Root cause analysis methods, apparatus, equipment, media, and procedures for financial provisions
By collecting multi-source heterogeneous data and utilizing cross-modal processing models and dynamic factor correlation analysis models, structured feature factors are generated, which solves the problem of insufficient utilization of unstructured information in financial provision management and enables more comprehensive, accurate and timely root cause analysis.
Patent Information
- Application Number
- CN202611095205.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in financial provision management do not make sufficient use of unstructured information, making it difficult to fully cover the key factors affecting provisions. Furthermore, they have limited ability to characterize the nonlinear relationships, time lag effects, and cascading transmission between factors, resulting in large deviations in contribution measurement and insufficient traceability and timeliness in the analysis process.
Collect multi-source heterogeneous data, perform semantic parsing on unstructured data through a cross-modal processing model to generate structured feature factors, combine them with structured data, and model them based on a dynamic factor association analysis model to generate factor contribution.
It improves the comprehensiveness, accuracy, timeliness, and interpretability of root cause analysis of changes in financial provisions, provides a more reliable data foundation, and supports risk assessment and provision management.
Smart Images

Figure CN123117318A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of financial technology or other related fields, and in particular to a method, apparatus, equipment, medium and program product for financial provision root cause analysis. Background Technology
[0002] In the field of financial provision management, existing technologies typically employ manual analysis or statistical models based on structured data to analyze the driving factors of provision changes.
[0003] However, such schemes do not make sufficient use of unstructured information such as policy documents and industry reports, making it difficult to fully cover the key factors affecting provisions. At the same time, they have limited ability to characterize nonlinear relationships, time lag effects and cascading transmission between factors, resulting in large deviations in contribution measurement and insufficient traceability and timeliness of the analysis process.
[0004] Therefore, improving the comprehensiveness, accuracy, timeliness, and interpretability of root cause analysis of provision changes under multi-source heterogeneous data conditions has become a technical problem that needs to be solved. Summary of the Invention
[0005] This application provides a method, apparatus, equipment, medium, and program product for root cause analysis of financial provisions, which addresses the problems of incomplete identification of driving factors of provision changes, inaccurate contribution measurement, and insufficient traceability and timeliness of analysis in financial provision management scenarios, thereby realizing the systematic mining and contribution assessment of the root causes of provision changes.
[0006] Firstly, this application provides a root cause analysis method for financial provisions, including:
[0007] Collect multi-source heterogeneous data, which includes structured and unstructured data;
[0008] Semantic parsing of unstructured data is performed using a cross-modal processing model to generate structured feature factors;
[0009] By combining structured feature factors with structured data, a unified feature set is formed;
[0010] A unified feature set is modeled based on a dynamic factor association analysis model to generate factor contribution.
[0011] In one possible embodiment, unstructured data is semantically parsed using a cross-modal processing model to generate structured feature factors, including:
[0012] Extract key entities from unstructured data and generate entity features;
[0013] Sentiment analysis is performed on unstructured data to generate sentiment features; these sentiment features are a quantitative representation of the sentiment tendency of the text.
[0014] By combining entity features with emotional features, structured feature factors are generated.
[0015] In one possible embodiment, a unified feature set is modeled based on a dynamic factor association analysis model to generate factor contributions, including:
[0016] Causal chain analysis of factor contributions is performed using Bayesian networks to generate factor correlation graphs.
[0017] In one possible embodiment, after generating the factor association map, the process includes:
[0018] Multi-hop path modeling is performed on the factor association graph using graph neural networks to generate multi-hop causal chains.
[0019] In one possible embodiment, multi-source heterogeneous data is collected, including at least one of the following:
[0020] Obtain structured data from the internal systems of financial institutions through application programming interfaces;
[0021] Unstructured data is preprocessed in a distributed manner using edge computing nodes.
[0022] In one possible embodiment, before performing semantic parsing on unstructured data via a cross-modal processing model, the following steps are included:
[0023] The parameters of the cross-modal processing model are updated through an incremental learning mechanism.
[0024] In one possible embodiment, the dynamic factor association analysis model includes an embedding layer, a multi-head attention layer, a feature learning layer, a parameter iterative optimization layer, and a factor inference output layer;
[0025] Among them: the embedding layer is used to convert sample data into standardized feature vectors. The sample data includes structured impact factor data and preprocessed unstructured impact factor data.
[0026] A multi-head attention layer is used to explore the correlation between different influencing factors and historical provision change data, and to calculate the attention weight of each factor.
[0027] The feature learning layer is used to iteratively learn the nonlinear mapping and correlation features between different influencing factors and historical provision change data based on sample data.
[0028] The parameter iteration optimization layer is used to adjust the model parameters based on the backpropagation algorithm and the error threshold. The model parameters include the learning rate and layer dimension parameters.
[0029] The factor inference output layer is used to calculate the magnitude of change of different influencing factors based on the model weight parameters.
[0030] In one possible embodiment, the method further includes:
[0031] The factor correlation graph and factor contribution are interactively displayed through a visual dashboard.
[0032] Secondly, embodiments of this application provide a financial provision root cause analysis device, comprising:
[0033] The acquisition module is used to acquire multi-source heterogeneous data, which includes structured data and unstructured data.
