CN-121981842-A - Financial risk early warning method and system based on social media information

CN121981842ACN 121981842 ACN121981842 ACN 121981842ACN-121981842-A

Abstract

The invention discloses a financial risk early warning method and a financial risk early warning system based on social media information, and relates to the technical field of enterprise risk management. The method comprises the steps of collecting multi-source heterogeneous data from a plurality of social media platforms, extracting and fusing features through a pre-training language model and a graph neural network, constructing an enterprise, event and user ternary dynamic causal graph based on the fused features, revealing a conducting path and an amplifying mechanism of a risk signal by adopting causal inference and anti-facts analysis, training an early warning model with dynamic adaptability based on a meta-learning and multi-task learning framework, realizing risk class classification, type identification and outburst time prediction, carrying out risk tracing based on the causal graph, and automatically generating an interpretable early warning report with a complete evidence chain. The method effectively solves the problems of insufficient interpretability of the early warning result and decision support value caused by lack of causal association analysis in the prior art, and improves the intelligent level and decision efficiency of enterprise risk management.

Inventors

XU FENG
CAO SHIYAN
ZHU PENGCHENG

Assignees

南京财经大学

Dates

Publication Date: 20260505
Application Date: 20260225

Claims (10)

1. The financial risk early warning method based on social media information is characterized by comprising the following steps: acquiring multi-source social media data of a target enterprise, preprocessing the multi-source social media data, and constructing a multi-mode data lake with aligned time sequences; Based on the multi-mode data lake, extracting text semantic features, information propagation network features and user authority features to generate a joint feature matrix; Constructing a ternary causal map comprising enterprises, events and users based on the joint feature matrix; generating a causal graph network by performing causal inference and anti-facts analysis on the ternary causal graph; Inputting a causal graph network, a joint feature matrix and a ternary causal graph into a pre-trained self-adaptive early warning model to obtain a predicted risk level, a predicted risk category and a predicted explosion time; based on the predictions of risk level, risk category, and burst time, interpretable early warning reports are generated.
2. The financial risk early warning method based on social media information according to claim 1, wherein the obtaining method of the adaptive early warning model comprises the following steps: The method comprises the steps of obtaining k groups of risk data sets, wherein the risk data sets comprise a joint feature matrix, a ternary causal map, a causal map network, corresponding risk levels, risk categories and explosion time, taking the risk data sets as sample sets, dividing the sample sets into training sets and test sets, constructing a multi-task classification and regression device, taking the joint feature matrix, the ternary causal map and the causal map network in the training sets as inputs, taking the risk levels, the risk categories and the explosion time corresponding to the training sets as outputs, carrying out iterative training on the multi-task classification and regression device, obtaining a preliminary early warning model, testing the preliminary early warning model by utilizing the test sets, taking the preliminary early warning model as an adaptive early warning model if a prediction result of the preliminary early warning model meets preset accuracy, wherein the preliminary early warning model is a graph neural network deep learning model, and k is a positive integer greater than 1.
3. The financial risk early warning method based on social media information according to claim 1, wherein the extraction method of the authority characteristics of the user comprises the following steps: based on the multi-modal data lake, extracting user interaction data, information propagation data and user portraits of a target enterprise, and constructing a directed weighted graph of user interaction and propagation; The authority values of all nodes in the directed weighted graph are calculated, and indexes of degree centrality, medium centrality and near centrality are calculated; extracting identity authentication labels, historical contribution quality and field correlation metadata characteristics from the user portrait; And (3) after the authority value, the degree centrality, the medium centrality, the near centrality index and the domain correlation metadata feature are subjected to standardized processing, splicing the metadata features into a multidimensional feature vector to serve as user authority features.
4. The financial risk early warning method based on social media information according to claim 1, wherein the text semantic feature extraction method comprises: extracting original text data of a target enterprise from the multi-modal data lake, and cleaning, word segmentation and stop word filtering the original text data to obtain a normalized text; semantic parsing and quantization of the normalized text includes: Calculating emotion tendency scores and emotion intensity values of normalized texts based on preset financial emotion rules; based on statistical distribution of the normalized text on the vocabulary and document layers, extracting topic distribution vectors of the normalized text on preset financial risk dimensions; identifying and counting the frequency of occurrence of risk keywords and financial entity names in the normalized text according to a preset financial risk keyword list; Carrying out standardized processing on the frequency, emotion tendency score, emotion intensity value and theme distribution vector of the risk keywords and the financial entity names, and fusing the risk keywords and the theme distribution vector into semantic feature vectors of standardized texts according to weights; And in a predefined time window, carrying out aggregation calculation on the semantic feature vectors of the normalized text to generate text semantic features.
