E-commerce financial statement generation system based on artificial intelligence

Abstract

The application belongs to the field of electric digital data processing, and particularly relates to an e-commerce financial statement generation system based on artificial intelligence. The system extracts order text, payment flow time sequence signal and logistics state image data and maps them to a unified vector space; a dynamic heterogeneous graph neural network is established with orders, accounts and commodities as nodes and funds and logistics circulation as edges with time sequence attributes; spatiotemporal coupling features of cross-platform fund flow and logistics are extracted through dynamic evolution and aggregation along time slices; an attention mechanism is introduced to dynamically allocate edge weights, identify abnormal subgraphs with unmatched fund flow and logistics, and locate error nodes based on abnormal subgraph topology analysis and correction, and decode into standard financial subject data flow to drive statement generation. The application solves the problem of hidden error missing identification of cross-platform heterogeneous data, breaks through the linear comparison limit, and improves the data checking accuracy.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing, specifically relating to an artificial intelligence-based e-commerce financial statement generation system. Background Technology

[0002] E-commerce financial statement generation typically involves order, payment, and logistics data from multiple platforms. Existing conventional report generation systems primarily employ a data matching and reconciliation mechanism based on static rules. Specifically, the system exports order text tables, payment transaction records, and logistics status tables from e-commerce platforms, payment gateways, and logistics tracking platforms, respectively. Using preset field mapping rules, it correlates and matches order numbers, payment transaction numbers, and waybill numbers from different tables. During the verification process, the system compares the monetary values ​​and status indicators under the correlated fields for consistency. If inconsistencies are found, the records are marked as abnormal and removed. Finally, the matching records are summarized according to fixed financial categories to generate financial statements. This approach heavily relies on the field matching logic of the rule engine, resulting in a linear processing flow.

[0003] The aforementioned static rule-based field matching reconciliation method suffers from a core technical problem when facing complex scenarios with cross-platform heterogeneous data and misaligned transaction times. This is because it fails to capture the deep topological relationships and spatiotemporal coupling characteristics between data points, leading to the missed detection of hidden errors. Due to differences in data update frequencies and transmission delays across different platforms, the records of cash flow and logistics often exhibit temporal misalignments. Static rules, relying solely on matching single slices or fields in a fixed order, cannot perceive the evolving coupling between cash flow and logistics over time. Furthermore, some financial errors are not manifested as numerical anomalies in a single field but are hidden within complex topological paths involving multiple nodes and edges. The linear comparison of static rules struggles to penetrate surface fields and extract deep topological anomalies causing mismatches between cross-platform cash flow and logistics, resulting in the missed detection of hidden errors. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based e-commerce financial statement generation system, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An AI-based e-commerce financial statement generation system includes: a multimodal data access layer that extracts order text, payment transaction time-series signals, and logistics status image data from the e-commerce platform, and maps them to a unified vector space through feature encoding; a dynamic heterogeneous graph construction layer that uses orders, accounts, and products as nodes and fund flow and logistics flow as edges with time-series attributes to establish a dynamic heterogeneous graph neural network; a spatiotemporal feature aggregation layer that controls the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along time slices, extracting the spatiotemporal coupling features of cross-platform fund flow and logistics; an anomaly detection layer that introduces an attention mechanism to dynamically allocate the weights of heterogeneous edges, identifying abnormal subgraph structures where fund flow and logistics do not match; a data correction and decoding layer that locates erroneous data nodes based on the topological analysis of the abnormal subgraphs, and decodes the corrected graph node features into standard financial subject data streams; and a report rendering layer that drives the report generation engine to output financial statements.

[0007] Preferably, the specific method by which the multimodal data access layer extracts multi-source heterogeneous data from the e-commerce platform and performs feature encoding mapping to a unified vector space is as follows: extracting the semantic feature vector of the order text through a text semantic parsing network, extracting the waveform feature vector of the payment transaction time-series signal through a one-dimensional temporal convolutional network, and extracting the visual feature vector of the logistics status image data through a two-dimensional visual convolutional network; using a cross-modal projection matrix to perform linear transformation on the semantic feature vector, the waveform feature vector, and the visual feature vector respectively, mapping them to a unified vector space of the same dimension, and performing orthogonalization processing on the linearly transformed multimodal features in the unified vector space to eliminate feature collinearity interference between different modal data.

[0008] Preferably, the specific method for establishing a dynamic heterogeneous graph neural network in the dynamic heterogeneous graph construction layer is as follows: the order, the account, and the product are initialized as heterogeneous graph nodes; the nodes corresponding to the fund flow and the logistics flow are determined based on the e-commerce transaction flow records; and the nodes are assigned timestamp labels to construct edges with time-series attributes. In the case where there are multiple edges of the same type with different timestamp labels between the same pair of nodes, multiple edges of the same type with timestamp labels within a preset time window are folded into a composite edge, and the weight of the composite edge is set to the frequency statistics value of the same type of edges within the preset time window. The timestamp label of the composite edge is updated to the median time value of the preset time window.

[0009] Preferably, the specific method by which the spatiotemporal feature aggregation layer controls the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along time slices is as follows: according to the temporal order of the time slices, the active nodes and active edges in the current time slice are activated sequentially; for the target node in the current time slice, the feature information of the heterogeneous neighbor nodes from the target node is aggregated, and the temporal decay factor is calculated in combination with the timestamp label of the active edge; the temporal decay factor and the feature information of the heterogeneous neighbor nodes are weighted and summed to generate the state embedding vector of the target node in the current time slice; the state embedding vector of the target node in the historical time slices is gated and fused with the state embedding vector of the current time slice to generate the final spatiotemporal coupling feature.

[0010] Preferably, the specific method by which the anomaly detection layer introduces an attention mechanism to dynamically allocate the weights of heterogeneous edges and identify abnormal subgraph structures is as follows: calculate the bidirectional attention score between the starting node features and the ending node features of the heterogeneous edge connection, and use the bidirectional attention score as the dynamic weight of the heterogeneous edge; traverse all edges in the dynamic heterogeneous graph neural network, and filter out heterogeneous edges whose dynamic weights are lower than a preset weight threshold as broken edges; extract the nodes at both ends of the broken edge and their first-order neighbor nodes to form a local subgraph, calculate the difference between the total inflow of funds flow nodes and the total confirmation of logistics nodes in the local subgraph, and if the difference exceeds the difference threshold, mark the local subgraph as the abnormal subgraph structure.