[0034] The parsing module is used to perform semantic parsing on unstructured data through a cross-modal processing model to generate structured feature factors;
[0035] The forming module is used to combine structured feature factors with structured data to form a unified feature set;
[0036] The generation module is used to model a unified feature set based on a dynamic factor association analysis model and generate factor contributions.
[0037] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;
[0038] The memory stores computer-executed instructions;
[0039] The processor executes computer execution instructions stored in the memory, causing the processor to perform the method described in any of the first aspects above.
[0040] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in any of the first aspects above.
[0041] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the first aspects above.
[0042] This application provides a method, apparatus, equipment, medium, and program product for root cause analysis of financial provisions. By collecting multi-source heterogeneous data, including structured and unstructured data, and performing semantic parsing on the unstructured data to generate structured feature factors, the structured feature factors are then combined with the structured data to form a unified feature set. Furthermore, based on a dynamic factor correlation analysis model, the unified feature set is modeled to generate factor contribution. This approach can comprehensively utilize multi-dimensional information and characterize the correlation effects of various influencing factors within the same analytical framework, thereby improving the comprehensiveness, accuracy, timeliness, and interpretability of root cause analysis of changes in financial provisions. Attached Figure Description
[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0044] Figure 1 A flowchart illustrating a root cause analysis method for financial provisions provided in this application embodiment;
[0045] Figure 2 A schematic diagram of the architecture of a financial provision change root cause analysis system provided in this application embodiment;
[0046] Figure 3 A schematic diagram of the structure of a financial provision root cause analysis device provided in this application embodiment;
[0047] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0048] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0050] In this application, the term "comprising" and its variations can refer to non-limiting inclusion; the term "or" and its variations can refer to "and / or". The terms "first", "second", etc., in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0051] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.
[0052] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.
[0053] First, let's explain some of the terms used in this application:
[0054] Application Programming Interface (API): A set of predefined functions, protocols, and tools used for interaction and communication between different software applications.
[0055] Financial provision management involves financial data processing, risk identification, and operational decision support. It is commonly used in institutions such as banks, insurance companies, and securities firms to analyze the reasons for increases or decreases in provisions. This scenario typically relies on internal credit systems, financial systems, and external policy information sources to continuously acquire data such as loan balances, default rates, regulatory documents, and industry reports to determine the driving factors behind changes in provisions.
[0056] Most existing analyses of provision changes rely on manual analysis or statistical models based on structured data. The former typically involves accounting or risk control personnel summarizing business data, reviewing policy documents, and combining experience to make attributions; the latter often utilizes models such as regression analysis and decision trees to calculate the relationship between indicators such as loan balance, non-performing loan ratio, and default rate and provisions, in order to output the main influencing factors and their degree of impact.
[0057] However, existing methods for analyzing provision changes fail to adequately utilize unstructured information such as policy documents and industry reports, often struggling to transform implicit policy guidance and industry risk signals into calculable factors, resulting in incomplete analysis. Furthermore, traditional analytical methods lack effective characterization of nonlinear relationships between factors, time-lag effects, and multi-factor cascading transmission, easily leading to distorted contribution quantification and difficulty in accurately distinguishing direct from indirect impacts. This consequently affects the timeliness of provision accrual, the targeted nature of risk management, and the reliability of operational judgments.
[0058] In view of the aforementioned existing problems, this application proposes a root cause analysis method for financial provisions. When provisions change or cause investigation is required, multi-source heterogeneous data, including structured and unstructured data, are collected. Then, the unstructured data is semantically parsed using a cross-modal processing model to generate structured feature factors. The structured feature factors are combined with the structured data to form a unified feature set. The unified feature set is modeled based on a dynamic factor association analysis model to generate factor contribution.
[0059] By using the above methods to organize and analyze multiple types of data in a unified manner, the comprehensiveness, accuracy and timeliness of identifying the root causes of provision changes can be improved, providing a more reliable data foundation for subsequent risk assessment and provision management.
[0060] The entity executing this root cause analysis method for financial provisions can be, for example, an analysis system. Optionally, the analysis system can be any existing electronic device with processing capabilities, such as a terminal or a server. In some embodiments, the analysis system can also be deployed in a server cluster or cloud environment. This application does not limit the deployment environment of the analysis system.
[0061] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0062] Figure 1 This is a flowchart illustrating a root cause analysis method for financial provisions provided in an embodiment of this application. Figure 1As shown, the method includes:
[0063] S101 collects multi-source heterogeneous data, which includes structured and unstructured data.
[0064] For example, multi-source heterogeneous data can be a collection of data from different sources and in different formats.
[0065] For example, structured data can be numerical or tabular data sources from multi-source heterogeneous data, providing basic business indicators, such as loan balance, default rate, non-performing loan rate, industry loan ratio, provision balance, provision coverage ratio, migration rate, and period financial item data related to provision changes.
[0066] For example, unstructured data can serve as a source of textual data in multi-source heterogeneous data, supplementing implicit information such as policy guidance and industry risks. For instance, it can include regulatory policy documents, industry research reports, market risk warnings, regional economic analysis materials, and internal risk notification texts.
[0067] In one possible implementation, the analytics system can acquire structured data from the financial institution's internal systems via an application programming interface (API), and can also perform distributed preprocessing of unstructured data via edge computing nodes.
