Financial information analysis method and system based on artificial intelligence

Abstract

The invention provides a financial information analysis method based on artificial intelligence, and the method comprises the steps: carrying out the multi-source data collection and fusion of financial information, and obtaining an original financial data set; performing heterogeneous data cleaning and structured processing on the original financial data set to obtain a standardized data set; performing clustering division operation on the standardized data set through a fuzzy high-order hybrid clustering algorithm to obtain a data cluster; constructing a knowledge graph among the data clusters, and extracting a data relationship and a data index according to the knowledge graph; performing feature extraction and prediction on the data clustering clusters through a neural network structure to obtain a key feature set and risk indexes; and performing risk identification on the financial information according to the data relationship, the key feature set and the risk index, and generating a risk report. According to the method, through data hybrid clustering, knowledge graph construction, neural network modeling and risk quantification, intelligent analysis of multi-source heterogeneous financial information is realized, and financial risk detection efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a financial information analysis method and system based on artificial intelligence. Background Technology

[0002] With the increasing complexity of business operations and the rapid changes in the financial environment, financial information is characterized by its multi-source heterogeneity, massive volume, and strong dynamic correlations, posing a severe challenge to the real-time and accurate identification and early warning of financial risks. Traditional analytical methods relying on human experience are inefficient, highly subjective, and unable to handle massive amounts of multimodal data; while existing automated analysis tools based on rules or simple statistical models have significant shortcomings in processing unstructured information, capturing deep data correlations, adapting to complex business logic, and handling dynamically evolving risks.

[0003] While artificial intelligence (AI) technologies applied to financial analysis have made some progress, they still face numerous bottlenecks. For example, the precision of deep cleaning of heterogeneous data during the data preprocessing stage is insufficient, leading to low data quality and significant information loss. Furthermore, in the data organization and analysis stages, existing clustering methods struggle to effectively analyze the high-dimensionality, fuzzy characteristics, and multimodal features of financial data, including numerical, textual, and graph relationships. They also cannot extract causal relationships from time-series data streams and financial flow, resulting in low prediction accuracy. Therefore, designing an AI-based financial information analysis method and system is essential. Summary of the Invention

[0004] The purpose of this invention is to provide a financial information analysis method and system based on artificial intelligence, which improves the efficiency and accuracy of financial risk detection through data hybrid clustering, knowledge graph construction, neural network modeling and risk quantification.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] An artificial intelligence-based financial information analysis method includes the following steps:

[0007] The original financial dataset is obtained by collecting and integrating financial information from multiple sources using online tools.

[0008] The original financial dataset is cleaned and structured using data processing algorithms to obtain a standardized dataset.

[0009] The standardized dataset is clustered using a fuzzy high-order hybrid clustering algorithm to obtain data clusters;

[0010] Construct a knowledge graph between data clusters, and extract data relationships and data indexes based on the knowledge graph;

[0011] By using a neural network structure to extract and predict features from data clusters, key feature sets and risk indicators are obtained; the risk indicators include: operational risk probability and financial health index.

[0012] Risks are identified in financial information based on data relationships, key feature sets, and risk indicators, and risk reports are generated.

[0013] Optionally, the original financial dataset is subjected to heterogeneous data cleaning and structuring processing using data processing algorithms to obtain a standardized dataset, including:

[0014] A semantic-numerical association mapping is performed on the original financial dataset to obtain a semantically anchored data matrix;

[0015] Entity-attribute alignment is performed on the unstructured text in the semantically anchored data matrix to obtain an entity set;

[0016] The entity set is segmented into event-driven segments using a dynamic temporal slicing algorithm to obtain an event slice dataset;

[0017] A third-order cleaning operation is performed on the event slice dataset to obtain a structured data cube;

[0018] The structured data cube is restructured and principal component analysis is performed to obtain a standardized dataset.

[0019] Optionally, a third-order cleaning operation is performed on the event slice dataset to obtain a structured data cube, including:

[0020] Subject mapping is performed on the event slice dataset to obtain unified subject code data;

[0021] Perform cross-table cross-checking on the unified subject code data to obtain corrected cross-checking relationship data;

[0022] The missing values ​​in the cross-referenced data were filled using the KNN interpolation method to obtain a structured data cube.

[0023] Optionally, a fuzzy high-order hybrid clustering algorithm is used to perform clustering operations on the standardized dataset to obtain data clusters, including:

[0024] The standardized dataset is subjected to feature space separation to obtain independent subspaces; the independent subspaces include: numerical feature subspace, text semantic subspace, and graph topology subspace.

