Self-evolvable intelligent fund analysis method

Abstract

The application provides a self-evolving intelligent fund analysis method, relates to the field of data processing, and comprises the following steps: receiving a natural language analysis instruction, combining a large language model decomposition task, extracting account fund flow characteristics and matching with historical cases, and generating a weighted execution instruction; identifying a transfer path based on a fund correlation graph, extracting a priority path subgraph structure according to a user identifier, and forming a directional analysis model; performing node risk scoring on the fund correlation graph, outputting a risk account identifier and an analysis report, and realizing intelligent and accurate fund risk identification.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to an intelligent financial analysis method that can evolve autonomously. Background Technology

[0002] Traditional fund analysis methods mainly rely on rule engines and simple statistical models. In recent years, machine learning and graph computing technologies have been gradually applied to the field of fund analysis. By constructing fund relationship networks and uncovering complex fund flow patterns, they provide new technical paths for financial risk prevention and control.

[0003] However, existing technologies lack flexible natural language interaction capabilities, making it difficult for analysts to customize analysis tasks through intuitive language descriptions. This leads to a disconnect between the analysis process and actual business needs, requiring analysts to master complex query syntax or rely on developers for customized development. Existing methods typically employ fixed feature weights and analysis models, failing to dynamically adjust analysis strategies based on different case types and historical interaction experience. This results in a lack of targeted analysis results and an inability to capture the unique fund flow characteristics of different case types. When facing complex fund networks, existing technologies struggle to effectively combine professional judgment from business personnel for precise targeted analysis. The lack of intelligent human-machine collaborative analysis mechanisms leads to inefficiency in complex multi-path fund flow scenarios, making it difficult to quickly identify key risk nodes and transfer paths. Summary of the Invention

[0004] This invention provides an intelligent financial analysis method that can evolve autonomously, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent financial analysis method capable of autonomous evolution, comprising:

[0006] Receive natural language analysis instructions for the target fund account, input them into the large language model, and output a task decomposition sequence;

[0007] Extract the fund flow characteristics of the target fund account and encode them into a case embedding vector. Select the historical case embedding vector with the smallest distance from the case embedding vector in the preset case profile vector library.

[0008] Historical case embedding vectors are decoded into feature weight distributions. The feature weight distributions are optimized based on historical interaction records. The optimized weight distributions are coupled to the task decomposition sequence to generate weighted execution instructions.

[0009] Based on the authorized execution instruction, the transaction data of the target fund account is correlated and calculated to generate a fund correlation graph. Based on the fund correlation graph, multiple transfer paths are identified and the flow characteristic difference is calculated. When the flow characteristic difference exceeds the preset difference threshold, the transfer path information is sent to the user and the priority path identifier is received.

[0010] Priority path identifiers are mapped to the fund association graph and the corresponding subgraph structure is extracted. The subgraph structure is fed back to the large language model to output the targeted task decomposition sequence. The subgraph structure is then subjected to depth traversal based on the targeted task decomposition sequence to extract node feature sequences and jointly train them with historical case embedding vectors to generate a targeted analysis model.

[0011] Input the fund relationship graph into the targeted analysis model to score node risk, and output risk account identifiers and relationship path analysis reports.

[0012] Receive natural language analysis instructions for the target fund account, input them into the large language model, and output a task decomposition sequence including:

[0013] The natural language analysis instructions of the target fund account are segmented to obtain the instruction word sequence. The contextual correlation between words in the instruction word sequence is calculated to generate the instruction semantic matrix. Transaction time sequence features, fund flow features and account relationship features are extracted from the instruction semantic matrix and fused to generate the instruction feature vector.

[0014] Construct a large language model processing structure that includes a word vector encoding layer, a multi-head attention layer, a feedforward computation layer, and a sequence decoding layer;

[0015] The instruction feature vector is input into the word vector encoding layer to generate the initial semantic code. The initial semantic code is input into the multi-head attention layer to calculate the feature weight distribution to obtain the weighted semantic vector. The weighted semantic vector is input into the feedforward calculation layer to perform dimensional transformation to obtain the semantic representation vector. The semantic representation vector is input into the sequence decoding layer to extract the time-series dependency information, account association information, and fund flow information, and combined into step-by-step analysis instructions. The information in the step-by-step analysis instructions is matched with the fund tracking sequence in historical fund analysis cases.

[0016] The analysis execution chain is constructed based on the degree of matching. The step-by-step analysis instructions are reorganized according to the node order in the analysis execution chain, and the task decomposition sequence is output.

[0017] Extract the fund flow characteristics of the target fund account and encode them into a case embedding vector. Select the historical case embedding vectors with the smallest distance from the pre-set case profile vector library, including:

[0018] Extract time-series transaction amounts, transaction frequencies, and counterparty information from the transaction records of the target fund account to construct a transaction attribute matrix;

[0019] The correlation coefficient matrix is ​​obtained by calculating the correlation coefficients between attributes in the transaction attribute matrix. The correlation coefficient matrix is ​​then converted into an adjacency matrix to construct a fund transaction association graph. Multi-layer convolution is then performed on the fund transaction association graph to obtain the node feature matrix.

[0020] The node feature matrix is ​​pooled and compressed to generate a capital flow feature vector.

[0021] The feature transformation of the fund flow feature vector is used to obtain the feature enhancement sequence. The distribution distance between the feature enhancement sequence and the fund flow feature vector is calculated. By minimizing the distribution distance, common features in the fund flow feature vector are extracted to generate the case embedding vector.

[0022] The cosine distance between the case embedding vector and the historical case embedding vector in the preset case profile vector library is calculated using the K-Means algorithm. The cluster center is updated based on the cosine distance, and the historical case embedding vector with the smallest distance from the case embedding vector is selected.

[0023] Historical case embedding vectors are decoded into feature weight distributions. These distributions are then optimized based on historical interaction records. The optimized weight distributions are coupled to the task decomposition sequence to generate weighted execution instructions, including:

[0024] Historical cases are embedded in vector decomposition to obtain a sequence of feature component values. The frequency distribution value is obtained by counting the number of times the feature component values ​​appear in historical successful cases. The information distribution value is obtained by calculating the information overlap between the feature component values ​​and transaction behavior. The association distribution value is obtained by calculating the temporal correlation of the feature values.

[0025] The frequency distribution value, information distribution value, and correlation distribution value are normalized and weighted to obtain the initial feature weight distribution;

[0026] Extract task interruption records and resource consumption records from historical interaction records. Identify interruption locations in task interruption records and peak points in resource consumption records based on the initial feature weight distribution. Calculate the occurrence probability of interruption locations and the utilization rate of resource peaks. Adjust the value of the initial feature weight distribution based on the interruption probability and utilization rate to obtain the optimized feature weight distribution.

[0027] The optimized feature weight distribution is sequentially mapped to the execution steps in the task decomposition sequence. The data transfer size and processing time between execution steps are calculated. The execution steps are prioritized and allocated computing resources according to the transfer size and processing time, and weighted execution instructions are generated.

[0028] Based on the authorized execution instruction, the transaction data of the target fund account is correlated and calculated to generate a fund correlation graph. Multiple transfer paths are identified based on the fund correlation graph, and the flow characteristic difference is calculated. When the flow characteristic difference exceeds a preset difference threshold, transfer path information is sent to the user, and a priority path identifier is received, including:

[0029] Extract the payment account identifier, receiving account identifier, transaction amount, and transaction time from the transaction record;

[0030] The payment account identifier and the receiving account identifier are constructed as graph nodes, and the total transaction amount on the node is marked as the node weight;

[0031] The rate of change of transaction amount between adjacent nodes is calculated according to the time window, and the rate of change of transaction amount is constructed as the energy transfer feature between nodes; a periodic energy curve is constructed based on the energy transfer feature, and the periodic energy curve is mapped as the connection edge weight between nodes; a capital correlation graph is generated through node weight and connection edge weight.

[0032] Identify the reachable path from the source node to the target node, calculate the ratio of the node weight to the edge weight on the reachable path to obtain the energy conversion rate, determine the path breakpoint based on the abrupt change in the energy conversion rate, split the path into sub-paths based on the path breakpoints, cluster the energy conversion rates of the sub-paths, and generate a set of transfer paths.

[0033] Calculate the energy conversion rate difference between the first and last nodes of the sub-path, construct the path feature matrix, and obtain the flow direction feature difference degree by calculating the loss difference between the sub-paths through the path feature matrix.