[0068] Optionally, the analysis system can combine structured data obtained through the application programming interface (API) with preprocessed results output from edge computing nodes to form multi-source heterogeneous data, which can be used for subsequent semantic parsing, feature construction, and factor association analysis. Since the structured data has been standardized and extracted through the interface, and the unstructured data has undergone preliminary screening and compression at the edge, the granularity and format of the data entering the central model are consistent.
[0069] By adopting the above approach, heterogeneous data can be aggregated under the condition that the internal systems of financial institutions and external text sources are accessed in parallel, and some unstructured processing is moved to the edge computing nodes, thereby reducing the instantaneous computing pressure and improving the timeliness of data, providing a complete data foundation for subsequent provision root cause identification.
[0070] Optionally, the analysis system can deploy dedicated computing nodes for analysis on the financial institution's data platform or risk management server, and establish data access links with the internal credit system, financial system, risk management system and external information platform respectively.
[0071] Optionally, the analytics system can acquire relevant business data from internal systems and collect structured data. For example, the analytics system can extract loan balances, overdue payments, and default status from the credit system based on customer, product, and industry dimensions; extract provision accrual results, profit and loss impact amounts, and accounting period identifiers from the financial system; and extract rating migration, warning levels, and default probability parameters from the risk management system.
[0072] Optionally, the analysis system can obtain relevant textual materials from external information sources or internal document sources. For example, the analysis system can aggregate regulatory documents by publication date, regulatory agency, and scope of application, and industry reports by industry classification, regional classification, and reporting period.
[0073] Optionally, the analysis system can uniformly register data from different sources after collection, forming a raw data index table that includes data source identifier, collection time, business period, applicable institution, data format, and subject tags, and establish the correspondence between structured and unstructured data with the provision analysis period as the main time axis.
[0074] S102 uses a cross-modal processing model to perform semantic parsing on unstructured data and generate structured feature factors.
[0075] For example, a cross-modal processing model can be an algorithmic model that processes both structured and unstructured data simultaneously, using techniques such as semantic parsing, entity recognition, and sentiment analysis to transform unstructured data into structured feature factors.
[0076] For example, structured feature factors can be semantic features extracted and quantified from unstructured data, combined with structured data to form a unified feature set, which can be transformed from the results of text semantic parsing.
[0077] Optionally, the analysis system can preprocess unstructured text data, including removing irrelevant content, text segmentation, time expression normalization, and name standardization, and then input the preprocessed text into the cross-modal processing model.
[0078] Optionally, the cross-modal processing model can perform semantic understanding of the text content, extract semantic information related to provision changes, and encode the semantic parsing results into computable structured feature factors. For longer policy documents and industry reports, the cross-modal processing model can perform semantic processing on each text segment before forming document-level features. Then, the semantic parsing results and time, region, and industry identifiers from the document metadata are encoded to obtain structured feature factors. For example, encoding methods may include one-hot encoding, label encoding, dimensionality reduction after vector representation, continuous value scoring, and rule mapping.
[0079] Optionally, for texts with ambiguous contexts or numerous industry terms, the analysis system can perform contextual semantic understanding using a pre-trained model, then convert it into computable numerical and categorical fields. By transforming textual information into computable structured feature factors, policy changes, industry risks, and regulatory guidance can be incorporated into the provisioning analysis process, thus supplementing root cause information that is difficult to cover when relying solely on tabular data. It should be understood that the above examples are merely illustrative and not limiting.
[0080] In one possible implementation, the analysis system can extract key entities from unstructured data, generate entity features, perform sentiment analysis on the unstructured data to generate sentiment features, which are quantitative representations of the text's sentiment tendency. Then, the entity features and sentiment features are combined to generate structured feature factors.
[0081] For example, key entities can be named entities with business significance in unstructured data, such as policy names, industry categories, regulatory indicators, etc.
[0082] For example, entity features can be a quantitative representation of key entities and their contextual semantic information that are not extracted through entity recognition.
[0083] For example, sentiment features can be a quantitative representation of the sentiment tendency of a text obtained through sentiment analysis.
[0084] Optionally, the analysis system can receive the unstructured text to be analyzed through a cross-modal processing model, and perform word segmentation, entity recognition, and semantic encoding on the text to obtain candidate key entities and their contextual representations. For unstructured data containing charts or scanned documents, the cross-modal processing model converts visual information into processable text through layout parsing and text extraction, and then uniformly enters the semantic parsing chain.
[0085] Optionally, the analysis system can perform sentiment analysis based on the output of sentence-level or paragraph-level sentiment scores from the pre-trained language model, and map positive, neutral, and negative results to preset numerical ranges to form calculable sentiment features.
[0086] Optionally, after obtaining entity features and sentiment features, the analysis system can use a cross-modal processing model to concatenate, weightedly fuse, or attentionally fuse them within the same feature space to generate structured feature factors. These factors are then written into a unified data structure for subsequent provision attribution modeling. During operation, the structured feature factors collectively represent policy names, regulatory orientations, risk statements, and their sentiment tendencies as computable variables, thereby enabling unstructured information to participate in unified modeling.
[0087] Using the above methods, business entities and their emotional orientation in the text can be quantified simultaneously, the expression of structured feature factors is more complete, and it is more suitable for joint analysis with structured financial indicators, thereby improving the accuracy and traceability of identifying the root causes of provision changes.