[0011] Preferably, the specific method by which the data correction and decoding layer locates erroneous data nodes based on the topological parsing of the abnormal subgraph and decodes the corrected graph node features into a standard financial subject data stream is as follows: The flow of funds is traced in reverse along the directed edges inside the abnormal subgraph, and extreme value nodes where the outflow of funds and the confirmed flow of goods are mismatched are located as the erroneous data nodes; the abnormal connections between the erroneous data nodes and surrounding nodes are severed, and the historical normal state embedding vector of the erroneous data nodes is used to replace the current abnormal state embedding vector to complete the data correction; the corrected graph node features are input into a fully connected decoding network, mapped to the lending direction and amount value under the corresponding financial subject system, and assembled to generate the standard financial subject data stream.

[0012] Preferably, the specific method for performing orthogonalization processing on the linearly transformed multimodal features in the unified vector space is as follows: construct an orthogonal reference basis based on the semantic feature vector, calculate the projection components of the waveform feature vector and the visual feature vector on the orthogonal reference basis; extract the orthogonal residual components perpendicular to the orthogonal reference basis from the projection components, and concatenate the orthogonal residual components with the semantic feature vector to form a cross-modal fusion feature; introduce a modality masking mechanism, randomly set the dimensions corresponding to some modalities in the cross-modal fusion feature to zero during the training phase, and reconstruct the zeroed modal features using the dimensions of the remaining modalities, forcing the cross-modal fusion feature to maintain complete financial semantic expression capability when any single modal input is missing.

[0013] Preferably, when dynamically evolving and aggregating according to the temporal order of the time slices, an adaptive time slice partitioning mechanism is adopted: The distribution density of e-commerce transaction timestamps is obtained; during periods where the transaction timestamp distribution density is higher than a high-density threshold, the time span of the time slice is reduced; during periods where the transaction timestamp distribution density is lower than a low-density threshold, the time span of the time slice is increased. In the gating fusion process, the cosine similarity between the state embedding vector of the historical time slice and the state embedding vector of the current time slice is calculated. The cosine similarity is input into a sigmoid function to generate a forgetting gate value. The forgetting gate value is used to attenuate the state embedding vector of the historical time slice, and the difference between 1 and the forgetting gate value is used as a retention gate value to weight and retain the state embedding vector of the current time slice.

[0014] Preferably, the specific method for calculating the bidirectional attention score between the starting node features and the ending node features of the heterogeneous edge connection is as follows: the starting node features and the ending node features are mapped to query vectors and key vectors respectively, and the dot product of the query vector and the key vector is calculated as the initial attention score; a time-series cumulative decay factor is introduced, and the initial attention score is multiplied by the time-series cumulative decay factor, wherein the time-series cumulative decay factor decreases as the duration of the transaction event corresponding to the heterogeneous edge increases; if the dynamic weight of the heterogeneous edge is lower than the preset weight threshold in multiple consecutive time slices, the abnormal score of the heterogeneous edge is exponentially accumulated, and when the accumulated abnormal score exceeds the abnormal outbreak threshold, the heterogeneous edge is directly marked as a broken edge.

[0015] Preferably, after severing the abnormal connections between the erroneous data node and surrounding nodes and replacing the current abnormal state embedding vector with the historical normal state embedding vector, a local topology blocking and re-decoding mechanism is executed: locking the first-order affected node directly connected to the erroneous data node and freezing the historical feature state of the first-order affected node before the anomaly occurred; when the fully connected decoding network decodes the corrected graph node features, the current feature input of the first-order affected node is skipped, and the historical feature state of the first-order affected node and the historical normal state embedding vector of the erroneous data node are forcibly jointly decoded to generate a locally corrected financial account data stream; the locally corrected financial account data stream is merged and spliced ​​with the normal financial account data stream of the unaffected nodes.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. This system constructs a dynamic heterogeneous graph neural network, using orders, accounts, and products as nodes, and fund flows and logistics flows as edges with temporal attributes. It dynamically evolves and aggregates along time slices to extract the spatiotemporal coupling features of cross-platform fund flows and logistics, overcoming the limitation of static rule matching in handling time-series misaligned data. An attention mechanism is introduced to dynamically allocate the weights of heterogeneous edges, identifying anomalous subgraph structures where fund flows and logistics do not match. Based on the topological analysis of these anomalous subgraphs, erroneous data nodes are located and corrected. This overcomes the limitation of linear field comparison in detecting deep topological association anomalies, solving the problem of missed identification of hidden errors in cross-platform heterogeneous data, and improving the accuracy of financial data verification and the depth of anomaly location.

[0018] 2. By using feature encoding and orthogonalization of multimodal data and modal masking mechanisms, the interference of feature collinearity between different modalities is eliminated, maintaining the complete expressive ability of features when a single modality is missing. By using an adaptive time slicing mechanism combined with gating fusion, time-series features are extracted when the transaction distribution density changes dynamically, suppressing the interference of historical noise. By calculating the time-series cumulative decay factor and accumulating the anomaly score to determine the break edge, misjudgment caused by occasional fluctuations is avoided. By adopting a local topology blocking and re-decoding mechanism, the current abnormal features of the affected nodes are skipped and the historical normal state is forced to be used for joint decoding, realizing the logical reconstruction of abnormal data and reducing the resource consumption of manual review. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multimodal data access layer feature encoding and unified mapping process of the present invention;

[0020] Figure 2 This is a flowchart of the temporal heterogeneous graph construction process for the dynamic heterogeneous graph construction layer of the present invention;

[0021] Figure 3 This is a flowchart illustrating the dynamic evolution and aggregation process of the spatiotemporal feature aggregation layer of the present invention.

[0022] Figure 4 This is a flowchart of the anomaly subgraph recognition process in the anomaly detection layer of the present invention;

[0023] Figure 5 This is a flowchart of the data correction and decoding layer error location and decoding process of the present invention;

[0024] Figure 6 This is a flowchart illustrating the financial statement generation process of the report rendering layer of this invention. Detailed Implementation

[0025] In one embodiment, the AI-based e-commerce financial statement generation system runs on the financial data processing server of an e-commerce enterprise, establishing data communication connections with multiple e-commerce platforms, payment gateways, and logistics tracking platforms. After system startup, the multimodal data access layer periodically acquires raw data from each platform through preset API interfaces, extracting three core data categories: order text, payment transaction time-series signals, and logistics status image data. The order text originates from the e-commerce platform's order management system and includes structured and semi-structured fields such as order number, product name, product specifications, quantity purchased, transaction price, order placement time, payment deadline, buyer nickname, seller store identifier, and order status. The payment transaction time-series signals originate from the payment gateway's transaction record system and are discrete event sequences arranged in ascending order of transaction occurrence time. Each event includes fields such as payment transaction number, transaction amount, transaction time, payment account identifier, receiving account identifier, transaction type, transaction status, and handling fee amount. The logistics status image data originates from the logistics tracking platform's image storage system and includes digital images in JPEG or PNG format, such as scanned images of logistics waybills, goods receipt certificates, transit node scanned images, and abnormal package photos.