[0034] The system compares the flow direction feature difference with a preset difference threshold. When the flow direction feature difference exceeds the preset difference threshold, it sends the transfer path set information to the user and receives the priority path identifier returned by the user.

[0035] Priority path identifiers are mapped to a funding relationship graph, and the corresponding subgraph structures are extracted. These subgraph structures are then fed back into the large language model, outputting a targeted task decomposition sequence including:

[0036] The priority path identifier is mapped to the fund association graph, and the transaction amount sequence of the target node set is extracted to obtain the fund flow curve;

[0037] Extract the fluctuation period and fluctuation amplitude from the capital flow curve, identify the position with the largest capital accumulation and the position with the largest capital dispersion within the fluctuation period, and construct a set of feature nodes;

[0038] Calculate the connection edge weight sequence between feature node sets to construct the node connection strength distribution, determine the subgraph structure boundary based on the variance of the node connection strength distribution, and generate the subgraph structure;

[0039] The transaction features of the feature node set and the transitive features of the connecting edges in the subgraph structure are used to construct node representation vectors and edge representation vectors, respectively. The node representation vectors and edge representation vectors are combined into a structural feature matrix. A scene description sequence is generated based on the structural feature matrix. The scene description sequence is input into a large language model and the output is a targeted task decomposition sequence.

[0040] Based on the targeted task decomposition sequence, a depth-first traversal of the subgraph structure is performed, node feature sequences are extracted, and jointly trained with historical case embedding vectors to generate a targeted analysis model, including:

[0041] Extract the node access order from the targeted task decomposition sequence to construct the node traversal order, and extract feature calculation rules to construct feature extraction rules;

[0042] The subgraph structure is subjected to a depth-first search according to the node traversal order. During the traversal, the rate of change of the frequency of transactions before and after a node is calculated to obtain the frequency change vector, the number of connected edges of a node is counted to obtain the connected edge vector, and the degree of overlap between a node and its neighboring nodes is calculated to obtain the overlap vector.

[0043] Based on the feature extraction rules, the frequency change vector, the connection edge vector, and the overlap vector are combined to obtain the node feature vector. The node feature sequence is constructed by concatenating all the node feature vectors on the traversal path in the traversal order.

[0044] A multi-layer neural network is constructed as the basic structure of the directional analysis model. The node feature sequence is encoded as the model input feature, and the historical case embedding vector is encoded as the model training label. The embedding loss is obtained by calculating the distance between the node feature sequence and the historical case embedding vector through forward propagation. The neural network parameters are updated by optimizing the embedding loss through backpropagation. The trained neural network is used as the directional analysis model.

[0045] Inputting the fund correlation graph into a targeted analysis model for node risk scoring, the output includes risk account identifiers and correlation path analysis reports:

[0046] Calculate the transaction frequency and transaction amount between nodes in the fund relationship graph to construct a transaction feature matrix, and calculate the fund flow direction and flow between nodes to construct a fund flow feature matrix;

[0047] Input the transaction feature matrix and the capital flow feature matrix into the targeted analysis model, extract the time series change pattern from the transaction feature matrix, extract the spatial distribution pattern from the capital flow feature matrix, and calculate the node risk score based on the time series change pattern and the spatial distribution pattern.

[0048] A risk distribution curve is constructed based on the node risk score. The mean and variance of the risk distribution curve are calculated to determine the risk threshold. Nodes with a risk score greater than the risk threshold are extracted as risk account identifiers.

[0049] Calculate the fund transfer path between nodes corresponding to risk account identifiers, and extract the risk scores and fund flow characteristics between nodes on the transfer path to construct a path score vector;

[0050] The risk level of a path is calculated based on the path score vector, and a related path analysis report is generated by sorting the paths according to their risk levels.

[0051] A second aspect of the present invention provides an electronic device, comprising:

[0052] processor;

[0053] Memory used to store processor-executable instructions;

[0054] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0056] In this embodiment, efficient interaction between the user and the system is achieved by receiving natural language analysis instructions and decomposing tasks using a large language model. Fund flow characteristics are encoded into case embedding vectors and matched with historical cases, effectively utilizing historical case experience and improving analysis efficiency and accuracy. By calculating the difference in flow characteristics and combining user feedback to select priority paths, the problem of easily overlooking non-mainstream fund transfer paths in traditional methods is solved. The user-identified priority paths are fed back to the model for targeted learning, enabling continuous optimization and autonomous evolution of the analysis model, constantly improving its analytical capabilities with use. The targeted analysis model scores the risks of nodes and generates analysis reports, providing intuitive risk identification results and decision-making basis, significantly improving the accuracy and interpretability of fund risk analysis. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the intelligent financial analysis method based on multi-agent collaboration according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic flowchart of the fund correlation graph analysis method according to an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0060] The technical solution of the present invention will be described in detail below with reference to 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.

[0061] Figure 1This is a flowchart illustrating an embodiment of the intelligent financial analysis method based on self-evolving technology of the present invention. Figure 1 As shown, the method includes:

[0062] Receive natural language analysis instructions for the target fund account, input them into the large language model, and output a task decomposition sequence;

[0063] Extract the fund flow characteristics of the target fund account and encode them into a case embedding vector. Select the historical case embedding vector with the smallest distance from the case embedding vector in the preset case profile vector library.

[0064] Historical case embedding vectors are decoded into feature weight distributions. The feature weight distributions are optimized based on historical interaction records. The optimized weight distributions are coupled to the task decomposition sequence to generate weighted execution instructions.

[0065] Based on the authorized execution instruction, the transaction data of the target fund account is correlated and calculated to generate a fund correlation graph. Based on the fund correlation graph, multiple transfer paths are identified and the flow characteristic difference is calculated. When the flow characteristic difference exceeds the preset difference threshold, the transfer path information is sent to the user and the priority path identifier is received.

[0066] Priority path identifiers are mapped to the fund association graph and the corresponding subgraph structure is extracted. The subgraph structure is fed back to the large language model to output the targeted task decomposition sequence. The subgraph structure is then subjected to depth traversal based on the targeted task decomposition sequence to extract node feature sequences and jointly train them with historical case embedding vectors to generate a targeted analysis model.

[0067] Input the fund relationship graph into the targeted analysis model to score node risk, and output risk account identifiers and relationship path analysis reports.

[0068] In one alternative implementation, a natural language analysis instruction for a target fund account is received, input into a large language model, and the output task decomposition sequence includes:

[0069] The natural language analysis instructions of the target fund account are segmented to obtain the instruction word sequence. The contextual correlation between words in the instruction word sequence is calculated to generate the instruction semantic matrix. Transaction time sequence features, fund flow features and account relationship features are extracted from the instruction semantic matrix and fused to generate the instruction feature vector.

[0070] Construct a large language model processing structure that includes a word vector encoding layer, a multi-head attention layer, a feedforward computation layer, and a sequence decoding layer;

[0071] The instruction feature vector is input into the word vector encoding layer to generate the initial semantic code. The initial semantic code is input into the multi-head attention layer to calculate the feature weight distribution to obtain the weighted semantic vector. The weighted semantic vector is input into the feedforward calculation layer to perform dimensional transformation to obtain the semantic representation vector. The semantic representation vector is input into the sequence decoding layer to extract the time-series dependency information, account association information, and fund flow information, and combined into step-by-step analysis instructions. The information in the step-by-step analysis instructions is matched with the fund tracking sequence in historical fund analysis cases.

[0072] The analysis execution chain is constructed based on the degree of matching. The step-by-step analysis instructions are reorganized according to the node order in the analysis execution chain, and the task decomposition sequence is output.

[0073] The natural language processing system based on fund account analysis aims to achieve intelligent parsing and task decomposition of financial transaction instructions. First, it receives a natural language analysis instruction for the target fund account, which can be text like "Analyze the fund flow of account A over the past three months and identify abnormal transactions." The received instruction undergoes word segmentation, dividing the continuous text into meaningful word units to form an instruction word sequence. The word segmentation process uses a bidirectional maximum matching algorithm combined with a financial-specific dictionary to improve the accuracy of financial terminology recognition. For example, "Analyze the fund flow of account A over the past three months" is segmented into ["analyze", "account A", "near", "three months", "of", "fund flow"].

