Business entity credit default probability prediction method and system based on deep learning
By constructing a heterogeneous graph of commercial entities and performing multi-relationship mapping weighting, and combining deep learning technology to train the model, the problem of insufficient accuracy in predicting the probability of credit default of commercial entities was solved, and higher prediction accuracy was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 宁波市金融发展服务中心(宁波市金融信息监测中心)
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to effectively model the complex relationships between business entities, resulting in insufficient accuracy in predicting credit default probabilities.
By constructing a heterogeneous graph network of business entities and performing multi-relationship mapping weighting, and combining deep learning technology to train a basic default probability prediction model, the accuracy of credit default probability prediction is improved.
It improves the accuracy of predicting the probability of credit default for commercial entities and solves the problem of insufficient modeling in existing technologies.
Smart Images
Figure CN122155745A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method and system for predicting the probability of credit default of commercial entities based on deep learning. Background Technology
[0002] In the context of commercial entity credit risk assessment, commercial entities often form complex and dynamically changing relationships through various events such as business registration changes, investment connections, administrative penalties, and litigation. Different types of data exhibit significant differences in structure, semantics, and time dimensions. Traditional data processing methods typically focus on static feature modeling of single entity attributes, making it difficult to uniformly characterize the multiple types of relationships and their interactions implied by multi-source heterogeneous data. This leads to the weakening or loss of correlation information during the modeling process, affecting the comprehensive expression of the credit risk status of commercial entities and reducing the accuracy and reliability of credit default probability prediction results. Summary of the Invention
[0003] This application provides a method and system for predicting the probability of credit default of commercial entities based on deep learning, which is used to address the technical problem that existing technologies are unable to effectively model the complex relationships between commercial entities, resulting in insufficient accuracy in predicting the probability of credit default.
[0004] In view of the above problems, this application provides a method and system for predicting the probability of credit default of commercial entities based on deep learning.
[0005] A first aspect of this application provides a method for predicting the probability of credit default of business entities based on deep learning, the method comprising: A multi-source heterogeneous data set of business entities is obtained, and field mapping, time alignment, and expiration date labeling are performed respectively to obtain a preprocessed multi-source heterogeneous data set. A heterogeneous graph network containing business entity nodes and event nodes is constructed based on the preprocessed multi-source heterogeneous data set. The heterogeneous graph network is weighted by multi-relation mapping to obtain a multi-relation weighted graph containing only business entity nodes. Using the multi-relation weighted graph as an index, a basic default probability prediction model is trained using deep learning technology, wherein the basic default probability prediction model is used to output the default probability value.
[0006] A second aspect of this application provides a deep learning-based system for predicting the probability of credit default of business entities, the system comprising: The data processing module is used to acquire a multi-source heterogeneous data set of business entities, and perform field mapping, time alignment, and validity period labeling processing to obtain a preprocessed multi-source heterogeneous data set; the network construction module is used to construct a heterogeneous graph network containing business entity nodes and event nodes based on the preprocessed multi-source heterogeneous data set; the weighting module is used to perform multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes; the training module is used to train a basic default probability prediction model using deep learning technology with the multi-relation weighted graph as an index, wherein the basic default probability prediction model is used to output default probability values.
[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application obtains a multi-source heterogeneous dataset of business entities, performs field mapping, time alignment, and expiration date labeling processing to obtain a preprocessed multi-source heterogeneous dataset; constructs a heterogeneous graph network containing business entity nodes and event nodes based on the preprocessed multi-source heterogeneous dataset; performs multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes; uses the multi-relation weighted graph as an index to train a basic default probability prediction model using deep learning technology, wherein the basic default probability prediction model is used to output default probability values. This invention solves the technical problem in the prior art where it is difficult to effectively model the complex relationships between business entities, leading to insufficient accuracy in credit default probability prediction. By constructing a heterogeneous graph of business entities and performing multi-relation mapping weighting, and introducing a deep learning model for structured feature learning, the technical effect of improving the accuracy of credit default probability prediction for business entities is achieved. Attached Figure Description
[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0009] Figure 1 A schematic diagram of a deep learning-based method for predicting the probability of credit default of a business entity, provided in an embodiment of this application. Figure 2 A schematic diagram of the structure of a deep learning-based business entity credit default probability prediction system provided in this application embodiment.
[0010] Figure labeling: Data processing module 11, network construction module 12, weighting module 13, training module 14. Detailed Implementation
[0011] This application provides a method and system for predicting the probability of credit default of commercial entities based on deep learning. It addresses the technical problem that existing technologies struggle to effectively model the complex relationships between commercial entities, leading to insufficient accuracy in predicting the probability of credit default. By constructing a heterogeneous graph of commercial entities and performing multi-relationship mapping and weighting, and introducing a deep learning model for structured feature learning, the application achieves the technical effect of improving the accuracy of predicting the probability of credit default of commercial entities.
[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.
[0014] Example 1, as Figure 1 As shown, this application provides a method for predicting the probability of credit default of commercial entities based on deep learning, the method comprising: Step S100: Obtain a multi-source heterogeneous data set of the business entity set, and perform field mapping, time alignment and validity period labeling processing respectively to obtain a preprocessed multi-source heterogeneous data set.
[0015] Furthermore, the method provided in the application embodiments also includes: The multi-source heterogeneous data includes business registration information, financial or tax summary information, administrative penalties and abnormal records, litigation records, environmental impact assessment records, qualification lists, and industry and sector indicators.