[0025] Based on the industry relative value matrix, an initial soft partitioning of the independent subspace is performed using an adaptive membership function to obtain a primary fuzzy cluster.

[0026] The initial clusters are obtained by dynamically aligning the primary fuzzy clusters using the distance matrix.

[0027] Calculate the silhouette coefficient of the initial clusters, and then screen the initial clusters for effectiveness based on the silhouette coefficients to obtain a set of effective clusters;

[0028] The correlation matrix is ​​obtained by aligning the probability distribution of the effective cluster set using the KL divergence metric.

[0029] Based on the correlation matrix, the optimal pairing of effective cluster sets is performed using the Hungarian algorithm to obtain data clusters.

[0030] Optionally, a knowledge graph is constructed between data clusters, and data relationships and data indexes are extracted from the knowledge graph, including:

[0031] The data clusters are sorted according to the time series to obtain a one-way time map;

[0032] Taking a specific fund as the tracking target, extract the target features from the one-way time graph, and form a fund operation chain based on the time sequence of the target features;

[0033] Based on the capital operation chain, business logic relationships are injected into the unidirectional time graph to obtain a directed causal graph.

[0034] Semantic recognition is performed on the unidirectional temporal graph based on the directed causal graph, and the driving factors in the unidirectional temporal graph are located based on the semantics to obtain the behavioral path;

[0035] Based on behavioral paths, a knowledge graph is obtained by mining causal relationships in behavioral paths using meta-path mining algorithms.

[0036] The frequent subgraph mining algorithm is used to locate subgraphs in the knowledge graph and obtain data relationships.

[0037] The SimRank algorithm is used to obtain the relationship similarity between different subgraphs, and a data index is constructed based on the relationship similarity.

[0038] Optionally, features of data clusters can be extracted and predicted using a neural network structure to obtain key feature sets and risk indicators, including:

[0039] A multi-head self-attention mechanism is used to perform cross-cluster feature interaction on data clusters to obtain an interaction feature matrix;

[0040] By decoupling the relationships between the interaction feature matrices using temporal and spatial convolution kernels, temporal and spatial subsets are obtained.

[0041] The temporal and spatial subsets are adaptively weighted and fused using the sigmoid function to obtain the fused feature tensor.

[0042] The long-short-term dependencies of the fused feature tensor are extracted using an LSTM network to obtain the dependency feature vector;

[0043] The feature contribution of the dependent feature vector is measured by the SHAP value, and the feature weight distribution matrix is ​​obtained.

[0044] The key feature set is obtained by filtering the feature weight distribution matrix using an orthogonal constraint loss function.

[0045] Optionally, by using a neural network structure to extract and predict features from data clusters to obtain key feature sets and risk indicators, the method further includes:

[0046] Risk calculations are performed on data clusters to obtain baseline risk values;

[0047] Calculate the risk deviation based on the benchmark risk value;

[0048] The operational risk probability and financial health index are calculated based on the benchmark risk value and risk deviation, respectively.

[0049] Early warning levels are classified based on the probability of operational risk and the financial health index.

[0050] Optionally, the benchmark risk value The calculation formula is: ,in, Let $\beta$ be the standard deviation of the earnings before interest in the i-th month. Let be the average earnings before interest in the i-th month. This represents the total liabilities for month i. Let t represent the total assets in month i, and t represent the current month.

[0051] Risk deviation The calculation formula is: ,in, This represents the monthly cash flow.

[0052] Operational risk probability The calculation formula is: ;in, The preset risk threshold;

[0053] Financial Health Index The calculation formula is: ,in, for The average over a year for The standard deviation over one year.

[0054] Optionally, the criteria for classifying warning levels are as follows:

[0055] When R d ≤ At that time, the warning level was 0;

[0056] when <R d ≤ At that time, the warning level was Level 1;

[0057] when <R d ≤ At that time, the warning level was Level 2;

[0058] When R d > At that time, the warning level was level 3.

[0059] An artificial intelligence-based financial information analysis system includes:

[0060] The data acquisition module is used to collect and integrate financial information from multiple sources through network tools to obtain the raw financial dataset.

[0061] The data cleaning module is used to perform heterogeneous data cleaning and structuring on the original financial dataset using data processing algorithms to obtain a standardized dataset.

[0062] The data clustering module is used to perform clustering operations on standardized datasets using a fuzzy high-order hybrid clustering algorithm to obtain data clusters.

[0063] The relationship building module is used to construct a knowledge graph between data clusters and extract data relationships and data indexes based on the knowledge graph.