[0088] In one possible implementation, the analysis system can update the parameters of the cross-modal processing model through an incremental learning mechanism before performing semantic parsing on the unstructured data via the cross-modal processing model.
[0089] Optionally, when updating parameters, the analysis system can input unstructured data samples from the most recent analysis period and their corresponding manual review labels into the model training environment. Then, based on the distribution differences between the samples and historical samples, it can determine the range of parameters that need to be updated and adjust the sample weights so that newly added policy texts, industry report texts, or risk warning texts have a higher response intensity during training.
[0090] Optionally, the analysis system can use the feedback results as a monitoring signal to perform online or small-batch incremental updates on the local parameters in the cross-modal processing model, so that the cross-modal processing model can enhance its ability to identify specific terms, implicit sentiments and policy orientations in the current financial provisioning scenario while maintaining its original general semantic representation capabilities.
[0091] The above method can correct existing parameter configurations, thus ensuring that the semantic parsing results are consistent with the current allocation analysis task. Since only local parameters are updated, the model update process consumes fewer computational resources compared to retraining the entire model, resulting in faster parameter convergence and maintaining high parsing stability and timeliness in allocation management scenarios.
[0092] S103 combines structured feature factors with structured data to form a unified feature set.
[0093] For example, a unified feature set can be a dataset that integrates structured data with structured feature factors, which is used as input to a dynamic factor association analysis model. Its formation process includes merging, alignment, standardization, encoding, and sample organization.
[0094] Optionally, the analysis system can determine the sample granularity based on the provision analysis target. The sample granularity can be institution-month, institution-industry-month, product-month, or region-month. After the granularity is determined, the structured data is summarized according to the same granularity to obtain a structured sample table containing the target provision change value and business indicators.
[0095] Optionally, the analysis system can incorporate structured feature factors into a structured sample table according to preset field mapping relationships. The mapping relationships include at least time mapping, industry mapping, regional mapping, and institutional applicability scope mapping.
[0096] For example, time mapping can map the text publication date, effective date, or reporting period to the provision analysis period. For text factors with time lag effects, a lag window can be set to continuously apply a text factor to the sample in one or more subsequent analysis periods.
[0097] For example, industry mapping could be mapping the industries covered by industry reports or policies to industry fields in a loan portfolio.
[0098] For example, regional mapping could be mapping regional policies or regional risk reports to regional credit samples.
[0099] For example, the scope of application mapping of an institution can be a distinction between national policies, regional policies, and internal policy texts that apply only to specific business lines.
[0100] Optionally, after mapping is completed, the analysis system can perform unified processing on data from different sources, at different scales, and at different time granularities. Specifically, it performs missing value imputation, outlier truncation, and standardization transformation on continuous structured data, encodes categorical text factors, and performs numerical processing on Boolean or directional factors. It also retains the source identifier, generation time, and mapping rule number of each factor to support traceability.
[0101] Optionally, for multiple texts corresponding to the same industry or region within the same analysis period, the analysis system can aggregate them based on the proximity of publication time, text authority level, and topic relevance to generate period-level text feature values. For example, the relevant text features parsed from multiple regulatory documents concerning the real estate industry within a certain month can be weighted and summarized, then concatenated with structured data such as the outstanding balance of real estate loans, non-performing loan ratio, and historical provision amounts for that month to form a unified sample vector for that industry in that month.
[0102] Optionally, the fields in the unified feature set may include at least the sample primary key, business period, provision change target value, structured business indicators, text parsing factors, factor source labels, and time lag labels, so that the subsequent model can identify both the direct impact of business indicators on provisions and the indirect impact of policy or industry risk factors on provisions after time transmission.
[0103] S104, a unified feature set is modeled based on a dynamic factor correlation analysis model to generate factor contribution.
[0104] For example, a dynamic factor correlation analysis model can be a model that learns nonlinear relationships, time lag effects, and multi-factor linkages between factors through a deep neural network structure. By modeling a unified feature set, factor contribution values are generated, and key driving factors affecting provision changes are identified.
[0105] For example, factor contribution can quantify the degree of influence of each influencing factor on provision changes, serving as the basis for subsequent root cause ranking and causal chain analysis. Factor contribution can be represented as a directional numerical value, with positive values indicating an increase in provisions and negative values indicating a decrease in provisions, and the absolute value indicating the strength of the influence.
[0106] Optionally, the analysis system can construct training samples based on a unified feature set for historical periods, where the input is a unified sample vector for each analysis period, and the output is the provision change value, provision increase amount, provision decrease amount, or provision change rate for the corresponding period.
[0107] In one possible implementation, the analysis system can use a Bayesian network to perform causal chain analysis on factor contributions and generate a factor correlation graph.
[0108] For example, a Bayesian network can be formed by performing structure learning and parameter learning based on historical provision change samples and corresponding feature data, thereby creating a graphical model that reflects the probabilistic dependencies between factors. For instance, structure learning can use score search or constraint testing to determine the candidate edge set, while parameter learning quantifies the node relationships based on sample frequency, conditional probability distribution, or posterior probability.
[0109] For example, a factor association map can be a visual representation of the relationships between factors generated through causal chain analysis.
