AI- and Knowledge Graph-Based Intelligent Business Data Analysis System

Abstract

This invention relates to the fields of artificial intelligence and knowledge graph technology, and discloses an intelligent business data analysis system based on AI and knowledge graphs. The system includes a hierarchical knowledge graph generation module, a cross-domain granularity-aware alignment module, an equivalent transmission operator generation module, a causal template verification and parameter correction module, a domain-invariant embedding generation module, and a predictive analysis report generation module. The system identifies the missing granularity levels of the target domain through hierarchical awareness alignment, compresses multi-step fine-grained causal paths into equivalent transmission operators using ordered multiplication of information transmission matrices, adaptively adjusts the Bayesian prior accuracy with information loss rate, and eliminates domain distribution shifts through adversarial training. Finally, it outputs an intelligent business data analysis report with granularity adaptation tracing and transfer confidence level labeling.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge graph technology, and more specifically, to a business data intelligent analysis system based on AI and knowledge graph. Background Technology

[0002] In enterprise multi-business-line operations, mature business lines have accumulated a multi-level, fine-grained business indicator system over a long period, storing rich causal analysis experience in the form of a time-series knowledge graph. This includes the hierarchical relationships of entity nodes, causal relationship edges, and their time delay parameters. When enterprises need to perform intelligent analysis and trend prediction on the business data of newly launched business lines, due to the short data accumulation cycle and sparse event samples in the new business domain, only coarse-grained aggregated business indicators are usually available for observation, making it difficult to independently construct a complete time-series knowledge graph and a reliable causal analysis link.

[0003] Existing cross-domain analysis transfer methods based on knowledge graphs map the causal analysis structure of a mature domain to a new business domain through semantic alignment. However, these methods assume that the indicator systems of the two domains are at similar levels of abstraction, performing a peer-to-peer mapping between nodes. These methods have the following drawbacks: Firstly, the intermediate transmission nodes in the causal analysis chain of the mature business domain are fine-grained indicators. Direct migration to the new business domain without corresponding fine-grained data leads to a break in the causal transmission chain and the inability to estimate intermediate parameters. Secondly, if fine-grained multi-step causal paths are compressed into coarse-grained single-step causal edges through granular aggregation, multi-step transmission information is lost, and the business meaning of the causal parameters changes. When subsequently using Bayesian methods combined with new domain observation data to correct parameters, the prior distribution of the source domain and the posterior distribution of the new domain observations are not in the same semantic space, resulting in systematic distortion in parameter correction and consequently, directional deviations in intelligent analysis conclusions and trend predictions. Summary of the Invention

[0004] This invention provides an intelligent business data analysis system based on AI and knowledge graphs, which solves the technical problems in related technologies such as sparse target domain business data, misaligned cross-domain granular structure leading to the inability to directly transfer causal analysis templates, and lack of traceable confidence labeling for prediction results.

[0005] This invention discloses a business data intelligent analysis system based on AI and knowledge graphs, comprising: The hierarchical knowledge graph generation module is configured to acquire business data from the source and target domains, generate time-series knowledge graphs for each domain, and perform hierarchical clustering based on the inclusion and aggregation relationships between nodes to divide business indicator nodes into multi-level granularity layers, generating hierarchical time-series knowledge graph structures for each of the two domains. The cross-domain granularity-aware alignment module is configured to extract the semantic embedding vector and topological feature vector of each node in the hierarchical temporal knowledge graph of two domains and concatenate them into a fusion representation. Within the same granularity layer, it calculates the alignment probability between the source domain node and the target domain node, generates a cross-domain entity alignment mapping table, and identifies the missing granularity level of the target domain by comparing the granularity hierarchical structure of the two domains. The equivalent transmission operator generation module is configured to extract multi-step causal transmission paths from the source domain temporal knowledge graph, passing through intermediate nodes with missing granularity levels, calculate the information transmission matrix between each adjacent node on the path based on historical data of the source domain, perform ordered matrix multiplication on all information transmission matrices to generate equivalent transmission operators, calculate the information loss rate of each path, and output a set of candidate causal analysis templates after granularity adaptation. The causal template verification and parameter correction module is configured to perform conditional independence tests on each equivalent transmission edge in the candidate causal analysis template using target domain data to filter incompatible causal edges, merge the equivalent transmission edges that pass the test with the sparse temporal knowledge graph of the target domain, adjust the Bayesian prior precision parameter with the reciprocal of the information loss rate, perform adaptive Bayesian parameter updates on each equivalent transmission edge, and output the target domain enhanced temporal knowledge graph. The domain-invariant embedding generation module is configured to extract node embedding vectors of the target domain enhanced temporal knowledge graph through a graph convolutional network, and eliminate the distribution offset between the source domain and target domain embeddings through a domain adversarial distribution alignment layer to generate a domain-invariant business node embedding sequence. The predictive analysis report generation module is configured to concatenate the time series of the target business indicators with the embedded sequence of business nodes with unchanged domains by time steps, generate the predicted value sequence of business indicators through the encoding and decoding prediction network, and label the confidence level and original transmission path traceability information of each equivalent transmission edge based on attention weights, and output an intelligent analysis report of business data.

[0006] Furthermore, in the cross-domain granularity-aware alignment module, the alignment probability is calculated as follows: ,in Source domain granularity layer Nodes within, For the target domain, the same granularity layer Nodes within, and These are the fused representation vectors of the source domain nodes and the target domain nodes, respectively. For temperature parameters, Target domain granularity layer The total number of nodes within, Traversing the target domain granularity layer All nodes within; the fusion representation vector is generated by concatenating the semantic embedding vector and the topological feature vector after their respective Z-score standardization, and the topological feature vector includes in-layer degree features, clustering coefficients and local community labels.

[0007] Furthermore, in the cross-domain granularity-aware alignment module, the granularity-level comparison submodule calculates the alignment success rate and the total number of nodes for each granularity layer in the hierarchical time-series knowledge graph of the target domain. When the total number of nodes in a certain granularity layer is lower than a preset minimum node threshold... Or the alignment success rate is lower than the preset minimum alignment rate. When this happens, the granularity level is marked as a granularity missing level.

[0008] Furthermore, in the equivalent transmission operator generation module, the information transmission matrix The element is defined as: ,in This represents the total number of time steps in the historical time series of the source domain. For nodes The time series in the first Always in a state of readiness The indicated quantity, For the node In state subsequence to node In state The conditional propagation entropy of the subsequence propagation. and These are the state range indices for the sending and receiving nodes, respectively. The number of discretized state intervals; each node's time series is Z-score standardized before calculating the information transmission matrix, and each node's time series independently and evenly divides its value range into . A number of equal-width intervals.