[0064] The risk detection module is used to extract and predict features from data clusters through a neural network structure to obtain key feature sets and risk indicators; the risk indicators include: operational risk probability and financial health index.

[0065] The visualization module is used to identify risks in financial information based on data relationships, key feature sets, and risk indicators, and generate risk reports.

[0066] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides an artificial intelligence-based financial information analysis method, which includes: collecting and fusing multi-source financial information through network tools to obtain an original financial dataset; cleaning and structuring the original financial dataset using a data processing algorithm to obtain a standardized dataset; performing clustering operations on the standardized dataset using a fuzzy high-order hybrid clustering algorithm to obtain data clusters; constructing a knowledge graph between data clusters and extracting data relationships and data indexes based on the knowledge graph; extracting and predicting features of the data clusters using a neural network structure to obtain key feature sets and risk indicators; the risk indicators include: operational risk probability and financial health index; identifying risks in financial information based on data relationships, key feature sets, and risk indicators, and generating a risk report. This method, through data hybrid clustering, knowledge graph construction, neural network modeling, and risk quantification, achieves intelligent analysis of multi-source heterogeneous financial information, improving the efficiency and accuracy of financial risk detection. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0068] Figure 1 This is a flowchart of the financial information analysis method of the present invention;

[0069] Figure 2 This is a flowchart of the heterogeneous data cleaning and structuring process of the present invention;

[0070] Figure 3 This is a flowchart of the clustering partitioning process of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0073] like Figure 1As shown, this invention provides a financial information analysis method based on artificial intelligence, comprising the following steps:

[0074] Step 100: Collect and integrate financial information from multiple sources using online tools to obtain the original financial dataset;

[0075] Specifically, the original financial dataset contains structured, semi-structured, and unstructured data. Structured data is extracted using a JDBC connector to retrieve the system's account balance sheet, voucher transaction records, and balance sheet. Bank transaction data is obtained through a SWIFT message parsing tool, and SSL / TLS encryption is used to ensure data transmission security. Semi-structured data is collected from electronic invoices on the tax platform using the Scrapy crawler framework. An OCR engine is used to parse the price and tax details from the invoice images, while the supply chain system's WebService interface is invoked to extract XML / JSON format data from purchase orders and logistics settlement documents. Unstructured data is accessed through the enterprise's email server, using NLP tools to identify payment terms in contract texts, and simultaneously scanning the archived voucher repository to locate key fields such as transaction amounts and contracting party names.

[0076] Step 200: Perform heterogeneous data cleaning and structuring on the original financial dataset using data processing algorithms to obtain a standardized dataset; specific steps are as follows: Figure 2 As shown, it includes:

[0077] Step 201: Perform semantic-numerical association mapping on the original financial dataset to obtain the semantic anchored data matrix.

[0078] Specifically, based on a pre-built accounting subject ontology library, which includes subject names, aliases, hierarchical relationships, and industry terms, the text descriptions in the original financial dataset are vectorized using a BERT pre-trained language model. By calculating the cosine similarity between the text vectors and the standardized subject names in the ontology library, the best association is automatically matched. In some embodiments, "Accounts Receivable - Customer A" is mapped to the subject code "1122," while simultaneously generating a confidence score. For key numerical values ​​such as amounts and dates in the original financial dataset, information is extracted using regular expressions combined with a conditional random field model to generate a semantically anchored data matrix, where rows represent business events and columns represent mapped subject-value pairs.

[0079] Step 202: Perform entity-attribute alignment on the unstructured text in the semantic anchoring data matrix to obtain the entity set.

[0080] Specifically, basic entities such as company name, date, and amount are first extracted from unstructured text using regular expressions. Then, a NER model is used to identify deeper entity types. In some embodiments, the guarantee relationship in the text "Guarantor: XX Group" is identified. For entities with the same name, equity relationship graphs and the Levenshtein edit distance algorithm are used, with the unified social credit code as the primary key, to perform disambiguation operations, establish "entity-attribute-value" triples, and output the entity set.

[0081] Step 203: Perform event-driven segmentation of the entity set using the dynamic temporal slicing algorithm to obtain the event slice dataset;

[0082] Specifically, key event types are first preset, such as loan disbursement and accounts receivable maturity. Then, the PELT algorithm is used to locate the timestamps of events occurring in the entity set. Next, time windows with preceding and following relationships are adaptively generated around each event. In some embodiments, the time window for loan disbursement events is a slice from 30 days before disbursement to 90 days after disbursement, ensuring that each slice contains complete business lifecycle data. For periodic businesses such as monthly settlements, seasonal fluctuation points are identified using Fourier transform, and the slice length is dynamically adjusted. In some embodiments, the peak season window is extended to 45 days. Finally, an event slice dataset is output, with each slice containing entities, values, and contextual semantic relationships within the time window.

