An enterprise big data mining method and system based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based enterprise big data mining method and system. The method includes: using a multimodal hypergraph neural network to perform dynamic modal alignment based on the enterprise's multi-source heterogeneous data to generate a spatiotemporally consistent multimodal joint embedding tensor; inputting the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced class separability; performing spatiotemporal causal association mining on the semantic embedding vector to output a spatiotemporal causal metapath graph containing implicit business logic; inputting the spatiotemporal causal metapath graph into a dynamic game adversarial interpretation framework to ultimately output an enterprise-level intelligent decision-making graph with counterfactual robustness. Utilizing embodiments of the present invention, multimodal data can be efficiently integrated and intelligently analyzed, improving the accuracy and effectiveness of mining results.

Description

Technical Field

[0001] The present invention belongs to the field of data mining technology, and in particular to an enterprise big data mining method and system based on artificial intelligence. Background Art

[0002] In today's era of information explosion and digital transformation, businesses face unprecedented data challenges. In the course of their operations, businesses generate and accumulate vast amounts of heterogeneous, multi-source data, including both structured and unstructured data, such as sales records, customer feedback, market research, social media interactions, and sensor data. This data holds significant commercial value and provides a crucial basis for business decision-making. However, effectively extracting useful information and insights from this vast amount of data has become a crucial challenge in business operations and strategy formulation. Traditional data mining methods rely heavily on linear models and single-type data analysis. However, in the face of increasingly complex business environments, these approaches are increasingly facing limitations. In particular, when dealing with multi-source, heterogeneous data, traditional methods often struggle to adapt to the dynamic nature of the data and the complexity of cross-modal features, significantly compromising the accuracy and effectiveness of mining results. Summary of the Invention

[0003] The purpose of this invention is to provide an enterprise big data mining method and system based on artificial intelligence to address the deficiencies in the existing technology, which can efficiently integrate multimodal data and perform intelligent analysis to improve the accuracy and effectiveness of mining results.

[0004] One embodiment of the present application provides an enterprise big data mining method based on artificial intelligence, the method comprising:

[0005] Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used for dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0006] The multimodal joint embedding tensor is input into an orthogonal adversarial manifold learning module. A manifold projection space for high-dimensional sparse data is constructed based on a transfer learning framework. The feature distribution is adversarially corrected through an orthogonal-constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced class separability. The orthogonal constraint enforces orthogonalization of the weight matrices of the generator and discriminator through Frobenius norm regularization, thereby suppressing mode collapse.

[0007] Performing spatiotemporal causal association mining on the semantic embedding vectors, extracting multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjusting causal thresholds through reinforcement learning, and outputting a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search;

[0008] The spatiotemporal causal meta-path graph is input into a dynamic game adversarial interpretation framework, and the optimal decision-making strategy is generated based on the attention-driven policy optimization mechanism. The rule interpretability and commercial value gain are balanced through the KL divergence-constrained adversarial perturbation compensation algorithm, and finally an enterprise-level intelligent decision graph with counterfactual robustness is output. Among them, the framework jointly optimizes the stability of the explanation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

[0009] Optionally, the method uses a multimodal hypergraph neural network to perform dynamic modal alignment based on the enterprise's multi-source heterogeneous data, fuses cross-modal features through a self-attention-driven dynamic hypergraph construction algorithm, and generates a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates the time drift noise of cross-domain data through tensor decomposition, including:

[0010] Based on the company's multi-source heterogeneous data, including structured transaction records, unstructured customer feedback text, and IoT time series data, the modal alignment tensor decomposition algorithm is used to align timestamps and disambiguate semantics of the multi-source heterogeneous data, generating an initial tensor for cross-modal time series alignment.