[0110] The optional factor contribution, after entering the Bayesian network, is mapped to the weight, probability increment, or influence intensity of the corresponding node, thereby identifying the direct action path and the indirect transmission path.
[0111] Optionally, the analysis system can generate a factor association graph based on the Bayesian network analysis results. The nodes and edges in the graph represent factors and their associations, and can be visualized and marked according to the factor contribution, path length, or transmission strength.
[0112] Optionally, for financial provision management scenarios, the graph can express a chain relationship such as "tightening policies increase financing difficulty, rising default rate, and increased provisions" as a continuous causal chain, thereby further transforming the quantitative results output by the dynamic factor association analysis model into a traceable relationship structure.
[0113] Optionally, the analysis system can implement the Bayesian network using a discrete node model or a continuous node model. The discrete node model is suitable for dividing factors into interval levels, while the continuous node model is suitable for preserving fine-grained information about the changes in factor values. In practical applications, the network can also be selected from other model forms, which are not limited in this application.
[0114] In this way, factor contribution can not only reflect the magnitude of the impact of a single factor on provision changes, but also clarify the transmission order and correlation path between different factors through causal chain analysis, making the root cause analysis results of provision changes more interpretable and traceable. The resulting factor correlation map can be directly used to help identify key driving factors and their cascading effects, providing a structured basis for provision accrual judgment and risk management decisions.
[0115] Based on the aforementioned embodiments, after generating the factor association graph, the analysis system can use a graph neural network to model multi-hop paths in the factor association graph and generate multi-hop causal chains.
[0116] Optionally, the analysis system can initialize the features of each node in the factor association graph using a graph neural network, and perform aggregation calculations based on the connections between nodes, passing the state information of adjacent nodes to the current node. Then, based on the graph neural network, the propagation range is further expanded outward along the graph structure, enabling nodes to fuse factor information from multiple neighborhoods, thereby forming encoded representations of two-hop, three-hop, and even longer propagation chains.
[0117] Optionally, the analysis system can introduce attention weights during message aggregation, assigning differentiated factor contribution coefficients to different intermediate factors, so that paths more strongly associated with provision changes receive higher representation weights, while weakly associated paths are suppressed.
[0118] Optionally, the graph neural network used in the analysis system can be a graph convolutional network, a graph attention network, or a combination of both. In practical applications, other models of this component can also be selected, and this application does not limit this.
[0119] Optionally, the analysis system can reconstruct the paths of initial factors, transitional factors, and target factors, and connect them in causal direction to form a multi-hop causal chain. This causal chain can reflect the process by which intermediate factors such as policy changes, increased industry risk, and default rate fluctuations are transmitted to provision changes through multiple levels.
[0120] This approach can elevate local correlations in the graph into traceable long-chain causal expressions, preserving and revealing the indirect transmission paths of provision changes, thereby improving the completeness and traceability of causal analysis and providing a more stable structured basis for subsequent factor contribution attribution and risk assessment.
[0121] In one possible implementation, the analysis system can also interactively display factor correlation maps and factor contributions through a visual dashboard.
[0122] Optionally, the analysis system can convert the Bayesian causal chains output by the dynamic factor association analysis model and the multi-hop causal chains generated by the graph neural network into visual coordinates, node styles, and edge connection parameters for display.
[0123] Optionally, the analysis system can generate factor contribution data based on the variation range or weight value of each factor, and synchronously map it to column height, color depth, or labeled value according to the normalization result.
[0124] For example, if a user clicks on a factor node in the dashboard, the analysis system can highlight the factor and its upstream and downstream related paths, while displaying the factor's historical factor contribution curve, the corresponding text factor source, and the associated provision change range; if the user selects a specific time period, business line, or risk category through filtering conditions, the analysis system can refresh the graph nodes and factor contribution charts based on the dashboard to display the analysis results within the corresponding range.
[0125] Optionally, the analysis system can also be configured with a dashboard cache update mechanism. When a new allocation data is received, the analysis system can redraw the graph and update the factor contribution chart simultaneously, thus forming a continuously iterative interactive analysis interface.
[0126] In this way, factor correlation maps and factor contributions are presented in the same visualization interface, which can output complex causal relationships, direct and indirect effects simultaneously, and support historical tracing and correlation positioning of individual factors, making it easier to verify, compare and apply the results of the provision root cause analysis to decision-making.
[0127] In this implementation, multi-source heterogeneous data was collected, including both structured and unstructured data, providing a data basis for subsequent provision root cause analysis. A cross-modal processing model was used to perform semantic parsing on the unstructured data, generating structured feature factors. This enabled the integration of internal business indicators with external and internal textual information into a single analytical framework. Combining structured feature factors with structured data to form a unified feature set allows textual semantic results to be converted into structured features for dynamic modeling, enabling provision change analysis to cover policy guidance, industry risks, business indicators, and their time-lag transmission relationships. Based on a dynamic factor correlation analysis model, the unified feature set was modeled to generate factor contribution rates, resulting in traceable provision root cause analysis results, improving the accuracy and adaptability of root cause analysis. It should be understood that the above example is for demonstration purposes only and is not limiting.
[0128] The dynamic factor association analysis model will be explained in detail below.