[0009] Furthermore, in the equivalent conduction operator generation module, the equivalent conduction operator is generated in the following way: ,in For the end node End node The equivalent propagation operator, For the path from segment 1 to segment 2 The information transmission matrix of the segment, The number of intermediate nodes in the granularity missing level; the equivalent time delay parameter is... ,in For the first The time delay parameter of the causal relationship edge; the equivalent influence strength parameter is... spectral norm Information loss rate: ,in It is the Frobenius norm.

[0010] Furthermore, in the causal template verification and parameter correction module, the Bayesian prior accuracy parameter is calculated as follows: ,in For prior accuracy parameters, This is a global scaling constant. The information loss rate of the equivalent propagation edge. To prevent extremely small positive numbers from being divided by zero; the prior distribution affecting the intensity parameter is as follows: The prior distribution of the time delay parameter is as follows Using the time series of node pairs corresponding to the target domain as likelihood data, the posterior parameter estimate is calculated through conjugate Bayes update.

[0011] Furthermore, the domain-invariant embedding generation module employs a relation-type-aware graph convolutional network, configuring independent weight matrices for different types of relation edges in the target domain-enhanced temporal knowledge graph. These different types of relation edges include causal edges, aggregation edges, decomposition edges, and equivalent propagation edges. The domain adversarial distribution alignment layer comprises a domain discriminator and a gradient inversion layer. The loss function of the domain discriminator is: ,in This represents the total number of nodes participating in the training. For nodes domain tags, For the domain discriminator to the node The output is the probability that the data belongs to the source domain; the gradient inversion layer inverts the gradient of the domain discriminator and passes it to the parameters of the graph convolutional network during backpropagation.

[0012] Furthermore, when the graph convolutional network and the domain discriminator are jointly optimized, the node representation task loss is the causal link reconstruction loss. The value of the inner product of the end node embedding vectors output by the graph convolutional network and mapped by the sigmoid function is used as the predicted value of the probability of the existence of causal edges, and the binary cross-entropy is calculated with the real edge labels.

[0013] Furthermore, in the predictive analysis report generation module, the encoder adopts a multi-layer Transformer encoder structure, and the decoder adopts a multi-layer Transformer decoder structure. The decoder uses cross-attention to access the embedding of the corresponding node in the target domain enhanced temporal knowledge graph of the future known planned event as a condition vector, and gradually generates the predicted value of the future time step in an autoregressive manner. The source tracing and annotation submodule combines the attention contribution of each equivalent transmission edge with the posterior confidence interval width of the equivalent transmission edge to divide the confidence level, and associates the original fine-grained transmission path information and information loss rate corresponding to each equivalent transmission edge in the analysis report.

[0014] This invention discloses a business data intelligent analysis method based on AI and knowledge graphs, used to execute the aforementioned knowledge graph-based business data intelligent analysis system, comprising: Step 1: Obtain business data from the source and target domains, generate time-series knowledge graphs for each domain, perform hierarchical clustering based on the inclusion and aggregation relationships between nodes, divide business indicator nodes into multi-level granularity layers, and generate hierarchical time-series knowledge graph structures for each of the two domains. Step 2: Extract the semantic embedding vector and topological feature vector of each node in the two-domain hierarchical temporal knowledge graph and concatenate them into a fusion representation. Calculate the alignment probability between source domain nodes and target domain nodes within the same granularity layer, generate a cross-domain entity alignment mapping table, and identify the missing granularity level of the target domain by comparing the granularity hierarchical structure of the two domains. Step 3: Extract multi-step causal transmission paths from the source domain time-series knowledge graph, passing through intermediate nodes of the missing granularity level. Calculate the information transmission matrix between each adjacent node on the path based on historical data from the source domain. Perform ordered matrix multiplication on all information transmission matrices to generate equivalent transmission operators. Calculate the information loss rate of each path and output a set of candidate causal analysis templates after granularity adaptation. Step 4: Use the target domain data to perform conditional independence tests on each equivalent transmission edge in the candidate causal analysis template to filter out incompatible causal edges. Merge the equivalent transmission edges that pass the test with the sparse temporal knowledge graph of the target domain, and adjust the Bayesian prior precision parameter with the reciprocal of the information loss rate. Perform adaptive Bayesian parameter updates on each equivalent transmission edge and output the enhanced temporal knowledge graph of the target domain. Step 5: Extract node embedding vectors of the target domain augmented temporal knowledge graph through a graph convolutional network, and eliminate the distribution offset between the source domain and target domain embeddings through a domain adversarial distribution alignment layer to generate a domain-invariant business node embedding sequence. Step 6: Concatenate the target business indicator time series with the domain-invariant business node embedding sequence by time step, generate the business indicator prediction value sequence through the encoding and decoding prediction network, and label the confidence level and original transmission path traceability information of each equivalent transmission edge based on attention weights, and output a business data intelligent analysis report.

[0015] The present invention has the following beneficial effects: The hierarchical knowledge graph generation module explicitly divides business indicators into three granularity levels: strategic, tactical, and operational, based on inclusion and aggregation relationships. This allows cross-domain alignment to be performed within the same granularity level and can identify missing granularity levels in the target domain. As a result, cross-domain analysis migration is extended from peer-level mapping to hierarchical perception mapping, overcoming the alignment failure problem caused by peer-level mapping under conditions of asymmetric granularity between mature and new business domain indicator systems.

[0016] The equivalent transmission operator generation module calculates the information transmission matrix for multi-step fine-grained causal transmission paths that have passed through the granularity missing level and performs ordered matrix product compression. While adapting the multi-step transmission link to a single-step equivalent transmission representation that can be supported by the coarse-grained data of the new business domain, it retains the directionality and intensity attenuation information of the original multi-step transmission in the form of matrix product, avoiding the causal link breakage caused by directly discarding intermediate transmission nodes.

[0017] The causal template verification and parameter correction module adjusts the Bayesian prior accuracy parameter by using the reciprocal of the information loss rate. This ensures that equivalent transmission edges with low information loss maintain strong prior constraints to fully reuse the causal transmission information from the source domain, while equivalent transmission edges with high information loss have reduced prior constraints so that the posterior is more dominated by the observation data from the target domain. This allows the weight ratio of prior and posterior in the parameter update process to adapt to the degree of granular compression, avoiding systematic distortion of the fixed-weight prior on equivalent transmission edges with different degrees of compression.