[0011] The initial tensor is input into the self-attention driven dynamic hypergraph construction module, the cross-modal feature association weights are calculated through the inter-modal causal reasoning mechanism, the time drift noise is filtered using the causal mask matrix, and the dynamic hypergraph adjacency matrix is output;

[0012] Performing spatiotemporal joint embedding learning on the dynamic hypergraph adjacency matrix, using a multi-head graph attention network to aggregate the spatiotemporal dependencies of cross-modal nodes to generate a graph embedding vector with multimodal feature fusion;

[0013] The graph embedding vector is input into the tensor rank constraint compression module, and redundant features are eliminated through non-negative matrix factorization and low-rank projection, and finally a spatiotemporally consistent multimodal joint embedding tensor is output.

[0014] Optionally, the multimodal joint embedding tensor is input into an orthogonal adversarial manifold learning module, a manifold projection space for high-dimensional sparse data is constructed based on a transfer learning framework, and adversarial correction is performed on the feature distribution through an orthogonal constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced category separability, wherein the orthogonal constraint forces the orthogonalization of the weight matrices of the generator and the discriminator through Frobenius norm regularization to suppress mode collapse, including:

[0015] Based on the multimodal joint embedding tensor, a manifold projection space based on transfer learning is constructed. The adversarial domain adaptation algorithm is used to align the feature distributions of the source domain and the target domain to generate the initial manifold projection vector.

[0016] Inputting the initial manifold projection vector into an orthogonal constrained generative adversarial network, forcing the weight matrices of the generator and the discriminator to be orthogonalized through Frobenius norm regularization, and outputting an adversarial feature distribution after mode collapse suppression;

[0017] Dynamically correct the adversarial feature distribution, use the Wasserstein distance to measure the difference in feature separability, optimize the discriminator decision boundary through the gradient penalty mechanism, and generate an intermediate feature vector with enhanced class separability;

[0018] The intermediate feature vector is input into the manifold compactification module, and a spectral clustering-driven dimensionality reduction algorithm is used to compress the high-dimensional sparse features, and finally a low-dimensional compact semantic embedding vector is output.

[0019] Optionally, the spatiotemporal causal association mining is performed on the semantic embedding vector, a deep probabilistic reasoning model with meta-path optimization is used to extract multi-hop association rules, causal thresholds are dynamically adjusted through reinforcement learning, and a spatiotemporal causal meta-path graph containing implicit business logic is output, wherein the deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search, including:

[0020] Based on the semantic embedding vector, a deep probabilistic reasoning model for meta-path optimization is constructed. The initial multi-hop association rules are extracted through the path walking algorithm to generate a set of candidate meta-paths.

[0021] Perform reinforcement learning on the candidate meta-path set, dynamically adjust the causal threshold based on the temporal difference error, select meta-paths with confidence higher than the preset value through the path reward mechanism, and output the optimized causal rule set;

[0022] Performing a Monte Carlo tree search on the causal rule set, evaluating the causal strength by simulating counterfactual scenarios, and correcting the path weights in combination with Bayesian posterior probabilities to generate a meta-path probability map of spatiotemporal causal associations;

[0023] The meta-path probability graph is input into the graph pruning module, and low-significance edges are removed using the information entropy threshold, and finally a spatiotemporal causal meta-path graph containing implicit business logic is output.

[0024] Optionally, the spatiotemporal causal metapath graph is input into a dynamic game adversarial interpretation framework, an attention-driven policy optimization mechanism is used to generate an optimal decision strategy, and a KL divergence-constrained adversarial perturbation compensation algorithm is used to balance rule interpretability and commercial value gain, ultimately outputting an enterprise-level intelligent decision graph with counterfactual robustness. The framework jointly optimizes the stability of the interpretation boundary and the decision confidence through an implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance decision robustness, including:

[0025] Based on the spatiotemporal causal metapath graph, an attention-driven policy optimization model is constructed. The priority weights of rule interpretations are dynamically assigned through a multi-head self-attention mechanism to generate an initial decision strategy vector.