[0016] In this embodiment of the application, to obtain a multi-source heterogeneous data set of a set of business entities, data records related to the business entities are first retrieved from multiple government information systems and business systems. These government information systems and business systems include industrial and commercial supervision systems, tax management systems, judicial disclosure systems, ecological and environmental management systems, and technology enterprise identification and industry statistics systems. Through interface reading, batch import, or database extraction, business registration information, financial or tax summary information, administrative penalties and abnormal records, litigation records, environmental assessment records, qualification lists, and industry and sector indicators are obtained respectively to form an initial multi-source heterogeneous data set. The different data sources in this multi-source heterogeneous data set have differences in field naming rules, data structure, and semantic expression.
[0017] Next, unified field processing is performed on various types of data in the multi-source heterogeneous dataset. During this process, based on preset feature field specifications, semantically consistent but differently named fields from different data sources are converted accordingly, and the data types of the fields are uniformly adjusted. Numeric fields are formatted according to a unified unit, enumerated fields are encoded according to a unified value-taking rule, and text fields are organized according to a predetermined structure, thus ensuring that the processed data has a consistent structure and semantics at the field level.
[0018] After the fields are processed uniformly, the time information in the data is organized and aligned. For data records containing occurrence or effective times, their original timestamps are parsed, and the data is mapped to the corresponding time interval according to a unified monthly time granularity. For data with inconsistent time granularity, time truncation or time merging is used to ensure that all data is expressed on the same time scale, thereby guaranteeing consistency in the time dimension across different data sources.
[0019] Building upon time alignment, validity period labels are applied to time-sensitive data. For data such as qualification lists and administrative licenses, a clear validity period identifier is generated for each data entry, based on its publication date, validity start and end dates, and revocation or cancellation status. This identifier characterizes the data's validity status across different time intervals. Simultaneously, missing field identifiers are generated for data records with incomplete fields, and a data source identifier is appended to each data entry to distinguish its origin and support subsequent processing.
[0020] By sequentially performing data acquisition, field unification, time alignment, and validity period labeling on multi-source heterogeneous data, the original data is unified in terms of structure, time, and validity, ultimately resulting in a preprocessed multi-source heterogeneous data set.
[0021] Step S200: Construct a heterogeneous graph network containing business entity nodes and event nodes based on a preprocessed multi-source heterogeneous data set.
[0022] In this embodiment, when constructing a heterogeneous graph network containing business entity nodes and event nodes based on a preprocessed multi-source heterogeneous data set, the preprocessed multi-source heterogeneous data set is first parsed to extract uniquely identified business entity information. Each business entity is then used as a basic node unit in the graph structure to generate a corresponding business entity node. Subsequently, using the business entity nodes as indexes, event records associated with each business entity are retrieved from the preprocessed multi-source heterogeneous data set, and each associated event record is mapped to an event node, thus forming a set of event nodes corresponding to the business entity nodes. Then, based on the correspondence between business entities and associated event records, association edges are established between business entity nodes and event nodes. The association edges and nodes are then type-labeled according to the event type corresponding to the event node, so that different event types correspond to different relational semantics. Finally, a heterogeneous graph network containing business entity nodes, event nodes, and multi-type relational edges is constructed.
[0023] Furthermore, the method provided in the application embodiments, which constructs a heterogeneous graph network containing business entity nodes and event nodes based on a preprocessed multi-source heterogeneous data set, further includes: Multiple business entities are extracted from a preprocessed multi-source heterogeneous data set, and a business entity node is generated for each business entity to obtain multiple business entity nodes. Using the multiple business entity nodes as indexes, associated event records are extracted from the preprocessed multi-source heterogeneous data set, and an event node is generated for each associated event record to obtain multiple event node sets. According to the correspondence between business entities and associated event records, at least one association edge is established between business entity nodes and event nodes. According to different event types, multiple business entity nodes and multiple event node sets are identified to construct a heterogeneous graph network containing multiple types of relationship edges.
[0024] In this embodiment, the entity identification information in the preprocessed multi-source heterogeneous data set is first parsed. Data from different sources, such as business registration information, financial or tax summary information, administrative penalties and abnormal records, litigation records, environmental impact assessment records, qualification lists, and industry and sector indicators, are aggregated and deduplicated using the unified social credit code, thereby identifying multiple distinguishable business entities. Then, for each identified business entity, a unique node identifier is assigned, and basic attribute information such as registered capital, number of insured persons, tax credit rating, enterprise type, and industry are aggregated and bound, thus generating multiple business entity nodes.
[0025] Next, using multiple business entity nodes as indexes, we perform correlation retrieval on the preprocessed multi-source heterogeneous data set. We match different types of event records corresponding to each business entity based on its unique identifier. For example, we extract administrative penalty information for the past year from administrative penalty and anomaly records; we extract litigation case information where the company is the defendant from litigation records; we extract environmental credit rating information from environmental assessment records; and we extract certification information such as high-tech enterprise and specialized and innovative enterprise from the qualification list. For each retrieved related event record, we extract key information such as event occurrence time, event category, event status, and data source, and assign a unique event identifier to each related event record, thereby generating corresponding event nodes. Based on event categories, we form multiple event node sets, such as administrative penalty event node set, litigation event node set, environmental assessment event node set, and qualification certification event node set, with each event node set corresponding to a specific event type.