[0074] The contextual association degree between words in the instruction word sequence is calculated to construct an instruction semantic matrix. This matrix reflects the semantic dependencies between words. The co-occurrence frequency between words is calculated using the sliding window method, and the semantic association matrix is ​​generated by combining the cosine similarity of word vectors. Taking a window size of 3 as an example, for each word, the semantic association between its three preceding and following words is examined, forming an n×n matrix, where n is the length of the word sequence. Each element in the matrix represents the semantic association strength between corresponding word pairs.

[0075] Three key features are extracted from the instruction semantic matrix: transaction timing features, fund flow features, and account relationship features. Transaction timing features are obtained by identifying time-related terms and their contextual relationships, such as time descriptive words like "last three months," "weekly," and "daily average." Fund flow features are extracted by identifying terms indicating fund movements, such as amount, transfer, and receipt / payment, and their related terms. Account relationship features focus on the dependencies between account identifiers, such as the transfer-in / transfer-out relationship in "transfer from account A to account B." The extracted features are combined into an instruction feature vector using a feature fusion algorithm. This vector contains the core semantic information of the instruction, and its dimension is typically set to 256 or 512 to retain sufficient semantic information.

[0076] A large language model processing structure is constructed, comprising four core layers: a word embedding layer, a multi-head attention layer, a feedforward computation layer, and a sequence decoding layer. The word embedding layer employs a pre-trained word embedding model, mapping each word to a 300-dimensional vector space, while combining positional encoding to preserve word order information. For specific financial terms, domain-adaptive fine-tuned word vectors are used to enhance expressive power.

[0077] The generated instruction feature vector is input into the word vector encoding layer. After the initial semantic encoding is formed, it is passed to the multi-head attention layer. The multi-head attention layer sets up 8 attention heads, each of which independently calculates the attention distribution between the query, key, and value, capturing semantic relationships at different levels. By calculating the feature weight distribution, a weighted semantic vector is generated, which strengthens the expression of key information.

[0078] The weighted semantic vector is then fed into a feedforward computation layer, which contains two linear transformations and a ReLU activation function. This layer reduces the dimensionality of the input vector and then increases it, achieving a non-linear transformation and enhancing the model's ability to express complex semantic relationships. The resulting semantic representation vector retains the core semantic information of the instruction while improving expressive efficiency.

[0079] The semantic representation vector enters the sequence decoding layer, which extracts time-series dependency information, account association information, and fund flow information through a bidirectional Long Short-Term Memory (LSTM) network structure. Time-series dependency information includes the analysis time range, frequency requirements, and chronological order; account association information includes transfer relationships and ownership relationships between accounts; and fund flow information includes the source, destination, and trend of fund changes. This information is combined into structured step-by-step analysis instructions, such as "Step 1: Extract the transaction records of account A for the most recent three months; Step 2: Analyze the distribution of counterparties...".

[0080] The generated step-by-step analysis instructions are matched with fund tracking sequences in the historical fund analysis case library. The matching process uses a combination of semantic similarity calculation and task pattern recognition to find the most similar historical analysis path for each analysis step. The similarity calculation comprehensively considers the cosine similarity between the instruction semantic vector and the historical case vector, the matching degree of key features, and the similarity of the task structure.

[0081] An analysis execution chain is constructed based on the degree of matching. The chain consists of multiple nodes, each corresponding to a specific analysis task. According to the dependencies and execution order between nodes, the step-by-step analysis instructions are reorganized and optimized to form the final task decomposition sequence. The optimization process considers the parallelism of tasks, resource dependencies, and execution efficiency, ensuring that the generated task sequence not only conforms to the semantic requirements of the original instructions but also has an efficient execution structure.

[0082] The final output task breakdown sequence contains clear execution steps, required data sources, analysis methods, and expected results, providing structured guidance for subsequent financial analysis work and significantly improving analysis efficiency and accuracy.

[0083] In one optional implementation, the fund flow characteristics of the target fund account are extracted and encoded into a case embedding vector. The historical case embedding vector with the smallest distance from the pre-defined case profile vector library is selected, including:

[0084] Extract time-series transaction amounts, transaction frequencies, and counterparty information from the transaction records of the target fund account to construct a transaction attribute matrix;

[0085] The correlation coefficient matrix is ​​obtained by calculating the correlation coefficients between attributes in the transaction attribute matrix. The correlation coefficient matrix is ​​then converted into an adjacency matrix to construct a fund transaction association graph. Multi-layer convolution is then performed on the fund transaction association graph to obtain the node feature matrix.

[0086] The node feature matrix is ​​pooled and compressed to generate a capital flow feature vector.

[0087] The feature transformation of the fund flow feature vector is used to obtain the feature enhancement sequence. The distribution distance between the feature enhancement sequence and the fund flow feature vector is calculated. By minimizing the distribution distance, common features in the fund flow feature vector are extracted to generate the case embedding vector.

[0088] The cosine distance between the case embedding vector and the historical case embedding vector in the preset case profile vector library is calculated using the K-Means algorithm. The cluster center is updated based on the cosine distance, and the historical case embedding vector with the smallest distance from the case embedding vector is selected.

[0089] To identify and analyze potential financial risks, the first step is to acquire complete transaction record data for the target fund account, including the time, amount, type, and counterparty of each transaction. This raw transaction data undergoes preprocessing, including missing value imputation, outlier handling, and data standardization. Following preprocessing, a transaction attribute matrix is ​​constructed, where rows represent transactions at different time points, and columns represent various transaction attribute characteristics. Specifically, this includes time-series transaction amount sequences, transaction frequency characteristics (e.g., daily and weekly transaction frequencies), and counterparty information (e.g., counterparty account type, region, etc.).

[0090] Based on the constructed transaction attribute matrix, the correlation coefficients between each attribute are calculated, forming a correlation coefficient matrix C. The Pearson correlation coefficient method is used to measure the strength of the linear relationship between attributes. If the absolute value of the correlation coefficient between two attributes exceeds a preset threshold (e.g., 0.6), then a significant association is considered to exist between the two attributes. The correlation coefficient matrix is ​​then converted into a binary adjacency matrix A. The specific conversion rule is: if the absolute value of the correlation coefficient between attribute i and attribute j is greater than the threshold, then A[i][j] = 1; otherwise, A[i][j] = 0.

[0091] A transaction association graph is constructed based on the adjacency matrix, where nodes represent transaction attributes and edges represent the relationships between attributes. Multi-layer graph convolution is performed on the association graph to extract node features. In the first convolution layer, for each node, the features of its neighboring nodes are aggregated, and combined with the node's own features to update the node features. For the h-th convolution layer, node feature updates are implemented through an aggregation function, integrating the feature information of neighboring nodes in the (h-1)-th layer. After H convolution operations, the final node feature matrix is ​​obtained.

[0092] For the node feature matrix, a pooling compression method is used to generate a fund flow feature vector. The pooling process includes two methods: global average pooling and global max pooling, which extract the average and most salient information of the features, respectively. The results of these two pooling methods are concatenated to form a comprehensive feature vector, which contains the core features of the target account's fund flow.

[0093] To enhance feature representation, a feature transformation is performed on the fund flow feature vector F to obtain the enhanced feature sequence E. A nonlinear transformation function, such as a multilayer perceptron network, is used to map the fund flow feature vector, while random noise is introduced to enhance feature diversity and robustness. The distribution distance between the enhanced feature sequence and the original fund flow feature vector is calculated, and the difference between the two is measured using JS divergence or KL divergence. By minimizing the distribution distance, common features in the fund flow feature vector are extracted, noise and redundant information are removed, and a lower-dimensional, more refined case embedding vector is generated.

[0094] The generated case embedding vectors are compared with the set of historical case embedding vectors in a pre-defined case profile vector library based on similarity. A K-Means clustering algorithm is used, initially selecting k random center points and calculating the cosine distance between the case embedding vector and each center point. Based on the minimum distance principle, the case embedding vector is assigned to the nearest cluster center, and the cluster centers are updated. After multiple iterations and optimizations, the set of historical case embedding vectors with the smallest distance to the case embedding vector is selected.

[0095] In specific application scenarios, this method can be used to quickly identify suspicious accounts similar to known cases. When abnormal transaction behavior of an account is detected, its fund flow characteristics are extracted and encoded into a case embedding vector, which is then matched with a labeled case library. If highly similar historical cases are found, the risk rating of the account is increased, triggering further investigation procedures.

[0096] The above technical solution enables the automatic extraction of key features from the transaction data of the target fund account and intelligent matching with historical cases, providing efficient and reliable technical support for financial risk identification and prevention.