[0026] refer to Figure 1The multimodal data access layer performs feature encoding on the three types of extracted raw data, mapping the data of different modalities to a unified vector space. For order text, the system extracts semantic feature vectors through a text semantic parsing network. The text semantic parsing network preprocesses the input order text, removing non-text characters and redundant spaces, converting the text into a token sequence, and then inputting it into a pre-trained language model, outputting a semantic feature vector containing order transaction semantics and attribute information. For payment transaction time-series signals, the system extracts waveform feature vectors through a one-dimensional temporal convolutional network. The one-dimensional temporal convolutional network performs sliding window convolution operations on the input time-series signal to capture the fund flow patterns at different time scales, outputting a waveform feature vector containing temporal fluctuation features and amount distribution features. For logistics status image data, the system extracts visual feature vectors through a two-dimensional visual convolutional network. After normalizing and resizing the input image, the two-dimensional visual convolutional network extracts the edge, texture, and semantic features of the image through multi-layer convolution and pooling operations, outputting a visual feature vector containing logistics node status and voucher information. The system uses a cross-modal projection matrix to linearly transform the feature vectors of the three different dimensions mentioned above, so that the transformed feature vectors have the same dimension, thereby mapping them to a unified vector space.

[0027] refer to Figure 2 The dynamic heterogeneous graph construction layer receives unified vector space features output from the multimodal data access layer and constructs a dynamic heterogeneous graph neural network. The system initializes three types of entities—orders, accounts, and products—as nodes in the heterogeneous graph. Each node corresponds to a unique entity identifier, and the initial features of the node are the multimodal fusion features of the corresponding entity. The system determines the edge relationships between nodes based on e-commerce transaction flow records. Fund flow edges connect payment account nodes and receiving account nodes, representing the transfer of funds from the payment account to the receiving account; logistics flow edges connect order nodes and product nodes, representing the delivery of goods from the seller's inventory to the buyer. The system assigns a timestamp label to each edge, corresponding to the occurrence time of the transaction event, giving the edges a temporal attribute.

[0028] refer to Figure 3The spatiotemporal feature aggregation layer controls the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along time slices, extracting the spatiotemporal coupling features of cross-platform capital flow and logistics. The system divides the entire financial statistics period into multiple consecutive time slices and processes the graph data within each time slice sequentially according to their temporal order. For each time slice, the system first activates the active nodes and active edges within the current time slice. Active nodes refer to nodes where transaction events occur within the current time slice, and active edges refer to edges whose timestamp labels fall within the current time slice. For each target node within the current time slice, the system aggregates the feature information of all heterogeneous neighbor nodes from that target node. Heterogeneous neighbor nodes refer to nodes connected to the target node through different types of edges. The system calculates a temporal decay factor based on the timestamp labels of active edges. The temporal decay factor decreases as the distance between the edge's timestamp and the center time of the current time slice increases, and is used to measure the degree of influence of historical transaction events on the current node's state. The system weightedly sums the temporal decay factor and the feature information of heterogeneous neighbor nodes to generate the target node's state embedding vector in the current time slice. The system further performs gated fusion of the state embedding vector of the target node in the historical time slice and the state embedding vector in the current time slice. The gated fusion process adjusts the weights of the two based on the correlation between the historical state and the current state to generate the final spatiotemporal coupling features of the target node.

[0029] refer to Figure 4 The anomaly detection layer introduces an attention mechanism to dynamically allocate the weights of heterogeneous edges, identifying anomalous subgraph structures where cash flow and logistics are mismatched. The system calculates a bidirectional attention score between the features of the starting and ending nodes connected by each heterogeneous edge. This score reflects the feature correlation between the starting and ending nodes, and is used as the dynamic weight of the heterogeneous edge. The system traverses all edges in the dynamic heterogeneous graph neural network, filtering out heterogeneous edges with dynamic weights below a preset threshold as broken edges. Broken edges represent anomalies in the transaction relationship between the two nodes they connect. The system extracts the nodes at both ends of the broken edge and their first-order neighbor nodes to form a local subgraph, which contains all direct transaction relationships related to the broken edge. The system calculates the difference between the total inflow of cash flow nodes and the total confirmation amount of logistics nodes within the local subgraph. The total inflow of cash flow nodes is the sum of the cash inflow amounts of all account nodes within the local subgraph, and the total confirmation amount of logistics nodes is the sum of the confirmed delivered goods value of all commodity nodes within the local subgraph. If the difference exceeds the preset difference threshold, it indicates that there is a mismatch between the flow of funds and the flow of goods in the local subgraph, and the system marks the local subgraph as an abnormal subgraph structure.

[0030] refer to Figure 5The data correction and decoding layer locates erroneous data nodes based on topological parsing of the anomaly subgraph and decodes the corrected graph node features into a standard financial subject data stream. The system traces the flow of funds backward along the directed edges within the anomaly subgraph, starting from the termination node of the broken edge and sequentially traversing all fund inflow edges pointing to that node, calculating the difference between the fund outflow and the corresponding logistics confirmation amount for each node. The system identifies the node with the largest absolute difference as the erroneous data node, which is the source of the mismatch between fund flow and logistics. The system cuts off all abnormal connections between the erroneous data node and surrounding nodes; abnormal connections refer to connections with dynamic weights below a preset weight threshold. The system obtains the historical normal state embedding vector of the erroneous data node before the anomaly occurred and replaces the current abnormal state embedding vector of the erroneous data node with this historical normal state embedding vector to complete the data correction. The system inputs all corrected graph node features into a fully connected decoding network, which maps the graph node features to the lending direction and amount values ​​under the corresponding financial subject system. The system assembles these mapping results according to financial subject classification to generate a standard financial subject data stream.

[0031] refer to Figure 6 The report rendering layer receives the standard financial account data stream output from the data correction and decoding layer, driving the report generation engine to output financial statements. The report generation engine, based on preset financial statement templates, fills the corresponding cells in the standard financial account data stream with the data from the standard financial account data stream, generating standard financial statements such as balance sheets, income statements, and cash flow statements. The system supports users to filter and drill down on financial statements by time, platform, and product dimensions. Users can view detailed and summarized financial data for different time periods, different e-commerce platforms, and different product categories. The system also supports exporting the generated financial statements to Excel, PDF, and other formats for easy archiving and sharing.