[0026] Next, based on the established correspondence between business entities and related event records in the preprocessed multi-source heterogeneous data set, a one-to-one or one-to-many mapping relationship between business entities and different event nodes is determined. At least one association edge is defined between the corresponding business entity node and event node to represent facts such as administrative penalties, litigation, qualification certification, or environmental impact assessment results. Simultaneously, based on the event category to which the event node belongs, a clear relationship type identifier is assigned to the association edge, such as administrative penalty relationship, litigation relationship, environmental impact assessment relationship, and qualification certification relationship, so that different event node sets correspond to different types of relationship edges. A consistent node type identifier is simultaneously assigned to the business entity nodes and event nodes.
[0027] After defining and identifying the business entity nodes, the multiple event node sets, and the associated edges with relation type identifiers, the business entity nodes, the multiple event node sets, and the multiple types of relation edges are organized in a unified manner to form a network structure that contains multiple node types and multiple relation types. Finally, a heterogeneous graph network containing business entity nodes, event nodes, and multiple types of relation edges is constructed.
[0028] Furthermore, the method provided in the application embodiments also includes: Based on the preprocessed multi-source heterogeneous data set, each event node is associated with its corresponding occurrence time, event category, and data source identifier to obtain the attribute information of each event node.
[0029] In this embodiment, the correspondence between event nodes and original event records is located based on a preprocessed multi-source heterogeneous data set. The event unique identifier contained in the event node is used to backtrack and match the corresponding event record in the preprocessed multi-source heterogeneous data set, thereby establishing a one-to-one mapping relationship between event nodes and source data.
[0030] After matching event records, the time field of the corresponding event records in the preprocessed multi-source heterogeneous dataset is parsed. The format of the event occurrence time, event effectiveness time, or record generation time is standardized and semantically determined. Time information that accurately reflects the actual occurrence or effectiveness of the event is selected and bound as the occurrence time attribute of the event node, representing the event's position in the time dimension. Next, the type field of the event records in the preprocessed multi-source heterogeneous dataset is parsed. Based on the data source and business meaning of the event record, the corresponding business category of the event is determined, and this result is used as the event category attribute of the event node to distinguish different types of events such as administrative penalty events, litigation events, environmental assessment events, and qualification certification events. Finally, the source field of the event records in the preprocessed multi-source heterogeneous dataset is parsed to identify the specific system or data channel from which the event data originates, and the corresponding information is bound as the data source identifier attribute of the event node.
[0031] By sequentially completing event record matching, occurrence time extraction, event category determination, and data source identifier association, the occurrence time, the event category, and the data source identifier are fully associated with each event node, thereby obtaining event node attribute information containing time attributes, semantic attributes, and source attributes.
[0032] Step S300: Perform multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes.
[0033] In this embodiment, when performing multi-relation mapping weighting on a heterogeneous graph network, the process first involves identifying pairs of business entity nodes connected by event nodes as intermediaries within the network according to a preset structural path. This yields a set of business entity node pairs that satisfy the path pattern of business entity node-event node-business entity node. Subsequently, the business entity node pairs are traversed to extract their corresponding mapping connection types, forming a set of mapping connection types. Then, based on this set, business entity node pairs connected by the same relationship mapping rule are uniformly identified as relationship edges of the same type, resulting in multiple sets of relationship edges and corresponding sets of business entity node pairs, where each set of relationship edges has the same relationship type. Finally, based on these multiple sets of relationship edges and business entity node pairs, a multi-relation weighted graph containing only business entity nodes and capable of representing multiple relationship types is constructed.
[0034] Furthermore, in the method provided in the application embodiments, performing multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes further includes: In the heterogeneous graph network, pairs of business entity nodes that satisfy a preset structural path are identified to obtain a set of business entity node pairs, wherein the preset structural path is business entity node-event node-business entity node; the business entity node pairs are traversed to extract mapping connection types to obtain a set of mapping connection types; based on the set of mapping connection types, business entity node pairs connected by the same relation mapping rule are identified as a type of relation edge to obtain multiple sets of relation edges and multiple sets of business entity node pairs, wherein each set of relation edges is of the same type; the multi-relation weighted graph is constructed based on the multiple sets of relation edges and the multiple sets of business entity node pairs.
[0035] In this embodiment, the business entity nodes, event nodes, and their connections within the heterogeneous graph network are first traversed. Starting with a business entity node, the process searches layer by layer along the edges connecting the business entity node and the event node to find event nodes directly associated with that business entity node. Then, each event node continues searching along its edges to find other connected business entity nodes. This process identifies node sequences in the heterogeneous graph network that connect a business entity node to another business entity node via an event node. All node sequences satisfying a preset structural path are then summarized, where the preset structural path is limited to business entity node-event node-business entity node. Through this path identification process, two business entity nodes indirectly associated through the same event node are grouped together, ultimately resulting in a set of business entity node pairs containing all such node combinations.
[0036] Next, the set of business entity node pairs is traversed one by one. By tracing back the intermediate event nodes corresponding to each pair of business entity nodes in the heterogeneous graph network, the attribute information such as the event category, occurrence time, and data source identifier bound to the event node is analyzed. Combined with the semantic meaning of the event node at the business level, it is determined whether the association between the business entity node pairs is formed due to administrative penalty events, litigation events, environmental assessment events, or qualification certification events. Then, for each pair of business entity node pairs, mapping connection types that can characterize their association mode and semantic features are extracted. The mapping connection types corresponding to all business entity node pairs are summarized to form a set of mapping connection types used to describe different association semantics.