[0026] Input the initial decision policy vector into the KL divergence constrained adversarial perturbation compensation module, optimize the geometric distribution of the interpretation boundary through the implicit policy gradient algorithm, and output an intermediate decision policy with enhanced perturbation resistance;

[0027] Perform dynamic game equilibrium solution on the intermediate decision strategy, use Nash equilibrium iterative algorithm to balance rule interpretability and commercial value gain, and generate a decision strategy map for counterfactual robustness verification;

[0028] The decision strategy map is input into the manifold projection adversarial enhancement module, and the abnormal samples are compressed by generating the discriminator feature space of the adversarial network, and finally an enterprise-level intelligent decision map is output.

[0029] Another embodiment of the present application provides an enterprise big data mining system based on artificial intelligence, the system comprising:

[0030] A fusion module is used to perform dynamic modal alignment based on the enterprise's multi-source heterogeneous data using a multimodal hypergraph neural network. This module fuses cross-modal features using a self-attention-driven dynamic hypergraph construction algorithm to generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm incorporates an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0031] A correction module is configured to input the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, construct a manifold projection space for high-dimensional sparse data based on a transfer learning framework, perform adversarial correction on the feature distribution through an orthogonal-constrained generative adversarial network, and generate a low-dimensional, compact semantic embedding vector with enhanced class separability. The orthogonal constraint enforces orthogonalization of the weight matrices of the generator and discriminator through Frobenius norm regularization, thereby suppressing mode collapse.

[0032] An adjustment module is configured to perform spatiotemporal causal association mining on the semantic embedding vector, extract multi-hop association rules using a deep probabilistic reasoning model optimized for meta-paths, dynamically adjust causal thresholds through reinforcement learning, and output a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of meta-paths through a path reward mechanism and Monte Carlo tree search;

[0033] The output module is used to input the spatiotemporal causal meta-path graph into a dynamic game adversarial interpretation framework, generate the optimal decision-making strategy based on the attention-driven policy optimization mechanism, balance the rule interpretability and commercial value gain through the KL divergence-constrained adversarial perturbation compensation algorithm, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. The framework jointly optimizes the stability of the explanation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

[0034] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0035] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0036] Compared with the existing technology, the present invention provides an enterprise big data mining method based on artificial intelligence. According to the multi-source heterogeneous data of the enterprise, a multimodal hypergraph neural network is used to perform dynamic modal alignment to generate a multimodal joint embedding tensor that is consistent in time and space; the multimodal joint embedding tensor is input into the orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced category separability; the semantic embedding vector is mined for spatiotemporal causal associations to output a spatiotemporal causal meta-path graph containing implicit business logic; the spatiotemporal causal meta-path graph is input into a dynamic game adversarial interpretation framework to finally output an enterprise-level intelligent decision-making graph with counterfactual robustness, thereby enabling efficient integration of multimodal data and intelligent analysis, thereby improving the accuracy and effectiveness of the mining results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A hardware structure block diagram of a computer terminal for an enterprise big data mining method based on artificial intelligence provided by an embodiment of the present invention;

[0038] Figure 2 A flowchart of an enterprise big data mining method based on artificial intelligence provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of the structure of an enterprise big data mining system based on artificial intelligence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0041] The embodiment of the present invention first provides an enterprise big data mining method based on artificial intelligence, which can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0042] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for an enterprise big data mining method based on artificial intelligence provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0043] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to execute any one of the artificial intelligence-based enterprise big data mining methods.

[0044] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0045] The internal memory provides an environment for the operation of computer programs in non-volatile storage media. When the computer program is executed by the processor, the processor can execute any enterprise big data mining method based on artificial intelligence.

[0046] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0047] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0048] See also Figure 2 , an embodiment of the present invention provides an enterprise big data mining method based on artificial intelligence, which may include the following steps:

[0049] S201: Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used to perform dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0050] This method integrates data from different sources within the enterprise (such as structured transaction records, unstructured text, time-series sensor data, etc.) through a hypergraph neural network, uses the self-attention mechanism to dynamically calculate the association weights between each modality, and combines causal reasoning to eliminate the noise interference caused by time asynchrony. Ultimately, a spatiotemporal aligned multimodal joint feature representation is generated, which solves the problem of difficult unified modeling of multi-source heterogeneous data in enterprise big data, ensures that information of different modalities remains consistent in time and semantics, and provides high-quality feature input for subsequent deep mining.