[0083] Step 204: Perform a third-order cleaning operation on the event slice dataset to obtain a structured data cube;

[0084] Specifically, based on the original account names in the event slice dataset, multilingual alias normalization is performed using an accounting account tree, fuzzy matching is performed using the Levenshtein-word vector hybrid similarity algorithm, and optimal account allocation is performed using the Hungarian algorithm, thus obtaining unified account coding data.

[0085] Then, cross-table reconciliation verification is performed to ensure logical consistency between different financial tables, resulting in corrected reconciliation data. First, the balance sheet, income statement, and cash flow statement are integrated into a reconciliation rule base. Then, the unified account coding data in this rule base undergoes double verification: The account association value is calculated using a dynamic programming algorithm, which in some embodiments represents the reconciliation relationship between "inventory" in the balance sheet and "purchase expenditures" in the cash flow statement. Data with a correlation value deviation exceeding ±5% is marked as abnormal. A dynamic time warping algorithm is used to align the accounting periods of different reports, which in some embodiments represents the monthly breakdown of quarterly income statement data to match monthly cash flows, thereby eliminating time granularity differences between coded data, ensuring logical self-consistency of financial data, and eliminating underlying data contradictions. Finally, the KNN imputation method is used to fill in missing values ​​in the corrected reconciliation data, resulting in a structured data cube, with the expression:

[0086] ;

[0087] ;

[0088] in, The imputed value of the j-th missing feature in the i-th sample. Let k be the set of the k nearest neighbors of the i-th sample. Let be the weights of sample i and its neighbor sample s. This is a weighted distance metric function. For sample i, the complete feature vector Let be the feature vector of the neighbor sample s. This is the distance attenuation factor, which is 2 in some embodiments. Let be the observation value of the neighbor sample s on the j-th feature.

[0089] Step 205: Perform structured recombination and principal component analysis on the structured data cube to obtain a standardized dataset.

[0090] Specifically, the structured data cube is reorganized according to a three-dimensional model of "enterprise entity - accounting period - account type". High-dimensional sparse accounts such as "sales expenses - travel expenses - city level" are folded according to business scenarios to obtain a merged representation of "sales expenses - travel expenses". Then, the covariance matrix of each slice of data is calculated by principal component analysis algorithm, and principal components with a cumulative contribution rate of ≥85% are retained to finally obtain a standardized dataset.

[0091] Step 300: Perform clustering operations on the standardized dataset using a fuzzy high-order hybrid clustering algorithm to obtain data clusters; specific steps are as follows: Figure 3 As shown, it includes:

[0092] Step 301: Perform feature space separation on the standardized dataset to obtain independent subspaces.

[0093] Specifically, the standardized dataset is decomposed into three independent orthogonal subspaces—numerical feature subspace, textual semantic subspace, and graph topology subspace—through CPD decomposition. The numerical feature subspace retains all continuous variables and eliminates the influence of dimensions through Z-score normalization. The textual semantic subspace extracts 300-dimensional word embedding vectors from texts such as account descriptions and contract terms, and uses an attention mechanism to focus them on key business terms. This achieves physical isolation of the multimodal characteristics of financial data, laying a spatial foundation for subsequent hybrid clustering.

[0094] Step 302: Based on the industry relative value matrix, perform initial soft partitioning of the independent subspace using an adaptive membership function to obtain a primary fuzzy cluster.

[0095] Specifically, based on the China Securities Regulatory Commission's industry classification standards, the relative offsets of each company's indicators to the industry quantiles are calculated. Then, fuzzy C-means clustering is performed in parallel on the three subspaces using a hyperbolic tangent membership function. The function expression is as follows: The slope k is dynamically adjusted according to industry volatility; k is 0.5 for high-volatility industries and 2 for stable industries. The threshold is... The value represents the industry median. After clustering, each data point obtains a three-dimensional membership vector, and candidate cluster pairs are generated through Cartesian product. Only combinations with a membership product greater than 0.25 are retained to generate primary fuzzy clusters.

[0096] Step 303: Perform dynamic inter-cluster alignment of the primary fuzzy clusters using the distance matrix to obtain the initial clusters.

