and its dimension remains consistent with the dimensions of other normally generated multimodal embedding vectors.

[0032] See Figure 7 In the graph, the horizontal axis represents the maximum cosine similarity, ranging from 0 to 1, and the vertical axis represents the probability density. The black solid line curve is the distribution density estimate of the maximum cosine similarity, and the gray bars filled with diagonal lines are the probability density histograms of the samples. The width of the bars corresponds to the grouping interval of the maximum cosine similarity. The graph uses a dashed line to mark the preset similarity threshold of 0.55, which is perpendicular to the horizontal axis and runs through the entire probability density range.

[0033] As can be seen from the probability density distribution and histogram, most of the maximum cosine similarity values ​​are concentrated in the range above 0.55, with the probability density peaking in the range of approximately 0.65 to 0.85. This indicates a high semantic matching degree between the semantic feature vectors of the financial statement text units and the distribution feature vectors of the corresponding transaction record time series. The probability density decreases significantly in the interval with similarity less than 0.55, and the number of samples to the left of this interval is relatively small, reflecting a low matching degree between the corresponding text units and any timestamp numerical units.

[0034] This figure verifies the rationality of the preset similarity threshold of 0.55 described in Example 1. Specifically, setting the threshold in the range of low similarity distribution density effectively filters out text units with insufficient matching degree, setting their multimodal embedding vectors to zero vectors, thereby avoiding interference from invalid information with low matching degree in subsequent risk feature fusion and risk prediction modules. Through this threshold segmentation, matching accuracy is guaranteed while avoiding the loss of effective matching samples due to setting the threshold too high.

[0035] In summary, this figure visually demonstrates the matching effect between semantic feature vectors and distributed feature vectors in the modality alignment module through the probability density distribution of the maximum cosine similarity, supporting the technical solution of multimodal embedding vector selection and generation based on a similarity threshold of 0.55 in Example 1.

[0036] Example 2 In specific implementation, please refer to Figure 3 The process of inputting multimodal embedding vectors into the first layer of a graph attention network, calculating the attention coefficients of any two adjacent nodes in the supply chain association graph, and updating the initial representation of each node in the supply chain association graph based on the attention coefficients to obtain the risk enhancement node representation includes the following steps.

[0037] A supply chain relationship graph is a directed or undirected graph where nodes represent entities within the supply chain and edges represent supply and sales relationships or equity relationships between enterprises. The first layer of a graph attention network consists of a learnable linear transformation matrix, an attention vector, and a non-linear activation function. All trainable parameters required for the first layer are learned through an offline supervised training phase. During this phase, batches of historically confirmed and non-confirmed risk events in financial data are used as training samples. Each training sample contains a complete multimodal embedding vector, the supply chain relationship graph structure, and a risk label (a binary label of 0 or 1). The cross-entropy loss function is used to measure the difference between the network's output risk prediction and the risk label. The Adam optimizer is used to update parameters with a learning rate of 0.001 and a batch size of 32. Training iteratively stops when the validation set loss no longer decreases, and the parameters of the first, second, and subsequent layers of the graph attention network are saved at this point.

[0038] During the generation of risk-enhancing node representations, each node in the supply chain graph is initialized as a vector fragment in the multimodal embedding vector that matches the corresponding node identifier. The total number of nodes in the supply chain graph is denoted as K, and the dimension of the multimodal embedding vector is D. Vectors of dimension D are retrieved from the multimodal embedding vector set according to node identifier index as the initial vectors for the corresponding nodes. For the first node p and the second node q connected by an edge in the supply chain graph, the current vector of the first node is obtained. The current vector of the second node The current vector is the initial vector during the first calculation. and Perform vector concatenation to obtain a concatenated vector with a dimension of 2D. The concatenated vector is input into a single-layer feedforward neural network, which consists of a learnable weight vector *a* and a linear transformation matrix *W*. The linear transformation matrix *W* has a dimension of 2D×F, where *F* is the dimension of the attention hidden layer, and its value is 128. This value is based on the stable performance obtained through grid search when the dimension *D* of the multimodal embedding vector after modal alignment is 1024. The single-layer feedforward neural network uses a leaky ReLU function to perform a nonlinear transformation on the intermediate results, outputting the original attention value of the first node *p* relative to the second node *q*. The calculation method is based on the following formula:

[0039] in, This represents the current vector of the first node p, with dimension D; Let represent the current vector of the second node q, with dimension D; represents the vector concatenation operation; W represents the linear transformation matrix of a single-layer feedforward neural network with dimension F×2D; a represents the attention weight vector of a single-layer feedforward neural network with dimension F; This represents a linear rectified function with leakage, and the negative slope is 0.2. During the offline supervised training phase, the attention weight vector *a* and the element values ​​of the linear transformation matrix *W* are automatically determined using the backpropagation algorithm.

[0040] The original attention value of the first node p relative to the second node q The normalized attention coefficient of the first node p relative to the second node q is calculated using the normalized exponential function. The calculation of the normalized exponential function is limited to the set of all neighboring nodes of the first node p. Let p represent the set of neighboring nodes of the first node, and let p represent the normalized attention coefficient. Depend on After exponential normalization, the sum of all normalized attention coefficients of the first node is equal to 1.