[0051] Specifically, based on the company's multi-source heterogeneous data, including structured transaction records, unstructured customer feedback text, and IoT time series data, the modal alignment tensor decomposition algorithm can be used to align timestamps and disambiguate semantics of the multi-source heterogeneous data, generating an initial tensor for cross-modal time series alignment.

[0052] This method uses tensor decomposition technology to unify the timestamps of different data sources and eliminate semantic ambiguity (such as the different definitions of "order volume" in financial and operational systems), generating a preliminary aligned multimodal data tensor. This solves the problems of temporal asynchrony and semantic inconsistency in multi-source data and lays the foundation for subsequent feature fusion.

[0053] Modal alignment of an enterprise's multi-source, heterogeneous data is fundamental to building a unified semantic space. First, for structured transaction records (such as daily sales figures and inventory changes), fields such as timestamps, transaction amounts, and product categories must be extracted. Then, clock skews across different business systems must be aligned using a dynamic time warping (DTW) algorithm. For example, if the transaction timestamp in the sales system differs from the shipping time in the logistics system by minutes, DTW uses nonlinear stretching to align the timelines of the two, ensuring temporal consistency across data modalities for the same order event.

[0054] For unstructured customer feedback text (such as customer service conversation transcripts and social media comments), a pre-trained language model (BERT) is used for semantic embedding. First, key information such as product name and sentiment polarity is extracted through entity recognition (NER). This is then combined with context-aware word vector mapping to eliminate ambiguity. For example, the semantic meaning of the word "apple" in feedback about electronics and food requires contextual analysis and mapping to different embedding subspaces. Furthermore, time expressions in the text (such as "last Wednesday" and "three months ago") are standardized and converted into unified timestamps to align with the timeline of the structured data.

[0055] Processing IoT time-series data (such as temperature and humidity from production line sensors) requires addressing sampling frequency discrepancies. Multi-scale wavelet transforms are used to downsample high-frequency sensor data (such as vibration signals sampled 1,000 times per second) to the same minute-level granularity as transaction records, while preserving key frequency band characteristics. For example, the low-frequency component of temperature data (reflecting daily cyclical variations) and the high-frequency component (reflecting transient equipment anomalies) are reconstructed separately to ensure temporal correlation with business events.

[0056] After completing the above preprocessing, the three types of data are integrated into a three-dimensional tensor (time × feature × modality) using a modality-aligned tensor decomposition algorithm. Specifically, each time slice contains structured numerical features (such as transaction amount), a text embedding vector (768-dimensional BERT output), and sensor statistics (mean, variance). The tensor decomposition uses the Tucker decomposition model, with the core tensor dimensions set to (time × 50 × 3). Alternating least squares (ALS) iterative optimization is used to eliminate temporal drift noise in the cross-modal data. The resulting initial tensor exhibits cross-modal temporal alignment. For example, a customer complaint can be precisely associated with abnormal sensor readings and sales declines recorded during the same time period.

[0057] The initial tensor is input into the self-attention driven dynamic hypergraph construction module, the cross-modal feature association weights are calculated through the inter-modal causal reasoning mechanism, the time drift noise is filtered using the causal mask matrix, and the dynamic hypergraph adjacency matrix is output;

[0058] This method uses the self-attention mechanism to dynamically learn the causal relationship between different modalities (such as the correlation strength between "customer complaint text" and "sales decline"), and uses causal masking to eliminate noise interference caused by time dislocation, enhance the rationality of cross-modal feature fusion, and avoid false correlations caused by time drift.

[0059] The goal of constructing a dynamic hypergraph is to capture nonlinear correlations between cross-modal features. First, the initial tensor is sliced along the time axis, and the feature vector for each time step is fed into a multi-head self-attention mechanism. The number of attention heads is set to 8, with each head independently calculating the correlation weights between different modalities. For example, one attention head might focus on the cross-modal correlation between negative customer sentiment and a sudden temperature rise on the production line, while another might capture the causal relationship between promotional activities and inventory fluctuations.