[0097] Specifically, Mahalanobis distance, Word Mover's Distance (WMD), and graph editing distance are merged into a composite distance matrix. The cluster pairs corresponding to the smallest element in the distance matrix are selected for merging and the new cluster centers are recalculated. The merged new cluster retains 20% of the clustering structure from the previous stage as a benchmark. The iteration is terminated when the distance change rate is less than 5% for three consecutive rounds, and the initial cluster is obtained.

[0098] Step 304: Calculate the silhouette coefficient of the initial clusters, and screen the initial clusters for effectiveness based on the silhouette coefficients to obtain a set of effective clusters.

[0099] Specifically, the formula for calculating the profile coefficient is: ;in The average distance within the cluster. This is the nearest neighbor cluster distance. (Through...) >0.3 and The top 70% are selected using a dual-threshold screening mechanism to determine the effectiveness of the clusters and obtain a set of effective clusters.

[0100] Step 305: Align the probability distribution of the effective cluster set using the KL divergence metric to obtain the association matrix.

[0101] Specifically, for numerical clusters in the effective cluster set, maximum likelihood estimation is used to calculate the degrees of freedom and scale parameters; for text clusters in the effective cluster set, LDA topic modeling is used to extract topic distribution vectors; for graph clusters in the effective cluster set, based on the ERGM model, Markov chain Monte Carlo method is used for sampling and estimating network structure parameters. For any two clusters, the KL divergence of their numerical subspace, the Jensen-Shannon distance of their text subspace, and the graph kernel function similarity of their graph subspace are calculated, and a weighted fusion is used to obtain the comprehensive association degree, with weight coefficients of 0.6, 0.25, and 0.15, respectively. All calculation results are used as matrix elements to construct an association matrix.

[0102] Based on this, the inter-cluster probability distribution differences are calculated using KL divergence.

[0103] Then, the KL divergence was used to measure the distributional differences between clusters. Then, three...

[0104] An asymmetric incidence matrix is ​​generated by calculating the KL divergence between clusters.

[0105] Step 306: Based on the correlation matrix, the effective cluster set is optimally paired using the Hungarian algorithm to obtain the data clusters.

[0106] Step 400: Construct a knowledge graph between data clusters, and extract data relationships and data indexes based on the knowledge graph;

[0107] Specifically, the data clusters are first transformed into an ordered structure along the time dimension, and the time series is sorted. Based on the accounting period timestamps annotated in the event slice dataset, the start / end points of each cluster are marked. A dynamic time warping algorithm is used to eliminate the effects of time series drift, thus obtaining a one-way time graph. Next, the fund operation chain is traced. Taking a specific fund as the tracking target and a user-specified fund identifier as the index, multi-hop tracking is performed in the one-way time graph to extract target features. Key attributes related to the target fund are identified based on the metadata of the clusters. In some embodiments, the metadata is a large amount of accounts payable, and the key attributes are the payee's account number and the settlement currency type. The feature nodes of these key attributes are then concatenated in chronological order to form the fund operation chain.

[0108] Then, the event nodes in the capital operation chain are matched with preset rules, thereby adding the causal relationships of business logic to the one-way time graph, forming a directed causal graph. In some embodiments, a causal relationship is established between "60-day delayed payment" and "deterioration of supplier credit". Simultaneously, gradient significance detection is performed on the edge weights in the causal graph to locate abrupt change nodes. In some embodiments, an abrupt change node represents a sudden increase in the accounts receivable collection period from 45 days to 70 days. The associated text in the one-way time graph is parsed using the BERT model to identify the textual descriptions of key events. The abrupt change nodes and associated text are integrated into driving factor triples, represented as: <Subject: Customer A, Action: Payment Delay, Degree: High Risk>, and the behavioral path is located based on the driving factors. Causal relationships are mined from the behavioral paths using a meta-path mining algorithm to form a knowledge graph.

[0109] Finally, a frequent subgraph mining algorithm is used to locate subgraphs in the knowledge graph. All connected subgraphs are traversed with a minimum support of 10%, and subgraphs containing risk labels are selected. High-frequency subgraphs are then encoded into signature vectors, resulting in a data relationship table containing signature vectors and subgraph mappings. Simultaneously, the SimRank algorithm is used to calculate the similarity of relationships between different subgraphs, and a data index is constructed based on the similarity matrix.