[0041] Normalized attention coefficients of all neighboring nodes q corresponding to the first node p. Multiply by the current vector of the corresponding neighboring node q respectively This yields multiple weighted adjacency vectors. For each node belonging to the set of adjacent nodes... For each neighboring node q in the array, calculate the weighted neighbor vector. The weighted adjacent vectors are summed element-wise, resulting in an aggregated vector of dimension D. This aggregated vector is then input into a nonlinear activation function, specifically the Exponential Linear Unit (ELU) activation function, to output a new representation of the first node p. The ELU activation function outputs a slow, smooth non-zero value when the input is less than 0, thus avoiding the problem of neuron death.

[0042] Traverse all nodes in the supply chain graph and perform the aforementioned attention-based vector update operation on each node to obtain new representations for each node. Stack all the new representations of the nodes in the order of their node numbers in the supply chain graph to form a risk enhancement node representation matrix of dimension K×D. Each row in this risk enhancement node representation matrix corresponds to the risk enhancement feature of a node.

[0043] In practice, the process of concatenating the risk enhancement node representation with the multimodal embedding vector in the second layer of the graph attention network to obtain the fused risk features includes the following steps.

[0044] The risk-enhancing node representation and the multimodal embedding vector are added element-wise according to the position of the corresponding node. For the k-th node in the supply chain relationship graph, let the vector of the k-th node in the risk-enhancing node representation be denoted as . The original vector of the k-th node in the multimodal embedding vector is The first intermediate residual is obtained by adding each element. The intermediate results of the first residuals of all nodes constitute the intermediate matrix of the first residuals.

[0045] The first residual intermediate matrix is ​​input into the first linear transformation layer, which is composed of a weight matrix. and bias vector Composition, weight matrix The dimension is D×D, and the bias vector is... The dimension is D. The first linear transformation layer performs a linear mapping on each row of the first residual intermediate matrix and outputs the first linear transformation vector. The dimension of the first linear transformation vector of each node is D.

[0046] The multimodal embedding vector matrix is ​​directly input into the second linear transformation layer, which is also composed of a weight matrix. and bias vector Composition, weight matrix The dimension is D×D, and the bias vector is... The dimension of the first linear transformation layer is D. The second linear transformation layer performs a linear mapping on each row of the multimodal embedding vector matrix, outputting a second linear transformation vector. The dimension of the second linear transformation vector of each node is D. All parameters of the first and second linear transformation layers are jointly learned with the parameters of the first layer of the graph attention network during the offline supervised training phase, and the initial values ​​are uniformly initialized using Xavier.

[0047] The first linear transformation vector and the second linear transformation vector are concatenated along the feature dimension. For the k-th node in the supply chain relationship graph, the first linear transformation vector of dimension D and the second linear transformation vector of dimension D are concatenated end to end to obtain a concatenated vector of dimension 2D. The concatenated vectors of all nodes constitute a concatenation matrix.

[0048] The concatenated matrix is ​​input to a batch normalization layer. During training, the batch normalization layer maintains the running mean and variance. During inference, it uses the global mean and variance collected during training to normalize each dimension of the concatenated matrix, resulting in a normalized concatenated vector. This normalized concatenated vector is 2D and is directly output as the fused risk feature. The fused risk feature simultaneously includes risk enhancement information after graph structure propagation and original multimodal embedding information, and its feature distribution is stabilized through residual paths and batch normalization.

[0049] See Figure 8 In the graph, the horizontal axis represents the edge index in the supply chain relationship graph, arranged in ascending order according to the corresponding attention coefficient, ranging from 0 to 800; the vertical axis represents the attention coefficient of the adjacent nodes corresponding to each edge, ranging from 0 to 1. The curve in the graph is a single black broken line, depicting the distribution of attention coefficients among all adjacent node pairs calculated by the first layer of the graph attention network.

[0050] As shown in the figure, the attention coefficient exhibits a slow overall upward trend. In the initial stage (edge ​​numbers 0 to approximately 100), the attention coefficient is low, ranging from 0 to 0.2, indicating that the risk correlation between the nodes connected by these edges is weak, and the weight of the transmitted risk information is small. In the middle stage (edge ​​numbers 100 to approximately 700), the attention coefficient gradually increases, concentrated in the range of 0.2 to 0.6, reflecting a moderate level of risk transmission between most adjacent nodes in the supply chain network. In the final stage (edge ​​numbers 700 to 800), the attention coefficient rises rapidly to approximately 0.78, indicating that the risk linkage between nodes connected by some key edges is strong, and they have a significant influence in the process of updating the representation of risk-enhancing nodes.

[0051] The trend of this curve aligns with the steps in the graph attention module that calculate the attention coefficients for any two adjacent nodes in the supply chain graph. It reflects the distribution characteristics of the attention weights calculated through a nonlinear mapping of node vectors concatenated using a single-layer feedforward neural network and the LeakyReLU activation function. The sorted attention coefficient curve reflects the model's differentiated assignment of weights to the risk propagation paths between nodes in the supply chain, which is beneficial for enhancing the expressive power of risk-enhancing node representations through weighted adjacency node vector aggregation.

[0052] In summary, the figure accurately illustrates the distribution characteristics and ranking trends of the attention coefficients of adjacent nodes calculated in the first layer of the graph attention module in Example 2. This verifies that the graph attention network can effectively distinguish the risk transmission intensity between different nodes in the supply chain, providing basic data support for the subsequent generation of risk enhancement node representations and fusion risk features.

[0053] Example 3 In specific implementation, please refer to Figure 4 The process of inputting the integrated risk features into the gating loop unit, which iteratively updates the hidden state according to time steps and outputs a risk gating value at each time step includes the following steps.