[0018] The domain-invariant embedding generation module eliminates the distribution offset between the source domain pre-trained embedding and the target domain embedding through gradient inversion adversarial training, so that the input embedding of the subsequent encoding-decoding prediction network is not affected by domain differences.

[0019] While outputting the predicted value sequence of business indicators, the predictive analysis report generation module marks the confidence level of each equivalent transmission edge on the target domain enhanced temporal knowledge graph based on attention weights and associates it with the original fine-grained transmission path information, so that the intelligent analysis conclusions of business data have a traceable causal explanation path and granularity-adapted credibility level. Attached Figure Description

[0020] Figure 1 This is a flowchart of the intelligent business data analysis method based on AI and knowledge graph provided in the embodiments of the present invention; Figure 2This is a schematic diagram of the overall architecture of the cross-domain causal knowledge transfer system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the number of nodes in the hierarchical temporal knowledge graph of the source domain and the target domain provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the cross-domain entity alignment probability distribution provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing key parameters of candidate equivalent conductive edges provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the equivalent conduction operator generation flowchart provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the influence intensity before and after Bayesian parameter correction according to an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the predicted transaction volume of a live-streaming e-commerce target domain over the next three months, provided by an embodiment of the present invention. Figure 9 This is a schematic diagram illustrating the relationship between information loss rate and Bayesian prior accuracy provided in an embodiment of the present invention. Detailed Implementation

[0021] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0022] At least one embodiment of the present invention discloses a business data intelligent analysis system based on AI and knowledge graphs, the method of which is described below. Figure 1 The system includes processes such as hierarchical knowledge graph generation, cross-domain granularity-aware alignment, equivalent transmission operator compression, adaptive Bayesian parameter correction, and domain-invariant embedding prediction to achieve reliable causal analysis migration and trend prediction of new business domains under granularity asymmetry conditions.

[0023] System main entrance: The business data intelligent analysis system receives historical business indicator time series and event records from mature source business domains, and short-term historical business data and event records from target new business domains as input. It outputs a business data intelligent analysis report with granular adaptation tracing and migration credibility grading annotations. The system includes a hierarchical knowledge graph generation module, a cross-domain granularity-aware alignment module, an equivalent transmission operator generation module, a causal template verification and parameter correction module, a domain-invariant embedding generation module, and a predictive analysis report generation module.

[0024] The hierarchical knowledge graph generation module is configured to acquire business data from two domains and generate a hierarchical time-series knowledge graph structure. The hierarchical knowledge graph generation module acquires historical business indicator time series and event records from one or more mature source business domains, extracts business indicator entities, business event entities, and their temporal relationships, and generates a complete temporal knowledge graph for each source domain. Nodes represent business indicator entities and business event entities, while edges represent causal relationships and carry influence strength and time delay parameters. Simultaneously, it acquires existing short-term historical business data and event records from the target new business domain and generates a sparse temporal knowledge graph for the target domain in the same manner.

[0025] The hierarchical granularity partitioning submodule performs hierarchical clustering based on inclusion and aggregation relationships on the two domain time-series knowledge graphs respectively. Specifically, the hierarchical granularity partitioning submodule identifies inclusion relationships (i.e., a certain indicator is composed of multiple sub-indicators) and aggregation relationships (i.e., a certain indicator is the statistical aggregation result of multiple similar indicators) between nodes in the time-series knowledge graph. Using inclusion and aggregation relationships as the distance metric for hierarchical clustering, the submodule divides the business indicator nodes in each time-series knowledge graph into three granularity layers according to their abstraction level: strategic layer, tactical layer, and operational layer. The strategic layer contains the highest level of abstraction business indicator nodes, such as overall customer value and total revenue growth rate; the tactical layer contains intermediate level business indicator nodes, such as channel conversion rate and customer retention rate; and the operational layer contains the finest level of abstraction business indicator nodes, such as functional module-level activity and single-channel daily clicks. The layers are connected by aggregation edges (from the operational layer to the tactical layer and from the tactical layer to the strategic layer) and decomposition edges (in the opposite direction), generating the hierarchical time-series knowledge graph structure for each domain.

[0026] The cross-domain granularity-aware alignment module is configured to generate a cross-domain entity alignment mapping table and identify missing granularity levels. The cross-domain granularity-aware alignment module extracts semantic embedding vectors of entity names and attribute descriptions for business entities and indicator entities in the two-domain hierarchical temporal knowledge graph using a pre-trained language model. It also granularizes the data at the source domain hierarchical temporal knowledge graph. Nodes within Layer with the same granularity as the target domain Nodes within Their semantic embedding vectors are denoted as follows: and ,in For granularity layer identification, For semantic embedding dimension, For source domain node indexing, Index the target domain node.

[0027] The topology feature extraction submodule performs graph topology encoding on the adjacency relationships of each node within its respective granularity layer, extracting the intra-layer degree features, clustering coefficients, and local community labels of each node, and converting the topology feature vector... and Concatenate with semantic embedding vectors to generate a fused representation and ,in For topological feature dimension, This represents a vector concatenation operation. Semantic embedding vectors. , Z-score standardization is performed on the inputs before concatenation, and the topological feature vectors are... , Z-score normalization is also performed to eliminate the impact of differences in the dimensions of semantic embedding and topological features on the fused representation.

[0028] Alignment probability calculation submodules at the same granularity level Within, compute source domain nodes With target domain node Alignment probability between: ; in Temperature is a parameter that controls the smoothness of the alignment probability distribution. The range of values ​​is , The smaller the distribution, the more concentrated it is on the node with the highest similarity. The larger the size, the more uniform the distribution; Traversing the target domain granularity layer All nodes within, Target domain granularity layer The total number of nodes within; Indicates transpose; Represents an exponential function; and Source domain nodes and target domain nodes The fused representation vector, where both vectors have dimensions. It can perform inner product operations. For each source domain node, the alignment probability exceeds a preset threshold. The target domain nodes form alignment node pairs, and the alignment node pairs of all granularity layers are aggregated to generate a cross-domain entity alignment mapping table.

[0029] The granularity-level comparison submodule compares the granularity-level structures of the two domains: for each granularity layer in the hierarchical knowledge graph of the target domain, it calculates the proportion of successfully aligned nodes in that layer to the total number of nodes in that layer, as well as the total number of nodes in that layer itself. When the total number of nodes in a certain granularity layer of the target domain falls below a preset minimum node threshold... Or the alignment success rate is lower than the preset minimum alignment rate. When the specified granularity level is reached, mark that granularity level as missing; mark the remaining granularity levels as symmetric. Output the identifiers for missing and symmetric granularity levels.