[0110] Step 500: Extract and predict features from data clusters using a neural network structure to obtain key feature sets and risk indicators; risk indicators include: operational risk probability and financial health index;

[0111] Specifically, by using a neural network structure to extract and predict features from data clusters, key feature sets and risk indicators are obtained, including:

[0112] First, a multi-head self-attention mechanism is used to perform cross-cluster feature interaction on the data clusters. The feature vector of each cluster is first mapped to 32 different subspaces, i.e., 32 attention heads. Each subspace corresponds to an independent Query, Key, and Value matrix. The Query matrix represents the feature requirements of the current cluster, the Key matrix describes the feature attributes of other clusters, and the Value matrix stores the feature values ​​of other clusters. Then, the scaled dot product attention score of the Query and Key in each attention head is calculated. This is obtained by dividing the inner product of the inter-cluster features by the square root of the feature dimension, and then normalizing using the SoftMax function to obtain the attention weight, thus measuring the influence of different clusters on the current cluster. Finally, the output features of the 32 attention heads are concatenated column-wise to obtain an interaction feature matrix containing inter-cluster association information. This overcomes the limitations of features from a single cluster and captures the implicit relationships between different clusters.

[0113] Next, the interaction feature matrix is ​​decoupled using temporal and spatial convolution kernels. A 1D temporal convolution kernel with a size of 5 and a stride of 1 is used to perform sliding convolution on the temporal dimension of the interaction feature matrix. This convolution kernel captures feature changes in adjacent time steps through local receptive fields and outputs a temporal subset reflecting the temporal trend. At the same time, a 1D spatial convolution kernel with a size of 3 and a stride of 1 is used to convolve the spatial dimension of the matrix. This convolution kernel focuses on the feature associations of different clusters at the same time point and outputs a spatial subset reflecting the structural relationship between clusters.

[0114] Then, the temporal and spatial subsets are adaptively weighted and fused using the sigmoid function. First, the feature variances of the temporal and spatial subsets are calculated. The larger the variance, the richer the effective information contained in the subset. Based on this, initial weight coefficients are generated. Then, the initial weight coefficients are input into the sigmoid function to obtain the temporal weight a1 and spatial weight 1-a1, which are in the range of 0-1. The temporal and spatial subsets are then summed element-wise according to the weights to obtain the fused feature tensor.

[0115] Then, a LSTM network is used to extract long- and short-term dependencies from the fused feature tensor. A two-layer LSTM structure is set up, with 64 hidden units in each layer. The fused feature tensor is input into the network structure in time step order. The first LSTM layer captures short-term feature dependencies, and the second LSTM layer further extracts long-term feature dependencies based on the output of the first layer. At the same time, a dropout mechanism with a dropout rate of 0.2 is introduced to prevent overfitting, and effective information is dynamically filtered through gating units of the input gate, forget gate, and output gate. Finally, the output is a dependency feature vector containing long- and short-term associations.

[0116] Next, the SHAP value is used to measure the feature contribution of the dependent feature vectors. The KernelExplainer is used, with the dependent feature vectors as model input and the risk prediction result as output target. The SHAP value of each feature to the prediction result is calculated. The SHAP values ​​are arranged according to the feature dimension to generate a feature weight distribution matrix, where the larger the absolute value of the matrix element, the more significant the influence of the corresponding feature on the risk prediction.

[0117] Finally, an orthogonal constraint loss function is used to filter features from the feature weight distribution matrix. This loss function measures the correlation by calculating the inner product between feature vectors and combines it with the feature weights as the loss term. Optimization stops when the loss value is less than a preset threshold. Subsequently, the top 30% of features with the highest weights and that are mutually orthogonal are retained to form the key feature set.

[0118] Specifically, by using a neural network structure to extract and predict features from data clusters, key feature sets and risk indicators are obtained, including:

[0119] Risk calculation is performed on the data clusters to obtain the baseline risk value. The calculation formula is:

[0120] ;

[0121] in, Let $\beta$ be the standard deviation of the earnings before interest in the i-th month. Let be the average earnings before interest in the i-th month. This represents the total liabilities for month i. Let be the total assets in month i, and t be the current month. Calculate the risk deviation based on the benchmark risk value. The calculation formula is:

[0122] ;

[0123] in, This represents the monthly cash flow. The operational risk probability and financial health index are calculated based on the benchmark risk value and risk deviation, respectively. The operational risk probability... and financial health index The calculation formulas are as follows:

[0124] ;

[0125] ;

[0126] in, As a preset risk threshold, for The average over a year for The standard deviation over one year. Finally, based on the probability of operational risk and the financial health index, a warning level is assigned, when R... d ≤ At that time, the warning level was 0; when <R d ≤ When, the warning level is Level 1; when <R d ≤ When R is in use, the warning level is Level 2; when R is in use, the warning level is Level 2. d > At that time, the warning level was level 3.