[0037] After forming a set of mapping connection types, the business entity node pairs are categorized based on this set. In this process, according to pre-defined relationship mapping rules, business entity node pairs connected by the same mapping connection type are uniformly identified as a single type of relationship edge, ensuring that each relationship edge corresponds to a specific mapping connection type. Through this categorization process, multiple sets of relationship edges and corresponding sets of business entity node pairs are obtained, where the relationship types within each set of relationship edges remain consistent, representing the association relationships between business entities of the same type.
[0038] Finally, a multi-relation weighted graph is constructed based on multiple sets of relation edges and multiple sets of business entity node pairs. In this process, business entity node pairs are first selected sequentially from the multiple sets of business entity node pairs, and different types of relation edges corresponding to the selected business entity node pairs are extracted from the multiple sets of relation edges, forming multiple sets of association type relation edges. Then, for cases where business entity node pairs are associated through the same type of relation edge in multiple sets of association type relation edges, the edge weight value corresponding to each type of relation edge is calculated, and the edge weight value is normalized to obtain the normalized edge weight value of the business entity node pair under different relation types. After completing the relation type identification and edge weight normalization value calculation for a single business entity node pair, the remaining business entity node pairs in the multiple sets of business entity node pairs are processed according to the same steps. Finally, the normalized edge weight values of each business entity node pair under different relation types are uniformly organized to construct a multi-relation weighted graph that contains only business entity nodes and has multiple relation types and corresponding weights.
[0039] Furthermore, in the method provided in the application embodiments, constructing the multi-relationship weighted graph based on the multiple sets of relationship edges and the multiple sets of business entity node pairs further includes: Extract the first business entity node pair from the multiple sets of business entity node pairs; extract the corresponding relationship edges of the first business entity node pair from the multiple sets of relationship edges to obtain multiple sets of relationship edges of association types; calculate the edge weight value of each type of relationship edge based on the fact that the first business entity node pair is associated with the same type of relationship edge in the multiple sets of relationship edges of association types, and perform normalization processing to obtain multiple edge weight normalization values of multiple types of relationship edges of the first business entity node pair; based on the same steps above, perform type relationship edge identification and edge weight normalization value calculation for each business entity node pair in the multiple sets of business entity node pairs to construct a multi-relationship weighted graph containing only business entity nodes.
[0040] In this embodiment of the application, the multiple sets of business entity node pairs are first sequentially traversed, and any set of business entity node pairs is selected as the first business entity node pair, wherein the first business entity node pair consists of two business entity nodes that have been identified and retained in the multi-relationship modeling process.
[0041] Next, index matching is performed on multiple sets of relation edges. By comparing the identifiers of the business entity nodes connected to both ends of the relation edges, relation edges that completely correspond to the first business entity node pair are selected. The selection results are then classified and organized according to the relation type identifier carried by the relation edges. Thus, the relation edges corresponding to the first business entity node pair under different relation types are assigned to different sets, thereby obtaining multiple sets of relation edges of association types. Each set of relation edges of association types represents all the associated records of the first business entity node pair under the same relation type.
[0042] After obtaining multiple sets of relational edges of various association types, for each pair of first business entity nodes, the edge weight value is calculated sequentially for each set of relational edges of different association types. Specifically, firstly, all relational edge records contained in the set of relational edges of that association type are read, and the intermediate event nodes corresponding to each relational edge are counted. The total number of times the first business entity node pair generates associations through event nodes under that relation type is counted, and this total number is recorded as the original association count value under that relation type. Subsequently, the original association count value is directly used as the initial edge weight value corresponding to that relation type, so that the first business entity node pair obtains an initial edge weight value calculated based on the actual number of associations under each relation type. This initial edge weight value is used to quantitatively reflect the actual association strength of the first business entity node pair under that relation type.
[0043] After calculating the initial edge weights of the first business entity node pair across all relational edge sets, normalization is performed on all initial edge weights corresponding to the first business entity node pair. In this process, all initial edge weights obtained for the first business entity node pair under different relational types are aggregated to form a weight vector, and the sum of all initial edge weights in this weight vector is calculated. Then, the initial edge weights of the first business entity node pair under a specific relational type are divided by the sum to obtain the normalized edge weights for that relational type. Through this item-by-item division operation, the sum of the normalized edge weights of the first business entity node pair under different relational types is made equal to 1, thus achieving a unified expression of edge weights for different relational types on the same numerical scale.
[0044] After identifying the association type relationship edge set, calculating the initial edge weight, and calculating the normalized edge weight for the first business entity node pair, the remaining business entity node pairs in multiple sets of business entity node pairs are processed sequentially according to the same calculation process. This involves extracting the corresponding relationship edges, constructing the association type relationship edge set, calculating the initial edge weight, and calculating the normalized edge weight for each pair. Once all business entity node pairs in multiple sets of business entity node pairs have completed the above calculations, all business entity nodes are treated as nodes in the graph structure. The normalized edge weights of each business entity node pair under different relationship types are then uniformly organized as the weight attributes of the relationship edges. Finally, a multi-relationship weighted graph is constructed that contains only business entity nodes and is connected between nodes through various relationship types and their corresponding normalized edge weights.
[0045] Step S400: Using the multi-relationship weighted graph as an index, train a basic default probability prediction model using deep learning technology, wherein the basic default probability prediction model is used to output the default probability value.