[0032] In this embodiment, the input and output parameters of the multimodal data access layer are compared as shown in Table 1.

[0033] Table 1. Comparison of Input and Output Parameters for Multimodal Data Access Layer

[0034] Order text Order number, product name, quantity, unit price, order time, buyer information, seller information, order status Text semantic parsing network 256 Includes transaction semantic information, product attribute information, transaction time information, and transaction status information for the order. Payment transaction timing signal Payment transaction number, transaction amount, transaction time, paying account, receiving account, transaction type, transaction status One-dimensional temporal convolutional networks 128 This includes the time-series fluctuation characteristics, amount distribution characteristics, transaction type characteristics, and transaction status characteristics of fund flows. Logistics status image data Images of shipping label, signed receipt, transit scan, and abnormal package. 2D visual convolutional network 128 Includes visual features of logistics nodes, voucher information features, package status features, and anomaly information features.

[0035] Table 1 shows the input and output parameters for the three core data types processed by the multimodal data access layer. The output vector dimension for order text is set to 256 to fully accommodate the complex semantic and multidimensional attribute information of orders; the output vector dimensions for payment transaction time-series signals and logistics status image data are both set to 128, ensuring feature representation capabilities while reducing the computational complexity of subsequent processing. The cross-modal projection matrix maps the feature vectors of the three different dimensions to a unified 256-dimensional vector space, providing a unified feature input for the subsequent construction of dynamic heterogeneous graphs.

[0036] This embodiment implements the basic functions of an AI-based e-commerce financial statement generation system. It achieves unified feature representation of different types of e-commerce data through a multimodal data access layer, captures the topological relationship and temporal evolution characteristics between capital flow and logistics through a dynamic heterogeneous graph neural network, identifies deep topological anomalies of mismatch between capital flow and logistics through an anomaly detection layer, completes the correction of abnormal data and mapping of financial items through a data correction and decoding layer, and finally generates standard financial statements through a report rendering layer.

[0037] In a preferred embodiment, the process of the multimodal data access layer extracting multi-source heterogeneous data from the e-commerce platform and mapping its features to a unified vector space is further refined. For order text, the text semantic parsing network adopts a pre-trained language model based on the Transformer architecture, specifically the BERT-base model. The system preprocesses the input order text, first removing special characters, emojis, and redundant spaces, and then using a token segmenter corresponding to the pre-trained model to convert the text into a token sequence. Tokens are added to the beginning and end of the sequence. Token sequences longer than 512 are truncated, and those shorter than 512 are padded with zeros. The preprocessed token sequence is input into the BERT-base model, which contains 12 Transformer encoder layers, each composed of a multi-head self-attention mechanism and a feedforward neural network. The system extracts the 768-dimensional vector corresponding to the token in the model output as the initial semantic feature vector of the order text.

[0038] For payment transaction time-series signals, the one-dimensional temporal convolutional network consists of three consecutive convolutional blocks. Each convolutional block comprises a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The first convolutional block has a kernel size of 3, 32 output channels, a stride of 1, and padding of 1. The second convolutional block has a kernel size of 5, 64 output channels, a stride of 1, and padding of 2. The third convolutional block has a kernel size of 7, 128 output channels, a stride of 1, and padding of 3. The system converts the payment transaction time-series signal into a two-dimensional tensor of shape (1,L), where L is the length of the time-series signal. This tensor is input into the one-dimensional temporal convolutional network. After processing through the three convolutional blocks, the output tensor is compressed into a 128-dimensional vector by a global average pooling layer, serving as the waveform feature vector of the payment transaction time-series signal.

[0039] For logistics status image data, the 2D visual convolutional network uses the ResNet-18 model, removing the final fully connected layer and softmax layer, retaining all network layers before the average pooling layer. The system preprocesses the input logistics status images, uniformly adjusting the image size to 224×224 pixels and normalizing the pixel values ​​to a mean of 0 and a standard deviation of 1. The preprocessed image is then input into the ResNet-18 model. After five convolutional stages, the output tensor is compressed into a 512-dimensional vector using the average pooling layer, serving as the visual feature vector for the logistics status image data.

[0040] The system utilizes a cross-modal projection matrix to linearly transform the feature vectors of the three different dimensions mentioned above, mapping them to a unified 256-dimensional vector space. Let the semantic feature vector be... The waveform feature vector is The visual feature vector is The corresponding cross-modal projection matrices are respectively , , The bias vectors are respectively , , The eigenvectors after linear transformation are calculated as follows:

[0041]

[0042]

[0043]

[0044] in, , , These are the semantic feature vector, waveform feature vector, and visual feature vector after linear transformation, respectively, all belonging to a unified vector space of 256 dimensions.

[0045] The system performs orthogonalization processing on the linearly transformed multimodal features within a unified vector space to eliminate feature collinearity interference between different modal data. The system constructs an orthogonal reference basis based on the semantic feature vectors. The orthogonal reference basis is a unit vector, and its calculation method is as follows:

[0046]

[0047] in, Let L2 be the norm of the semantic feature vector. The system calculates the projection components of the waveform feature vector and the visual feature vector onto the orthogonal reference basis, respectively.

[0048]

[0049]

[0050] in, These are the projection components of the waveform eigenvectors onto the orthogonal reference basis. This represents the projection component of the visual feature vector onto the orthogonal reference base. The system extracts the orthogonal residual component perpendicular to the orthogonal reference base from the projection component:

[0051]

[0052]

[0053] in, These are the orthogonal residual components of the waveform eigenvectors. These are the orthogonal residual components of the visual feature vector. The orthogonal residual components contain unique information from the waveform and visual features that is irrelevant to the semantic features, eliminating collinearity interference with the semantic features. The system concatenates the semantic feature vector with the two orthogonal residual components to form a cross-modal fused feature:

[0054]

[0055] in, It is a cross-modal fusion feature with a dimension of 768.