[0054] The core architecture of the gated recurrent unit consists of a reset gate, an update gate, a candidate hidden state calculation unit, and a risk-gated output layer. The hidden state dimension of the gated recurrent unit is set to 256. This dimension value is determined through cross-validation of the candidate hidden state dimension set {128, 256, 512} during the offline supervised training phase, with the dimension that minimizes the loss being used as the final setting. All parameters of the weight matrices and bias vectors in the gated recurrent unit architecture are jointly trained with the parameters of the first and second layers of the graph attention network during the offline supervised training phase. The training configuration remains consistent with the aforementioned offline supervised training phase. The loss function uses the risk gating value and the binary cross-entropy loss of the risk label from the last time step, backpropagated.

[0055] The fused risk features are divided into multiple time slices according to time steps, with each time slice corresponding to a subset of fused risk features for one time step. The total number of time steps for the fused risk features is denoted as T. The subset of fused risk features for each time step is the row vector in the fused risk feature matrix corresponding to that time step, with the dimension being the dimension of the fused risk features, denoted as T. The splitting process extracts the corresponding row vectors sequentially from time step index 1 to T to form a time step sequence.

[0056] At the t-th time step of the gated recurrent unit, a subset of fused risk features from the previous hidden state and the current time slice is obtained. The previous hidden state is denoted as... The dimension is 256; the subset of fused risk features of the current time slice is denoted as... , dimension When t=1, the previous hidden state Set it to a vector of all zeros. and Perform vector concatenation to obtain a vector with dimension . The concatenated vector is input into the reset gate.

[0057] The reset gate consists of a fully connected layer and an sigmoid growth curve activation function. The weight matrix of the fully connected layer is denoted as... , dimension The bias vector of a fully connected layer is denoted as The dimension is 256. The reset gate outputs the reset gate vector. The calculation formula is:

[0058] in, Let represent the reset gate vector at time step t, with a dimension of 256; This represents the activation function for the S-shaped growth curve, with an output range of (0,1). This represents the weight matrix of the reset gate; This represents the previous hidden state vector, with a dimension of 256; This represents a vector concatenation operation; This represents a subset of the fusion risk features of the current time slice, with dimensions of [missing information]. ; This represents the bias vector of the reset gate. Optionally, during offline supervised training, the weight matrix of the reset gate... and bias vector The initial values ​​are initialized using a Glorot uniform distribution.

[0059] Reset the gate vector Compared to the previous hidden state Element-wise multiplication yields the reset hidden state. The element-wise multiplication operation resets the gate vector. The i-th component multiplied by the previous hidden state The i-th component is used to selectively retain some information from the previous hidden state, and the dimension of the reset hidden state is 256.

[0060] The risk feature subset of the fusion of the reset hidden state with the current time slice. Perform vector concatenation to obtain a vector with dimension . The candidate concatenation vector is then input into the candidate hidden state computation unit. The candidate hidden state computation unit consists of a fully connected layer and a hyperbolic tangent activation function. The weight matrix of the fully connected layer is denoted as... , dimension The bias vector of a fully connected layer is denoted as The dimension is 256. The candidate hidden state calculation unit outputs the candidate hidden state. The dimension is 256.

[0061] Previous hidden state Fusing risk feature subset with the current time slice After vector concatenation, the input is fed into the update gate. The update gate also consists of a fully connected layer and an sigmoid growth curve activation function. The weight matrix of the fully connected layer is denoted as... , dimension The bias vector of a fully connected layer is denoted as The dimension is 256. The update gate outputs the update gate vector. With a dimension of 256, update the gate vector. Each component takes a value in the interval (0,1).

[0062] Based on the updated gate vector For the previous hidden state and candidate hidden state Perform linear interpolation to obtain the current hidden state at time step t. The linear interpolation is performed by updating the gate vector. Element-wise multiplication by candidate hidden state ,Will Element-wise multiplication by the previous hidden state The current hidden state is obtained by adding the two parts element by element. The dimension is 256. Currently hidden. It serves as the final hidden state at time step t and is passed to time step t+1 as the previous hidden state.

[0063] Set the current hidden state at time step t. The input is fed to the risk-gated output layer, which consists of a fully connected layer and an S-shaped growth curve activation function. The weight matrix of the fully connected layer has a dimension of 256×1, and the bias vector has a dimension of 1. The risk-gated output layer will... A linear mapping is performed to a scalar value, which is then mapped to the (0,1) interval using an S-shaped growth curve activation function. The mapping result serves as the risk gate value for the t-th time step. The closer the risk gate value is to 1, the higher the risk level at the corresponding time step.

[0064] See Figure 9In the graph, the horizontal axis represents the time step, ranging from 0 to 600, and the vertical axis represents the risk threshold value, ranging from 0 to 1. The legend indicates that the curve is the "risk threshold value," corresponding to the risk threshold scalar output by the gated loop unit at each time step. The risk threshold value generally shows a fluctuating upward trend. In the initial stage (0 to 100 time steps), the risk threshold value fluctuates between 0.2 and 0.45, indicating a low and relatively stable risk level. In the middle stage (100 to 400 time steps), the risk threshold value fluctuates significantly, roughly between 0.2 and 0.6, with multiple short-term peaks reflecting periodic fluctuations in the risk status. In the later stage (400 to 600 time steps), the risk threshold value continues to rise, especially after 550 time steps, rapidly rising to above 0.8, indicating a significant increase in risk level, which may trigger a risk warning mechanism.