[0046] In this embodiment, when training the basic default probability prediction model using a multi-relationship weighted graph as an index, the multi-relationship weighted graph is first used as a structural index to associate historical default data, constructing a model input sample set containing the structural information of business entity nodes and their corresponding default labels. Then, under the structural constraints of the multi-relationship weighted graph, deep learning technology is used to learn features from the model input sample set. By jointly modeling the structural features and association features of business entity nodes, the trained basic default probability prediction model is obtained. The basic default probability prediction model outputs prediction results in a continuous numerical form within the 0-1 interval, representing the probability of a business entity defaulting on its credit.
[0047] Furthermore, in the method provided in the application embodiments, a basic default probability prediction model is trained using deep learning technology, with the multi-relationship weighted graph as an index. The basic default probability prediction model is used to output default probability values and further includes: Using a multi-relationship weighted graph as a structural index, historical default data is obtained to construct a model input sample set. Under the constraints of the multi-relationship weighted graph, deep learning technology is used to perform feature learning on the model input sample set to obtain a trained basic default probability prediction model. The basic default probability prediction model outputs a continuous value between 0 and 1 as the default probability prediction result.
[0048] In this embodiment, when constructing the model input sample set using a multi-relation weighted graph as a structural index, default information is first extracted from historical default data to determine the corresponding defaulting business entity. Based on the multi-relation weighted graph, a set of relationship edges of various relationship types associated with the business entity node corresponding to the defaulting business entity is obtained. Then, intermediate representation vectors are extracted from each relationship type set in conjunction with the default information, and the obtained intermediate representation vectors are weighted and fused to form a default node embedding representation that characterizes the structural features of the defaulting business entity. Next, the default information and the default node embedding representation are mapped and associated to generate corresponding model input samples. These generated model input samples are then collected one by one to finally construct the model input sample set used for model training.
[0049] Next, under the structural constraints of a multi-relationship weighted graph, the model input sample set is used as the model input. Samples containing default node embeddings and default information are fed into a deep neural network. A multi-layer network structure performs nonlinear mapping on the input features, achieving high-order representation learning of the credit risk characteristics of commercial entities. During model training, a numerical loss function is constructed based on the deviation between the predicted output of the samples and the actual default labels. The backpropagation algorithm is used to iteratively update the model parameters in the deep neural network to continuously reduce the value of the loss function. When the model parameter update process meets the preset convergence condition, model training is completed, obtaining a trained basic default probability prediction model. This basic default probability prediction model outputs prediction results in a continuous numerical form within the 0-1 interval, representing the probability of a commercial entity defaulting on its credit.
[0050] Furthermore, the method provided in the application embodiments, which uses a multi-relationship weighted graph as a structural index to obtain historical default data and construct a model input sample set, also includes: First default information is extracted from historical default data, and first default business entity is extracted from the first default information. Based on the multi-relationship weighted graph, multiple sets of first default type relationship edges with different relationship types associated with the business entity node corresponding to the first default business entity are extracted. Intermediate representation vectors are extracted from the first default information and the first default type relationship edge sets to obtain a first intermediate representation vector set. The first intermediate representation vector set is weighted to obtain the first default node embedding representation. The first default information and the first default node embedding representation are mapped and associated to obtain a first model input sample, and the first model input sample is added to the model input sample set.
[0051] In this embodiment, when constructing the model input sample set using a multi-relationship weighted graph as a structural index, historical default data is first sequentially read or randomly sampled. Any default record is selected as the first default information, and this first default information is parsed to extract the default identifier, the time of default, and the information of the defaulting entity. Then, based on the defaulting entity information, the business entity that committed the default is identified, and this business entity is designated as the first defaulting business entity. The selection of the first default information and the first defaulting business entity is not limited to specific records; both are arbitrary samples from the historical default data.
[0052] Next, based on the unique identifier of the business entity node in the multi-relation weighted graph, the business entity node corresponding to the first defaulting business entity is located in the multi-relation weighted graph. Taking this business entity node as the central node, the connection structure formed by the relation edges of different relation types in the multi-relation weighted graph is extracted. According to the relation type identifier of the relation edge, various relation edges directly connected to the business entity node are retrieved respectively, thereby obtaining multiple sets of first default type relation edges of different relation types corresponding to the first defaulting business entity. Each set of first default type relation edges contains all relation edges of the same relation type that are associated with the first defaulting business entity node.
[0053] Next, intermediate representation vector extraction is performed for each set of relationship edges for the first default type. In this process, firstly, the node feature vectors of the adjacent business entity nodes connected to each relationship edge in the set of relationship edges for the first default type are read, and simultaneously, the normalized edge weight value of the corresponding relationship edge in the multi-relationship weighted graph is read. Then, for each relationship edge, its corresponding normalized edge weight value is used to numerically weight the feature vectors of the adjacent business entity nodes connected to it; that is, each dimension of the adjacent business entity node feature vector is multiplied by the normalized edge weight value to obtain the weighted node feature vector. After completing the weighting process for the feature vectors of adjacent business entity nodes corresponding to all relationship edges in the set of relationship edges for the first default type, a summation operation is performed on all weighted node feature vectors according to their dimensions. This aggregates the feature information of multiple adjacent business entity nodes under this relationship type into a unified vector representation, which serves as the intermediate representation vector corresponding to the set of relationship edges for the first default type. This completes the extraction of intermediate representation vectors for this relationship type, and corresponding intermediate representation vectors are obtained for different relationship types, forming a first set of intermediate representation vectors.