[0030] The equivalent propagation operator generation module is configured to compress multi-step causal paths that pass through granularity-deficient levels into equivalent propagation operators: The equivalent propagation operator generation module, based on a cross-domain entity alignment mapping table, extracts a causal analysis subgraph associated with the alignment nodes in the target domain from the complete temporal knowledge graph of the source domain. The causal analysis subgraph contains a multi-step causal propagation path formed by the alignment nodes as endpoints and fine-grained intermediate index nodes within the source domain.

[0031] The path identification submodule traverses all causal propagation paths in the causal analysis subgraph, identifying paths that pass through nodes at the granularity missing level. For a path starting from a granularity symmetric level endpoint... Starting from, passing through the intermediate node sequence of the missing granularity level. Reaching another granularity-symmetric level end node The multi-step causal transmission path, in which This represents the number of intermediate nodes. The intermediate nodes of the sequentially arranged granularity missing hierarchy, along with the causal relationship edges between adjacent nodes on this path, carry influence strength parameters. and time delay parameters , ,in For the first on the path The sequence number of the edge of the causal relationship.

[0032] The information transmission matrix calculation submodule uses historical business data from the source domain to calculate the information transmission matrix based on the causal transmission relationships between adjacent nodes along the path. For adjacent nodes... arrive (in , Extracting nodes from historical business data in the source domain time series and nodes time series ,in To index the time steps, Z-score standardization is performed on the time series of each node before calculating the information transmission matrix, eliminating the impact of differences in units and numerical ranges of different business indicators on the calculation of transmission entropy. The time series... The range of values ​​is uniformly discretized as Each state interval will be used to divide the time series. The range of values ​​is also uniformly discretized as follows: There are 1 state intervals, among which To determine the number of discretized state intervals, generate the information transmission matrix for that edge. Its elements Defined as: ; in This represents the total number of time steps in the historical time series of the source domain. For nodes The time series in the first Always in a state of readiness The indicator value (a value of 1 indicates that it is in state) A value of 0 indicates that the state is not in the specified state. ), and These are the state range indices for the sending and receiving nodes, respectively. For the node In state subsequence to node In state The conditional transfer entropy of the subsequence propagation is calculated using the information gain of the sending sequence to the next state of the receiving sequence, given the history of the receiving sequence itself, and is measured by the standard transfer entropy algorithm. Represents a node In state Time-direction node status The intensity of the transmitted information. (The above) The discretization method for each state interval is uniformly applied to all nodes on the path, meaning that the time series of each node independently and evenly divides its value range into [variables]. Each node has an equal-width interval, and the information transmission matrix of each node is... The dimension ensures the consistency of the dimension between adjacent information transmission matrices along the path, making subsequent matrix multiplication operations valid.

[0033] The matrix compression submodule covers all paths. Perform ordered matrix multiplication on each information transmission matrix to generate the equivalent transmission operator for that path: ; in Compress multi-step, fine-grained transmission links into slave nodes. End node The single-step equivalent propagation representation, For the path from segment 1 to segment 2 The information transmission matrix of the segment. The equivalent time delay parameter of this path is... ,in For the first The time delay parameter of the causal relationship edge. The total number of causal edges on the path; the equivalent influence strength parameter is... spectral norm .

[0034] The information loss rate calculation submodule calculates the information loss rate for each compressed path: ; in It is the Frobenius norm. For the equivalent propagation operator, For the first Information transmission matrix of the segment, This represents the total number of information transmission matrices along the path. The range of values ​​is A larger value indicates that more transmission details are lost during the compression process.

[0035] The above compression process is performed on all multi-step causal propagation paths passing through the missing granularity level in the causal analysis subgraph, generating a granularity-adapted candidate causal analysis template set. Each equivalent propagation edge in the candidate causal analysis template set contains end node pairs and equivalent propagation operators. Equivalent time delay parameter Equivalent influence intensity parameters and the corresponding information loss rate At the same time, the original fine-grained transmission path information corresponding to the equivalent edge is preserved.

[0036] The causal template validation and parameter correction module is configured to filter incompatible causal edges and perform adaptive Bayesian parameter updates. The conditional independence test submodule uses short-term business data from the target new business domain to perform conditional independence tests on each equivalent propagation edge in the candidate causal analysis template set. For equivalent propagation edges... Extract the time series of the corresponding aligned nodes from the short-term business data of the target domain. Under the condition of controlling for other adjacent node variables besides the equivalent propagation edge, perform a conditional independence test based on partial correlation. If the test... The value is higher than the preset significance level. If the equivalent propagation edge is found to be statistically inconsistent with the target domain observation data, it will be filtered out and removed from the candidate causal analysis template set.

[0037] The graph structure merging submodule merges the equivalent propagation edges and their corresponding granularity-symmetric hierarchical nodes from the candidate causal analysis templates that have passed the conditional independence test with the target domain sparse temporal knowledge graph. During merging, nodes with existing correspondences in the cross-domain entity alignment mapping table are directly merged into the same node, and newly introduced equivalent propagation edges are added to the target domain temporal knowledge graph as new causal relationship edges.

[0038] The Bayesian parameter correction submodule performs adaptive Bayesian parameter updates on the influence strength parameters and time delay parameters of each equivalent propagation edge in the merged target domain temporal knowledge graph. For equivalent propagation edges... Using source domain equivalent influence intensity parameters and equivalent time delay parameter These are used as the means of the prior distributions for influence intensity and time delay, respectively. The accuracy of the prior distributions is adjusted by the information loss rate. ; in For prior accuracy parameters, This is a global scaling constant, taking the value of a positive real number, used to control the overall magnitude of prior accuracy. The information loss rate of this equivalent propagation edge. To prevent division by zero of extremely small positive numbers. When the information loss rate... At lower levels, Larger size, higher prior distribution accuracy, smaller information loss in the equivalent propagation operator of the source domain, and stronger prior constraints; when At higher levels, The smaller size and lower prior distribution accuracy indicate a large loss of information in the source domain equivalent transmission operator, weak prior constraints, and a greater reliance on the target domain observation data in the posterior domain.

[0039] The priors affecting the intensity parameters are set to a normal distribution. The prior setting of the time delay parameter is a normal distribution. Using the time series of corresponding node pairs in the coarse-grained operational observation data of the target domain as likelihood data, posterior parameter estimates are calculated through conjugate Bayesian updates, generating posterior influence strength parameters and posterior time delay parameters for each equivalent transmission edge.