[0129] As one possible implementation, the dynamic factor association analysis model structure can adopt a neural network architecture that includes a time-series encoding layer, a feature interaction layer, and an output interpretation layer. The time-series encoding layer is used to receive a unified feature set of the current period and several historical periods and learn the time dependency relationship. The feature interaction layer is used to learn the nonlinear interaction between structured business indicators and text feature factors. The output interpretation layer is used to output the provision change prediction results and the factor contribution of each input factor.
[0130] For example, the temporal coding layer can employ a recurrent neural network, a long short-term memory network, a gated recurrent unit, or a temporal attention network, while the feature interaction layer can employ a multilayer perceptron, a cross network, or an attention mechanism module.
[0131] Optionally, during training, the analysis system can use the error between the predicted provision change value and the actual provision change value as the loss function to update the parameters of the dynamic factor correlation analysis model. The error can be in the form of mean squared error, mean absolute error or a combination thereof. The parameters are iteratively optimized through the backpropagation algorithm until convergence.
[0132] Optionally, during the inference phase, the analysis system can input the unified feature set of the current time point into the trained dynamic factor correlation analysis model. When the dynamic factor correlation analysis model outputs the predicted value of provision changes, the analysis system can calculate the influence degree of each factor based on attention weights, gradient attribution, and perturbation sensitivity, thus obtaining the factor contribution. If there are factors with time-delay propagation, the dynamic factor correlation analysis model can also simultaneously output the current period factor contribution and the lag period cumulative factor contribution to distinguish between immediate and delayed effects.
[0133] For example, during a certain analysis period, if the outstanding loan balance in the real estate industry increases, the non-performing loan ratio rises, and negative risk signals are obtained from the analysis of regulatory texts, the dynamic factor correlation analysis model can quantify and output the factor contributions of relevant industry risk factors, default rate factors, and text risk factors, and form a list of root causes of provision changes by sorting them according to absolute values.
[0134] As another possible implementation, the dynamic factor association analysis model includes an embedding layer, a multi-head attention layer, a feature learning layer, a parameter iterative optimization layer, and a factor inference output layer.
[0135] The embedding layer is used to convert sample data into standardized feature vectors. The sample data includes structured impact factor data and preprocessed unstructured impact factor data.
[0136] A multi-head attention layer is used to explore the correlation between different influencing factors and historical provision change data, and to calculate the attention weight of each factor.
[0137] The feature learning layer is used to iteratively learn the nonlinear mapping and correlation features between different influencing factors and historical provision change data based on sample data.
[0138] The parameter iteration optimization layer is used to adjust the model parameters based on the backpropagation algorithm and the error threshold. The model parameters include the learning rate and layer dimension parameters.
[0139] The factor inference output layer is used to calculate the magnitude of change of different influencing factors based on the model weight parameters.
[0140] Optionally, the embedding layer can adopt a fully connected mapping structure to implement vectorized encoding, and the multi-head attention layer can be composed of multiple parallel attention heads, with the number of heads matching the input dimension, which facilitates the improvement of the subspace representation capability of different factors. In practical applications, other models of this component can also be selected, and this application does not limit them.
[0141] Optionally, the feature learning layer can be composed of a multilayer perceptron or a residual neural network to enhance the fitting ability to nonlinear relationships. The parameter iterative optimization layer can use gradient descent with momentum to perform parameter updates and stop iterating when the error continuously falls below a preset threshold. The factor inference output layer jointly calculates the variation magnitude of different influencing factors by combining the trained weight parameters with the attention weights, and outputs the results for ranking and interpretation.
[0142] By adopting this implementation method, collaborative modeling of structured and preprocessed unstructured factors is achieved through unified embedding, attention association, nonlinear learning, and iterative optimization. Multi-source influencing factors are converted into a unified feature representation, and key driving relationships in historical provision changes are quantitatively identified, resulting in complete factor contribution results. Due to the introduction of multi-head attention and nonlinear feature learning, the dynamic factor association analysis model more fully represents complex linkages and time-lag effects, and the quantitative results of factor contributions are more consistent and interpretable.
[0143] Figure 2 This is a schematic diagram of the architecture of a financial provision change root cause analysis system provided in an embodiment of this application. Figure 2 As shown, the root cause analysis system for changes in financial provisions includes:
[0144] (1) Data layer: As the basic support, it is responsible for data collection and preprocessing, and transmits the processed data to the model layer.
[0145] Optionally, the data layer can enable comprehensive acquisition and standardized processing of multi-source heterogeneous data. The data layer includes a data acquisition interface and a data preprocessing unit.
[0146] Optionally, the data acquisition interface can connect to the core business systems (such as credit management systems and risk management systems) and financial systems (such as provision accrual ledgers and profit and loss statements) within financial institutions via Application Programming Interfaces (APIs), as well as external data sources (such as macroeconomic databases, industry report libraries of industry associations, and policy release platforms of regulatory agencies) to capture structured and unstructured data in real time or periodically. Structured data can include, for example, monthly loan balances, quarterly default rates, and Gross Domestic Product (GDP) growth rates, while unstructured data can include, for example, policy documents in Portable Document Format (PDF), industry analysis Word documents, and customer credit rating texts.
