Decision parameter generation method and system based on multi-source heterogeneous data

Abstract

The application discloses a kind of decision parameter generation method and system based on multi-source heterogeneous data, according to multi-source heterogeneous data construction graph structure, generate space feature vector and calculate area correction coefficient;From multi-source heterogeneous data, extract entity and the relationship between entity and construct directed knowledge graph, learn the node feature of directed knowledge graph, determine the conduction path for the entity of conduction relationship and calculate conduction score;According to the feature similarity of adjacent node under competitive relationship, calculate competitive relationship score;Obtain time series feature for time series data in multi-source heterogeneous data, calculate growth score;With growth score, risk conduction score, competitive relationship score and area correction coefficient as condition vector, decision parameter generation is carried out.Path length is determined along the conduction relationship path to determine the conduction path intensity, and the conduction score is obtained according to the conduction path intensity, which represents the conduction path of risk from upstream to downstream node, and provides risk conduction quantization basis for downstream decision.

Description

Technical Field

[0001] This invention relates to data processing, and more particularly to a method and system for generating decision parameters based on multi-source heterogeneous data. Background Technology

[0002] Financial leasing, as a flexible financing method, is widely used in various industries, providing enterprises with the right to use equipment and assets while reducing their capital expenditures. However, the financial leasing process involves multiple participants and data from various channels and modalities. Current technologies cannot effectively integrate this data, which affects the accuracy and timeliness of risk assessment. If effective information cannot be accurately extracted from the data, it will lead to inaccurate decision-making results.

[0003] Decision-making faces the systemic impacts of a complex interplay of factors. These impacts are not isolated; they exhibit intricate nonlinear relationships and evolve dynamically over time. Traditional decision-making methods often rely on simplistic assumptions and static parameters, failing to capture dynamic correlations and the contagion effects of risks. For instance, a sudden adjustment in international trade policy can indirectly affect the financing costs and feasibility of a seemingly unrelated project through multiple channels, including supply chains, exchange rates, and market sentiment. Existing analytical methods lack the ability to effectively simulate and track such risk transmission paths, failing to dynamically update parameters based on upstream shocks. Consequently, generated parameters lag significantly when facing supply chain shocks, leading to incomplete risk assessments, insufficient decision-making basis, and inaccurate decision results. Summary of the Invention

[0004] Purpose of the invention: To address the above-mentioned shortcomings, this invention provides a method for generating decision parameters based on multi-source heterogeneous data to improve the accuracy of decision parameters. This invention also provides a system for generating decision parameters based on multi-source heterogeneous data.

[0005] Technical solution: The method for generating decision parameters based on multi-source heterogeneous data described in this invention includes the following steps:

[0006] Acquire multi-source heterogeneous data;

[0007] A graph structure is constructed based on the regional information in multi-source heterogeneous data. Features are extracted from the graph structure to generate spatial feature vectors. Regional correction coefficients are then calculated based on the spatial feature vectors.

[0008] This method extracts entities and relationships between entities from multi-source heterogeneous data, constructs a directed knowledge graph based on the entities and their relationships, learns the node features of the directed knowledge graph using a graph attention network, determines the transmission path for entities with transmission relationships, and determines the transmission path strength based on attention weights and path lengths, calculating a transmission score based on the transmission path strength. The node features include the static attributes of entity nodes, comprehensive semantic features, and positional encoding; the comprehensive semantic features are obtained by feature fusion of different modalities related to entities in the multi-source heterogeneous data.

[0009] Based on the node features of the directed knowledge graph learned by the graph attention network, the competition relationship score is calculated according to the feature similarity of adjacent nodes under the competition relationship.

[0010] Temporal features are obtained from time series data in multi-source heterogeneous data, and growth scores are calculated based on these temporal features;

[0011] Decision parameters are generated using growth potential score, risk transmission score, competitive relationship score, and regional correction coefficient as condition vectors.

[0012] Further, the specific steps for generating spatial feature vectors are as follows: Based on the regional information in the multi-source heterogeneous data, the administrative region associated with the geographical location of the target object is identified, the hierarchical structure of administrative divisions is mapped to a tree graph structure, and spatial adjacent edges are added to the spatially adjacent division nodes at the same level to construct an extended graph; a graph convolutional network is used to aggregate the features of spatially adjacent nodes for each node in the extended graph to obtain the spatially aggregated node features of each node; the features of child nodes are aggregated from bottom to top for each node; the spatially aggregated node features of all child nodes of each node are weighted and aggregated to obtain the intermediate features of the child node aggregation; the intermediate features of the child node aggregation are residually fused with the spatially aggregated node features of the parent node to obtain the spatial feature vector of each node; based on the division chain corresponding to the target object, a set of hierarchical spatial feature vectors is obtained.

[0013] Furthermore, the formula for calculating the regional correction coefficient is as follows:

[0014] ;

[0015] in, This is the regional correction factor. It is the hyperbolic tangent function. For the weight vector, This is the transpose of the vector. For bias terms, The region feature vector is obtained by concatenating the spatial feature vectors of all levels in hierarchical order.

[0016] Furthermore, the formula for calculating the conduction score is as follows:

[0017] ;

[0018] in, For the target node Transmission score at the location, The source node in the path. To reach the target node along the supply relation edge The set of all source nodes For path, For the source node To the target node The set of all feasible paths along the supply relation edges. For the source node The rating, For the source node To the target node Along the path Path conduction strength, The calculation formula is:

[0019]

[0020] in, For nodes and nodes exist Attention weights on relation edges This is the distance attenuation coefficient. Let be the path length, and exp be an exponential function with the natural constant e as its base.