[0065] The curve's trend reflects the gating loop unit's ability to dynamically capture the temporal risk of the fused risk feature sequence in Example 3. The gating loop unit dynamically outputs risk gate values ​​by iteratively updating the hidden states, reflecting subtle changes in risk level over time. The overall rise of the curve and its rapid climb at the end demonstrate the system's timely identification of potential high-risk events, providing a quantitative basis for triggering risk warnings in Example 4. The overall curve change fully illustrates that after processing by the gating loop unit, the fused risk features can effectively reflect the dynamic characteristics of financial risk over time, providing accurate risk gate signals for the financial risk control system.

[0066] Example 4 In specific implementation, please refer to Figure 5 The process of triggering a risk warning when the risk threshold exceeds the dynamic threshold and storing the hidden state at the trigger moment as a risk anchor vector includes the following steps.

[0067] Initialize a sliding window queue of fixed size, storing the most recent preset number of historical risk threshold values ​​into the sliding window queue. The fixed window size is set to 60, which is based on the decay period of the autocorrelation function of the risk threshold values ​​obtained from historical risk data statistics, taken from the number of lag steps when the autocorrelation function value first drops below 0.2. The sliding window queue adopts a first-in-first-out linear queue structure, with each slot storing a scalar risk threshold value. When processing the t-th time step, the sliding window queue stores 60 historical risk threshold values ​​corresponding to time steps t-60 to t-1.

[0068] Calculate the arithmetic mean of all historical risk gate values ​​in the sliding window queue, and multiply the arithmetic mean by a preset relaxation coefficient to obtain the dynamic threshold. The m-th historical risk gate value in the sliding window queue is denoted as... The total number of historical risk gate values ​​in the sliding window queue is denoted as W, where W equals the fixed window size of 60. (Arithmetic mean) Calculated as The preset relaxation coefficient is denoted as... The value range is set between 1.05 and 1.50, and the initial value is set at 1.20. This initial value is obtained by analyzing the risk threshold values ​​of confirmed risk events and risk threshold values ​​of periods without risk occurrence in historical risk data using receiver operating characteristic curve analysis, and taking the threshold amplification factor corresponding to the maximum value of the Youden index.

[0069] Preset relaxation coefficient It is adaptively adjusted based on the frequency of historical risk events. The adaptive adjustment method is as follows: maintain a risk event counter with a length of 90 days, record the total number of risk warnings triggered in the most recent 90 days, and if the total number of risk warnings triggered in the most recent 90 days is greater than or equal to 5, adjust the slack coefficient. Subtracting 0.05 from the current value, the lower limit of the adjusted relaxation coefficient is 1.05; if the total number of risk warnings triggered in the last 90 days is 0, the relaxation coefficient will be adjusted accordingly. Increase the current value by 0.05, and the adjusted relaxation coefficient upper limit is 1.50; if the total number of risk warnings triggered in the last 90 days is between 1 and 4, maintain the relaxation coefficient. Unchanged. Dynamic threshold. Calculated according to the formula:

[0070] in, The dynamic threshold at time step t is a scalar value. This represents the preset relaxation coefficient, with a value range of [1.05, 1.50]. This represents the arithmetic mean of all historical risk gating values ​​in the sliding window queue.

[0071] Real-time comparison of the risk gate value output at the current time step with the dynamic threshold. If the risk threshold is greater than the dynamic threshold. If the risk threshold is less than or equal to the dynamic threshold, a risk trigger flag is generated and set to 1; If so, the risk trigger flag will remain at 0.

[0072] In response to the risk trigger flag being set to 1, the current hidden state output by the gating recurrent unit at the current time step is extracted. The current hidden state is the hidden state vector obtained by the gating recurrent unit after iterative updates at time step t, denoted as... The dimension is 256. Read from the hidden state cache of the gated loop unit. ,Will Copy to the risk anchor point storage area. The risk anchor point storage area is a separate memory region used to persistently store the hidden state vector at the moment the warning is triggered, and the storage format is a floating-point array.

[0073] The risk trigger flag and risk anchor vector are associated and stored in the risk event log. The risk event log is a structured data table containing a timestamp field, a risk trigger flag field, a risk gating value field, and a risk anchor vector field. After saving, all historical risk gating values ​​in the sliding window queue are immediately cleared, and all slots in the sliding window queue are reset to zero.

[0074] After triggering a risk warning, the process of generating explanatory text for the risk warning includes the following steps. The risk anchor vector is input to the back attention module. The input to the back attention module is the risk anchor vector and the set of historical hidden states of the gated recurrent unit at each historical time step before the triggering time. The triggering time is denoted as the T-th triggering time, and the range of historical time steps is from the 1st time step to the T-th triggering time - 1. The back attention module consists of a trainable attention query vector, a key mapping matrix, and a value mapping matrix. The key mapping matrix and the value mapping matrix are both 256×128 in dimension, and the attention query vector is 128 in dimension. The parameters of the back attention module are jointly trained with the risk warning explanation generation task during offline supervised training. During training, manually annotated risk factor explanation text is used as the supervision signal. The parameters of the back attention module are updated by minimizing the cross-entropy between the probability distribution output by the explanation generation module and the annotated text.