[0040] The Bayesian parameter correction submodule outputs a target-domain enhanced temporal knowledge graph with adapted granularity and corrected parameters. Each equivalent propagation edge carries the posterior parameter estimate, posterior confidence interval, and corresponding information loss rate. And the original fine-grained conduction path information.

[0041] The domain-invariant embedding generation module is configured to generate business node embedding sequences that eliminate domain distribution offsets through adversarial training. The domain-invariant embedding generation module extracts node embedding vectors from business nodes in the target domain augmented temporal knowledge graph using a relation-type-aware graph convolutional network (R-GCN). The input to the R-GCN is the adjacency structure and initial node features of the target domain augmented temporal knowledge graph. The initial node features include the posterior influence strength parameters of each equivalent propagation edge, the posterior time delay parameters, and node attributes. The output is the embedding vector of each business node. ,in For embedded dimensions, For node indexing, R-GCN assigns independent weight matrices to different types of relation edges (causal edges, aggregation edges, decomposition edges, and equivalent propagation edges) in the target domain augmented temporal knowledge graph. It aggregates information from each type of neighboring node for each node, generating relation-aware node embedding vectors. The posterior influence strength parameter and posterior time delay parameter in the initial node features are Z-score standardized before being input into R-GCN to eliminate the influence of different parameter scales on embedding learning.

[0042] The domain adversarial distributed alignment layer receives node embedding vectors output by R-GCN and pre-trained embedding vectors of source domain temporal knowledge graph nodes encoded by the same R-GCN as references. The domain adversarial distributed alignment layer comprises a domain discriminator and a gradient inversion layer. The domain discriminator is a multi-layer fully connected network, with node embedding vectors as input. The output layer is a single-neuron fully connected layer connected to a sigmoid activation function, outputting the probability that the node belongs to the source domain. The binary cross-entropy is used as the loss function for the domain discriminator: ; in This represents the total number of nodes participating in the training. For nodes The domain label (source domain is 1, target domain is 0). For the domain discriminator to the node The output is the probability that the value belongs to the source domain. For node indexing. The gradient inversion layer inverts the gradient of the domain discriminator during backpropagation and passes it to the R-GCN parameters, so that R-GCN can optimize the loss of node representation task while adversarially reducing the distribution difference between the source domain embedding and the target domain embedding.

[0043] During training, R-GCN is jointly optimized with the domain discriminator: the domain discriminator uses... To maximize domain classification accuracy, R-GCN uses a gradient inversion layer to adversarially minimize the classification accuracy of the domain discriminator, while simultaneously constraining the business semantic fidelity of node embeddings with a node representation task loss. The node representation task loss is the causal link reconstruction loss, which involves predicting the existence of causal edges in the target domain-enhanced temporal knowledge graph using binary cross-entropy: the dot product of the end node embedding vectors output by R-GCN, mapped by the sigmoid function, is used as the predicted probability of the causal edge existence, and this is used to calculate the binary cross-entropy with the real edge labels; parameter updates employ the Adam optimization algorithm. After convergence of adversarial training, the node embedding vectors output by R-GCN are domain-invariant representations.

[0044] To enhance the target domain, the domain-invariant embedding vectors of each business node in the temporal knowledge graph are arranged according to time steps, generating a domain-invariant business node embedding sequence. ,in The number of existing time steps in the target domain. For time step indexing.

[0045] The predictive analysis report generation module is configured to generate a series of predicted business indicator values ​​and an intelligent analysis report with source traceability annotations. The time-step splicing submodule splices the coarse-grained numerical time series of the target business metrics. Business node embedding sequence with domain immutability Concatenate the sequences according to time steps to generate a fused input sequence. ,in , For the first Coarse-grained business metric values ​​for each time step. For the first The domain-invariant node embedding vector at each time step. This represents the existing time steps in the target domain. Before splicing, the coarse-grained numerical time series... Z-score standardization is performed to eliminate the impact of dimensional differences between business metric values ​​and embedded vector components on subsequent encoding; the scalar prediction value of the prediction output is then used. The business indicator values ​​are restored to their original dimensions through inverse transformation.

[0046] The encoder-decoder prediction network consists of an encoder and a decoder. During training, the real business metrics from historical time steps in the target domain are used as supervision signals. The mean squared error loss function is employed, and the parameters are updated using the Adam optimization algorithm. The encoder uses a multi-layer Transformer encoder structure, and the input is a fused input sequence. The fused input sequence undergoes self-attention computation and feedforward transformation, outputting a context-coded sequence. .

[0047] The decoder enhances the temporal knowledge graph embedding by using cross-attention to access the target domain of future known planned events, thus providing conditional guidance. For future time steps... ( To predict the time domain length, the corresponding nodes of known planned events (such as product launch dates and marketing campaign dates) in the target domain augmented time-series knowledge graph are embedded as condition vectors. ,in , This provides an index for future time steps. The decoder employs a multi-layer Transformer decoder structure, progressively generating predictive representations for future time steps in an autoregressive manner. Within each decoding layer, cross-attention is used to link the decoder's hidden state with the conditional vector. and context encoding sequence Attention interaction is performed. The decoder output layer is a fully connected layer, which stores the final hidden states of the decoder at each future time step. ( The hidden state dimension of the decoder is mapped to a scalar prediction value, i.e. ,in This is the output layer weight matrix. The output layer bias parameters output a sequence of predicted business metrics for the target new business domain. And intelligent analysis conclusions.

[0048] The source annotation submodule calculates the attention contribution of each node and causal edge in the target domain augmented temporal knowledge graph to the output of each prediction time step based on the attention weights of each cross-attention layer of the decoder. For each equivalent transmission edge in the target domain augmented temporal knowledge graph, the attention contribution of each equivalent transmission edge is combined with the width of its posterior confidence interval to classify the confidence level: equivalent transmission edges with narrow posterior confidence intervals and high attention contributions are labeled as high-confidence edges, while equivalent transmission edges with wide posterior confidence intervals or low attention contributions are labeled as low-confidence edges. Simultaneously, each equivalent transmission edge is associated with its original fine-grained transmission path information and information loss rate. The data is annotated in the business data intelligent analysis report. The output includes a business data intelligent analysis report with granularity-adaptive source tracing and migration reliability grading annotations.