[0097] In one optional implementation, the historical case embedding vector is decoded into a feature weight distribution, the feature weight distribution is optimized based on historical interaction records, and the optimized weight distribution is coupled to the task decomposition sequence to generate weighted execution instructions, including:

[0098] Historical cases are embedded in vector decomposition to obtain a sequence of feature component values. The frequency distribution value is obtained by counting the number of times the feature component values ​​appear in historical successful cases. The information distribution value is obtained by calculating the information overlap between the feature component values ​​and transaction behavior. The association distribution value is obtained by calculating the temporal correlation of the feature values.

[0099] The frequency distribution value, information distribution value, and correlation distribution value are normalized and weighted to obtain the initial feature weight distribution;

[0100] Extract task interruption records and resource consumption records from historical interaction records. Identify interruption locations in task interruption records and peak points in resource consumption records based on the initial feature weight distribution. Calculate the occurrence probability of interruption locations and the utilization rate of resource peaks. Adjust the value of the initial feature weight distribution based on the interruption probability and utilization rate to obtain the optimized feature weight distribution.

[0101] The optimized feature weight distribution is sequentially mapped to the execution steps in the task decomposition sequence. The data transfer size and processing time between execution steps are calculated. The execution steps are prioritized and allocated computing resources according to the transfer size and processing time, and weighted execution instructions are generated.

[0102] First, historical case data needs to be acquired, and the case information is transformed into a multi-dimensional vector representation using a vector embedding model. The vector embedding model employs a deep learning architecture, including an input layer, multiple hidden layers, and an output layer. The hidden layers use an attention mechanism to capture the correlations between various elements of the case. When decomposing the embedded vectors to obtain a sequence of feature component values, principal component analysis is used to decompose the vectors and extract the main feature components. Taking a financial analysis task as an example, the embedding vector of a transaction case may include dimensions such as transaction amount, transaction frequency, and transaction object; after decomposition, the feature component values ​​of each dimension are obtained.

[0103] For processing the feature component value sequence, the system calculates three types of distribution values. When calculating the frequency distribution value by counting the occurrences of feature component values ​​in historical successful cases, a sliding window statistical method is used, with a window size of 30 days and a step size of 1 day, to count the frequency of each feature component in successfully identified cases of abnormal fund flows. The information distribution value is calculated using the mutual information entropy method, which measures the degree of information overlap by calculating the mutual information between feature components and transaction behavior.

[0104] For calculating the temporal correlation degree, a temporal correlation analysis method based on a vector autoregression (VAR) model is adopted. First, the feature component values ​​are arranged in chronological order to form a time series dataset. For each feature component, a VAR model is constructed, which describes how the feature component value at the current moment depends on the feature component values ​​at several past moments. In the specific implementation, an appropriate order is chosen, generally ranging from 1 to 5, determined by the minimum information criterion. Then, the model parameters, i.e., the coefficient matrix, are estimated. The element values ​​in this matrix reflect the temporal correlation strength between different feature components. For each pair of feature components, their influence is calculated, i.e., by taking the weighted sum of the absolute values ​​of corresponding elements in all coefficient matrices, where the weights decrease with increasing time intervals. A correlation strength matrix is ​​constructed, where each element represents the degree of influence between feature components. Row normalization is performed on the correlation strength matrix to obtain the temporal correlation degree matrix of the feature components. The summation and normalization of each column of the temporal correlation degree matrix yields a correlation distribution vector reflecting the importance of the temporal correlation of each feature component.

[0105] Assume three feature components: transaction amount, transaction frequency, and anomaly marker. By establishing a vector autoregressive model, it is found that the anomaly marker at the current time step has a strong correlation with the transaction amount and transaction frequency at the previous time step, with coefficients of 0.7 and 0.5, respectively, while its correlation with the anomaly marker itself at the previous time step is 0.3. After applying time-decay weights to these coefficients, a correlation strength matrix is ​​obtained. After normalization, the temporal correlation distribution values ​​for the anomaly marker are 0.6, for transaction frequency 0.3, and for transaction amount 0.1, indicating that when predicting abnormal behavior, the historical value of the anomaly marker is the most important, followed by transaction frequency, and lastly, transaction amount.

[0106] The initial feature weight distribution is generated using a normalized weighting method. Normalization functions are applied to the frequency distribution, information distribution, and association distribution values ​​respectively to ensure that the sum of each distribution value is 1. Then, the three types of distribution values ​​are weighted and fused according to preset weight coefficients. These weight coefficients are determined through cross-validation: the frequency distribution value has a weight of 0.4, the information distribution value has a weight of 0.35, and the association distribution value has a weight of 0.25. The weighted distribution values ​​are then normalized to form the initial feature weight distribution.

[0107] To optimize the feature weight distribution, task interruption records and resource consumption records are extracted from historical interaction records. Task interruption records include information such as the interruption time, interruption reason, and processing status at the time of interruption. Resource consumption records contain time-series data on resource usage, such as CPU utilization, memory usage, and network bandwidth consumption. Interruption locations in the task interruption records are identified based on the initial feature weight distribution, and clustering algorithms are used to divide these locations into high-frequency and low-frequency interruption regions. For resource consumption records, a peak detection algorithm is used to identify peak resource usage points, and the resource utilization rate at each peak point is calculated. The interruption probability is calculated using maximum likelihood estimation, statistically analyzing the frequency of task interruptions within a specific feature weight value range. The resource utilization rate is calculated based on the ratio of peak resource usage points to the upper limit of available system resources.

[0108] The optimization of the feature weight distribution employs gradient descent to construct the objective function, using interruption probability and resource utilization as penalty terms. The initial feature weight distribution is iteratively optimized to minimize the objective function value. During iteration, the initial learning rate is set to 0.01, and an adaptive adjustment strategy is used, gradually decreasing the learning rate as the number of iterations increases. The iteration process terminates when the objective function value changes by less than a preset threshold of 0.001 after five consecutive iterations, yielding the optimized feature weight distribution.

[0109] In coupling the optimized feature weight distribution to the task decomposition sequence, the task is first decomposed into multiple execution steps. For a financial analysis task, these steps include data preprocessing, feature extraction, pattern recognition, and risk assessment. The optimized feature weight distribution is mapped to each execution step according to feature relevance, and the data transfer size and processing time between steps are calculated. The data transfer size is calculated based on the amount of data transferred between steps, and the processing time is obtained through statistical analysis of historical execution data. Based on the transfer size and processing time, the execution steps are prioritized using a critical path algorithm, allocating more computing resources to higher-priority steps. Resource allocation is based on a weighted proportional method, dynamically allocating resources according to step priority and resource requirements. The final weighted execution instruction includes the execution step sequence, the priority of each step, the resource allocation scheme, and the dependencies between steps.

[0110] In this embodiment, the feature weight distribution can be automatically adjusted based on successful experiences from historical cases, improving the accuracy and efficiency of financial analysis. The optimized feature weight distribution can identify and avoid potential risks during task execution, reducing the probability of task interruption and improving system stability. Intelligent adjustment of resource allocation enables efficient use of computing resources, reduces resource waste, and adapts to financial analysis tasks of varying complexity. The generation of weighted execution instructions makes the task execution process more flexible and efficient, dynamically adjusting execution strategies based on real-time conditions, achieving continuous optimization and autonomous evolution of financial analysis capabilities.

[0111] like Figure 2 The diagram illustrates the flow of the fund correlation graph analysis method in this embodiment.

[0112] In one optional implementation, transaction data of the target fund account is correlated and calculated according to the authorized execution instruction to generate a fund correlation graph. Multiple transfer paths are identified based on the fund correlation graph, and the flow characteristic difference is calculated. When the flow characteristic difference exceeds a preset difference threshold, transfer path information is sent to the user, and a priority path identifier is received, including:

[0113] Extract the payment account identifier, receiving account identifier, transaction amount, and transaction time from the transaction record;

[0114] The payment account identifier and the receiving account identifier are constructed as graph nodes, and the total transaction amount on the node is marked as the node weight;

[0115] The rate of change of transaction amount between adjacent nodes is calculated according to the time window, and the rate of change of transaction amount is constructed as the energy transfer feature between nodes; a periodic energy curve is constructed based on the energy transfer feature, and the periodic energy curve is mapped as the connection edge weight between nodes; a capital correlation graph is generated through node weight and connection edge weight.