[0075] The reverse attention module calculates the reverse attention weights of the historical hidden state and the risk anchor vector for each of the Z historical time steps prior to the triggering time of the gated recurrent unit. For the historical hidden state at the z-th historical time step... The risk anchor vector is mapped to a key vector through a key mapping matrix. The query vector is mapped to a value vector using a value mapping matrix. The inner product of the query vector and the key vector is calculated, and the inverse attention weights are obtained using a normalized exponential function. A scalar inverse attention weight value is obtained at each historical time step, and all weights constitute an inverse attention weight vector of dimension Z.

[0076] For each historical hidden state at each historical time step, apply the corresponding reverse attention weights and perform a weighted summation, then sum the results for each historical hidden state. Multiply by its reverse attention weight, and sum all the weighted vectors element by element to obtain the contribution feature vector with a dimension of 256.

[0077] The contributing feature vector is input into the interpretation generation module. The core architecture of this module is the encoder portion of a vector quantization variational autoencoder, consisting of three fully connected layers with a hidden layer dimension of 128 and an output layer dimension of 32. After layer-by-layer nonlinear mapping through the three fully connected layers, the contributing feature vector outputs a continuous semantic vector. This continuous semantic vector is then discretized by searching for the codebook vector with the highest cosine similarity to the continuous semantic vector within a pre-constructed learnable embedding codebook. The learnable embedding codebook contains 256 codebook entries, each with a dimension of 32. The index number corresponding to the codebook vector serves as the discrete risk factor index, outputting a discrete risk factor index sequence. The length of this sequence corresponds to the number of risk factors identified by the interpretation generation module.

[0078] Each discrete risk factor index in the discrete risk factor index sequence is mapped to a pre-defined risk explanation vocabulary. The pre-defined risk explanation vocabulary contains 256 entries, each consisting of a risk explanation phrase and a corresponding index number. The risk explanation phrase is Chinese text with a length not exceeding 10 characters, covering types such as "sharp drop in accounts receivable turnover," "debt ratio exceeding the warning line," and "abnormal growth in related-party transactions." The corresponding risk explanation phrase is retrieved from the pre-defined risk explanation vocabulary based on the discrete risk factor index.

[0079] All retrieved risk explanation phrases are concatenated in chronological order. The chronological order is determined by arranging the timestamps of the original historical time steps corresponding to each discrete risk factor index in ascending order. Each risk explanation phrase is separated by a semicolon. After concatenation, a risk warning explanation text is generated. The risk warning explanation text is associated with the risk trigger flag and stored in the risk event log, specifically in the risk warning explanation text field.

[0080] After storing the risk anchor vectors, the process of risk trend prediction is performed. Multiple risk anchor vectors stored from consecutive historical risk warnings are retrieved. The scope of historical risk anchor vector retrieval includes complete risk event records within the last three years. The risk anchor vectors corresponding to each risk event are extracted to form a set, and the total number of risk anchor vectors in the set is denoted as S.

[0081] Calculate the average Euclidean distance between each risk anchor vector and the remaining risk anchor vectors to obtain the clustering feature value for each risk anchor vector. For the s-th risk anchor vector... ,calculate With the u-th risk anchor vector Euclidean distance , where u iterates through all risk anchor vector indices except s, resulting in a total of S-1 distance values, and the clustering feature value is the arithmetic mean of the S-1 distance values.

[0082] The clustering feature values ​​of all risk anchor vectors are sorted in ascending order, and the anchor vector with the smallest clustering feature value is selected as the reference anchor. The reference anchor represents the most representative vector in the geometric center neighborhood of the risk anchor vector set.

[0083] Perform vector difference operations between each of the remaining risk anchor vectors and the reference anchor vector to obtain multiple difference vectors. For the s-th risk anchor vector... and Calculate the difference vector instead of using a reference anchor point. ,in, As the reference anchor point, the difference vector The dimension is 256.

[0084] Multiple difference vectors are input into the encoding part of the autoencoder. The encoding part of the autoencoder has a three-layer fully connected network structure with hidden layer dimensions of 128 and 64, and the output layer latent evolution vector has a dimension of 16. The encoding part maps each difference vector to a latent evolution vector of dimension 16. The parameters of the autoencoder are pre-trained offline under supervised training. The training data used during pre-training is the set of difference vectors of all historical risk anchor vectors. The decoding part of the autoencoder is symmetrical to the encoding part, and the reconstruction loss is the mean squared error. The parameters of the encoding part are saved after training until the reconstruction loss converges.

[0085] The latent evolution vectors are input into the trend prediction layer, which consists of a Long Short-Term Memory (LSTM) network with a hidden state dimension of 32 and two layers. The trend prediction layer reads the time series sequence of all latent evolution vectors up to the current moment and outputs the predicted risk gating sequence within the future time window, with the future time window length set to 30 time steps.

[0086] Whether to generate an early warning signal depends on whether any predicted value in the predicted risk gating sequence exceeds the dynamic threshold. If any predicted value in the predicted risk gating sequence over the next 30 time steps exceeds the current dynamic threshold, an early warning signal is generated and stored in the early warning field of the risk event log; if none of the predicted values ​​in the predicted risk gating sequence exceed the dynamic threshold, no early warning signal is generated.

[0087] See Figure 10 In the figure, the horizontal axis represents the time step, and the vertical axis represents the range of risk gate values ​​and dynamic thresholds, ranging from 0 to 1. The solid curves in the figure represent the risk gate values ​​output by the gate control loop unit at each time step, the dashed curves represent the dynamic thresholds calculated based on the historical risk gate values ​​in the sliding window queue, and the △ symbol represents the risk warning triggered at the corresponding time step.