[0054] Next, the first intermediate representation vector set is weighted. In this process, based on the normalized edge weights of the edges of different relation types in the multi-relation weighted graph, a cross-relation type weighted fusion process is performed on each intermediate representation vector in the first intermediate representation vector set. Each intermediate representation vector is multiplied by the normalized edge weight of its corresponding relation type, and all weighted intermediate representation vectors are accumulated according to their dimensions to obtain a vector representation that integrates the influence of multiple relation types. This vector is the embedded representation of the first default node corresponding to the first defaulting business entity.
[0055] Finally, the first default node is embedded and mapped to the default status identifier in the first default information to construct a first model input sample containing the correspondence between node structural features and default labels, and the first model input sample is added to the model input sample set.
[0056] Furthermore, in the method provided in the application embodiments, under the constraints of the multi-relationship weighted graph, deep learning technology is used to perform feature learning on the model input sample set to obtain a trained basic default probability prediction model, which further includes: Using the set of input samples for the model as the model input, a nonlinear mapping is performed through a deep neural network; the model parameters are updated through the backpropagation algorithm to minimize the numerical loss function; when the training converges, the basic default probability prediction model is obtained.
[0057] In this embodiment, the model input sample set is used as the overall training data and input into a deep neural network for training. The deep neural network consists of an input layer, at least one hidden layer, and an output layer. The input layer receives the default node embedding representation contained in the model input samples. During the forward computation of the model, at least one hidden layer processes the input data using a local feature extraction method based on weight parameters. The input default node embedding representation is divided into several adjacent feature regions according to a predetermined dimension, and each feature region is weighted using a corresponding weight parameter matrix. The weight parameter matrix is functionally equivalent to a convolution kernel, and its size can be set according to the actual feature dimension. For example, a 3×3 local weight structure can be used to jointly model adjacent feature dimensions to extract local correlation features in the default node embedding representation. After the local feature weighting is completed, the weighted results are summed item by item to obtain the linear computation result of the hidden layer neurons. Then, deterministic numerical processing steps are performed on the linear computation result. When the linear computation result is negative, it is adjusted to zero; when the linear computation result is positive, its value remains unchanged. This filters the feature responses so that the network retains only the feature information that contributes positively to the judgment of default risk. By repeatedly performing feature extraction, weighted summation, and numerical filtering based on local weight parameters in the hidden layers, the model can progressively change the numerical distribution and representation of features, gradually transforming the original default node embedding representation into a high-level feature representation capable of distinguishing different credit risk levels. Finally, the output layer calculates the output of the last hidden layer. In this process, the feature vector output from the last hidden layer is first multiplied term by term with the corresponding weight parameters in the output layer and then summed to obtain the original numerical value representing the prediction result. Subsequently, a numerical range constraint process is performed on the original value. The original value is compared with a preset lower limit of 0. If the original value is less than 0, it is adjusted to 0. The original value is compared with a preset upper limit of 1. If the original value is greater than 1, it is adjusted to 1. If the original value is between 0 and 1, its original value remains unchanged. Through the above numerical clipping and constraint process, the calculation result of the output layer is limited to the numerical range of 0 to 1, thereby obtaining the final prediction result corresponding to each model input sample. The prediction result represents the probability of the business entity defaulting in the form of continuous numerical values. The closer the value is to 1, the higher the default risk, and the closer the value is to 0, the lower the default risk.
[0058] After completing the forward computation of the deep neural network and obtaining the prediction results, the predicted probabilities are compared one by one with the corresponding actual default indicators in the model input samples. The actual default indicator indicates whether a business entity has actually defaulted; a value of 1 corresponds to a default, and a value of 0 corresponds to no default. Subsequently, the numerical difference between the predicted probability and the actual default indicator is calculated for each model input sample. The differences for all samples in the model input sample set are then summed or averaged to obtain a numerical loss function that measures the overall error of the current model prediction results. This numerical loss function reflects the degree of deviation between the deep neural network's predicted default probability and the actual default situation under the current parameter conditions. A larger value indicates a larger prediction error, while a smaller value indicates that the predicted result is closer to the actual default state.
[0059] Then, the backpropagation algorithm is used to update the model parameters in the deep neural network. Starting from the output layer, based on the calculation results of the numerical loss function, the influence of each layer's weight parameters on the change in the numerical loss function value is calculated layer by layer. The corresponding weight parameters are then adjusted according to this influence, ensuring that the updated model parameters further reduce the value of the numerical loss function in the next round of forward computation. This achieves parameter optimization with the goal of minimizing the numerical loss function. By repeatedly executing the forward computation, numerical loss function calculation, and model parameter update steps during training, the parameters of the deep neural network gradually converge towards the direction that minimizes the numerical loss function value.
[0060] As training iterations continue, when the numerical loss function gradually decreases and eventually stabilizes over multiple training iterations, it indicates that adjusting the model parameters is no longer sufficient to further reduce the prediction error. At this point, the training process of the deep neural network is considered to have converged, and updates to the model parameters cease, thus obtaining the trained basic default probability prediction model. With fixed parameters, the basic default probability prediction model can stably output continuous values between 0 and 1 based on the input default node embedding representation, through layer-by-layer weighted calculations and numerical filtering. These values serve as the prediction results for the credit default probability of commercial entities.