[0056] The system introduces a modality masking mechanism. During the training phase, it randomly sets the dimensions of some modalities in the cross-modal fusion features to zero, and then reconstructs the zero-dimension features using the dimensions of the remaining modalities. The system randomly generates a 3D mask vector. Each element in the mask vector independently follows a Bernoulli distribution, with a probability of 0.2 for each value. The three elements of the mask vector correspond to the dimensions of the semantic, waveform, and visual features, respectively. If any element in the mask vector is 0, all dimensions of the corresponding modality in the cross-modal fusion feature are set to zero. The system inputs the masked cross-modal fusion feature into a reconstruction network, which is a neural network containing two fully connected layers. The output is a reconstructed feature with the same dimensions as the original cross-modal fusion feature. The system calculates the mean squared error between the reconstructed feature and the original cross-modal fusion feature as the loss function and updates the network parameters through a backpropagation algorithm. The modal masking mechanism forces the cross-modal fusion feature to reconstruct a complete financial semantic expression using the information of the remaining modalities even when any single modality input is missing, thus improving the system's robustness to data with missing modalities.

[0057] The process of building a dynamic heterogeneous graph neural network in the dynamic heterogeneous graph construction layer is further refined. The system initializes three types of entities—orders, accounts, and products—as nodes in the heterogeneous graph. Each node has a unique entity identifier and attribute set. The unique identifier of an order node is the order number, and the attribute set includes order time, transaction amount, product quantity, order status, platform identifier, etc.; the unique identifier of an account node is the account number, and the attribute set includes account type, opening bank, account balance, registration time, etc.; the unique identifier of a product node is the product number, and the attribute set includes product name, product category, unit price, inventory quantity, supplier identifier, etc. The system determines the edge relationships between nodes based on e-commerce transaction flow records. Fund flow edges connect payment account nodes and receiving account nodes, with the edge direction pointing from the payment account to the receiving account, and the attribute set includes transaction amount, transaction time, payment serial number, transaction status, handling fee amount, etc.; logistics flow edges connect product nodes and order nodes, with the edge direction pointing from the product node to the order node, representing the product being shipped from inventory and associated with the corresponding order, and the attribute set includes logistics tracking number, shipping time, receipt time, logistics status, transportation company identifier, etc. The system assigns a timestamp label to each connection edge. The timestamp label for the fund transfer edge is the time the transaction occurred, and the timestamp label for the logistics transfer edge is the time the goods were shipped.

[0058] For cases where multiple edges of the same type with different timestamp labels exist within the same node pair, the system collapses multiple edges of the same type with timestamp labels within a preset time window into a single composite edge. The preset time window is 24 hours long, from 00:00 to 24:00 each day. For node pairs... All edges of the same type between Their timestamp labels are respectively The system first sorts these edges in ascending order by timestamp labels. Then, starting with the first edge, it groups edges whose timestamp labels differ from the timestamp labels of the first edge by no more than 24 hours. For each group of edges, the system collapses it into a composite edge. The starting and ending nodes of the composite edge are the same as those of the original edge, and the edge type is also the same. The weight of the composite edge is set to the number of edges in the group, i.e., the frequency statistics of the same type of transactions within the time window. The timestamp label of the composite edge is updated to the average of the timestamp labels of all edges in the group, i.e., the median time.

[0059]

[0060] in, For the timestamp label of the composite edge. This represents the number of edges within the group. If the time span of similar edges between the same pair of nodes exceeds 24 hours, they are divided into multiple groups, each corresponding to a non-overlapping 24-hour time window, and each group is folded into multiple composite edges. This composite edge folding mechanism reduces the number of edges in the dynamic heterogeneous graph, lowers the computational complexity of subsequent spatiotemporal feature aggregation, and preserves the transaction frequency information within the time window.

[0061] In this embodiment, the node and edge attributes of the dynamic heterogeneous graph are defined as shown in Table 2.

[0062] Table 2 Definition of Node and Edge Attributes in Dynamic Heterogeneous Graphs

[0063] node Order Node Order Number String Unique identifier of the order node Order Node Order time Timestamp Order creation time node Order Node Transaction amount floating-point number Total transaction amount of the order node Order Node Quantity of goods Integer Total number of items included in the order node Order Node Order status Enumeration value The current status of the order, including pending payment, paid, shipped, signed for, cancelled, etc. node Order Node Platform logo String The e-commerce platform identifier of the order node Account Node Account Number String Unique identifier of the account node Account Node Account Type Enumeration value Account types include buyer accounts, seller accounts, platform accounts, payment institution accounts, etc. node Account Node Bank of account String Account opening bank name node Account Node Account Balance floating-point numbers Current account balance node Account Node Registration time Timestamp Account registration time node Product Node Product Number String Unique identifier of a product node Product Node Product Name String Product Name node Product Node Product Category String The category to which the product belongs node Product Node unit price floating-point numbers The selling price of the product node Product Node Inventory Quantity Integer Current inventory quantity of the product node Product Node Supplier Identifier String Supplier identification of the product side Capital Flow Transaction amount floating-point numbers Amount of funds transferred in this transaction side Capital Flow Trading hours Timestamp Timing of fund transfer side Capital Flow Payment serial number String The unique serial number for this payment side Capital Flow Transaction status Enumeration value The status of fund transfers, including pending, completed, failed, and refunded. side Capital Flow Transaction fee amount floating-point number The amount of transaction fees generated in this payment side Logistics circulation edge tracking number String The unique tracking number for this shipment side Logistics circulation edge Delivery time Timestamp Product shipping time side Logistics circulation edge Delivery time Timestamp Goods receipt time side Logistics circulation edge Logistics status Enumeration value The current status of the logistics, including pending shipment, in transit, signed for, and returned. side Logistics circulation edge Transportation company logo String Company logo responsible for this logistics transportation

[0064] Table 2 defines in detail the attributes of nodes and edges in a dynamic heterogeneous graph. Node attributes include basic entity information and transaction-related information, while edge attributes include detailed information and temporal sequence information of transaction events. These attributes provide rich input information for feature learning in dynamic heterogeneous graph neural networks, enabling the model to accurately capture the flow patterns of funds and goods.

[0065] This embodiment refines the feature encoding and orthogonalization process for multimodal data, as well as the construction process of the dynamic heterogeneous graph. Orthogonalization eliminates collinearity interference between different modalities, modality masking improves the system's robustness to missing modal data, and composite edge folding reduces the complexity of the dynamic heterogeneous graph, thereby improving system efficiency.