[0088] As shown in the graph, the risk threshold curve exhibits fluctuations, generally ranging from 0.1 to 0.7. The dynamic threshold curve is relatively smooth, with a value range of approximately 0.2 to 0.65. The dynamic threshold is adaptively adjusted by multiplying the arithmetic mean of 60 historical risk threshold values ​​by a relaxation coefficient, effectively reflecting the long-term trend and volatility of the risk threshold.

[0089] Between time steps approximately 30 and 150, the risk threshold value repeatedly exceeded the dynamic threshold, triggering multiple risk warnings. The dense distribution of △ symbols indicates a persistently high risk level during this phase. Subsequently, between time steps 150 and 250, the risk threshold value decreased significantly and remained below the dynamic threshold, with fewer risk warning triggers, indicating a stable phase with a lower risk level.

[0090] Between time steps 250 and 500, the risk threshold repeatedly approached or exceeded the dynamic threshold, especially between approximately time steps 400 and 480, where the risk threshold showed a significant peak, and the dynamic threshold also increased accordingly, triggering multiple risk warnings, indicating a high level of risk activity during this stage. After 500 time steps, the risk threshold gradually fell back below the dynamic threshold, the number of risk warnings triggered decreased, and the risk level tended to stabilize.

[0091] Whenever the risk threshold exceeds the dynamic threshold, the system immediately triggers a risk warning and stores the hidden state vector at that moment as a risk anchor vector, serving as the basis for subsequent risk interpretation and trend prediction. The dynamic comparison mechanism between the risk threshold and the dynamic threshold, as well as the design of the sliding window and relaxation coefficient, ensure the timeliness and dynamic adaptability of the risk warning, effectively avoiding false alarms caused by short-term fluctuations in the risk threshold.

[0092] In summary, this figure fully illustrates the dynamic process in Example 4 where the risk warning module compares the risk gate value output by the gated loop unit with the dynamic threshold in real time, triggers a risk warning, and stores the risk anchor point. It clearly shows the changing trends of the risk gate value and the dynamic threshold over time and the triggering distribution characteristics of the risk warning.

[0093] Example 5 In specific implementation, please refer to Figure 6 After generating the multimodal embedding vector, the multimodal embedding vector is also input into a temporal convolutional network to enhance the temporal expressive power of the multimodal embedding vector. The process of obtaining the enhanced multimodal embedding vector includes the following steps.

[0094] The temporal convolutional network consists of three stacked dilated convolutional layers. The kernel size of each dilated convolutional layer is set to 3, and the dilation rates of the three layers are set to 1, 2, and 4, respectively. The dilation rates are set according to powers of 2, allowing higher-level dilated convolutional layers to cover an exponentially increasing receptive field. The number of kernels in each dilated convolutional layer is consistent with the feature dimension of the multimodal embedding vector, denoted as _____. All trainable convolutional kernel parameters of the temporal convolutional network are jointly trained with the parameters of the first layer of the graph attention network, the second layer of the graph attention network, and the gated recurrent unit during the offline supervised training phase. The training configuration is consistent with that of the aforementioned offline supervised training phase.

[0095] The multimodal embedding vectors are arranged into a temporal feature matrix according to time steps. The number of rows in the temporal feature matrix is ​​the total number of time steps, and the number of columns is... The temporal feature matrix is ​​input into a temporal convolutional network. The first dilated convolutional layer uses a dilated kernel with a dilation rate of 1 and a kernel size of 3 to perform a sliding window convolution operation along the time dimension of the input matrix. The sliding window stride is 1, and zero padding is used at the boundaries of the time dimension to keep the output time step count consistent with the input time step count. The output of the first dilated convolutional layer is fed into the second dilated convolutional layer, which uses a dilated kernel with a dilation rate of 2 for convolution. The third dilated convolutional layer uses a dilated kernel with a dilation rate of 4 for convolution. The output channels of all three dilated convolutional layers are... The output of the third dilated convolutional layer serves as a multi-scale temporal feature map. The number of rows in the multi-scale temporal feature map is equal to the total number of time steps, and the number of columns is... .

[0096] The multi-scale temporal feature map is input into the channel attention module, which performs two pooling operations on each channel of the multi-scale temporal feature map. The channel attention module consists of a global average pooling branch and a global max pooling branch operating in parallel. The global average pooling branch calculates the arithmetic mean of all elements in the c-th channel of the multi-scale temporal feature map across all time steps, outputting a scalar value. The global max pooling branch extracts the maximum value of all elements in the c-th channel of the multi-scale temporal feature map across all time steps, outputting a scalar value. Each channel performs the two pooling operations described above, and the global average pooling branch output dimension is... The average pooling description vector, and the output dimension of the global max pooling branch are... The max-pooling descriptor vector is obtained by concatenating the average-pooling descriptor vector and the max-pooling descriptor vector along the channel dimension. The channel descriptor vector has the following dimension: .

[0097] The channel description vector is input into a shared multilayer perceptron, which consists of two fully connected layers. The input dimension of the first fully connected layer is... The output dimension is The input dimension of the second fully connected layer is The output dimension is The first fully connected layer is activated using a linear rectified function. The output of the second fully connected layer is mapped to the (0,1) interval using a sigmoid growth curve activation function to obtain the channel attention weights. Each channel attention weight is a dimensional... The vectors. All trainable weight parameters of the shared multilayer perceptron are jointly trained with the parameters of other modules in the system during the offline supervised training phase.