[0116] Identify the reachable path from the source node to the target node, calculate the ratio of the node weight to the edge weight on the reachable path to obtain the energy conversion rate, determine the path breakpoint based on the abrupt change in the energy conversion rate, split the path into sub-paths based on the path breakpoints, cluster the energy conversion rates of the sub-paths, and generate a set of transfer paths.

[0117] Calculate the energy conversion rate difference between the first and last nodes of the sub-path, construct the path feature matrix, and obtain the flow direction feature difference degree by calculating the loss difference between the sub-paths through the path feature matrix.

[0118] The system compares the flow direction feature difference with a preset difference threshold. When the flow direction feature difference exceeds the preset difference threshold, it sends the transfer path set information to the user and receives the priority path identifier returned by the user.

[0119] The fund relationship graph construction and transfer path identification process first receives a weighted execution instruction, which includes the target fund account identifier and calculation parameters. These parameters include the time window size, preset difference threshold, and energy conversion rate threshold. Historical transaction data related to the target fund account is then obtained via an interface.

[0120] Basic information such as payment account identifier, receiving account identifier, transaction amount, and transaction time are extracted from transaction data. For each transaction record, the account identifier is extracted as a graph node, and the amount and time information related to the transaction are also recorded. For example, in a transaction record where account A transfers 1,000 yuan to account B, account A and account B are extracted as graph nodes respectively.

[0121] When constructing graph nodes, nodes with the same account identifier are merged, and the total transaction amount of that account across all transactions is calculated. This total amount is used as the node's weight. For example, if account A transferred out 5,000 yuan and transferred in 3,000 yuan in multiple transaction records, the weight of the node corresponding to account A would be 8,000 yuan. The node weight reflects the account's activity level and the amount of funds involved in the entire fund flow network.

[0122] The energy transfer characteristics between nodes are calculated based on time windows. The entire transaction timeframe is divided into several time windows, such as 7 days per window. Within each time window, the rate of change in transaction amount between adjacent nodes is calculated. Specifically, for nodes A and B, the ratio of the transaction amount from A to B to the weight of node A is calculated to obtain the fund transfer rate from A to B. Similarly, the fund transfer rate from B to A is calculated. These fund transfer rates are then sorted according to the time windows to form the temporal characteristics of energy transfer between nodes.

[0123] The energy transfer time-series characteristics are converted into periodic energy curves using Fourier transform, allowing for the identification of frequency characteristics and periodic fluctuations. Based on the characteristic parameters of the periodic energy curves, such as amplitude, frequency, and phase difference, the strength of the connections between nodes is calculated, serving as the weights of the connection edges. For example, if there are regular financial transactions between two nodes, the connection edge has a higher weight; if the financial flows exhibit random characteristics, the connection edge has a lower weight.

[0124] The calculated node weights and edge weights are combined to generate a complete fund relationship graph. In the graph, nodes represent fund accounts, and the node size reflects the total amount of funds in the account; edges represent the fund flow relationship between accounts, and the thickness of the edges reflects the intensity and regularity of the fund flow.

[0125] Based on the generated capital relationship graph, all reachable paths from the source node to the target node are identified using depth-first search or breadth-first search algorithms. For each reachable path, the energy conversion rate between adjacent nodes on the path is calculated, which is the ratio of the weight of the next node to the weight of the previous node divided by the weight of the connecting edge. The energy conversion rate reflects the increase or decrease in capital during the transfer process.

[0126] The energy conversion efficiency along the path is analyzed to identify locations where abrupt changes occur (i.e., where the change in efficiency exceeds the energy conversion efficiency threshold), and these locations are marked as path breakpoints. Based on the identified path breakpoints, the complete path is divided into multiple sub-paths. The division into sub-paths ensures that the energy transfer characteristics within each sub-path are relatively consistent.

[0127] K-means clustering is applied to the energy conversion rate sequences of all sub-paths to group sub-paths with similar characteristics into one class. Each cluster represents a typical transfer path pattern, and these patterns constitute a set of transfer paths. For example, different types of transfer paths, such as "fast direct," "multi-level intermediate transfer," and "cyclical," may be identified.

[0128] For each sub-path, the energy conversion rate difference between its first and last nodes is calculated to construct a path feature matrix. Each element in the path feature matrix represents the difference in energy conversion characteristics between two sub-paths. By calculating the standard deviation or variance of the elements in the path feature matrix, the overall flow characteristic difference is obtained. The flow characteristic difference reflects the degree of inconsistency in the fund flow characteristics between different transfer paths.

[0129] The calculated flow characteristic difference is compared with a preset difference threshold. When the flow characteristic difference exceeds the preset difference threshold, it indicates that there are significantly different fund flow patterns, which may suggest abnormal fund flows. At this time, the transfer path set information is sent to the user, including key information such as the node sequence, energy conversion characteristics, and fund flow of each path.

[0130] After receiving the set of transfer path information, users select a preferred path based on their business needs and risk preferences and return a preferred path identifier. The system receives the preferred path identifier returned by the user and marks that path as a recommended path for subsequent fund monitoring and risk management.

[0131] In practical applications, this method can be used in multiple fields such as optimizing cross-border payment paths for banks, monitoring abnormal fund flows by financial regulators, and analyzing the fund efficiency of corporate finance departments, effectively improving the transparency and controllability of fund flows.

[0132] In one optional implementation, priority path identifiers are mapped to a funding relationship graph and the corresponding subgraph structure is extracted. The subgraph structure is then fed back to a large language model, and the output targeted task decomposition sequence includes:

[0133] The priority path identifier is mapped to the fund association graph, and the transaction amount sequence of the target node set is extracted to obtain the fund flow curve;

[0134] Extract the fluctuation period and fluctuation amplitude from the capital flow curve, identify the position with the largest capital accumulation and the position with the largest capital dispersion within the fluctuation period, and construct a set of feature nodes;

[0135] Calculate the connection edge weight sequence between feature node sets to construct the node connection strength distribution, determine the subgraph structure boundary based on the variance of the node connection strength distribution, and generate the subgraph structure;

[0136] The transaction features of the feature node set and the transitive features of the connecting edges in the subgraph structure are used to construct node representation vectors and edge representation vectors, respectively. The node representation vectors and edge representation vectors are combined into a structural feature matrix. A scene description sequence is generated based on the structural feature matrix. The scene description sequence is input into a large language model and the output is a targeted task decomposition sequence.

[0137] First, priority path identifiers are mapped to a fund relationship graph, and the transaction amount sequence of the target node set is extracted to form a fund flow curve. Specifically, fund relationship graph data containing multiple account nodes and transaction relationships is received, along with priority path identifiers containing identifiers of key monitored accounts. An identifier matching algorithm is used to locate all account nodes on the priority path in the graph, and these nodes are marked as the target node set. For the target node set, all transaction records are extracted in chronological order, including transaction time, transaction amount, and transaction direction. These transaction records are sorted according to the time dimension to generate a fund flow curve showing the change in transaction amount over time. For example, for a certain account group, the daily transaction amount changes over six months are extracted to form a fund flow curve containing 180 data points.

[0138] This method extracts the fluctuation cycle and amplitude from the capital flow curve, identifying the locations of maximum capital accumulation and maximum capital dispersion within each fluctuation cycle to construct a set of feature nodes. Specifically, wavelet transform or Fourier transform is applied to the capital flow curve to decompose its frequency components and identify the main cyclical fluctuations. The first and second derivatives of the capital flow curve at each time point are calculated to determine the curve's rate of ascent and acceleration. Within each identified fluctuation cycle, the accumulated capital is calculated using a cumulative sum function to pinpoint the time point when the accumulated capital reaches its maximum value and the corresponding account node. Similarly, by calculating capital dispersion indicators (such as entropy or the Gini coefficient), the time point with the maximum capital dispersion and the corresponding account node are located. The set of account nodes corresponding to these specific time points is defined as the feature node set. For example, in a capital flow cycle analysis, a fluctuation pattern with a 15-day cycle might be identified, and two to three key account nodes in each cycle could be found as feature nodes.

[0139] By analyzing the frequency, amount, and time interval of direct transactions between characteristic nodes, a weighted directed graph structure is constructed. For each pair of characteristic nodes, a comprehensive weight value is calculated, taking into account factors such as transaction frequency, cumulative amount, and time correlation. A sequence distribution of all edge weights is generated, and the statistical characteristics of this distribution, including mean, variance, and quantiles, are calculated. A threshold rule is set, and significant edges are selected based on the variance or specific quantiles of the weight distribution. The selected significant edges and their associated characteristic nodes form a subgraph structure, representing the core path and key nodes of fund flow. For example, when the variance of connection strength exceeds a preset threshold, edges with weights above the upper quartile are retained, potentially resulting in a subgraph structure containing 20 nodes and 35 edges.