[0021] Furthermore, the different modal data include time series data, graph structure data, and semi-structured text data. For the time series data, a sliding window method is used to extract statistical features to obtain temporal features. For the graph structure data, a semi-supervised algorithm is used to obtain the embedding vector of each node based on the adjacency matrix and edge weight matrix of the data to obtain graph embedding features. For the semi-structured text data, a pre-trained language model is used for semantic encoding to obtain text semantic features. The temporal features, graph embedding features, and text semantic features are projected onto a unified semantic space through three sets of projection networks, respectively. The average representation of the multimodal data after projection is calculated, and the average representation is retained or attenuated based on the consistency score between modalities to obtain a comprehensive semantic representation.

[0022] Furthermore, the formula for calculating the competition relationship score is as follows:

[0023] ;

[0024] in, For nodes Competition rating For nodes The set of neighbors under the compete relationship, where sim is the cosine similarity function. For nodes Node characteristics, For nodes The node features; the competition relationship score is calculated based on the cosine similarity of the graph attention network features of the competing neighbors.

[0025] Furthermore, the acquisition of time series features from time series data in multi-source heterogeneous data includes extracting time series features using LSTM and Transformer dual-channel time encoders respectively, fusing the time series features extracted by LSTM and Transformer dual channels through a gating fusion mechanism, and outputting a growth score after passing the fused time series features through a fully connected network.

[0026] Furthermore, the generation of decision parameters using growth score, risk transmission score, competition relationship score, and regional correction coefficient as conditional vectors specifically includes: using growth score, risk transmission score, competition relationship score, and regional correction coefficient as conditional inputs, generating a decision parameter distribution using a conditional variational autoencoder, searching for the optimal decision parameter combination through multi-objective Pareto optimization, and selecting the optimal decision parameter combination based on a knee detection method.

[0027] Furthermore, the selected optimal decision parameters are validated according to rules, and the final decision parameters are adjusted based on preset rules to obtain the adjusted decision parameters. The optimal decision parameter combination obtained by multi-objective Pareto optimization search is optimized and selected based on a multi-armed slot machine to obtain the optimized decision parameters. The adjusted decision parameters and the optimized decision parameters are weighted and fused to obtain the final decision parameters.

[0028] The decision parameter generation system based on multi-source heterogeneous data described in this invention includes:

[0029] The data acquisition module is used to acquire heterogeneous data from multiple sources.

[0030] The spatial feature extraction module is used to construct a graph structure based on regional information in multi-source heterogeneous data, extract features from the graph structure, generate spatial feature vectors, and calculate regional correction coefficients based on the spatial feature vectors.

[0031] The transmission feature extraction module is used to extract entities and relationships between entities from multi-source heterogeneous data, construct a directed knowledge graph based on the entities and their relationships, learn the node features of the directed knowledge graph using a graph attention network, determine the transmission path for entities with transmission relationships, and determine the transmission path strength based on attention weights and path lengths, and calculate the transmission score based on the transmission path strength; the node features include the static attributes of entity nodes, comprehensive semantic features, and positional encoding; the comprehensive semantic features are obtained by feature fusion of different modalities related to entities in multi-source heterogeneous data;

[0032] The competition feature extraction module is used to extract node features from a directed knowledge graph learned by a graph attention network and to calculate a competition relationship score based on the feature similarity between adjacent nodes under a competition relationship.

[0033] The growth feature extraction module is used to obtain time series features from time series data in multi-source heterogeneous data and calculate growth scores based on the time series features;

[0034] The parameter generation module is used to generate decision parameters using growth score, risk transmission score, competitive relationship score, and regional correction coefficient as condition vectors.

[0035] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0036] (1) Based on the graph attention network, attention weights are learned between nodes with transmission relationship, and the transmission path strength is determined by superimposing the path length along the transmission relationship path. The transmission score is obtained according to the transmission path strength, so that the score result can locate the specific transmission path (source node, intermediate node sequence, path length). This realizes the dynamic transmission modeling with traceability of each path and intensity with path length, rather than the weighted summation of static feature values. It represents the transmission path and intensity decay law of risk from upstream to downstream nodes, and provides downstream decision-making with traceable and interpretable quantitative basis for risk transmission.

[0037] (2) By using graph convolutional networks to aggregate features of spatially adjacent nodes in a tree structure of spatial hierarchical structures (such as administrative division levels), the feature representation of each region includes the correlation information of adjacent regions. Compared with the existing technology of independent encoding of each region, the two-layer aggregation of GCN increases the spatial autocorrelation coefficient of regional features from 0.32 to 0.78, effectively eliminating the problem of information silos between regions and substantially enhancing the representation ability of spatial features. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the decision parameter generation method of the present invention.

[0039] Figure 2This is a flowchart of the process for obtaining the comprehensive semantic representation in this invention. Detailed Implementation

[0040] Example 1

[0041] like Figure 1 As shown in this embodiment, a method for generating decision parameters based on multi-source heterogeneous data includes the following steps:

[0042] S1: Multi-granularity spatial feature encoding, constructing a graph structure based on regional information in multi-source heterogeneous data, extracting features from the graph structure, generating spatial feature vectors, and calculating regional correction coefficients based on the spatial feature vectors;

[0043] The process involves identifying the administrative regions associated with the geographical location of the target entity (the enterprise making the decision), mapping the hierarchical structure of administrative divisions to a tree graph structure, encoding the features of each division node, aggregating the features of spatially adjacent regions through a graph convolutional network, and finally generating multi-granularity spatial feature vectors through cascaded downsampling. The specific steps are as follows:

[0044] S1.1 Construction of Administrative Division Tree Structure and Node Initialization

[0045] Determine the five-level administrative divisions associated with the geographical location of the target object, and map these five levels of administrative divisions (province, city, district / county, street / township, industrial park / community) as a directed tree structure T=(V,E), where V is the set of nodes and E is the set of parent-child edges. Each node... Includes initial feature vectors related to policy parameters. Based on the national administrative division coding database, the codes, hierarchical relationships, and geographic coordinates of each administrative division are obtained. Regional indicators for each administrative division are obtained from multi-source heterogeneous data, and a tree structure T is constructed with nodes... initial feature vector In this embodiment, d0=14, corresponding to 14 categories of regional indicators (total GDP, per capita disposable income, industrial structure concentration, fixed asset investment growth rate, fiscal self-sufficiency rate, urbanization rate, logistics accessibility index, financial activity, technological innovation index, labor cost index, land cost index, environmental carrying capacity index, policy incentive index, and infrastructure completeness index). Depending on the specific decision to be made, the corresponding regional indicator can be selected as the initial feature vector for each node. The tree structure T has a tree depth L=5, and the maximum fan-out number of leaf nodes B=30.