[0098] The channel attention weights are multiplied channel by channel in the multi-scale temporal feature map. For the feature values ​​of all time steps corresponding to the c-th channel of the multi-scale temporal feature map, the result is multiplied by the scalar value of the c-th element in the channel attention weight vector to obtain the weighted temporal feature. The number of rows in the weighted temporal feature is equal to the total number of time steps, and the number of columns is... .

[0099] The weighted temporal features are residually concatenated with the temporal feature matrix of the original input multimodal embedding vector. The residual concatenation method is element-wise addition, resulting in an enhanced multimodal embedding vector with the same dimension as the original multimodal embedding vector. The enhanced multimodal embedding vector replaces the original multimodal embedding vector and is input into the first layer of the graph attention network for further processing.

[0100] In practice, after obtaining the risk-enhanced node representation, the risk-enhanced node representation is input into the graph pooling layer to extract graph-level risk summary information. The process of fusing the graph-level risk summary vector with the fused risk features includes the following steps.

[0101] The graph pooling layer consists of trainable projection vectors and a TopK node selection mechanism. Risk-enhancing node representations are input to the graph pooling layer as K×D matrices, where K is the total number of nodes in the supply chain graph and D is the dimension of the risk-enhancing node vector. The trainable projection vector is denoted as p, with dimension D. The initial values ​​of the trainable projection vector p are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.1, and updated during offline supervised training using the backpropagation algorithm.

[0102] For the k-th node in the supply chain relationship graph, extract the vector corresponding to the k-th node from the risk enhancement node representation matrix. The dimension is D. Calculate the score of the k-th node, the score is... The score is the absolute value of the inner product of the trainable projection vector p. The calculation formula is:

[0103] in, The score of the k-th node is represented by a scalar value. Let represent the risk enhancement node representation vector of the k-th node, with dimension D; p represents the trainable projection vector, with dimension D. This represents the dot product operation of vectors. The node score reflects the response strength of the risk enhancement feature of the k-th node in the projection direction.

[0104] Sort all nodes in the supply chain graph in descending order based on their scores, and then arrange the K nodes by score from highest to lowest. Select the K nodes with the highest scores as retained nodes, where K is a preset pooling retention number. The preset pooling retention number K is [value missing]. ,in This is the pooling ratio. The value is 0.6, which is determined by cross-validating the pooling ratio candidate set {0.4, 0.6, 0.8} during the offline supervised training phase, to verify the pooling ratio with the lowest loss as the final setting; This indicates the rounding up operation.

[0105] For each retained node, its first-order neighbors are located in the original supply chain graph. First-order neighbors are all nodes in the original supply chain graph that have a direct edge connection to the retained node. Nodes in the first-order neighbor set that also belong to the retained node set are selected. In the generated pooled subgraph structure, an edge is added between the retained node and the selected first-order neighbors. If there is an edge connection between retained nodes in the original supply chain graph, this edge connection is retained in the pooled subgraph structure; if there is an edge connection in the original supply chain graph but one end of the edge is not retained, this edge connection is discarded in the pooled subgraph structure.

[0106] The risk-enhanced node representations of all retained nodes in the subgraph structure are weighted and summed according to the scores of the retained nodes. For the u-th retained node in the subgraph structure, its score is... The risk enhancement node representation vector is The weighted vector is The weighted vectors of all retained nodes are summed element-wise, and the resulting sum has dimension D. This sum is used as the graph-level risk summary vector.

[0107] The graph-level risk summary vector and the fused risk feature are concatenated and then input into the gated recurrent unit. The fused risk feature is a 2D vector. The D-dimensional graph-level risk summary vector and the 2D-dimensional fused risk feature are concatenated end-to-end along the feature dimension to obtain a 3D concatenated feature vector. This 3D concatenated feature vector is used as the input to the gated recurrent unit at each time step, replacing the input method that only contains the fused risk feature. The dimension of the fully connected matrix of the input layer of the gated recurrent unit is adjusted accordingly. The adjusted input layer parameters are reinitialized and jointly trained during the offline supervised training phase.

[0108] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A financial risk control system based on big data and artificial intelligence, characterized in that: include: The data acquisition module acquires heterogeneous financial data, including text sequences of financial statements, time-series sequences of transaction records, and supply chain relationship graphs. The modality alignment module performs modality alignment operations on the heterogeneous financial data, semantically mapping each text unit in the financial statement text sequence to the numerical units with the same timestamp in the transaction record time sequence, and generating a multimodal embedding vector. The graph attention module inputs the multimodal embedding vector into the first layer of the graph attention network, calculates the attention coefficients of any two adjacent nodes in the supply chain association graph, and updates the initial representation of each node in the supply chain association graph according to the attention coefficients to obtain the risk enhancement node representation. The residual fusion module performs residual concatenation of the risk enhancement node representation and the multimodal embedding vector in the second layer of the graph attention network to obtain fused risk features; The gated loop module inputs the fused risk features into the gated loop unit, which iteratively updates the hidden state according to time steps and outputs a risk gate value at each time step. The risk warning module triggers a risk warning when the risk threshold value exceeds the dynamic threshold, and stores the hidden state at the trigger time as a risk anchor vector.