[0061] In summary, the embodiments of this application have at least the following technical effects: This application obtains a multi-source heterogeneous dataset of business entities, performs field mapping, time alignment, and expiration date labeling processing to obtain a preprocessed multi-source heterogeneous dataset; constructs a heterogeneous graph network containing business entity nodes and event nodes based on the preprocessed multi-source heterogeneous dataset; performs multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes; uses the multi-relation weighted graph as an index to train a basic default probability prediction model using deep learning technology, wherein the basic default probability prediction model is used to output default probability values. This invention solves the technical problem in the prior art where it is difficult to effectively model the complex relationships between business entities, leading to insufficient accuracy in credit default probability prediction. By constructing a heterogeneous graph of business entities and performing multi-relation mapping weighting, and introducing a deep learning model for structured feature learning, the technical effect of improving the accuracy of credit default probability prediction for business entities is achieved.
[0062] Example 2, based on the same inventive concept as the deep learning-based business entity credit default probability prediction method in the previous examples, such as... Figure 2 As shown, this application provides a business entity credit default probability prediction system based on deep learning. The system and method embodiments in this application are based on the same inventive concept. The system includes: The data processing module 11 is used to acquire a multi-source heterogeneous data set of business entities, and perform field mapping, time alignment and validity period labeling processing respectively to obtain a preprocessed multi-source heterogeneous data set; the network construction module 12 is used to construct a heterogeneous graph network containing business entity nodes and event nodes based on the preprocessed multi-source heterogeneous data set; the weighting module 13 is used to perform multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph containing only business entity nodes; the training module 14 is used to train a basic default probability prediction model using deep learning technology with the multi-relation weighted graph as an index, wherein the basic default probability prediction model is used to output default probability values.
[0063] Furthermore, the system is also used to implement the following functions: The multi-source heterogeneous data includes business registration information, financial or tax summary information, administrative penalties and abnormal records, litigation records, environmental impact assessment records, qualification lists, and industry and sector indicators.
[0064] Furthermore, the system is also used to implement the following functions: Multiple business entities are extracted from a preprocessed multi-source heterogeneous data set, and a business entity node is generated for each business entity to obtain multiple business entity nodes. Using the multiple business entity nodes as indexes, associated event records are extracted from the preprocessed multi-source heterogeneous data set, and an event node is generated for each associated event record to obtain multiple event node sets. According to the correspondence between business entities and associated event records, at least one association edge is established between business entity nodes and event nodes. According to different event types, multiple business entity nodes and multiple event node sets are identified to construct a heterogeneous graph network containing multiple types of relationship edges.
[0065] Furthermore, the system is also used to implement the following functions: Based on the preprocessed multi-source heterogeneous data set, each event node is associated with its corresponding occurrence time, event category, and data source identifier to obtain the attribute information of each event node.
[0066] Furthermore, the system is also used to implement the following functions: In the heterogeneous graph network, pairs of business entity nodes that satisfy a preset structural path are identified to obtain a set of business entity node pairs, wherein the preset structural path is business entity node-event node-business entity node; the business entity node pairs are traversed to extract mapping connection types to obtain a set of mapping connection types; based on the set of mapping connection types, business entity node pairs connected by the same relation mapping rule are identified as a type of relation edge to obtain multiple sets of relation edges and multiple sets of business entity node pairs, wherein each set of relation edges is of the same type; the multi-relation weighted graph is constructed based on the multiple sets of relation edges and the multiple sets of business entity node pairs.
[0067] Furthermore, the system is also used to implement the following functions: Extract the first business entity node pair from the multiple sets of business entity node pairs; extract the corresponding relationship edges of the first business entity node pair from the multiple sets of relationship edges to obtain multiple sets of relationship edges of association types; calculate the edge weight value of each type of relationship edge based on the fact that the first business entity node pair is associated with the same type of relationship edge in the multiple sets of relationship edges of association types, and perform normalization processing to obtain multiple edge weight normalization values of multiple types of relationship edges of the first business entity node pair; based on the same steps above, perform type relationship edge identification and edge weight normalization value calculation for each business entity node pair in the multiple sets of business entity node pairs to construct a multi-relationship weighted graph containing only business entity nodes.
[0068] Furthermore, the system is also used to implement the following functions: Using a multi-relationship weighted graph as a structural index, historical default data is obtained to construct a model input sample set. Under the constraints of the multi-relationship weighted graph, deep learning technology is used to perform feature learning on the model input sample set to obtain a trained basic default probability prediction model. The basic default probability prediction model outputs a continuous value between 0 and 1 as the default probability prediction result.
[0069] Furthermore, the system is also used to implement the following functions: First default information is extracted from historical default data, and first default business entity is extracted from the first default information. Based on the multi-relationship weighted graph, multiple sets of first default type relationship edges with different relationship types associated with the business entity node corresponding to the first default business entity are extracted. Intermediate representation vectors are extracted from the first default information and the first default type relationship edge sets to obtain a first intermediate representation vector set. The first intermediate representation vector set is weighted to obtain the first default node embedding representation. The first default information and the first default node embedding representation are mapped and associated to obtain a first model input sample, and the first model input sample is added to the model input sample set.
[0070] Furthermore, the system is also used to implement the following functions: Using the set of input samples for the model as the model input, a nonlinear mapping is performed through a deep neural network; the model parameters are updated through the backpropagation algorithm to minimize the numerical loss function; when the training converges, the basic default probability prediction model is obtained.
[0071] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for predicting the probability of credit default of commercial entities based on deep learning, characterized in that, The method includes: Obtain a multi-source heterogeneous data set of business entities, and perform field mapping, time alignment and validity period labeling processing respectively to obtain a preprocessed multi-source heterogeneous data set; A heterogeneous graph network containing business entity nodes and event nodes is constructed based on a preprocessed multi-source heterogeneous data set; The heterogeneous graph network is weighted by multi-relation mapping to obtain a multi-relation weighted graph that contains only business entity nodes. Using the aforementioned multi-relationship weighted graph as an index, a basic default probability prediction model is trained using deep learning technology, wherein the basic default probability prediction model is used to output default probability values.
2. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 1, characterized in that, The multi-source heterogeneous data includes business registration information, financial or tax summary information, administrative penalties and abnormal records, litigation records, environmental assessment records, qualification lists, and industry and sector indicators.
3. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 1, characterized in that, A heterogeneous graph network containing business entity nodes and event nodes is constructed based on a preprocessed multi-source heterogeneous data set, including: Multiple business entities are extracted from a preprocessed multi-source heterogeneous data set, and a business entity node is generated for each business entity, resulting in multiple business entity nodes. Using the multiple business entity nodes as indexes, related event records are extracted from the preprocessed multi-source heterogeneous data set, and an event node is generated for each related event record to obtain multiple event node sets; Based on the correspondence between business entities and associated event records, at least one association edge is established between business entity nodes and event nodes. Based on different event types, multiple business entity nodes and multiple event node sets are identified to construct a heterogeneous graph network containing multiple types of relationship edges.
4. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 3, characterized in that, Based on the preprocessed multi-source heterogeneous data set, each event node is associated with its corresponding occurrence time, event category, and data source identifier to obtain the attribute information of each event node.
5. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 1, characterized in that, The heterogeneous graph network is weighted by multi-relation mapping to obtain a multi-relation weighted graph containing only business entity nodes, including: In the heterogeneous graph network, identify pairs of business entity nodes that satisfy a preset structural path to obtain a set of business entity node pairs, wherein the preset structural path is business entity node-event node-business entity node; The mapping connection type is extracted by traversing the business entity node pairs to obtain a set of mapping connection types. Based on the set of mapping connection types, pairs of business entity nodes connected by the same relation mapping rule are identified as a type of relation edge, resulting in multiple sets of relation edge sets and multiple sets of business entity node pairs, wherein each set of relation edge sets has the same type. Based on the multiple sets of relation edges and the multiple sets of business entity node pairs, the multi-relation weighted graph is constructed.
6. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 5, characterized in that, Based on the multiple sets of relation edges and the multiple sets of business entity node pairs, the multi-relationship weighted graph is constructed, including: Extract the first business entity node pair from the set of multiple sets of business entity node pairs; From the multiple sets of relation edges, extract the relation edges corresponding to the first business entity node pairs to obtain multiple sets of relation edges of association types; Based on the fact that the first business entity node pair is associated with the same type of relation edge in multiple association type relation edge sets, calculate the edge weight value of each type of relation edge, and perform normalization processing to obtain the normalized values of multiple edge weights of multiple types of relation edges of the first business entity node pair. Based on the same steps described above, type relationship edge identification and edge weight normalization value calculation are performed for each business entity node pair in the set of multiple business entity node pairs to construct a multi-relation weighted graph containing only business entity nodes.
7. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 1, characterized in that, Using the aforementioned multi-relationship weighted graph as an index, a basic default probability prediction model is trained using deep learning techniques. This basic default probability prediction model outputs default probability values, including: Using a multi-relationship weighted graph as a structural index, historical default data is obtained to construct the model input sample set; Under the constraints of the multi-relationship weighted graph, deep learning technology is used to learn features from the model input sample set to obtain a trained basic default probability prediction model. The basic default probability prediction model outputs a continuous value between 0 and 1 as the default probability prediction result.
8. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 7, characterized in that, Using a multi-relationship weighted graph as the structural index, historical default data is obtained to construct the model input sample set, including: Extract the first default information from historical default data, and extract the first defaulting business entity from the first default information; Based on the multi-relation weighted graph, extract a set of multiple first default type relation edges associated with different relation types of the business entity node corresponding to the first defaulting business entity; By combining the first default information with the intermediate representation vector extraction of the first default type relation edge set, a first intermediate representation vector set is obtained; The first intermediate representation vector set is weighted to obtain the first default node embedding representation; The first default information and the first default node embedding representation are mapped and associated to obtain the first model input sample, and the first model input sample is added to the model input sample set.
9. The method for predicting the probability of credit default of commercial entities based on deep learning as described in claim 8, characterized in that, Under the constraints of the multi-relationship weighted graph, deep learning techniques are used to learn features from the model input sample set to obtain a trained basic default probability prediction model, including: Using the set of input samples for the model as input, a nonlinear mapping is performed through a deep neural network; The model parameters are updated using the backpropagation algorithm to minimize the numerical loss function; Once the training converges, the basic default probability prediction model is obtained.
10. A deep learning-based system for predicting the probability of credit default of commercial entities, characterized in that, The system is used to execute the deep learning-based business entity credit default probability prediction method as described in any one of claims 1-9, the system comprising: The data processing module is used to acquire a multi-source heterogeneous data set of business entities, and perform field mapping, time alignment and validity period labeling processing to obtain a preprocessed multi-source heterogeneous data set. The network construction module is used to build a heterogeneous graph network containing business entity nodes and event nodes based on a preprocessed multi-source heterogeneous data set; The weighting module is used to perform multi-relation mapping weighting on the heterogeneous graph network to obtain a multi-relation weighted graph that only contains business entity nodes; The training module is used to train a basic default probability prediction model using deep learning technology, with the multi-relationship weighted graph as an index, wherein the basic default probability prediction model is used to output default probability values.