[0147] Optionally, the data preprocessing unit can clean and transform different types of data. For structured data, this includes deduplication, filling in missing values, and outlier detection (e.g., based on...). To improve data quality, operations such as identifying values that exceed reasonable ranges are performed. For unstructured data, the natural language processing capabilities of large models are utilized for processing. For example, key information such as "policy name", "regulatory indicator" and "industry name" are extracted through entity recognition. The direction of the policy document's impact on the provision is determined by text classification (positive / negative). The processed unstructured data is then transformed into structured feature factors containing weight values to ensure that the data can be directly called by the model layer.
[0148] (2) Model layer: Based on the information provided by the data layer, the model is trained and reasoned, and the analysis results are transmitted to the analysis layer.
[0149] Optionally, the model layer can be a provisioning factor analysis model fine-tuned from a pre-trained large model (such as a large model specifically for the financial field). In the actual analysis process, after receiving real-time data transmitted from the data layer, the model layer can use the inference capabilities formed through pre-training to calculate the magnitude of changes in each factor, and based on the learned correlations, quantify the contribution of each factor to the provisioning change (e.g., factor A causes a provisioning increase of X yuan, and factor B causes a provisioning decrease of Y yuan), and identify the top N (e.g., N=5) key factors in terms of contribution through real-time inference.
[0150] For example, this application also provides a method for constructing a provision factor analysis model, which includes three steps: training data preparation, model fine-tuning, and validation.
[0151] Step 1: Training data preparation stage. Collect provision change data and corresponding influencing factor data of financial institutions over the past 5-10 years (including structured and preprocessed unstructured data), and divide them into training set, validation set and test set according to time series.
[0152] The second step is the model fine-tuning stage. The factor data in the training set is used as input and the provision change value is used as output. The parameters of the large model are adjusted through the backpropagation algorithm so that the model learns the mapping relationship between factors and provision changes. At the same time, an attention mechanism is introduced so that the model can automatically focus on factors that have a significant impact on provision changes (such as default rate changes in specific industries, major policy adjustments, etc.).
[0153] The third step is the validation phase. The validation set data is used to evaluate the analytical accuracy of the model. If the error exceeds the preset threshold (e.g., mean absolute error rate > 5%), the model is optimized by increasing the amount of training data or adjusting the model hyperparameters (e.g., learning rate, hidden layer dimension) until the model performance meets the target.
[0154] (3) Analysis layer: Deeply analyze and integrate the model output results to form an interpretable root cause reasoning path.
[0155] Optionally, the analysis layer includes a factor association resolution unit and a root cause ranking unit. The factor association resolution unit, based on the factor contribution data output by the large model, identifies the cascading influence relationships between factors through causal inference algorithms (such as causal discovery based on Bayesian networks) and generates factor transmission paths. For example, a factor transmission path could be "macroeconomic downturn, increased default rate in industry A, increased loan loss provisions in that industry, and overall increase in loan loss provisions," generating a visualized factor association graph.
[0156] Optionally, the root cause ranking unit can combine the magnitude of contribution and the length of the impact path to prioritize key factors, distinguish between direct root causes (such as a sudden increase in the default rate of a certain type of loan) and indirect root causes (such as macroeconomic policy adjustments that lead to an increase in the default rate), and generate a root cause priority list to ensure that accounting personnel can quickly focus on the core influencing factors.
[0157] (4) Application layer: Presents a visual root cause analysis report to accounting personnel.
[0158] Optionally, the application layer, serving as the user interface, can present the results of the analysis layer to accounting personnel in an intuitive and easy-to-understand manner.
[0159] Optionally, the application layer offers diverse display functions: First, an automatic root cause analysis report generation module that can output text reports containing a list of key factors, contribution percentages, and descriptions of impact paths; second, a visualization dashboard that uses bar charts to compare the contributions of each factor, line charts to show the time correlation between factor changes and provision changes, and network diagrams to demonstrate the transmission relationships between factors; and third, an interactive query function that allows accounting personnel to click on a factor to view its historical change data, relevant policy basis, and the specific calculation logic of the model's impact on it, enhancing the traceability and credibility of the analysis results.
[0160] The above are the method embodiments provided in this application. The apparatus provided in this application will be described below.
[0161] Figure 3 This is a schematic diagram of the structure of a financial provisioning root cause analysis device provided in an embodiment of this application, as shown below. Figure 3 As shown, the financial provision root cause analysis device 300 includes: a data acquisition module 301, a data parsing module 302, a data formation module 303, and a data generation module 304. Among them,
[0162] The acquisition module 301 is used to acquire multi-source heterogeneous data, which includes structured data and unstructured data.
[0163] The parsing module 302 is used to perform semantic parsing on the unstructured data through a cross-modal processing model to generate structured feature factors;
[0164] The forming module 303 is used to combine the structured feature factors with the structured data to form a unified feature set;
[0165] The generation module 304 is used to model the unified feature set based on the dynamic factor association analysis model and generate factor contribution.
[0166] Optionally, the acquisition module 301 is also used to acquire structured data from the internal system of the financial institution through an application programming interface;
[0167] Unstructured data is preprocessed in a distributed manner using edge computing nodes.
[0168] Optionally, the parsing module 302 is also used to extract key entities from the unstructured data and generate entity features;
[0169] Sentiment analysis is performed on the unstructured data to generate sentiment features; these sentiment features are a quantitative representation of the sentiment tendency of the text.