[0049] This invention explicitly divides business indicators into three granularity levels—strategic, tactical, and operational—based on inclusion and aggregation relationships through a hierarchical knowledge graph generation module. This allows cross-domain alignment to be performed within the same granularity level and identifies missing granularity levels in the target domain. This expands cross-domain analysis migration from peer-to-peer mapping to hierarchical-aware mapping, overcoming alignment failures caused by peer-to-peer mapping under conditions of granularity asymmetry between mature and new business domain indicator systems. The equivalent transmission operator generation module calculates information transmission matrices for multi-step fine-grained causal transmission paths passing through missing granularity levels and performs ordered matrix product compression. While adapting multi-step transmission links to a single-step equivalent transmission representation supported by coarse-grained data in the new business domain, the original multi-step transmission's directionality and intensity attenuation information are preserved in the form of matrix products. This avoids causal link breaks and loss of diagnostic depth caused by directly discarding intermediate transmission nodes. By adjusting the Bayesian prior precision parameter with the reciprocal of the information loss rate, equivalent transmission edges with low information loss maintain strong prior constraints to fully reuse causal transmission information from the source domain, while equivalent transmission edges with high information loss have reduced prior constraints so that the posterior is more dominated by observation data from the target domain. This allows the weight ratio of prior and posterior to adapt to the degree of granular compression during parameter updates, solving the problem of systematic distortion caused by fixed-weight priors on equivalent transmission edges with different degrees of compression. The domain-invariant embedding generation module eliminates the distribution offset between the source domain pre-trained embeddings and the target domain embeddings through gradient inversion adversarial training, ensuring that the input embeddings of subsequent encoding, decoding, and prediction networks are not affected by domain differences. The prediction analysis report generation module, while outputting the predicted value sequence of business indicators, labels the confidence level of each equivalent transmission edge on the target domain augmented temporal knowledge graph based on attention weights and associates it with the original fine-grained transmission path information, enabling the intelligent analysis conclusions of business data to have a traceable causal explanation path and granularity-adapted credibility classification.

[0050] See Figures 2-9 The following application examples are proposed: A mature e-commerce business line (source domain) of an internet platform has been operating for many years, accumulating a complete multi-level business indicator system and time-series knowledge graph. In year A, the platform launched a new live-streaming e-commerce business line (target domain). Having only been online for three months, the observable data primarily consists of coarse-grained aggregated indicators (such as total transaction volume and monthly active users), lacking fine-grained indicator data at the channel and functional module levels. The goal is to leverage the causal analysis experience from the mature e-commerce domain to predict and intelligently analyze the monthly transaction volume trend of the live-streaming e-commerce business.

[0051] Hierarchical knowledge graph generation module: For mature e-commerce source domains, the module extracts business indicator entities and event entities from its 36 months of historical business data, identifies the hierarchical structure based on inclusion and aggregation relationships, and divides the indicator nodes into three levels of granularity. The strategic layer includes high-abstract granularity indicators such as overall customer value and total revenue growth rate; the tactical layer includes intermediate-granularity indicators such as channel conversion rate and customer retention rate; and the operational layer includes fine-granularity indicators such as functional module-level activity and single-channel daily click volume. A complete time-series knowledge graph with 47 nodes and 83 parameterized causal edges is generated for the mature e-commerce source domain.

[0052] For the target domain of live-streaming e-commerce, the module generates a sparse time-series knowledge graph from its short-term historical data of 3 months. The observable indicators of the target domain only cover the strategic layer (total transaction volume, monthly active users) and the tactical layer (anchor conversion rate, user retention rate). Due to insufficient data accumulation, the operational layer has only 2 nodes. The overall structure is a sparse structure with 12 nodes and 9 causal edges.

[0053] Table 1. Distribution of nodes in the two-domain hierarchical temporal knowledge graph

[0054] Cross-domain granularity-aware alignment module: The cross-domain granularity-aware alignment module uses a pre-trained language model to extract 512-dimensional semantic embedding vectors from the names and attribute descriptions of each node in both domains. The topology feature extraction submodule extracts 64-dimensional topology feature vectors. After splicing, a 576-dimensional fused representation is generated. Temperature parameter. Alignment probability threshold .

[0055] Taking the alignment probability calculation of the source domain tactical layer node "channel conversion rate" (S-T04) and the target domain tactical layer node "anchor conversion rate" (T-T02) as an example, the inner product of its fused representation vector after Z-score normalization is 1.83. The number of target domain nodes at this granularity level... The alignment probability is: ; The probability exceeds the threshold of 0.65, forming an aligned node pair. Granularity-level alignment results show that the target domain of the operational layer has only two nodes, and the successful alignment rate is only 8%, both below the minimum node threshold. and minimum alignment ratio The operational layer is marked as a granularity-deficient layer; the strategic and tactical layers are marked as granularity-symmetric layers.

[0056] Table 2 Cross-Domain Entity Alignment Mapping Table (Partial)

[0057] Equivalent conduction operator generation module: A typical multi-step causal transmission path was identified from the source domain causal analysis subgraph: the tactical layer node "channel conversion rate" ( → Operation layer node "Single channel daily click volume" → Operation layer node "Module Activity" → Tactical layer node "Customer retention rate" The path passes through a granularity-deficient level (operational layer), with the number of intermediate nodes... There are 3 causal relationship edges in total, and the time delay parameters for each edge are as follows: sky, sky, sky.

[0058] The information transmission matrix calculation submodule calculates the 36-month historical time series of the source domain ( After Z-score standardization, the time series of each node are uniformly discretized. For each state interval, calculate the information transmission matrix of the three causal relationship edges. The matrix compression submodule performs ordered matrix multiplication: ; The equivalent time delay parameter is Heaven, equivalent influence intensity parameter (spectral norm) The information loss rate is calculated as follows: ; Table 3. Set of candidate causal analysis templates (summary of equivalent transmission edges)

[0059] Causal template verification and parameter correction module: The conditional independence test submodule uses three months of data from the target domain to test three equivalent propagation edges. Perform a conditional independence test based on partial correlation. Test the equivalent propagation side of the "page popularity → channel conversion rate" relationship. The value is 0.11, which is higher than 0.05, indicating a statistical contradiction with the target domain observation data. It was filtered out and removed. The remaining two equivalent transmission edges passed the test and were included in the graph structure merging.

[0060] After the graph structures are merged, the Bayesian parameter correction submodule performs adaptive prior accuracy calculation on the equivalent propagation edges that have passed the test. , Taking "channel conversion rate → customer retention rate" as an example, its prior precision parameter is: ; The prior distribution of influence intensity is The prior distribution of time delay is Using the time series corresponding to the corresponding nodes of the target domain over three months as likelihood data, after conjugate Bayes update, the posterior influence strength parameter is 0.63, the posterior time delay parameter is 21 days, and the posterior confidence intervals (95%) are as follows: and sky.