[0140] Multidimensional transaction features are extracted from each feature node, including transaction frequency, amount statistics, time patterns, and diversity of transaction objects, forming a node representation vector. Transmission features of each connecting edge are extracted, including transaction directionality, stability, periodicity, and anomaly degree, forming an edge representation vector. All node and edge representation vectors are combined into a structural feature matrix, which comprehensively describes the structure and behavioral characteristics of the subgraph. Based on the structural feature matrix, a text sequence describing the fund flow scenario is constructed using template generation or natural language generation techniques. This scenario description sequence covers elements such as the scale, cycle, key node roles, and anomaly patterns of fund flows. The scenario description sequence is input into a large language model, which outputs a targeted task decomposition sequence containing specific investigation steps based on the identified fund flow patterns. For example, the scenario description might include patterns such as "funds circulate among three core accounts, with large inflows every Friday and dispersed outflows on Mondays," and the large language model generates specific task instructions such as "verify the relationship between core account holders" and "analyze the source of funds on Fridays."

[0141] By following the steps above, the method of mapping priority path identifiers to the fund relationship graph and extracting the corresponding subgraph structure can effectively identify key nodes and paths in complex fund networks and generate targeted investigation task instructions, thereby improving the efficiency and accuracy of financial supervision.

[0142] In one alternative implementation, the subgraph structure is depth-traversed according to the directional task decomposition sequence, node feature sequences are extracted, and jointly trained with historical case embedding vectors to generate a directional analysis model, including:

[0143] Extract the node access order from the targeted task decomposition sequence to construct the node traversal order, and extract feature calculation rules to construct feature extraction rules;

[0144] The subgraph structure is subjected to a depth-first search according to the node traversal order. During the traversal, the rate of change of the frequency of transactions before and after a node is calculated to obtain the frequency change vector, the number of connected edges of a node is counted to obtain the connected edge vector, and the degree of overlap between a node and its neighboring nodes is calculated to obtain the overlap vector.

[0145] Based on the feature extraction rules, the frequency change vector, the connection edge vector, and the overlap vector are combined to obtain the node feature vector. The node feature sequence is constructed by concatenating all the node feature vectors on the traversal path in the traversal order.

[0146] A multi-layer neural network is constructed as the basic structure of the directional analysis model. The node feature sequence is encoded as the model input feature, and the historical case embedding vector is encoded as the model training label. The embedding loss is obtained by calculating the distance between the node feature sequence and the historical case embedding vector through forward propagation. The neural network parameters are updated by optimizing the embedding loss through backpropagation. The trained neural network is used as the directional analysis model.

[0147] First, two key pieces of information are extracted from the targeted task decomposition sequence: node access order and feature calculation rules. The node access order is the specified fund flow path in the task sequence, determining the traversal order of the subgraph structure; the feature calculation rules are a set of calculation methods for node attributes, used in the subsequent feature extraction process. In the context of fund analysis, the targeted task decomposition sequence can be described as a task description such as "tracking fund flows from the starting account, prioritizing the analysis of large-value fund transaction paths, and focusing on nodes with a transaction frequency change rate greater than 50%". By parsing the task description, "starting account → large-value transaction nodes → nodes with high transaction frequency change rates" is extracted as the node traversal order, and "calculating transaction frequency change rates, counting the number of connected edges, and calculating node overlap" are extracted as feature extraction rules.

[0148] A depth-first search (DFS) is performed on the subgraph structure, a subset of the fund flow network, based on the extracted node traversal order. The DFS is implemented recursively, starting from a specified initial node and proceeding deeper along the fund flow direction until a preset search depth is reached or a termination condition is encountered. The search depth is typically set to 5 to 10 levels, and termination conditions include reaching a specific type of node or the fund amount falling below a threshold. During the traversal, three types of feature vectors are calculated: frequency change vector, connection edge vector, and overlap vector. The frequency change vector reflects the temporal change in node transaction activity, specifically calculated as the ratio of the number of transactions in the current period to the number of transactions in the previous period, with a period length of 7 days. For example, if an account node had 10 transactions in the previous week and 20 transactions in the current week, the frequency change rate is 1.0. The connection edge vector describes the connection complexity of a node, statistically calculated by counting the number of incoming and outgoing edges, and the ratio of incoming to outgoing edges. The overlap vector measures the degree of business association between a node and its neighboring nodes. It is calculated by the proportion of repeated occurrences among the node's trading partners, specifically by the ratio of the intersection to the union of the sets of neighboring nodes.

[0149] Based on the feature extraction rules, the three types of vectors are combined to construct node feature vectors. The combination process involves two steps: vector concatenation and weighted fusion. Vector concatenation sequentially combines the frequency change vector, connection edge vector, and overlap vector into a single long vector. Weighted fusion weights each feature component according to the weight configuration in the feature extraction rules. The weight configuration is determined based on the task's priorities, with a default weight ratio of 40% for frequency change, 30% for connection edge count, and 30% for node overlap. The node feature vector typically has 15 to 30 dimensions, containing information on transaction behavior, relationship structure, and business attributes. All node feature vectors along the traversal path are concatenated in traversal order to form a node feature sequence. If the traversal depth is 5 and each node feature vector has a dimension of 20, the final node feature sequence will have a dimension of 100.

[0150] A multi-layer neural network is constructed as the basic structure of the directional analysis model. This network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the node feature sequence, with the number of nodes equal to the dimension of the feature sequence. The hidden layers employ a multi-layer perceptron structure, typically containing 3 to 5 layers, with the number of nodes decreasing layer by layer, from half the number of nodes in the input layer to twice the number of nodes in the output layer. The number of nodes in the output layer is the same as the dimension of the historical case embedding vector, used to generate an output that matches the format of the historical case embedding vector. The activation function uses the ReLU function in the hidden layer to enhance the non-linear expressive power, and the Tanh function in the output layer to map the output to the same value range as the historical case embedding vector.

[0151] When encoding node feature sequences as model input features, positional encoding techniques are employed to enhance the representation of positional information within the sequences. Positional encoding uses sine and cosine functions to generate fixed-pattern positional information vectors, which are then added to the node feature vectors to form an enhanced feature vector containing positional information. Historical case embedding vectors serve as model training labels, guiding model learning through similarity calculations. The similarity calculation uses a cosine similarity metric to calculate the similarity between the output vector of the node feature sequence after neural network transformation and the historical case embedding vector.

[0152] During model training, the loss function is calculated via forward propagation. The loss function is designed as a weighted sum of the embedding loss and the regularization term. The embedding loss uses mean squared error or cross-entropy loss to calculate the difference between the model output and the target embedding vector. L2 regularization is used to prevent overfitting. The Adam optimizer is employed for optimization, with an initial learning rate of 0.001 and a learning rate decay strategy, multiplying the learning rate by 0.95 every 100 batches. The batch size is set to 64, and the number of training iterations ranges from 100 to 300. An early stopping strategy is used based on validation set performance to avoid overfitting. After each training round, the model performance is evaluated using the validation set. Training stops when the validation loss no longer decreases for five consecutive rounds, and the optimal model parameters are saved as a model for targeted analysis.

[0153] In this embodiment, node features can be adaptively extracted and organized according to specific analysis tasks, enabling accurate tracking and analysis of fund flows. The targeted analysis model can learn the characteristic patterns of historical successful cases, continuously optimizing its ability to identify suspicious financial activities. A depth-first search-based traversal strategy makes the analysis process more efficient, avoiding the waste of computational resources in full-graph analysis. The construction of multi-dimensional feature vectors achieves a comprehensive characterization of fund flow behavior, improving the detection rate of abnormal patterns.

[0154] In one optional implementation, the fund correlation graph is input into a targeted analysis model for node risk scoring, and the output includes risk account identifiers and correlation path analysis reports:

[0155] Calculate the transaction frequency and transaction amount between nodes in the fund relationship graph to construct a transaction feature matrix, and calculate the fund flow direction and flow between nodes to construct a fund flow feature matrix;

[0156] Input the transaction feature matrix and the capital flow feature matrix into the targeted analysis model, extract the time series change pattern from the transaction feature matrix, extract the spatial distribution pattern from the capital flow feature matrix, and calculate the node risk score based on the time series change pattern and the spatial distribution pattern.