5. The financial risk early warning method based on social media information according to claim 1, wherein the extraction method of the information propagation network characteristics comprises the following steps: Based on the multi-mode data lake, constructing a directional propagation network which takes a user as a node and takes an information propagation path as an edge; Calculating topological structure characteristics of the directional propagation network, wherein the topological structure characteristics comprise the total number of nodes, the total number of edges, the network density, the average clustering coefficient and the maximum connected subgraph scale of the network; Calculating propagation dynamic characteristics of a directional propagation network, wherein the propagation dynamic characteristics comprise maximum depth of information propagation, average path length, node degree distribution on a key propagation path, and cascade diffusion scale and diffusion rate of information in a preset time window; and (3) carrying out standardization processing on the topological structure features and the propagation dynamic features, and carrying out dimension splicing and fusion to generate information propagation network features.
6. The financial risk early warning method based on social media information according to claim 1, wherein the step of generating the joint feature matrix is: After aligning the extracted text semantic features, information propagation network features and user authority feature time sequences, carrying out feature scaling and normalization processing to eliminate dimension differences; And splicing and fusing the text semantic features, the information propagation network features and the user authority features subjected to the feature scaling and normalization processing on feature dimensions to generate a joint feature matrix.
7. The financial risk early warning method based on social media information according to claim 1, wherein the step of constructing a ternary causal map of enterprises, events and users is: defining a business entity, a risk event entity and a user entity as three core nodes; Based on the joint feature matrix, defining and extracting time sequence association, semantic association and influence association among three core nodes, and constructing multi-type side relations of user-release-event, event-related-enterprise and user-influence-user; And integrating and correlating the three core nodes with the multi-type side relationship to form a ternary causal map.
8. The social media information-based financial risk early warning method according to claim 1, wherein the step of generating a causal graph network is: Carrying out causal intensity calculation on the multi-type side relation in the ternary causal map to obtain causal intensity weight of each side; Combining the ternary causal graph and the causal intensity weight, and performing inverse fact analysis on key nodes and edges of the ternary causal graph to obtain an inverse fact deduction conclusion; Ternary causal graph with causal intensity weights and negative fact deduction conclusions is defined as a causal graph network.
9. The method of claim 1, wherein the step of generating an interpretable early warning report is: Screening the first N paths with the causal intensity weight sum higher than a preset path threshold value based on the causal intensity weight in the causal graph network, and taking the first N paths as key causal paths; In the key causal path, locating user entity nodes with user authority characteristics higher than a preset authority threshold as core users, and locating event nodes with diffusion scale higher than a preset scale threshold in the propagation dynamic characteristics as core events; And filling and correlating the predicted risk level, risk category, explosion time, key causal path and causal intensity weight thereof, core user and user authority characteristics thereof, core event and propagation dynamic characteristics thereof according to a preset report template to obtain an interpretable early warning report.
10. A financial risk early warning system based on social media information for implementing the method of any one of claims 1 to 9, comprising: The data acquisition module acquires multi-source social media data and outputs multi-mode data lakes through preprocessing; the feature extraction module extracts text semantic features, information propagation network features and user authority features from the multi-mode data lake and fuses the text semantic features, the information propagation network features and the user authority features into a joint feature matrix; The causal analysis module is used for constructing a ternary causal map based on the joint feature matrix, and generating a causal map network by carrying out causal inference and anti-facts analysis on the ternary causal map; The early warning module takes a joint feature matrix, a ternary causal map and a causal map network as input, and obtains a predicted risk level, a predicted risk category and a predicted explosion time through a pre-trained self-adaptive early warning model; the report generation module is used for filling and correlating according to a preset report template based on the risk level, the risk category and the explosion time, the key causal path and causal intensity weight thereof, the authority characteristics of the core user and the user, and the core event and the propagation dynamic characteristics thereof, and generating an interpretable early warning report.