[0066] In another preferred embodiment, the process of the spatiotemporal feature aggregation layer controlling the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along time slices is further refined. The system adopts an adaptive time slice partitioning mechanism, dynamically adjusting the time span of the time slices according to the distribution density of e-commerce transaction timestamps. The system first obtains the timestamps of all transaction events within the entire financial statistics period, and uses the kernel density estimation method to calculate the transaction density at each moment on the time axis. The kernel function used for kernel density estimation is a Gaussian kernel function, with a bandwidth set to 1 hour. A high-density threshold is set in the system. With low density threshold The high-density threshold is the 90th quantile of the transaction density distribution, and the low-density threshold is the 10th quantile of the transaction density distribution. The system adjusts the time span of the time slice based on the transaction density: when the transaction density at a certain moment... At that time, the time span of the time slice containing that moment is set to the minimum value. That is, 1 hour; when the transaction density at a certain moment At that time, the time span of the time slice containing that moment is set to the maximum value. That is, 24 hours; the transaction density at a certain moment. At that time, the time span of the time slice containing that moment is calculated using linear interpolation:

[0067]

[0068] The adaptive time slicing mechanism reduces the time slice span during periods of high transaction density, thereby improving the temporal resolution of time series features; and increases the time slice span during periods of low transaction density, thereby reducing computational load and balancing the system's accuracy and operational efficiency.

[0069] The system processes the graph data within each time slice sequentially according to the temporal order of the adaptively divided time slices. For each time slice... Its time range is ,in This represents the time span of the time slice. The system first activates the active nodes and active edges within the current time slice. Active edges refer to all connecting edges whose timestamp labels fall within the current time slice range, and active nodes refer to the start and end nodes of all active edges. For inactive nodes, their state embedding vectors remain unchanged from the previous time slice.

[0070] For each target node within the current time slice The system aggregates feature information from all heterogeneous neighbor nodes of the target node. (Heterogeneous neighbor node set) Including the target node All account nodes are connected via the fund flow edge, and all order nodes or product nodes are connected via the logistics flow edge. The system calculates a time-series decay factor by combining the timestamp labels of the active edges. The time-series decay factor is used to measure the degree of influence of historical transaction events on the current node state. The calculation method is as follows:

[0071]

[0072] in, For the edge The time decay factor, This is the attenuation coefficient, with a value of 0.1 / hour. For the edge timestamp tags, The center time of the current time slice. The time decay factor decays exponentially as the distance between the edge's timestamp and the center time of the current time slice increases; the farther away a transaction event is, the smaller its impact on the current node's state.

[0073] The system performs a weighted summation of the temporal decay factor and the feature information of heterogeneous neighbor nodes to generate the initial state embedding vector of the target node in the current time slice:

[0074]

[0075] in, For the target node Embed the vector in the initial state of the current time slice. For neighboring nodes The state embedding vector from the previous time slice The weight matrix is... For bias vectors, This is the ReLU activation function.

[0076] The system further performs gated fusion of the target node's state embedding vector in historical time slices with the initial state embedding vector in the current time slice to generate the final spatiotemporal coupling feature. The system first calculates the historical state embedding vector. Embedded vector with the current initial state Cosine similarity between them:

[0077]

[0078] in, Cosine similarity, with a value ranging from [-1, 1], reflects the correlation between historical and current states. The system inputs the cosine similarity into the sigmoid function to generate the forgetting gate value:

[0079]

[0080] in, The forgetting gate value ranges from [0,1]. A larger forgetting gate value indicates a higher correlation between the historical state and the current state, and a greater contribution of the historical state to the final state. The system uses the forgetting gate value to attenuate the historical state embedding vector, and uses the difference between 1 and the forgetting gate value as the retention gate value to weight and retain the current initial state embedding vector. The final spatiotemporal coupling feature after fusion is:

[0081]

[0082] The gating fusion mechanism dynamically adjusts the weights of the historical and current states based on their correlation, which can effectively suppress the interference of historical noise while retaining useful historical information, thus improving the accuracy of spatiotemporal coupling features.

[0083] The anomaly detection layer further refines the process of dynamically allocating heterogeneous edge weights and identifying anomalous subgraph structures by introducing an attention mechanism. The system calculates the features of the starting node connected by each heterogeneous edge. Features of the Termination Node The system calculates bidirectional attention scores between nodes. It maps the starting node features to a query vector and the ending node features to a key vector.

[0084]

[0085]

[0086] in, For query vector, For key vectors, and This is a learnable weight matrix. The system calculates the dot product of the query vector and the key vector as the initial attention score:

[0087]

[0088] in, This is a scaling factor used to prevent the gradient of the softmax function from vanishing due to an excessively large dot product. The system further introduces a temporal cumulative decay factor, which decreases as the duration of the transaction event corresponding to the heterogeneous edge increases. The calculation method is as follows:

[0089]

[0090] in, This is the time-series cumulative decay factor. This is the cumulative attenuation coefficient, with a value of 0.05 / day. The duration of the transaction event is the interval from the transaction initiation time to the current processing time, in days. The system multiplies the initial attention score by the temporal cumulative decay factor to obtain the final dynamic weights of the heterogeneous edges:

[0091]

[0092] The calculation of the bidirectional attention score also includes the direction from the termination node to the start node; the system calculates the direction from the termination node in the same way. To the starting node Dynamic weights Then, the average of the two values ​​is taken as the final bidirectional attention score for the heterogeneous edge.

[0093] The system iterates through all edges in a dynamic heterogeneous graph neural network and filters out edges whose dynamic weights are below a preset weight threshold. Heterogeneous edges are used as candidate break edges. A preset weight threshold is applied. The value is determined based on the 5th percentile of the dynamic weight distribution of normal edges in the training set, ranging from 0.1 to 0.3. The system accumulates anomaly scores for candidate broken edges. If a heterogeneous edge is in continuous... If the dynamic weights within each time slice are all below a preset weight threshold, then their abnormal scores are accumulated exponentially.

[0094]

[0095] in, heterogeneous edges The anomaly score is initially set to 0. When the accumulated anomaly score exceeds the anomaly outbreak threshold... At that time, the system directly marks the heterogeneous edge as a broken edge. Abnormal outbreak threshold. Adjusted according to the system's false positive rate requirements, the value ranges from 5 to 10. The abnormal score accumulation mechanism can distinguish between occasional weight fluctuations and persistent anomalies, avoiding false positives caused by network jitter or data latency.

[0096] The system extracts the nodes at both ends of the break edge and their first-order neighbor nodes to form a local subgraph. This local subgraph contains all direct transaction relationships related to the break edge. The system calculates the difference between the total inflow of funds to the nodes in the local subgraph and the total confirmation volume of logistics nodes.

[0097]

[0098] in, For the degree of difference, The set of all account nodes within the local subgraph. For account nodes Total capital inflow within the time frame corresponding to the local sub-chart The set of all product nodes within the local subgraph. For product nodes The total value of confirmed delivered goods within the time frame corresponding to the local subgraph. If the difference... Exceeding the preset difference threshold If so, the system marks that local subgraph as an anomalous subgraph structure. Difference threshold. The value is determined based on the transaction amount, ranging from 100 to 10,000 yuan.