[0157] A risk distribution curve is constructed based on the node risk score. The mean and variance of the risk distribution curve are calculated to determine the risk threshold. Nodes with a risk score greater than the risk threshold are extracted as risk account identifiers.

[0158] Calculate the fund transfer path between nodes corresponding to risk account identifiers, and extract the risk scores and fund flow characteristics between nodes on the transfer path to construct a path score vector;

[0159] The risk level of a path is calculated based on the path score vector, and a related path analysis report is generated by sorting the paths according to their risk levels.

[0160] A transaction feature matrix is ​​constructed by calculating the transaction frequency and transaction amount between nodes in the fund relationship graph. This matrix is ​​a two-dimensional matrix, where rows and columns represent account nodes, and matrix elements represent the transaction characteristics between nodes. Transaction frequency is obtained by counting the number of transactions between two accounts within a specified time window, typically set to 30 days. Transaction amount is obtained by calculating the total transaction amount between two accounts within the specified time window. If there is no direct transaction between two accounts, the corresponding matrix element is zero. The elements in the transaction feature matrix are calculated as a weighted sum of transaction frequency and transaction amount. The weight ratio is determined according to business needs, generally with a transaction frequency weight of 0.4 and a transaction amount weight of 0.6. Taking account A and account B as an example, if A and B transact 3 times within 30 days, with a cumulative amount of 50,000 yuan, then the transaction feature value is 0.4×3+0.6×50000=30001.2.

[0161] A fund flow feature matrix is ​​constructed by calculating the direction and volume of fund flows between nodes. The direction of fund flow is represented by a positive or negative sign; fund flow from node i to node j is positive, and vice versa. Fund flow refers to the net flow of funds between two accounts, calculated as the net inflow or outflow of all transaction amounts between the two accounts. The fund flow feature matrix is ​​also a two-dimensional matrix, where rows and columns represent account nodes, and matrix elements represent the fund flow characteristics between nodes. The fund flow feature value is calculated by multiplying the direction of fund flow by the volume of fund flow. For example, if account A transfers three times to account B, totaling 80,000 yuan, and account B transfers once to account A, totaling 30,000 yuan, then the net fund flow between A and B is 50,000 yuan, the direction of fund flow is from A to B, and the fund flow feature value is 50,000.

[0162] The constructed transaction feature matrix and capital flow feature matrix are input into the targeted analysis model to extract temporal variation patterns and spatial distribution patterns. The targeted analysis model employs a dual-channel structure: one channel processes the transaction feature matrix, and the other processes the capital flow feature matrix. Temporal variation pattern extraction utilizes a sliding window technique, sliding a 30-day data window along the time axis to form a sequence of feature matrices for multiple time segments. By comparing the feature differences between different time segments, temporal variation features are extracted. Specifically, a Long Short-Term Memory (LSTM) network is used to process the time-series data. The network structure includes an input layer, an LSTM layer, and an output layer. The LSTM layer contains 64 hidden units and captures temporal dependencies through forget gates, input gates, and output gates. Spatial distribution pattern extraction employs graph convolutional network technology, treating the capital flow feature matrix as an adjacency matrix of a graph. Graph convolutional operations are used to extract the spatial distribution features of nodes. The graph convolutional network contains two graph convolutional layers: the first layer has 32 convolutional kernels, and the second layer has 16 convolutional kernels. Each layer is followed by a ReLU activation function and batch normalization.

[0163] The node risk score is calculated based on the fusion features of temporal variation patterns and spatial distribution patterns. The fusion method combines feature concatenation and a multilayer perceptron. The temporal feature vector output from the LSTM is concatenated with the spatial feature vector output from the graph convolutional network, and then input into a three-layer perceptron. The number of nodes in each layer of the perceptron is 48, 24, and 1, respectively. The last layer outputs a single risk score. The risk score ranges from 0 to 100, with higher scores indicating greater risk. During the risk score calculation, an attention mechanism is used to weight the temporal features of different time windows. The attention weights are calculated using a normalized exponential function, which transforms the input vector into a probability distribution, highlighting the impact of abnormal temporal patterns. The normalized exponential function is implemented by calculating an exponential value for each element of the input vector and then dividing by the sum of all element exponential values ​​to ensure that the sum of all elements in the output vector is 1. For spatial features, node centrality indices are used for weighting, including degree centrality, betweenness centrality, and proximity centrality, to strengthen the role of key nodes in the scoring.

[0164] A risk distribution curve is constructed based on the calculated node risk scores. All node risk scores are sorted by score, and a cumulative distribution function curve is constructed. The mean and variance of the risk distribution curve are calculated, and a risk threshold is determined based on the mean and variance. The risk threshold is calculated by adding the square root of the mean and variance to the variance, multiplying by an adjustment coefficient. The adjustment coefficient is set according to risk sensitivity, typically between 1.5 and 2.5. For example, in a financial relationship graph, if the mean risk score of all nodes is 30, the variance is 100, and the adjustment coefficient is 2, then the risk threshold is 30 + 100. 1 / 2 ×2=50. Nodes with risk scores greater than the risk threshold are extracted as risk account identifiers, forming a risk account set.

[0165] Calculate the fund transfer paths between nodes corresponding to risk account identifiers. An improved breadth-first search algorithm is used, starting from each risk account and searching for fund transfer paths between it and other risk accounts. The improvement lies in considering edge weights during the search process; the weights are determined by the fund flow, prioritizing paths with larger fund flows. The search depth is limited to 5 to avoid excessively long paths. Risk scores and fund flow features between nodes along the transfer path are extracted to construct a path score vector. The path score vector includes elements such as path length, average risk score of nodes along the path, total fund flow along the path, maximum single fund flow along the path, and the time span of the path.

[0166] Path risk level is calculated based on path score vectors. The calculation employs a weighted summation method, weighting each element in the path score vector with the following weights: average risk score 0.3, total fund flow 0.25, maximum single fund flow 0.2, path length 0.15, and time span 0.1. The calculated path risk level is used for path ranking, with higher-risk paths ranked higher. When generating the associated path analysis report, the top 10 paths by risk level are selected. The report includes a list of account nodes along the path, node risk scores, fund flow diagrams, descriptions of abnormal features, and a risk level assessment.

[0167] In this embodiment, the transaction and fund flow characteristics in the fund relationship graph can be comprehensively analyzed to identify abnormal account activity from both temporal and spatial dimensions. The dual-channel structure of the targeted analysis model allows the risk assessment process to consider both changes in the account's own transaction behavior and its position within the overall fund network, improving the accuracy of risk identification. The adaptive threshold determination mechanism for the risk distribution curve enables the risk account screening process to dynamically adjust according to the risk distribution in different scenarios, avoiding false positives and false negatives caused by fixed thresholds. The multi-dimensional assessment of path risk levels enables a comprehensive risk judgment of fund flow paths, helping to uncover hidden fund transfer chains.

[0168] A second aspect of the present invention provides an electronic device, comprising:

[0169] processor;

[0170] Memory used to store processor-executable instructions;

[0171] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0172] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0173] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-evolving intelligent financial analysis method, characterized in that: include: Receive natural language analysis instructions for the target fund account, input them into the large language model, and output a task decomposition sequence; Extract the fund flow characteristics of the target fund account and encode them into a case embedding vector. Select the historical case embedding vector with the smallest distance from the case embedding vector in the preset case profile vector library. Historical case embedding vectors are decoded into feature weight distributions. The feature weight distributions are optimized based on historical interaction records. The optimized weight distributions are coupled to the task decomposition sequence to generate weighted execution instructions. Based on the authorized execution instruction, the transaction data of the target fund account is correlated and calculated to generate a fund correlation graph. Based on the fund correlation graph, multiple transfer paths are identified and the flow characteristic difference is calculated. When the flow characteristic difference exceeds the preset difference threshold, the transfer path information is sent to the user and the priority path identifier is received, including: extracting the payment account identifier, receiving account identifier, transaction amount and transaction time from the transaction record. The payment account identifier and the receiving account identifier are constructed as graph nodes, and the total transaction amount on the node is marked as the node weight; The rate of change of transaction amount between adjacent nodes is calculated according to the time window, and the rate of change of transaction amount is constructed as the energy transfer feature between nodes; a periodic energy curve is constructed based on the energy transfer feature, and the periodic energy curve is mapped as the connection edge weight between nodes; a capital correlation graph is generated through node weight and connection edge weight. Identify the reachable path from the source node to the target node, calculate the ratio of the node weight to the edge weight on the reachable path to obtain the energy conversion rate, determine the path breakpoint based on the abrupt change in the energy conversion rate, split the path into sub-paths based on the path breakpoints, cluster the energy conversion rates of the sub-paths, and generate a set of transfer paths. Calculate the energy conversion rate difference between the first and last nodes of the sub-path, construct the path feature matrix, and obtain the flow direction feature difference degree by calculating the loss difference between the sub-paths through the path feature matrix. The flow direction feature difference is compared with a preset difference threshold. When the flow direction feature difference exceeds the preset difference threshold, the transfer path set information is sent to the user, and the priority path identifier returned by the user is received. Priority path identifiers are mapped to the fund association graph and the corresponding subgraph structure is extracted. The subgraph structure is fed back to the large language model to output the targeted task decomposition sequence. The subgraph structure is then subjected to depth traversal based on the targeted task decomposition sequence to extract node feature sequences and jointly train them with historical case embedding vectors to generate a targeted analysis model. Input the fund relationship graph into the targeted analysis model to score node risk, and output risk account identifiers and relationship path analysis reports.