2. The financial risk control system based on big data and artificial intelligence according to claim 1, characterized in that, The modal alignment module performs modal alignment operations including: Extract the semantic feature vector of each text unit in the financial statement text sequence, and simultaneously extract the distribution feature vector of each numerical unit in the transaction record time sequence; Calculate the cosine similarity matrix between the semantic feature vector and the distribution feature vector, where the rows of the cosine similarity matrix correspond to text units and the columns correspond to timestamps; Select the column index corresponding to the maximum value of each row in the cosine similarity matrix, and determine the timestamp represented by the column index as the matching timestamp of the text unit; The semantic feature vector of the text unit is concatenated with the distribution feature vector of the numerical unit corresponding to the matching timestamp to obtain the multimodal embedding vector.

3. The financial risk control system based on big data and artificial intelligence according to claim 2, characterized in that, When the maximum value of each row of the cosine similarity matrix is ​​lower than a preset similarity threshold, the multimodal embedding vector corresponding to that text unit is set to a zero vector.

4. The financial risk control system based on big data and artificial intelligence according to claim 1, characterized in that, The multimodal embedding vector is input into the first layer of the graph attention network. The attention coefficients of any two adjacent nodes in the supply chain association graph are calculated, and the initial representation of each node in the supply chain association graph is updated according to the attention coefficients to obtain the risk enhancement node representation, including: Each node in the supply chain association graph is initialized as a vector fragment in the multimodal embedding vector corresponding to that node; For any first node and second node connected by an edge in the supply chain association graph, the current vector of the first node and the current vector of the second node are concatenated and then input into a single-layer feedforward neural network, which outputs the original attention value. Applying a normalized exponential function to the original attention value yields the normalized attention coefficient of the first node relative to the second node; The normalized attention coefficients of all neighboring nodes corresponding to the first node are multiplied by the current vector of the corresponding neighboring node to obtain multiple weighted neighbor vectors. The sum of the multiple weighted neighbor vectors is then input into a nonlinear activation function to output a new representation of the first node. The new representations of all nodes are stacked according to the node order of the supply chain association graph to obtain the risk enhancement node representation.

5. The financial risk control system based on big data and artificial intelligence according to claim 4, characterized in that, The single-layer feedforward neural network uses the LeakyReLU activation function to perform a nonlinear transformation on the original attention value.

6. The financial risk control system based on big data and artificial intelligence according to claim 1, characterized in that, The risk enhancement node representation and the multimodal embedding vector are residually concatenated at the second layer of the graph attention network to obtain the fused risk features, including: The risk enhancement node representation and the multimodal embedding vector are added element-wise according to the position of the corresponding node to obtain the first residual intermediate result; The first residual intermediate result is input into the first linear transformation layer, and the first linear transformation layer outputs the first linear transformation vector. The multimodal embedding vector is input to the second linear transformation layer, and the second linear transformation layer outputs the second linear transformation vector; The first linear transformation vector and the second linear transformation vector are concatenated along the feature dimension to obtain the concatenated vector; The concatenated vector is input to the batch normalization layer, which outputs a normalized concatenated vector. The normalized concatenated vector is then used as the fusion risk feature.

7. The financial risk control system based on big data and artificial intelligence according to claim 1, characterized in that, The fused risk features are input into a gated loop unit, which iteratively updates the hidden state according to time steps and outputs a risk gating value at each time step, including: The fusion risk features are divided into multiple time slices according to time steps, and each time slice corresponds to a subset of fusion risk features of a time step; At the t-th time step of the gated loop unit, a subset of the fused risk features of the previous hidden state and the current time slice is obtained. The subset of the fused risk features of the previous hidden state and the current time slice is concatenated and then input into the reset gate. The reset gate outputs a reset gate vector. The reset gate vector is multiplied element by element with the previous hidden state to obtain the reset hidden state. The reset hidden state is then concatenated with the fusion risk feature subset of the current time slice and input into the candidate hidden state calculation unit. The candidate hidden state calculation unit outputs the candidate hidden state. The risk feature subset of the previous hidden state and the current time slice is concatenated and input into the update gate, and the update gate outputs the update gate vector. The previous hidden state and the candidate hidden state are linearly interpolated based on the update gate vector to obtain the current hidden state at the t-th time step; The current hidden state at the t-th time step is input to the risk gating output layer, and the risk gating output layer outputs the risk gating value at the t-th time step.

8. The financial risk control system based on big data and artificial intelligence according to claim 7, characterized in that, The risk-gated output layer uses the Sigmoid activation function to map the current hidden state at time step t to the (0,1) interval as the risk gating value.

9. The financial risk control system based on big data and artificial intelligence according to claim 1, characterized in that, When the risk threshold value exceeds the dynamic threshold, a risk warning is triggered, and the hidden state at the trigger time is stored as a risk anchor vector, including: Initialize a sliding window queue with a fixed window size, and store the most recent preset number of historical risk gate values ​​into the sliding window queue; Calculate the arithmetic mean of all historical risk gate values ​​in the sliding window queue, and multiply the arithmetic mean by a preset relaxation coefficient to obtain the dynamic threshold; The risk threshold value output at the current time step is compared with the dynamic threshold in real time. If the risk threshold value is greater than the dynamic threshold, a risk trigger flag is generated. In response to the risk trigger flag, the current hidden state output by the gated loop unit at the current time step is extracted, and the current hidden state is copied to the risk anchor storage area. The risk trigger flag is associated with the risk anchor vector and stored in the risk event log, and all historical risk gating values ​​in the sliding window queue are cleared.

10. The financial risk control system based on big data and artificial intelligence according to claim 9, characterized in that, The preset relaxation coefficient is adaptively adjusted based on the frequency of historical risk events.

Images