[0046] S1.2 Normalization of initial eigenvectors

[0047] For nodes The 14-dimensional initial feature vector is Z-score standardized to eliminate dimensional differences, resulting in a standardized feature vector. The standardized formula is:

[0048]

[0049] in, For nodes Initial values ​​on the k-th dimension feature, and Let be the mean and standard deviation of the k-th feature across all nodes at the same level, respectively. To prevent division by zero, a smoothing constant is used. A statistic is employed. and The calculation is performed by hierarchical grouping and is updated periodically in practical applications. In this embodiment, the update cycle is once per quarter.

[0050] S1.3 Neighboring Region Feature Aggregation Based on Graph Convolutional Networks

[0051] Based on the tree structure T constructed according to administrative divisions, an extended graph is constructed by adding spatially adjacent edges between division nodes that are at the same level and are spatially adjacent. ,in , This is the set of spatial adjacency edges. The rule for constructing spatial adjacency edges is: for any two administrative division nodes at the same level (i.e., without administrative affiliation)... and When the administrative geographical boundaries of the two regions have a common boundary segment, and the length of the boundary segment is greater than a preset threshold. =5 kilometers, indicating the administrative division node and Spatially adjacent, at zoning nodes and Spatially adjacent edges are established between nodes. The graph structure in this embodiment is based on an administrative division hierarchical tree. Supplementary edges are established based on the common boundary segments of the geographical boundaries of administrative divisions, used to support the graph convolutional network in aggregating features of adjacent regions along the spatial topology. A two-layer graph convolutional network (GCN) is used to aggregate the feature information of its spatially adjacent nodes for each node. The propagation formula of the graph convolutional network (GCN) is:

[0052]

[0053] in, Let A be the adjacency matrix with self-loops, and let A be the adjacency matrix of the extended graph G'. It is an N-order identity matrix. for degree matrix and , , This is the normalized node feature matrix of the tree structure. For the first The layer is a learnable weight matrix, where σ is the ReLU activation function.

[0054] The first layer weight matrix of a two-layer graph convolutional network (GCN) is: The input to the first layer is the adjacency matrix à of the extended graph G', and the node feature matrix after tree structure normalization. ,Right now The output of the first layer is Simultaneously, as the input to the second layer, the weight matrix of the second layer is... The output of the second layer is Node feature matrix after GCN aggregation The features of each node in the graph convolutional network have been fused with information from its spatial neighbors. In this embodiment, the parameter settings of the graph convolutional network include the number of GCN layers K=2, and the hidden layer dimension... =32, output dimension =64, learning rate η=0.001, training epochs =100, Dropout ratio =0.3.

[0055] S1.4 Multi-granularity cascaded downsampling

[0056] For the node features of each node in the tree structure aggregated by GCN, aggregation is performed layer by layer from the finest granularity level (park / community level) to the coarsest granularity level (provincial level). At each layer, an attention mechanism is used to allocate aggregation weights from child nodes to parent nodes. For the ... parent node of the layer First based on the first The set of child nodes of a layer Calculate intermediate features of child node aggregation The formula is as follows:

[0057]

[0058] in, child node For the parent node Normalized attention weights For the first Layer child nodes Feature vectors aggregated by GCN, attention weights The calculation is as follows:

[0059]

[0060] in, , For the first Layer-learnable query and key projection matrices, Let be the attention vector, and || denote the vector concatenation operation. The negative slope parameter of LeakyReLU is 0.2. parent node Node features aggregated by GCN, superscript This represents the parent node's own initial GCN characteristics, not its administrative level number. For the first Layer child nodes The feature vector after GCN aggregation.

[0061] Aggregate the intermediate values ​​of the calculated child nodes. With parent node The node features aggregated by GCN are subjected to residual fusion to obtain the spatial feature vectors of each node, ultimately resulting in a set of spatial feature column vectors for each level of the partition chain corresponding to the target object. These correspond to the provincial, municipal, district / county, street / township, and park / community levels, respectively.

[0062] S1.5 Multi-granularity spatial feature stitching and region correction coefficient calculation

[0063] The spatial feature vectors of each level (node) in the five-level hierarchical spatial feature vector set corresponding to the target object (province, city, district / county, street / township, and park / community) are concatenated. Specifically, the spatial feature vectors of each level are concatenated vertically along the feature dimension in hierarchical order to obtain a complete regional feature vector, based on which the regional correction coefficient is calculated. Regional Feature Vector for:

[0064]

[0065] in, For the spatial feature vector of the province, For city-level spatial feature vectors, For the spatial feature vector at the district / county level, This represents the spatial feature vector at the street / township level. This is a spatial feature vector at the park / community level.

[0066] Calculation of the regional correction factor γ:

[0067]

[0068] in The weight vector is a 320-dimensional learnable column vector. This is its transposed 1×320 row vector. As a scalar bias term, the tanh function constrains the correction coefficient within the interval (0, 2).

[0069] Furthermore, in this embodiment, the spatial feature vectors of each level are 64-dimensional column vectors, and the concatenated region feature vectors are 320-dimensional column vectors. The weight column vector is a learnable weight column vector. The transpose of this weight column vector is a row vector. The product of the weights of the row vectors and the region feature vectors of the column vectors is a scalar. This scalar is activated by tanh and incremented by 1 to obtain the one-dimensional region correction coefficient γ. During the training phase... and Other learnable parameters of the network are generated along with the parameters and updated via backpropagation.