2. The method according to claim 1, characterized in that, Receive natural language analysis instructions for the target fund account, input them into the large language model, and output a task decomposition sequence including: The natural language analysis instructions of the target fund account are segmented to obtain the instruction word sequence. The contextual correlation between words in the instruction word sequence is calculated to generate the instruction semantic matrix. Transaction time sequence features, fund flow features and account relationship features are extracted from the instruction semantic matrix and fused to generate the instruction feature vector. Construct a large language model processing structure that includes a word vector encoding layer, a multi-head attention layer, a feedforward computation layer, and a sequence decoding layer; The instruction feature vector is input into the word vector encoding layer to generate the initial semantic code. The initial semantic code is input into the multi-head attention layer to calculate the feature weight distribution to obtain the weighted semantic vector. The weighted semantic vector is input into the feedforward calculation layer to perform dimensional transformation to obtain the semantic representation vector. The semantic representation vector is input into the sequence decoding layer to extract the time-series dependency information, account association information, and fund flow information, and combined into step-by-step analysis instructions. The information in the step-by-step analysis instructions is matched with the fund tracking sequence in historical fund analysis cases. The analysis execution chain is constructed based on the degree of matching. The step-by-step analysis instructions are reorganized according to the node order in the analysis execution chain, and the task decomposition sequence is output.

3. The method according to claim 1, characterized in that, Extract the fund flow characteristics of the target fund account and encode them into a case embedding vector. Select the historical case embedding vectors with the smallest distance from the pre-set case profile vector library, including: Extract time-series transaction amounts, transaction frequencies, and counterparty information from the transaction records of the target fund account to construct a transaction attribute matrix; The correlation coefficient matrix is ​​obtained by calculating the correlation coefficients between attributes in the transaction attribute matrix. The correlation coefficient matrix is ​​then converted into an adjacency matrix to construct a fund transaction association graph. Multi-layer convolution is then performed on the fund transaction association graph to obtain the node feature matrix. The node feature matrix is ​​pooled and compressed to generate a capital flow feature vector. The feature transformation of the fund flow feature vector is used to obtain the feature enhancement sequence. The distribution distance between the feature enhancement sequence and the fund flow feature vector is calculated. By minimizing the distribution distance, common features in the fund flow feature vector are extracted to generate the case embedding vector. The cosine distance between the case embedding vector and the historical case embedding vector in the preset case profile vector library is calculated using the K-Means algorithm. The cluster center is updated based on the cosine distance, and the historical case embedding vector with the smallest distance from the case embedding vector is selected.

4. The method according to claim 1, characterized in that, Historical case embedding vectors are decoded into feature weight distributions. These distributions are then optimized based on historical interaction records. The optimized weight distributions are coupled to the task decomposition sequence to generate weighted execution instructions, including: Historical cases are embedded in vector decomposition to obtain a sequence of feature component values. The frequency distribution value is obtained by counting the number of times the feature component values ​​appear in historical successful cases. The information distribution value is obtained by calculating the information overlap between the feature component values ​​and transaction behavior. The association distribution value is obtained by calculating the temporal correlation of the feature values. The frequency distribution value, information distribution value, and correlation distribution value are normalized and weighted to obtain the initial feature weight distribution; Extract task interruption records and resource consumption records from historical interaction records. Identify interruption locations in task interruption records and peak points in resource consumption records based on the initial feature weight distribution. Calculate the occurrence probability of interruption locations and the utilization rate of resource peaks. Adjust the value of the initial feature weight distribution based on the interruption probability and utilization rate to obtain the optimized feature weight distribution. The optimized feature weight distribution is sequentially mapped to the execution steps in the task decomposition sequence. The data transfer size and processing time between execution steps are calculated. The execution steps are prioritized and allocated computing resources according to the transfer size and processing time, and weighted execution instructions are generated.

5. The method according to claim 1, characterized in that, Priority path identifiers are mapped to a funding relationship graph, and the corresponding subgraph structures are extracted. These subgraph structures are then fed back into the large language model, outputting a targeted task decomposition sequence including: The priority path identifier is mapped to the fund association graph, and the transaction amount sequence of the target node set is extracted to obtain the fund flow curve; Extract the fluctuation period and fluctuation amplitude from the capital flow curve, identify the position with the largest capital accumulation and the position with the largest capital dispersion within the fluctuation period, and construct a set of feature nodes; Calculate the connection edge weight sequence between feature node sets to construct the node connection strength distribution, determine the subgraph structure boundary based on the variance of the node connection strength distribution, and generate the subgraph structure; The transaction features of the feature node set and the transitive features of the connecting edges in the subgraph structure are used to construct node representation vectors and edge representation vectors, respectively. The node representation vectors and edge representation vectors are combined into a structural feature matrix. A scene description sequence is generated based on the structural feature matrix. The scene description sequence is input into a large language model and the output is a targeted task decomposition sequence.

6. The method according to claim 1, characterized in that, Based on the targeted task decomposition sequence, a depth-first traversal of the subgraph structure is performed, node feature sequences are extracted, and jointly trained with historical case embedding vectors to generate a targeted analysis model, including: Extract the node access order from the targeted task decomposition sequence to construct the node traversal order, and extract feature calculation rules to construct feature extraction rules; The subgraph structure is subjected to a depth-first search according to the node traversal order. During the traversal, the rate of change of the frequency of transactions before and after a node is calculated to obtain the frequency change vector, the number of connected edges of a node is counted to obtain the connected edge vector, and the degree of overlap between a node and its neighboring nodes is calculated to obtain the overlap vector. Based on the feature extraction rules, the frequency change vector, the connection edge vector, and the overlap vector are combined to obtain the node feature vector. The node feature sequence is constructed by concatenating all the node feature vectors on the traversal path in the traversal order. A multi-layer neural network is constructed as the basic structure of the directional analysis model. The node feature sequence is encoded as the model input feature, and the historical case embedding vector is encoded as the model training label. The embedding loss is obtained by calculating the distance between the node feature sequence and the historical case embedding vector through forward propagation. The neural network parameters are updated by optimizing the embedding loss through backpropagation. The trained neural network is used as the directional analysis model.

7. The method according to claim 1, characterized in that, Inputting the fund correlation graph into a targeted analysis model for node risk scoring, the output includes risk account identifiers and correlation path analysis reports: Calculate the transaction frequency and transaction amount between nodes in the fund relationship graph to construct a transaction feature matrix, and calculate the fund flow direction and flow between nodes to construct a fund flow feature matrix; Input the transaction feature matrix and the capital flow feature matrix into the targeted analysis model, extract the time series change pattern from the transaction feature matrix, extract the spatial distribution pattern from the capital flow feature matrix, and calculate the node risk score based on the time series change pattern and the spatial distribution pattern. A risk distribution curve is constructed based on the node risk score. The mean and variance of the risk distribution curve are calculated to determine the risk threshold. Nodes with a risk score greater than the risk threshold are extracted as risk account identifiers. Calculate the fund transfer path between nodes corresponding to risk account identifiers, and extract the risk scores and fund flow characteristics between nodes on the transfer path to construct a path score vector; The risk level of a path is calculated based on the path score vector, and a related path analysis report is generated by sorting the paths according to their risk levels.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images