[0061] For "anchor activity level → user retention rate" ), the prior accuracy parameter is The prior constraints are stronger, the posterior influence strength is 0.73, the posterior time delay is 17 days, the posterior confidence interval is narrower, and the confidence level is higher.

[0062] Table 4 Bayesian parameter correction results

[0063] Domain-invariant embedding generation module: The domain-invariant embedding generation module takes the target domain-enhanced temporal knowledge graph (14 nodes after merging) as input, and uses R-GCN to configure independent weight matrices for causal edges, aggregation edges, decomposition edges, and equivalent propagation edges, generating a 128-dimensional ( Node embedding vectors. The posterior influence strength parameter and posterior time delay parameter in the initial features of the nodes are normalized by Z-score and then input into the R-GCN.

[0064] The domain adversarial distribution alignment layer introduces source domain pre-trained embedding vectors as a reference, and is jointly trained through a gradient reversal layer. The domain discriminator loss function is: ; The training involved After convergence of adversarial training with 14 target domain nodes and 47 source domain nodes, the domain discriminator accuracy dropped to 54% (close to random guessing), indicating that the node embedding vectors output by R-GCN have eliminated the distribution offset between the source and target domains. The domain-invariant embedding vectors of each service node in the target domain are arranged according to a 3-month time step to generate an embedding sequence. ( (with months as the time step).

[0065] Predictive analysis report generation module: The time-step splicing submodule will display the time series of the overall transaction volume of the target domain of live-streaming e-commerce over three months. (After Z-score normalization) and a 128-dimensional domain-invariant embedding sequence are concatenated step-by-step to generate a 129-dimensional fused input sequence. ( In the encoder-decoder prediction network, the encoder performs self-attention computation on the fused sequence and outputs a context-encoded sequence. The decoder receives the condition vector of a known planned event (e.g., the node embedding corresponding to the "Platform Anniversary Live Stream Event" planned for the 4th month of year A). Perform cross-attention guidance to predict the next 3 months in an autoregressive manner. The transaction amount is predicted and then restored to its original dimensions using the inverse Z-score transformation.

[0066] The source tracing and annotation submodule assigns confidence levels to two equivalent transmission edges based on the attention weights of each decoding cross-attention layer to the target domain enhanced temporal knowledge graph nodes, combined with the posterior confidence interval width: "Anchor activity → User retention rate" (posterior confidence interval width of 0.18, attention contribution of 0.41) is labeled as high confidence; "Channel conversion rate → Customer retention rate" (posterior confidence interval width of 0.24, attention contribution of 0.27) is labeled as medium confidence. Each equivalent transmission edge is also associated with the corresponding original multi-step fine-grained transmission path information and information loss rate.

[0067] Table 5. Output Results of the Predictive Analysis Report

[0068] The entire data flow process embodies a complete link logic: the hierarchical knowledge graph generation module divides the two-domain indicator nodes into three granularity layers and identifies the target domain operation layer as the granularity missing layer. This label directly drives the equivalent transmission operator generation module to perform matrix compression and quantify the information loss rate on the multi-step causal path passing through the operation layer; the information loss rate is passed as a key intermediate quantity to the causal template verification and parameter correction module, adaptively adjusting the Bayesian prior accuracy (an information loss rate of 0.22 corresponds to a prior accuracy of 2.27, and an information loss rate of 0.36 corresponds to a prior accuracy of 1.39), so that the source domain prior and... The weighting and compression levels of the observations in the target domain are matched; the corrected posterior parameters are used as the initial features of the nodes and enter the domain-invariant embedding generation module. After adversarial training to eliminate the domain distribution offset, a domain-invariant embedding sequence is generated; this embedding sequence is concatenated with the coarse-grained transaction amount sequence of the target domain and input into the encoding-decoding prediction network. Under the guidance of known planned event conditions, the network outputs the predicted values ​​for the next 3 months. The source tracing and annotation submodule associates the prediction conclusions with the confidence levels of each equivalent transmission edge and the original fine-grained transmission path, and finally generates a live-streaming e-commerce business data intelligent analysis report with complete causal source tracing and granular adaptive annotation.

[0069] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A business data intelligent analysis system based on AI and knowledge graphs, characterized in that, include: The hierarchical knowledge graph generation module is configured to acquire business data from the source and target domains, generate time-series knowledge graphs for each domain, and perform hierarchical clustering based on the inclusion and aggregation relationships between nodes to divide business indicator nodes into multi-level granularity layers, generating hierarchical time-series knowledge graph structures for each of the two domains. The cross-domain granularity-aware alignment module is configured to extract the semantic embedding vector and topological feature vector of each node in the hierarchical temporal knowledge graph of two domains and concatenate them into a fusion representation. Within the same granularity layer, it calculates the alignment probability between the source domain node and the target domain node, generates a cross-domain entity alignment mapping table, and identifies the missing granularity level of the target domain by comparing the granularity hierarchical structure of the two domains. The equivalent transmission operator generation module is configured to extract multi-step causal transmission paths from the source domain temporal knowledge graph, passing through intermediate nodes with missing granularity levels, calculate the information transmission matrix between each adjacent node on the path based on historical data of the source domain, perform ordered matrix multiplication on all information transmission matrices to generate equivalent transmission operators, calculate the information loss rate of each path, and output a set of candidate causal analysis templates after granularity adaptation. The causal template verification and parameter correction module is configured to perform conditional independence tests on each equivalent transmission edge in the candidate causal analysis template using target domain data to filter incompatible causal edges, merge the equivalent transmission edges that pass the test with the sparse temporal knowledge graph of the target domain, adjust the Bayesian prior precision parameter with the reciprocal of the information loss rate, perform adaptive Bayesian parameter updates on each equivalent transmission edge, and output the target domain enhanced temporal knowledge graph. The domain-invariant embedding generation module is configured to extract node embedding vectors of the target domain enhanced temporal knowledge graph through a graph convolutional network, and eliminate the distribution offset between the source domain and target domain embeddings through a domain adversarial distribution alignment layer to generate a domain-invariant business node embedding sequence. The predictive analysis report generation module is configured to concatenate the time series of the target business indicators with the embedded sequence of business nodes with unchanged domains according to time steps, generate the predicted value sequence of business indicators through the encoding and decoding prediction network, and label the confidence level and original transmission path traceability information of each equivalent transmission edge based on attention weights, and output an intelligent analysis report of business data.

2. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the cross-domain granularity-aware alignment module, the alignment probability is calculated as follows: ； in Source domain granularity layer Nodes within, For the target domain, the same granularity layer Nodes within, and These are the fused representation vectors of the source domain nodes and the target domain nodes, respectively. For temperature parameters, Target domain granularity layer The total number of nodes within, Traversing the target domain granularity layer All nodes within; The fusion representation vector is generated by concatenating the semantic embedding vector and the topological feature vector after their respective Z-score standardization. The topological feature vector includes intra-layer degree features, clustering coefficients, and local community labels.

3. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the cross-domain granularity-aware alignment module, the granularity-level comparison submodule calculates the alignment success rate and the total number of nodes for each granularity layer in the hierarchical time-series knowledge graph of the target domain. When the total number of nodes in a certain granularity layer is lower than a preset minimum node threshold, the alignment is terminated. Or the alignment success rate is lower than the preset minimum alignment rate. When this happens, the granularity level is marked as a granularity missing level.

4. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the equivalent transmission operator generation module, the information transmission matrix The element is defined as: ； in This represents the total number of time steps in the historical time series of the source domain. For nodes The time series in the first Always in a state of readiness The indicated quantity, For the node In state subsequence to node In state The conditional propagation entropy of the subsequence propagation. and These are the state range indices for the sending and receiving nodes, respectively. The number of discretized state intervals; Before calculating the information transmission matrix, each node's time series is Z-score standardized, and each node's time series independently and evenly divides its value range into... A number of equal-width intervals.

5. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the equivalent conduction operator generation module, the equivalent conduction operator is generated in the following way: ； in For the end node End node The equivalent propagation operator, For the path from segment 1 to segment 2 The information transmission matrix of the segment, This represents the number of intermediate nodes in the missing granularity level. The equivalent time delay parameter is ,in For the first The time delay parameter of the causal relationship edge; the equivalent influence strength parameter is... spectral norm ; Information loss rate: ； in It is the Frobenius norm.

6. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the causal template verification and parameter correction module, the Bayesian prior accuracy parameter is calculated as follows: ； in For prior accuracy parameters, This is a global scaling constant. The information loss rate of the equivalent propagation edge. To prevent extremely small positive numbers from being divided by zero; The prior distribution of the influence on the intensity parameter is as follows The prior distribution of the time delay parameter is as follows Using the time series of node pairs corresponding to the target domain as likelihood data, the posterior parameter estimate is calculated through conjugate Bayes update.

7. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, The domain-invariant embedding generation module employs a relation type-aware graph convolutional network to configure independent weight matrices for different types of relation edges in the target domain-enhanced temporal knowledge graph. These different types of relation edges include causal edges, aggregation edges, decomposition edges, and equivalent propagation edges. The domain adversarial distribution alignment layer includes a domain discriminator and a gradient inversion layer. The loss function of the domain discriminator is: ； in This represents the total number of nodes participating in the training. For nodes domain tags, For the domain discriminator to the node The output is the probability that the value belongs to the source domain; The gradient inversion layer inverts the gradient of the domain discriminator and passes it to the parameters of the graph convolutional network during backpropagation.

8. The business data intelligent analysis system based on AI and knowledge graphs according to claim 7, characterized in that, When the graph convolutional network and the domain discriminator are jointly optimized, the node representation task loss is the causal link reconstruction loss. The value of the inner product of the end node embedding vectors output by the graph convolutional network and mapped by the sigmoid function is used as the predicted value of the probability of the existence of causal edges, and the binary cross-entropy is calculated with the real edge labels.

9. The business data intelligent analysis system based on AI and knowledge graphs according to claim 1, characterized in that, In the predictive analysis report generation module, the encoder adopts a multi-layer Transformer encoder structure, and the decoder adopts a multi-layer Transformer decoder structure. The decoder uses cross-attention to access the embedding of the corresponding node in the target domain augmented temporal knowledge graph of the future known planned event as a condition vector, and gradually generates the predicted value of the future time step in an autoregressive manner. The source tracing and annotation submodule combines the attention contribution of each equivalent transmission edge with the posterior confidence interval width of that equivalent transmission edge to classify the confidence level, and associates the original fine-grained transmission path information and information loss rate corresponding to each equivalent transmission edge in the analysis report.

10. A business data intelligent analysis method based on AI and knowledge graphs, used to execute the business data intelligent analysis system based on knowledge graphs as described in any one of claims 1 to 9, characterized in that, include: Step 1: Obtain business data from the source and target domains, generate time-series knowledge graphs for each domain, perform hierarchical clustering based on the inclusion and aggregation relationships between nodes, divide business indicator nodes into multi-level granularity layers, and generate hierarchical time-series knowledge graph structures for each of the two domains. Step 2: Extract the semantic embedding vector and topological feature vector of each node in the two-domain hierarchical temporal knowledge graph and concatenate them into a fusion representation. Calculate the alignment probability between source domain nodes and target domain nodes within the same granularity layer, generate a cross-domain entity alignment mapping table, and identify the missing granularity level of the target domain by comparing the granularity hierarchical structure of the two domains. Step 3: Extract multi-step causal transmission paths from the source domain time-series knowledge graph, passing through intermediate nodes of the missing granularity level. Calculate the information transmission matrix between each adjacent node on the path based on historical data from the source domain. Perform ordered matrix multiplication on all information transmission matrices to generate equivalent transmission operators. Calculate the information loss rate of each path and output a set of candidate causal analysis templates after granularity adaptation. Step 4: Use the target domain data to perform conditional independence tests on each equivalent transmission edge in the candidate causal analysis template to filter out incompatible causal edges. Merge the equivalent transmission edges that pass the test with the sparse temporal knowledge graph of the target domain, and adjust the Bayesian prior precision parameter with the reciprocal of the information loss rate. Perform adaptive Bayesian parameter updates on each equivalent transmission edge and output the enhanced temporal knowledge graph of the target domain. Step 5: Extract node embedding vectors of the target domain augmented temporal knowledge graph through graph convolutional network, and eliminate the distribution offset between the source domain and target domain embeddings through domain adversarial distribution alignment layer to generate domain-invariant business node embedding sequence. Step 6: Concatenate the target business indicator time series with the domain-invariant business node embedding sequence by time step, generate the business indicator prediction value sequence through the encoding and decoding prediction network, and label the confidence level and original transmission path source information of each equivalent transmission edge based on attention weights, and output the business data intelligent analysis report.

Images