[0070] The regional correction coefficient γ is used to quantitatively characterize the direction and degree of deviation of the overall environment of the target area from the regional benchmark. γ>1 indicates that the environmental conditions in the area are better than the benchmark level, while γ<1 indicates that the environmental conditions in the area are worse than the benchmark level. The regional correction coefficient γ is used as a component of the condition vector to determine the strategy parameters, making the generated strategy parameter distribution aware of regional environmental differences: the parameter schemes generated for areas with γ>1 can be matched with more aggressive configurations, while the parameter schemes generated for areas with γ<1 tend to be conservative configurations, thereby achieving adaptive correction of parameter generation to the regional environment.

[0071] Existing methods, when processing data with hierarchical structures (such as administrative divisions), typically treat each node as an independent entity for feature extraction, or only consider the features of its parent / child nodes. When dealing with multi-source heterogeneous data, nodes at the same level may have spatial relationships. Existing tree-structured feature processing ignores the feature associations and information transmission between spatially adjacent regions, resulting in an island effect in regional feature representation. The parameter generation model cannot perceive the influence between regions. For example, in this embodiment, the same rating company ultimately obtains the exact same parameter output in different regions, which is inconsistent with the actual business scenario.

[0072] S2: Heterogeneous Modal Preprocessing Pipeline

[0073] like Figure 2 As shown, the required multi-source heterogeneous data is preprocessed according to modality type. Through contrastive learning, features from different modalities are mapped to a unified semantic space, and multi-modal consistency is verified. The data from different modalities include time-series data, graph-structured data, and semi-structured text. In this embodiment, the acquired multi-source heterogeneous data includes the credit relationship network graph structure of the decision-making entity, business registration information text, annual report text, and operational data (equipment usage data, cost data, etc.).

[0074] Statistical features are extracted from time series data (such as equipment operating time series, electricity consumption series, and API call volume series in this embodiment).

[0075] For time series The sliding window method is used to extract statistical features. The 6-dimensional statistical feature vector corresponding to each time window is as follows: The statistical feature vector of the time window t is:

[0076]

[0077] in: This represents the mean of the time series within the window. The standard deviation of the time series in the window The skewness of the time series within the window. This represents the excess kurtosis of the time series within the window. The increment of the time series within the window. The slope of the linear regression of the time series within the window. This refers to the window size.

[0078] For graph-structured data such as credit relationship graphs and mortgage network graphs, the Node2Vec algorithm is used to generate node embedding vectors to obtain graph embedding features.

[0079] For semi-structured text data such as business information texts and SaaS activity description texts, semantic encoding is performed using the pre-trained language model BERT-base-chinese to obtain text semantic features.

[0080] Three projection networks (MLPs) are constructed to project temporal features, graph embedding features, and textual semantic features onto a unified semantic space. The projection networks are trained using a contrastive learning loss function (InfoNCE Loss) to bring different modal features from the same entity closer together in the unified space.

[0081] Consistency verification is performed on the multimodal representations of the same entity in a unified semantic space. When the semantic representations of multiple modalities point in the same direction in the unified space, they are judged as strong signals; when the directions are inconsistent, the overall signal strength of the entity is reduced.

[0082] Intermodal consistency score The calculation formula is:

[0083]

[0084] Where M is the number of available modes ( ), For similarity function, for Features after modal feature projection for Features projected from modal features. When When, it is determined to be a strong consistent signal; when At this time, the overall signal strength is attenuated, and the overall semantic representation after consistency verification is as follows:

[0085]

[0086] in, For multimodal average characterization, This is the consistency threshold.

[0087] By projecting different modal features into the same semantic space through contrastive learning, a directional consistency judgment is performed on the multimodal representation of the same entity. When the consistency score reaches the threshold, the comprehensive semantic representation of the entity is retained; when it does not reach the threshold, the comprehensive representation of the entity is attenuated as a whole. The output is a binary judgment result of strong signal / weak signal at the entity level, which suppresses the interference of intermodal semantic conflict on downstream transmission modeling and growth assessment.

[0088] S3: Supply Chain Transmission Modeling Based on Graph Attention Networks

[0089] Entities and relationships between entities are extracted from multi-source heterogeneous data. A directed knowledge graph is constructed based on the entities and their relationships. In this embodiment, an industrial chain knowledge graph is automatically constructed from enterprise business registration data. Graph attention networks are used to learn node features, and risk transmission path reasoning is performed based on attention weights.

[0090] S3.1 Construction of Industry Chain Knowledge Graph

[0091] This paper extracts enterprise entities and their supply, competition, and cooperation relationships from enterprise registration data (including enterprise name, industry code, business scope text, shareholder relationships, and external investment information). An attributed directed knowledge graph KG=(E,R,A) is constructed, where E is the set of enterprise entities, R is the set of relationship types (supply, compete, cooperate), and A is the set of entity attributes. The knowledge graph KG includes the number of entities |E|, the set of relation triples {(h,r,t) | h,t ∈ E, r ∈R}, and the entity attribute matrix. Here, h is the head entity, representing the starting enterprise node of the relationship; t is the tail entity, representing the ending enterprise node of the relationship; and r is the relationship type, ranging from {supply, compete, cooperate}. The triple (h,r,t) indicates that the head entity h points to the tail entity t through the relation r. For example, (A, supply, B) indicates that enterprise A supplies products to enterprise B.

[0092] This process extracts enterprise entities and their relationships from business registration data. Specifically, it involves: using named entity recognition (NAME) models to identify enterprise entities and categorize them into their respective industry chain segments (upstream raw materials, midstream manufacturing, downstream applications, and services) based on industry codes and business scope text in the registration data; extracting supply relationships through equity investment information (investment direction is supply-side); extracting compete relationships through industry code overlap and business scope text similarity (cosine similarity ≥ 0.8); and extracting cooperate relationships through common shareholders or joint bidding records. Other entities and entity relationships can be extracted from other data as needed.