[0170] The entity feature is combined with the sentiment feature to generate a structured feature factor.
[0171] Optionally, the generation module 304 is also used to perform causal chain analysis on the factor contribution through a Bayesian network to generate a factor association graph.
[0172] Optionally, the generation module 304 is also used to generate a multi-hop path model of the factor association graph after generating the factor association graph, and generate a multi-hop causal chain by using a graph neural network.
[0173] Optionally, the parsing module 302 is also used to update the parameters of the cross-modal processing model through an incremental learning mechanism before performing semantic parsing on the unstructured data through the cross-modal processing model.
[0174] Optionally, the generation module 304 is also used to include an embedding layer, a multi-head attention layer, a feature learning layer, a parameter iterative optimization layer, and a factor inference output layer in the dynamic factor association analysis model.
[0175] Wherein: the embedding layer is used to convert sample data into standardized feature vectors, the sample data including structured impact factor data and preprocessed unstructured impact factor data;
[0176] This multi-head attention layer is used to explore the correlation between different influencing factors and historical provision change data, and to calculate the attention weight of each factor;
[0177] This feature learning layer is used to iteratively learn the nonlinear mapping correlation features between the different influencing factors and the historical provision change data based on the sample data.
[0178] This parameter iterative optimization layer is used to adjust the model parameters based on the backpropagation algorithm and the error threshold. These model parameters include the learning rate and layer dimension parameters.
[0179] This factor inference output layer is used to calculate the magnitude of change of the different influencing factors based on the model weight parameters.
[0180] Optionally, the generation module 304 is also used to interactively display the correlation graph of the factor and the contribution of the factor through a visualization dashboard.
[0181] The financial provision root cause analysis device provided in this embodiment can execute the methods provided in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0182] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4As shown, the electronic device 400 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 400 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.
[0183] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.
[0184] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0185] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0186] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0187] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0188] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0189] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.
[0190] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.
[0191] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.
[0192] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.
[0193] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.
[0194] When an integrated unit / module is implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.
[0195] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.
[0196] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.
[0197] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0198] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A financial provision root cause analysis method, characterized by, include: Collect multi-source heterogeneous data, including structured data and unstructured data; The unstructured data is semantically parsed using a cross-modal processing model to generate structured feature factors; The structured feature factors are combined with the structured data to form a unified feature set; The unified feature set is modeled based on a dynamic factor association analysis model to generate factor contribution.
2. The method of claim 1, wherein, The step of semantically parsing the unstructured data using a cross-modal processing model to generate structured feature factors includes: Extract key entities from the unstructured data and generate entity features; Sentiment analysis is performed on the unstructured data to generate sentiment features; the sentiment features are a quantitative representation of the sentiment tendency of the text. The entity features are combined with the emotional features to generate structured feature factors.
3. The method of claim 1, wherein, The process of modeling the unified feature set based on the dynamic factor association analysis model to generate factor contribution values includes: The causal chain analysis of the factor contributions is performed using a Bayesian network to generate a factor association graph.
4. The method of claim 3, wherein, Following the generating factor association map, the following is included: Multi-hop path modeling is performed on the factor association graph using a graph neural network to generate multi-hop causal chains.
5. The method of claim 1, wherein, The collection of multi-source heterogeneous data includes at least one of the following: Obtain structured data from the internal systems of financial institutions through application programming interfaces; Unstructured data is preprocessed in a distributed manner using edge computing nodes.
6. The method of claim 1, wherein, Before performing semantic parsing on the unstructured data using a cross-modal processing model, the following steps are included: The parameters of the cross-modal processing model are updated through an incremental learning mechanism.
7. The method of claim 1, wherein, The dynamic factor association analysis model includes an embedding layer, a multi-head attention layer, a feature learning layer, a parameter iterative optimization layer, and a factor inference output layer. Wherein: the embedding layer is used to convert sample data into standardized feature vectors, the sample data including structured impact factor data and preprocessed unstructured impact factor data; The multi-head attention layer is used to mine the correlation between different influencing factors and historical provision change data, and to calculate the attention weight of each factor. The feature learning layer is used to iteratively learn the nonlinear mapping correlation features between the different influencing factors and the historical provision change data based on the sample data. The parameter iterative optimization layer is used to adjust the model parameters according to the backpropagation algorithm and the error threshold. The model parameters include the learning rate and the layer dimension parameters. The factor inference output layer is used to calculate the variation magnitude of the different influencing factors based on the model weight parameters.
8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: The factor correlation graph and the contribution of the factors are interactively displayed through a visual dashboard.
9. A financial provision root cause analysis apparatus characterized by comprising: include: The acquisition module is used to acquire multi-source heterogeneous data, which includes structured data and unstructured data. The parsing module is used to perform semantic parsing on the unstructured data through a cross-modal processing model to generate structured feature factors; A forming module is used to combine the structured feature factors with the structured data to form a unified feature set; The generating module is configured to model the unified feature set based on a dynamic factor correlation analysis model to generate a factor contribution degree.
10. An electronic device, comprising: The method comprises: a processor, and a memory connected to the processor in communication; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 8.
11. A computer readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by the processor to implement the method according to any one of claims 1 to 8.
12. A computer program product, characterised in that, The computer program is executed by the processor to implement the method according to any one of claims 1 to 8.