Description

Financial risk early warning method and system based on social media information Technical Field The invention relates to the technical field of enterprise risk management, in particular to a financial risk early warning method and system based on social media information. Background Financial risk early warning techniques underwent evolution from traditional statistical analysis methods to machine learning models. In recent years, social media data is introduced in the field, and potential risks of enterprises are mainly identified through emotion analysis, propagation monitoring, user influence evaluation and other modes. However, the prior art solutions have a general key drawback in that the lack of systematic analysis of causal links between risk signals makes it difficult to reveal the risk conduction paths and amplification mechanisms. The disadvantage causes weak interpretation of the early warning result, and cannot provide an inference basis with logic support for risk management decision, so that the practical application value of the early warning system is limited. Therefore, developing a financial risk early warning model with causal reasoning capability has become an important technical direction for improving early warning accuracy and decision support effectiveness. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide a financial risk early warning method and a financial risk early warning system based on social media information, which are used for deeply revealing a causal conduction path and an amplification mechanism of a risk signal by fusing multi-source social media data and combining with the construction and analysis of a causal graph network, realizing multi-task risk prediction by training a self-adaptive early warning model and driving the generation of an interpretable early warning report with a complete evidence chain, and realizing the accurate identification and the foresight prediction of potential financial risks so as to solve the problems of insufficient interpretability and decision support value of early warning results caused by lack of causal correlation analysis in the prior art. In order to achieve the above purpose, the invention provides a financial risk early warning method based on social media information, which comprises the following steps: acquiring multi-source social media data of a target enterprise, preprocessing the multi-source social media data, and constructing a multi-mode data lake with aligned time sequences; Based on the multi-mode data lake, extracting text semantic features, information propagation network features and user authority features to generate a joint feature matrix; Constructing a ternary causal map comprising enterprises, events and users based on the joint feature matrix; generating a causal graph network by performing causal inference and anti-facts analysis on the ternary causal graph; Inputting the causal graph network into a pre-trained self-adaptive early warning model to obtain a predicted risk level, a predicted risk category and a predicted explosion time; based on the predictions of risk level, risk category, and burst time, interpretable early warning reports are generated. Preferably, the method for acquiring the adaptive early warning model comprises the following steps: The method comprises the steps of obtaining k groups of risk data sets, wherein the risk data sets comprise a joint feature matrix, a ternary causal map, a causal map network, corresponding risk levels, risk categories and explosion time, taking the risk data sets as sample sets, dividing the sample sets into training sets and test sets, constructing a multi-task classification and regression device, taking the joint feature matrix, the ternary causal map and the causal map network in the training sets as inputs, taking the risk levels, the risk categories and the explosion time corresponding to the training sets as outputs, carrying out iterative training on the multi-task classification and regression device, obtaining a preliminary early warning model, testing the preliminary early warning model by utilizing the test sets, taking the preliminary early warning model as an adaptive early warning model if a prediction result of the preliminary early warning model meets preset accuracy, wherein the preliminary early warning model is a graph neural network deep learning model, and k is a positive integer greater than 1. Preferably, the method for extracting the authority degree characteristics of the user comprises the following steps: based on the multi-modal data lake, extracting user interaction data, information propagation data and user portraits of a target enterprise, and cons