[0093] S3.2 Initialization of Industry Chain Node Characteristics

[0094] An initial feature vector is constructed for each enterprise node in the knowledge graph, containing the enterprise's static attributes and a comprehensive semantic representation output from feature fusion in step S2. The initial feature vector is represented as follows:

[0095]

[0096] in, The static attribute vector of the enterprise (including the logarithm of registered capital, years of establishment, logarithm of employee size, and logarithm of number of patents, totaling 4 dimensions). For comprehensive semantic representation, It is a location encoding vector based on the links in the industrial chain (4-dimensional one-hot encoding).

[0097] S3.3 Node Feature Learning Based on Graph Attention Network

[0098] A two-layer graph attention network (GAT) is applied to the knowledge graph. Each node aggregates information from its upstream and downstream neighbors through an attention mechanism. Different attention parameters are used for different types of relation edges, resulting in a node feature matrix learned by GAT. .

[0099] S3.4 Risk Transmission Path Reasoning

[0100] Using the attention weight matrix learned by GAT, the transmission path of risk from upstream nodes to downstream nodes is identified. In this embodiment, since changes in upstream nodes in the supply relationship can affect downstream nodes, this embodiment takes the supply relationship path as an example. For a given source node s, a depth-first search is performed to traverse along the supply relationship edges, and the attention weight decay on the path is accumulated.

[0101] The risk propagation intensity along path P from source node s to target node t is:

[0102]

[0103] in For nodes and nodes The attention weights learned by GAT on the supply relationship edges (normalized to the [0,1] interval), |P| is the path length (number of hops), and the path length is the number of edges traversed in the path. This is the distance attenuation coefficient, which is a preset parameter.

[0104] Comprehensive risk transmission score for target node t:

[0105]

[0106] in The risk score for the source node s itself (derived from the features of the data in step S2).

[0107] The aforementioned risk transmission score is based on the attention weights learned end-to-end in the graph attention network. The attention weights of each edge along the supply relationship edge are multiplied together, and an exponential decay factor of the path length and a maximum search depth control are superimposed. This allows the score to pinpoint the specific transmission path (source node, intermediate node sequence, path length), achieving dynamic transmission modeling that is traceable along each path and whose intensity decays explicitly with the path length and edge weights.

[0108] S3.5 Competitive Situation Score Calculation

[0109] Based on the node features output by GAT, the competitive position score of the target object in its respective industry chain segment is calculated, and the competitive intensity signal is extracted from the GAT feature space. Based on the node features aggregated through two layers of GAT, the node features are fused with the multimodal semantic representation of the target node (based on the comprehensive semantic representation output in step S2), static attributes, and structural information of upstream and downstream neighbors in the industry chain. The feature difference between the target node and its competing neighbor nodes is calculated, resulting in a measure of competitive difference that incorporates the industry chain structural context. This score serves as a conditional vector to constrain the positioning of the generated parameters within the market competition landscape. The formula for calculating the competitive position score is as follows:

[0110]

[0111] in For nodes In the neighbor set under the competing relationship, sim is the cosine similarity function. The closer the CTS value is to 1, the greater the difference between the target node and its competing neighbors in the GAT feature space (the lower the competition intensity); the closer it is to 0, the higher the feature similarity (the higher the competition intensity).

[0112] In this embodiment, the competitive situation score is calculated for neighboring nodes in a competitive relationship. For other non-transferable relationships (such as parallel relationships), the cosine similarity can also be used to calculate the difference between the two nodes to obtain a difference score, which is used to fully consider the difference between the target object and related entities when determining decision parameters.

[0113] S4: Enterprise Growth Assessment Based on Time-Sequence Encoder

[0114] By using a dual-channel LSTM and Transformer time encoder, the digital signal time series of enterprises is encoded to capture trend, periodic and abrupt change features. Then, through a gating fusion mechanism, the dynamic features of the time series are fused with the static snapshot features to generate enterprise growth assessment results.

[0115] S4.1 Time Series Data Preprocessing and Sequence Construction

[0116] For each type of digital signal (equipment operating time, electricity consumption, SaaS API call volume, logistics throughput), a monthly time series of nearly N months is constructed, and missing values ​​are imputed and standardized.

[0117] S4.2 LSTM Timing Coding

[0118] A bidirectional LSTM encoder is used to extract time series features for each time series, capturing long-term dependencies and trend information in the series.

[0119] For the standardized time series The forward and backward hidden states of the bidirectional LSTM are as follows:

[0120]

[0121]

[0122] The concatenation of the forward and backward hidden states from the last time step is used as the LSTM encoded output:

[0123]

[0124] S4.3 Transformer Timing Coding

[0125] For the standardized time series The Transformer encoder is used to encode the time series in parallel, and the self-attention mechanism is used to capture the periodicity and abrupt change points in the sequence.

[0126] First, add positional encoding to the time series:

[0127]

[0128]

[0129] in, For location index, It is an embedded dimension index. The size of the embedded dimension.

[0130] The input embedding is:

[0131]

[0132] After passing through the H-layer Transformer encoder, average pooling is performed on the output of all time steps:

[0133]

[0134] S4.4 Dual-channel feature fusion and growth potential scoring

[0135] For each signal type, LSTM and Transformer encoded features are fused using a gated fusion unit (GRU-like structure). Then, the fused features of each signal type are concatenated with static snapshot features (based on the comprehensive semantic representation output from step S2 and the risk transmission score and competitive situation score output from step S3), and then processed through a fully connected network. Output a growth potential score.