[0127] Step 600: Identify risks in financial information based on data relationships, key feature sets, and risk indicators, and generate a risk report.

[0128] This invention also provides an artificial intelligence-based financial information analysis system, comprising:

[0129] The data acquisition module is used to collect and integrate financial information from multiple sources through network tools to obtain the raw financial dataset.

[0130] The data cleaning module is used to perform heterogeneous data cleaning and structuring on the original financial dataset using data processing algorithms to obtain a standardized dataset.

[0131] The data clustering module is used to perform clustering operations on standardized datasets using a fuzzy high-order hybrid clustering algorithm to obtain data clusters.

[0132] The relationship building module is used to construct a knowledge graph between data clusters and extract data relationships and data indexes based on the knowledge graph.

[0133] The risk detection module is used to extract and predict features from data clusters through a neural network structure to obtain key feature sets and risk indicators; the risk indicators include: operational risk probability and financial health index.

[0134] The visualization module is used to identify risks in financial information based on data relationships, key feature sets, and risk indicators, and generate risk reports.

[0135] The beneficial effects of this invention are as follows:

[0136] 1) Multi-source collection and fusion of structured, semi-structured and unstructured data were achieved through network tools. Combined with cleaning techniques such as semantic-numerical association mapping and entity-attribute alignment, the problem of traditional methods being unable to integrate heterogeneous data was solved, and comprehensive coverage and deep integration of financial data were achieved.

[0137] 2) The three-stage cleaning operation eliminated data contradictions and missing data, and combined with structured recombination and principal component analysis, the data was standardized, which improved the accuracy and quality of data preprocessing.

[0138] 3) Fuzzy high-order hybrid clustering algorithm is used to separate the data into three independent subspaces: numerical, textual semantic, and graph topology. Accurate clustering is achieved through industry relative value matrix and adaptive membership function, which overcomes the limitation of difficulty in collaborative analysis of multimodal features of financial data and improves the accuracy and relevance of data clustering.

[0139] 4) By constructing a knowledge graph, combined with time series sorting, capital operation chain tracking and meta-path mining, scattered financial data are linked into a structured causal network, accurately capturing time-series dynamic correlations and business logic relationships.

[0140] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0141] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A financial information analysis method based on artificial intelligence, characterized in that, Includes the following steps: The original financial dataset is obtained by collecting and integrating financial information from multiple sources using online tools. The original financial dataset is subjected to heterogeneous data cleaning and structuring processing using data processing algorithms to obtain a standardized dataset; The standardized dataset is clustered using a fuzzy high-order hybrid clustering algorithm to obtain data clusters; Construct a knowledge graph between the data clusters, and extract data relationships and data indexes based on the knowledge graph; By using a neural network structure to extract and predict features from the data clusters, a key feature set and risk indicators are obtained; the risk indicators include: operational risk probability and financial health index. Based on the data relationships, the key feature set, and the risk indicators, the financial information is risk-identified, and a risk report is generated.

2. The financial information analysis method based on artificial intelligence according to claim 1, characterized in that, The original financial dataset is subjected to heterogeneous data cleaning and structuring processing using data processing algorithms to obtain a standardized dataset, including: The original financial dataset is subjected to semantic-numerical association mapping to obtain a semantically anchored data matrix; Entity-attribute alignment is performed on the unstructured text in the semantic anchoring data matrix to obtain an entity set; The entity set is segmented into event-driven segments using a dynamic temporal slicing algorithm to obtain an event slice dataset; A third-order cleaning operation is performed on the event slice dataset to obtain a structured data cube; The structured data cube is restructured and subjected to principal component analysis to obtain the standardized dataset.

3. The financial information analysis method based on artificial intelligence according to claim 2, characterized in that, A third-order cleaning operation is performed on the event slice dataset to obtain a structured data cube, including: Subject mapping is performed on the event slice dataset to obtain unified subject code data; Cross-table cross-checking is performed on the unified subject code data to obtain cross-checking relationship correction data; The missing values ​​in the cross-reference data are filled using the KNN interpolation method to obtain the structured data cube.

4. The financial information analysis method based on artificial intelligence according to claim 1, characterized in that, The standardized dataset is clustered using a fuzzy high-order hybrid clustering algorithm to obtain data clusters, including: The standardized dataset is subjected to feature space separation to obtain independent subspaces; the independent subspaces include: numerical feature subspace, text semantic subspace, and graph topology subspace. Based on the industry relative value matrix, the independent subspace is initially softly partitioned using an adaptive membership function to obtain a primary fuzzy cluster. The initial clusters are obtained by dynamically aligning the primary fuzzy clusters using a distance matrix. Calculate the silhouette coefficient of the initial cluster, and perform effectiveness screening on the initial cluster based on the silhouette coefficient to obtain a set of effective clusters; The correlation matrix is ​​obtained by aligning the probability distribution of the effective cluster set using the KL divergence metric. Based on the correlation matrix, the effective cluster set is optimally paired using the Hungarian algorithm to obtain the data cluster.