[0099] The data correction and decoding layer further refines the process of locating erroneous data nodes based on topological parsing of the anomaly subgraph and decoding the corrected graph node features into standard financial account data flows. The system traces the flow of funds backward along the directed edges within the anomaly subgraph. Starting from the terminating node of a broken edge, it sequentially traverses all fund inflow edges pointing to that node, calculating the difference between the fund outflow and the corresponding logistics confirmation amount for each node. The fund outflow is the total amount of funds outflow from the node within the time range corresponding to the anomaly subgraph, and the logistics confirmation amount is the total value of all confirmed delivered goods associated with the node. The system identifies the node with the largest absolute value of the difference as the erroneous data node, which is the source of the mismatch between fund flow and logistics.

[0100] The system disconnects all abnormal edges between the erroneous data node and its surrounding nodes. Abnormal edges are defined as those with dynamic weights below a preset weight threshold. The system obtains the normal state embedding vectors of the erroneous data node for the three consecutive time slices prior to the anomaly, and calculates the average of these vectors as the historical normal state embedding vector.

[0101]

[0102] in, For error data nodes The historical normal state embedding vector, For the erroneous data node before the anomaly occurs The system uses the state embedding vectors of each time slice to replace the current abnormal state embedding vector of the erroneous data node with the historical normal state embedding vectors, thus completing the data correction.

[0103] The system further implements a local topology blocking and re-decoding mechanism to reduce the impact of abnormal data on surrounding nodes. The system locks all first-order affected nodes directly connected to the erroneous data node, freezing their historical feature states before the anomaly occurred. When the fully connected decoding network decodes the corrected graph node features, the system skips the current feature input of the first-order affected nodes and forces joint decoding by embedding the historical feature states of the first-order affected nodes with the historical normal states of the erroneous data node into the embedding vector. The fully connected decoding network consists of three fully connected layers. The first fully connected layer has an input dimension of 256, an output dimension of 512, and uses ReLU activation. The second fully connected layer has an input dimension of 512, an output dimension of 256, and uses ReLU activation. The third fully connected layer has an input dimension of 256 and an output dimension of... ,in For the number of financial items, the first Each dimension corresponds to The lending direction of each financial item is activated using a softmax activation function, and the last dimension, corresponding to the transaction amount, is activated using a linear activation function. The system assembles the output of the fully connected decoding network according to the financial item classification to generate a locally corrected financial item data stream. The system then merges and splices this locally corrected financial item data stream with the normal financial item data stream from unaffected nodes to generate the final standard financial item data stream.

[0104] In this embodiment, the key parameters of the anomaly detection and correction process are shown in Table 3.

[0105] Table 3 Key Parameters for Anomaly Detection and Correction Process

[0106] Attenuation coefficient 0.05-0.2 / hour Determined based on the temporal correlation of transaction data Controlling the decay rate of the timing decay factor Cumulative attenuation coefficient 0.02-0.1 / day Determined based on the duration characteristics of the abnormal event Controlling the decay rate of the timing cumulative decay factor Preset weight threshold 0.1-0.3 5th percentile of the dynamic weight distribution of normal edges in the training set Threshold for distinguishing normal edges from candidate broken edges Abnormal outbreak threshold 5-10 Adjust according to system misjudgment rate requirements Threshold for distinguishing between sporadic fluctuations and persistent anomalies Difference threshold 100-10000 yuan Determined based on the magnitude of the transaction amount Threshold for distinguishing normal subgraphs from abnormal subgraphs Number of historical time slices 2-5 Determined based on the stability of historical data Number of time slices used to calculate the historical normal state embedding vector High density threshold - 90th quantile of the transaction density distribution Transaction density threshold that triggers the minimum time slice span Low density threshold - 10th percentile of the transaction density distribution Transaction density threshold that triggers the maximum time slice span Minimum time slice span 0.5-2 hours Determined based on the time resolution requirements during peak trading periods Minimum span of time slice during peak trading periods Maximum time slice span 12-48 hours Determined based on computational efficiency requirements during periods of sparse transactions. Maximum span of time slices during sparse trading periods

[0107] Table 3 lists the key parameters, their value ranges, and calculation methods in the anomaly detection and correction process. These parameters are determined through grid search and cross-validation on the training set, ensuring the accuracy of anomaly detection while controlling the system's false positive rate and operational efficiency.

[0108] This embodiment refines the adaptive time-slicing mechanism and gating fusion mechanism for spatiotemporal feature aggregation, the bidirectional attention mechanism and anomaly score accumulation mechanism for anomaly detection, and the local topology blocking and re-decoding mechanism for data correction. The adaptive time-slicing mechanism balances the system's accuracy and operational efficiency, the gating fusion mechanism suppresses interference from historical noise, the anomaly score accumulation mechanism avoids misjudgments caused by occasional fluctuations, and the local topology blocking and re-decoding mechanism reduces the impact of anomalous data on surrounding nodes, further improving the overall performance of the system.

Claims

1. An AI-based e-commerce financial statement generation system, characterized in that: include: The multimodal data access layer extracts order text, payment transaction time-series signals, and logistics status image data from e-commerce platforms, and maps them to a unified vector space through feature encoding; The dynamic heterogeneous graph construction layer uses orders, accounts, and products as nodes, and fund flow and logistics flow as edges with temporal attributes to establish a dynamic heterogeneous graph neural network. The spatiotemporal feature aggregation layer controls the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along time slices, extracting the spatiotemporal coupling features of cross-platform capital flow and logistics; The anomaly detection layer introduces an attention mechanism to dynamically allocate the weights of heterogeneous edges, identifying abnormal subgraph structures where capital flow and logistics do not match. The data correction and decoding layer locates the erroneous data nodes based on the topology analysis of the abnormal subgraph and decodes the corrected graph node features into a standard financial account data stream. The report rendering layer drives the report generation engine to output financial statements.

2. The AI-based e-commerce financial statement generation system according to claim 1, characterized in that, The specific method by which the multimodal data access layer extracts multi-source heterogeneous data from the e-commerce platform and performs feature encoding mapping to a unified vector space is as follows: the semantic feature vector of the order text is extracted through a text semantic parsing network, the waveform feature vector of the payment transaction time-series signal is extracted through a one-dimensional temporal convolutional network, and the visual feature vector of the logistics status image data is extracted through a two-dimensional visual convolutional network. The semantic feature vector, the waveform feature vector, and the visual feature vector are linearly transformed using a cross-modal projection matrix and mapped to a unified vector space of the same dimension. Orthogonalization is then performed on the linearly transformed multimodal features within the unified vector space to eliminate feature collinearity interference between different modal data.