[0136] Gated fusion unit:

[0137]

[0138]

[0139] in, As a feature of fusion, Let be the gated vector, σ be the Sigmoid function, and ⊙ be the element-wise multiplication.

[0140] Growth rating:

[0141]

[0142] The trend characteristics, periodic characteristics, and mutation characteristics of time series parameters are used as growth scores to reflect the data, so as to fully consider the time series characteristics of the target object when determining decision parameters.

[0143] S5: Joint Generation of Multi-Objective Parameters Based on Conditional Variational Autoencoder

[0144] The obtained risk assessment results, growth score, competitive situation score and regional correction coefficient are used as the conditional vector of CVAE. The parameter distribution is generated by the conditional variational autoencoder (CVAE), and the optimal parameter combination is searched through multi-objective Pareto optimization.

[0145]

[0146] CreditLevel is the credit rating code (the 20-level credit rating is mapped to a 16-dimensional vector through the Embedding layer).

[0147] The CVAE encoder (inference network) uses a conditional vector c and an observation parameter vector. Using historical observations of financing decision parameters as input, the posterior distribution of latent variables is inferred. The input consists of a conditional vector and a historical observation parameter vector. These vectors are concatenated, and the mean of the posterior distribution is output via a CVAE encoder. and logarithmic variance The posterior distribution is obtained as follows: .

[0148] The CVAE decoder (generative network) takes a conditional vector *c* and *z* sampled from the latent variable space of the posterior distribution as input to generate a conditional distribution of the parameter vector. The input conditional vector and the sampled latent variables... Condition variables and hidden variables The concatenation is performed, and the mean of the conditional distribution is output through a CVAE decoder. and logarithmic variance Sample parameter vectors from conditional distribution:

[0149]

[0150] The training objective of CVAE is to maximize the lower bound of evidence (ELBO) while balancing reconstruction accuracy and latent variable regularization.

[0151] Using the CVAE decoder as a generator, a set of candidate parameters is generated. At the same time, three objective functions are optimized. The NSGA-II algorithm is used to search for the Pareto optimal solution set. The Pareto optimal solution set P* contains K non-dominated parameter combinations, K ∈ [5, 20].

[0152] The first component of the parameter vector For example, the three optimization objectives are:

[0153] ;

[0154] ;

[0155] ;

[0156] ;

[0157] in: For risk loss function, For the comprehensive scoring of risk transmission, The first component of the parameter vector. The risk sensitivity coefficient is set to 0.5 in this embodiment; To achieve a probabilistic prediction model (a pre-trained MLP classifier). For the weights of the prediction model, This is the bias term for the prediction model; For competitiveness scoring functions, For reference benchmark value, Competitive landscape assessment.

[0158] In the formulas for the three objective functions mentioned above, the first component of the parameter vector is used. For example, in the actual optimization process, the parameter vector =( , , , , All components of the risk loss function participate in parallel optimization as decision variables. RiskLoss, DealProb, and MarketComp have an overall effect on the parameter vector during actual computation, which can be equivalently expressed as a joint optimization of each component. When expressed in L1 norm form, the risk loss function can be written as... .

[0159] The final recommended parameters are selected from the Pareto optimal solution set using a knee detection-based method.

[0160] For the K solutions on the Pareto front, calculate the curvature of each solution:

[0161] For three adjacent solutions on the Pareto front , , (i=2,...,K-1), each solution =(f1( ),f2( Let f1 and f2 be the coordinates of the objective function value, respectively, and f1 and f2 be combinations of RiskLoss and MarketComp or RiskLoss and DealProb in multi-objective optimization. The curvature of the i-th solution is... Based on the radius of the circumcircle of three adjacent points Approximate calculation of the reciprocal: = 1 / = 4· / (| |·| |·| |), where It is a triangle The area is calculated by the cross product, where |·| represents the Euclidean distance.

[0162] Select the solution with the largest curvature (knee point) as the final output parameter vector x*. Output the final parameter vector. The business meaning of each component is defined in specific practical applications.

[0163] By considering spatial characteristics, transmission characteristics, competition characteristics, and temporal characteristics, multi-source heterogeneous data is extracted to achieve comprehensive information extraction from the data and improve the accuracy of decision-making results. In this embodiment, the generated decision parameters can be technical parameters such as risk monitoring threshold parameters, scoring model weight parameters, and target asset operation monitoring threshold parameters required by the financial leasing risk control system, which can be used as input for downstream decision-making systems.

[0164] S6: Rules Engine and Conflict Resolution

[0165] The final output parameter vector is validated by rules. Business constraint rules are pre-encoded into a decision table. Rule matching is accelerated by hash index. Conflict detection and resolution are performed on multiple rules that are triggered at the same time (such as conflicting operations of reducing down payment ratio and increasing down payment ratio among the rules that are triggered at the same time). The marginal impact of each rule on the final output is quantified by SHAP value.

[0166] S6.1 Rule Decision Table Coding

[0167] Business rules are encoded into structured decision tables, with each rule represented as a condition-action pair.

[0168] Rule set definition (three core rules):

[0169] Lower bound protection constraint for decision parameters: ,if Then the decision parameters Pick , Decision parameters The preset lower threshold;

[0170] Parameter upper limit constraints in low-growth scenarios: for decision parameters If growth score ,and Then the decision parameters Pick ;in, To determine the threshold for growth potential scoring, Decision parameters The preset upper limit threshold;

[0171] Lower bound constraints on parameters in high-risk scenarios: for decision parameters If the risk transmission comprehensive score ,and Then the decision parameters Pick , The threshold for the comprehensive risk transmission assessment can be set at 0.7. Decision parameters The preset lower threshold.

[0172] S6.2 Accelerated Rule Matching Based on Hash Index

[0173] A hash index is constructed from the attribute fields in the rule conditions to achieve rule matching lookup with O(1) time complexity. For the rules... The set of conditional attributes is used to construct a hash map. Based on each attribute value of the input parameter vector, the subset of rules involving that attribute is directly located through hash lookup. Only the subset is used for conditional judgment to obtain the set of triggering rules F.