5. The financial information analysis method based on artificial intelligence according to claim 1, characterized in that, Constructing a knowledge graph among the data clusters, and extracting data relationships and data indexes based on the knowledge graph, including: The data clusters are sorted according to the time series to obtain a one-way time map; Taking a specific fund as the tracking target, the target features are extracted from the one-way time graph, and a fund operation chain is formed according to the time sequence of the target features; Based on the aforementioned capital operation chain, business logic relationships are injected into the unidirectional time graph to obtain a directed causal graph. Semantic recognition is performed on the unidirectional temporal graph based on the directed causal graph, and the driving factors in the unidirectional temporal graph are located based on the semantics to obtain the behavioral path; Based on the behavioral path, the knowledge graph is obtained by mining causal relationships of the behavioral path using a meta-path mining algorithm. The knowledge graph is used to locate subgraphs using a frequent subgraph mining algorithm to obtain the data relationships; The SimRank algorithm is used to obtain the relationship similarity of different subgraphs, and the data index is constructed based on the relationship similarity.

6. The financial information analysis method based on artificial intelligence according to claim 1, characterized in that, By using a neural network structure to extract and predict features from the data clusters, key feature sets and risk indicators are obtained, including: The data clusters are subjected to cross-cluster feature interaction through a multi-head self-attention mechanism to obtain an interaction feature matrix; The interaction feature matrix is ​​decoupled using temporal and spatial convolution kernels to obtain temporal and spatial subsets; The temporal and spatial subsets are adaptively weighted and fused using the sigmoid function to obtain a fused feature tensor. The long-short-term dependencies of the fused feature tensor are extracted using an LSTM network to obtain the dependency feature vector. The feature contribution of the dependent feature vector is quantified by the SHAP value to obtain the feature weight distribution matrix; The key feature set is obtained by filtering the feature weight distribution matrix using an orthogonal constraint loss function.

7. The financial information analysis method based on artificial intelligence according to claim 6, characterized in that, The data clusters are subjected to feature extraction and prediction using a neural network structure to obtain key feature sets and risk indicators, and the method also includes: Risk calculations are performed on the data clusters to obtain baseline risk values; Calculate the risk deviation based on the benchmark risk value; The operational risk probability and the financial health index are calculated based on the benchmark risk value and the risk deviation, respectively. Early warning levels are classified based on the operational risk probability and the financial health index.

8. The financial information analysis method based on artificial intelligence according to claim 7, characterized in that, The benchmark risk value The calculation formula is: ,in, Let $\beta$ be the standard deviation of the earnings before interest in the i-th month. Let be the average earnings before interest in the i-th month. This represents the total liabilities for month i. Let t represent the total assets in month i, and t represent the current month. The risk deviation The calculation formula is: ,in, This represents the monthly cash flow. The probability of operational risks The calculation formula is: ;in, The preset risk threshold; The financial health index The calculation formula is: ,in, for The average over a year for The standard deviation over one year.

9. The financial information analysis method based on artificial intelligence according to claim 8, characterized in that, The criteria for classifying the warning levels are as follows: When R d ≤ At that time, the warning level is level 0; when <R d ≤ At that time, the warning level is Level 1; when <R d ≤ At that time, the warning level is Level 2; When R d > At that time, the warning level was level 3.

10. A financial information analysis system based on artificial intelligence, characterized in that, include: The data acquisition module is used to collect and integrate financial information from multiple sources through network tools to obtain the raw financial dataset. The data cleaning module is used to perform heterogeneous data cleaning and structuring on the original financial dataset using data processing algorithms to obtain a standardized dataset; The data clustering module is used to perform clustering operations on the standardized dataset using a fuzzy high-order hybrid clustering algorithm to obtain data clusters; The relationship building module is used to construct a knowledge graph between the data clusters and extract data relationships and data indexes based on the knowledge graph. The risk detection module is used to extract and predict features from the data clusters through a neural network structure to obtain a key feature set and risk indicators; the risk indicators include: operational risk probability and financial health index. The visualization module is used to identify risks in the financial information based on the data relationships, the key feature set, and the risk indicators, and generate a risk report.

Images