3. The AI-based e-commerce financial statement generation system according to claim 1, characterized in that, The specific way to establish a dynamic heterogeneous graph neural network in the dynamic heterogeneous graph construction layer is as follows: the order, the account and the product are initialized as heterogeneous graph nodes, the nodes corresponding to the fund flow and the logistics flow are determined based on the e-commerce transaction flow records, and the nodes are assigned timestamp labels to construct edges with time sequence attributes. For cases where there are multiple edges of the same type with different timestamp labels between the same pair of nodes, multiple edges of the same type with timestamp labels within a preset time window are folded into a composite edge, and the weight of the composite edge is set to the frequency statistics value of the edges of the same type within the preset time window. The timestamp label of the composite edge is then updated to the median time value of the preset time window.

4. The AI-based e-commerce financial statement generation system according to claim 1, characterized in that, The specific way in which the spatiotemporal feature aggregation layer controls the dynamic heterogeneous graph neural network to dynamically evolve and aggregate along the time slice is as follows: according to the temporal order of the time slice, the active nodes and active edges in the current time slice are activated sequentially. For the target node in the current time slice, the feature information of the heterogeneous neighbor nodes from the target node is aggregated, and the time decay factor is calculated by combining the timestamp label of the active edge. The time decay factor is weighted and summed with the feature information of the heterogeneous neighbor nodes to generate the state embedding vector of the target node in the current time slice. The state embedding vector of the target node in the historical time slice is gated and fused with the state embedding vector of the current time slice to generate the final spatiotemporal coupling feature.

5. The AI-based e-commerce financial statement generation system according to claim 1, characterized in that, The specific method by which the anomaly detection layer introduces an attention mechanism to dynamically allocate the weights of heterogeneous edges and identify abnormal subgraph structures is as follows: calculate the bidirectional attention score between the features of the starting node and the features of the ending node connected by the heterogeneous edge, and use the bidirectional attention score as the dynamic weight of the heterogeneous edge. Traverse all edges in the dynamic heterogeneous graph neural network and filter out heterogeneous edges whose dynamic weights are lower than a preset weight threshold as broken edges. Extract the nodes at both ends of the fracture edge and their first-order neighbor nodes to form a local subgraph. Calculate the difference between the total inflow of funds and the total confirmation of logistics within the local subgraph. If the difference exceeds the difference threshold, mark the local subgraph as the abnormal subgraph structure.

6. The AI-based e-commerce financial statement generation system according to claim 1, characterized in that, The specific method by which the data correction and decoding layer locates the erroneous data nodes based on the topological analysis of the abnormal subgraph and decodes the corrected graph node features into a standard financial subject data stream is as follows: the fund flow is traced in reverse along the directed edges inside the abnormal subgraph, and the extreme value nodes where the fund outflow and logistics confirmation do not match are located as the erroneous data nodes. The abnormal connections between the erroneous data node and its surrounding nodes are severed, and the current abnormal state embedding vector is replaced with the historical normal state embedding vector of the erroneous data node to complete the data correction. The corrected graph node features are input into a fully connected decoding network and mapped to the lending direction and amount values ​​under the corresponding financial subject system, and then assembled to generate the standard financial subject data stream.

7. The AI-based e-commerce financial statement generation system according to claim 2, characterized in that, The specific method for performing orthogonalization processing on the linearly transformed multimodal features in the unified vector space is as follows: construct an orthogonal reference basis based on the semantic feature vector, and calculate the projection components of the waveform feature vector and the visual feature vector on the orthogonal reference basis; Extract the orthogonal residual components that are perpendicular to the orthogonal reference base from the projection components, and concatenate the orthogonal residual components with the semantic feature vector to form cross-modal fusion features; A modality masking mechanism is introduced, in which the dimensions corresponding to some modalities in the cross-modal fusion features are randomly set to zero during the training phase, and the zeroed modal features are reconstructed using the dimensions of the remaining modalities, thus forcing the cross-modal fusion features to maintain complete financial semantic expression capabilities when any single modal input is missing.

8. The AI-based e-commerce financial statement generation system according to claim 4, characterized in that, When dynamically evolving and aggregating according to the temporal order of the time slices, an adaptive time slice division mechanism is adopted: the distribution density of e-commerce transaction timestamps is obtained, and the time span of the time slice is reduced when the transaction timestamp distribution density is higher than the high density threshold, and the time span of the time slice is increased when the transaction timestamp distribution density is lower than the low density threshold. In the gating fusion process, the cosine similarity between the state embedding vector of the historical time slice and the state embedding vector of the current time slice is calculated. The cosine similarity is input into the sigmoid function to generate a forgetting gate value. The forgetting gate value is used to attenuate the state embedding vector of the historical time slice. The difference between 1 and the forgetting gate value is used as the retention gate value to weight and retain the state embedding vector of the current time slice.

9. The AI-based e-commerce financial statement generation system according to claim 5, characterized in that, The specific method for calculating the bidirectional attention score between the starting node features and the ending node features of the heterogeneous edge connection is as follows: the starting node features and the ending node features are mapped to query vector and key vector respectively, and the dot product of the query vector and the key vector is calculated as the initial attention score; A time-series cumulative decay factor is introduced, and the initial attention score is multiplied by the time-series cumulative decay factor, wherein the time-series cumulative decay factor decreases as the duration of the transaction event corresponding to the heterogeneous edge increases; If the dynamic weight of the heterogeneous edge is lower than the preset weight threshold in multiple consecutive time slices, the abnormal score of the heterogeneous edge is accumulated exponentially. When the accumulated abnormal score exceeds the abnormal outbreak threshold, the heterogeneous edge is directly marked as a broken edge.

10. The AI-based e-commerce financial statement generation system according to claim 6, characterized in that, After severing the abnormal connections between the erroneous data node and surrounding nodes and replacing the current abnormal state embedding vector with the historical normal state embedding vector, a local topology blocking and re-decoding mechanism is executed: the first-order affected node directly connected to the erroneous data node is locked, and the historical feature state of the first-order affected node before the anomaly occurred is frozen. When the fully connected decoding network decodes the corrected graph node features, it skips the current feature input of the first-order affected node and forces the historical feature state of the first-order affected node and the historical normal state of the error data node to be jointly decoded in the embedding vector to generate a locally corrected financial account data stream. The locally modified financial account data stream is merged and spliced ​​with the normal financial account data stream of the unaffected nodes.

Images