[0174] S6.3 Rule Conflict Detection and Resolution

[0175] When the set of triggering rules |F|>1, detect whether there is a conflict between the rules (applying contradictory adjustment directions to the same parameter), and sort them based on the priority matrix and constraint satisfaction.

[0176] Define rules and Conflict degree between rules: when rules With rules When the same parameter is modified in opposite directions, the conflict degree is 1; otherwise, the conflict degree is 0.

[0177] When resolving rule conflicts, the resolution ranking score of the rules is calculated:

[0178]

[0179] in, For rules The resolution sorting score, For regulatory compliance indicator functions (regulatory mandatory rules) Custom business rules ), For regulatory compliance, a weight of 0.5 is assigned. The severity of rule triggering (the distance of the parameter from the constraint boundary). The severity weight for rule triggering is set to 0.3. The constraint coverage of the rule (involving the number of parameter dimensions / total number of parameter dimensions). To constrain the coverage weight, we set it to 0.2.

[0180] The conflicting rules are sorted in descending order based on the calculated resolution ranking score, executed sequentially, and rules that conflict with already executed rules are skipped.

[0181] S6.4 SHAP Quantitative Evaluation of Rule Effectiveness

[0182] The marginal contribution of each triggering rule to the final parameter output is quantified using SHAP (SHapley Additive exPlanations) values.

[0183] rule The SHAP value for parameter dimension j is:

[0184]

[0185] in, It is the set of triggering rules, and S is the set of rules that do not contain rules. a subset of This indicates the output value of parameter dimension j after executing only the rules in the subset S of rules.

[0186] rule Overall impact:

[0187]

[0188] in, The parameters are defined as the total dimension, and the final output is a compliance parameter and rule impact report.

[0189] The conflict resolution algorithm based on the priority matrix ensures consistent execution when multiple rules are triggered simultaneously, and the SHAP value quantification provides a traceable numerical basis for the impact of each rule.

[0190] S7: Online parameter optimization based on multi-armed slot machines

[0191] The obtained decision parameters are further optimized by modeling the parameter adjustment as a multi-armed slot machine problem. The upper confidence bound (UCB) strategy is adopted to achieve a balance between utilizing known optimal parameters and exploring new parameters. The parameter selection strategy is continuously optimized through Bayesian posterior update.

[0192] Modeling the Multi-Armed Slot Machine Problem (S7.1)

[0193] The K parameter combinations in the Pareto optimal solution set generated in step S5 are modeled as K arms. Each decision selects one arm to execute, and a reward signal is obtained based on the positive / negative feedback signals from the external environment.

[0194] At time step t, the system selects the arm. ∈ {1, 2, ..., K}, receive a reward. :

[0195]

[0196] in, Based on positive feedback rewards, For the arm The corresponding gain coefficient, The negative feedback penalty coefficient is set to 0.1.

[0197] S7.2 Exploration-Exploitation Strategy Based on UCB

[0198] The UCB1 algorithm is used to select the arm at the current time step, balancing the use of the arm with the best historical performance with the arm that has fewer exploration attempts.

[0199]

[0200] in The historical average reward for arm A. Let be the number of times arm a is selected, and t be the current total number of time steps. The first term in the UCB formula... The first item is the utilization item, and the second item is the exploration item.

[0201] S7.3 Bayesian Post-hoc Update

[0202] A Bayesian model is built for the reward distribution of each arm, and the posterior distribution parameters are updated using the reward observed at each time step. Specifically, this is based on the reward observation at the current time step. The chosen arm We obtain the updated mean of the posterior distribution parameters and the variance of the observation noise.

[0203] The cumulative regret of the UCB strategy is theoretically bounded to ensure that the online optimization process converges to the near-optimal arm within a finite number of time steps. After convergence, the optimized parameters are output.

[0204] S7.4 Final Parameter Output and Feedback Closed Loop

[0205] The parameters constrained by the rules in step S6 are weighted and fused with the parameters optimized in step S7 to output the final decision parameters, and a feedback closed-loop mechanism is established.

[0206] The final decision parameter calculation formula is as follows:

[0207]

[0208] in These are the parameters constrained by the rules in step S6. The parameters selected for UCB optimization Rule weighting coefficient (regulatory mandatory rules) =1, meaning full compliance with the rules; business guidance rules =0.6).

[0209] Feedback loop: Each time the external environment feedback result is asynchronously sent back to the system through the message queue, the reward observation and Bayesian posterior distribution of the corresponding arm in S7 are updated, and the incremental update of the CVAE model in S5 is triggered (a gradient update is performed after every 100 new samples are accumulated).

[0210] The parameter tuning is modeled as a multi-armed slot machine problem and a UCB1 strategy is adopted to ensure that the accumulated regret converges within the theoretical bound. Compared with the existing method of periodically tuning parameters based on manual experience, the online Bayesian update mechanism enables the system to automatically converge to the near-optimal parameter combination within multiple feedback observations. The parameter tuning response time is shortened from 30 days for manual operation to real-time updates, and the positive feedback rate of the converged parameters during continuous operation is increased by more than 15% compared with the initial parameters.

Claims

1. A method for generating decision parameters based on multi-source heterogeneous data, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data; A graph structure is constructed based on the regional information in multi-source heterogeneous data. Features are extracted from the graph structure to generate spatial feature vectors. Regional correction coefficients are then calculated based on the spatial feature vectors. Entities and relationships between entities are extracted from multi-source heterogeneous data. A directed knowledge graph is constructed based on the entities and their relationships. Graph attention networks are used to learn the node features of the directed knowledge graph. For entities with transmission relationships, transmission paths are determined, and the transmission path strength is determined based on attention weights and path lengths. Transmission scores are calculated based on the transmission path strength. The node features include the static attributes, comprehensive semantic features, and positional encoding of the entity nodes. The comprehensive semantic features are obtained by fusing features from different modalities of entity-related data in multi-source heterogeneous data; Based on the node features of the directed knowledge graph learned by the graph attention network, the competition relationship score is calculated according to the feature similarity of adjacent nodes under the competition relationship. Temporal features are obtained from time series data in multi-source heterogeneous data, and growth scores are calculated based on these temporal features; Decision parameters are generated using growth potential score, risk transmission score, competitive relationship score, and regional correction coefficient as condition vectors.

2. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, The specific steps for generating spatial feature vectors are as follows: Based on the regional information in the multi-source heterogeneous data, identify the administrative regions associated with the geographical location of the target object, map the hierarchical structure of administrative divisions into a tree graph structure, and supplement spatially adjacent edges for the division nodes at the same level and spatially adjacent to each other to construct an extended graph; use a graph convolutional network to aggregate the features of spatially adjacent nodes for each node in the extended graph to obtain the spatially aggregated node features of each node; aggregate the features of child nodes for each node from bottom to top; perform weighted aggregation of the spatially aggregated node features of all child nodes of each node to obtain the intermediate features of the child node aggregation; perform residual fusion between the intermediate features of the child node aggregation and the spatially aggregated node features of the parent node to obtain the spatial feature vectors of each node; and obtain a set of hierarchical spatial feature vectors based on the division chain corresponding to the target object.

3. The method for generating decision parameters based on multi-source heterogeneous data according to claim 2, characterized in that, The formula for calculating the regional correction coefficient is as follows: ； in, This is the regional correction factor. It is the hyperbolic tangent function. For the weight vector, This is the transpose of the vector. For bias terms, The region feature vector is obtained by concatenating the spatial feature vectors of all levels in hierarchical order.

4. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, The formula for calculating the conduction score is: ； in, For the target node Transmission score at the location, The source node in the path. To reach the target node along the supply relation edge The set of all source nodes For path, For the source node To the target node The set of all feasible paths along the supply relation edges. For the source node The rating, For the source node To the target node Along the path Path conduction strength, The calculation formula is: ； in, For nodes and nodes exist Attention weights on relation edges This is the distance attenuation coefficient. Let be the path length, and exp be an exponential function with the natural constant e as its base.

5. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, Different modal data include time series data, graph structure data, and semi-structured text data. For the time series data, the sliding window method is used to extract statistical features to obtain time series features. For the graph structure data, a semi-supervised algorithm is used to obtain the embedding vector of each node based on the adjacency matrix and edge weight matrix of the data, thus obtaining graph embedding features. For the semi-structured text data, a pre-trained language model is used for semantic encoding to obtain text semantic features. The temporal features, graph embedding features, and text semantic features are projected onto a unified semantic space through three sets of projection networks, respectively. The average representation of the multimodal features after projection is calculated. The average representation is retained or attenuated based on the consistency score between modalities to obtain a comprehensive semantic representation.

6. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, The formula for calculating the competition relationship score is as follows: ； in, For nodes Competition rating For nodes The set of neighbors under the compete relationship, where sim is the cosine similarity function. For nodes Node characteristics, For nodes The node characteristics.

7. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, The process of obtaining time series features from time series data in multi-source heterogeneous data includes extracting time series features using LSTM and Transformer dual-channel time encoders respectively, fusing the time series features extracted by LSTM and Transformer dual channels through a gating fusion mechanism, and outputting a growth score after passing the fused time series features through a fully connected network.

8. The method for generating decision parameters based on multi-source heterogeneous data according to claim 1, characterized in that, The specific steps for generating decision parameters using growth score, risk transmission score, competition relationship score, and regional correction coefficient as conditional vectors include: using growth score, risk transmission score, competition relationship score, and regional correction coefficient as conditional inputs, generating a decision parameter distribution using a conditional variational autoencoder, searching for the optimal decision parameter combination through multi-objective Pareto optimization, and selecting the optimal decision parameter combination based on a knee detection method.

9. The method for generating decision parameters based on multi-source heterogeneous data according to claim 8, characterized in that, The selected optimal decision parameters are validated according to rules, and the final decision parameters are adjusted based on preset rules to obtain the adjusted decision parameters. The optimal decision parameter combination obtained by multi-objective Pareto optimization search is optimized and selected based on multi-armed slot machine to obtain the optimized decision parameters. The adjusted decision parameters and the optimized decision parameters are weighted and fused to obtain the final decision parameters.

10. A decision parameter generation system based on multi-source heterogeneous data, characterized in that, include: The data acquisition module is used to acquire heterogeneous data from multiple sources. The spatial feature extraction module is used to construct a graph structure based on regional information in multi-source heterogeneous data, extract features from the graph structure, generate spatial feature vectors, and calculate regional correction coefficients based on the spatial feature vectors. The transmission feature extraction module is used to extract entities and relationships between entities from multi-source heterogeneous data, construct a directed knowledge graph based on the entities and their relationships, learn the node features of the directed knowledge graph using a graph attention network, determine the transmission path for entities with transmission relationships, and determine the transmission path strength based on attention weights and path lengths, and calculate the transmission score based on the transmission path strength; the node features include the static attributes of entity nodes, comprehensive semantic features, and positional encoding; The comprehensive semantic features are obtained by fusing features from different modalities of entity-related data in multi-source heterogeneous data; The competition feature extraction module is used to extract node features from a directed knowledge graph learned by a graph attention network and to calculate a competition relationship score based on the feature similarity between adjacent nodes under a competition relationship. The growth feature extraction module is used to obtain time series features from time series data in multi-source heterogeneous data and calculate growth scores based on the time series features; The parameter generation module is used to generate decision parameters using growth score, risk transmission score, competitive relationship score, and regional correction coefficient as condition vectors.

Images