[0060] To enhance the robustness of causal inference, an inter-modal causal mask matrix is introduced. Based on the principle of temporal precedence, this matrix prohibits the causal influence of subsequent events on previous events. For example, a customer complaint at time t cannot affect the sensor reading at time t-1, so the mask weight at the corresponding position is set to 0. The mask matrix is generated using a time-lagged Granger causality test. Causal paths with a significance greater than p=0.05 are screened using the F statistic; the remaining paths are suppressed.

[0061] The adjacency matrix of a dynamic hypergraph is updated using a gated recurrent unit (GRU). The attention weights for each time step are fed into the GRU along with the historical hypergraph state. The forget gate in the update formula controls the proportion of historical information retained (for example, a setting of 0.7 means 70% of the historical association weights are inherited). For example, if the causal strength of "declining customer satisfaction → declining sales" is detected to have increased continuously over three consecutive time steps, the GRU will dynamically increase the weight of that path while attenuating occasional noisy paths (such as spurious associations caused by a single sensor false alarm).

[0062] The final output dynamic hypergraph adjacency matrix contains weight information for nodes (features) and hyperedges (cross-modal connections). For example, a hyperedge might connect three nodes: "Social media negative sentiment score ≥ 0.8," "Production line temperature standard deviation > 5°C," and "Same-day return rate increased by 20%." Its weight of 0.92 indicates a strong causal relationship. This matrix serves as the foundation for subsequent graph embedding learning.

[0063] Performing spatiotemporal joint embedding learning on the dynamic hypergraph adjacency matrix, using a multi-head graph attention network to aggregate the spatiotemporal dependencies of cross-modal nodes to generate a graph embedding vector with multimodal feature fusion;

[0064] This method uses a graph attention network to capture the temporal and spatial dependencies of nodes in different modalities (such as the spatiotemporal correlation between "equipment failure in a certain area" and "customer churn in the area"), generates a fused graph embedding representation, realizes deep interactive modeling of multimodal data, and extracts more business-significant spatiotemporal features.

[0065] The core of joint spatiotemporal embedding learning is to simultaneously capture the spatial correlations and temporal evolution patterns between features. First, the dynamic hypergraph adjacency matrix is input into the spatiotemporal graph convolutional network (ST-GCN), which consists of two branches: spatial convolution and temporal convolution. The spatial convolution layer uses Chebyshev polynomial approximation (order K = 3) to aggregate neighborhood features on the hypergraph structure; the temporal convolution layer uses dilated causal convolution with a dilation factor of 2 to capture long-term dependencies. For example, for an abnormal temperature pattern in a piece of production equipment, spatial convolution identifies the nodes associated with upstream and downstream processes, while temporal convolution traces cyclical fluctuations over the past 24 hours.

[0066] To enhance cross-modal feature fusion, a heterogeneous multi-head graph attention mechanism (HMGAT) is designed. Each attention head focuses on the feature interactions of a specific modality combination:

[0067] Head 1: Structured data → Text modality (e.g., transaction amount and sentiment correlation of customer service texts);

[0068] Head 2: Text modality → sensor data (e.g., correlation between complaint keywords and the spectrum of device vibration);

[0069] Head 3: Sensor data → structured data (e.g. correlation of temperature spikes with inventory depletion rates).

[0070] The attention score for each head is calculated using cosine similarity weighting and activated with a LeakyReLU (negative slope 0.2). For example, in head 1, the cosine similarity between a product's daily sales vector and the corresponding customer service text embedding vector is 0.85, which is normalized to become the attention weight for that edge.

[0071] Finally, the outputs of each head are combined through a gated fusion mechanism, with gate weights dynamically learned by the fully connected layer. For example, if the importance of the text modality in recent data increases significantly (e.g., a large number of customer complaints), the gating mechanism will automatically increase the fusion weights of heads 1 and 2. The resulting graph embedding vector has a dimension of 256 and contains a compressed representation of cross-modal spatiotemporal dependencies. For example, an embedding vector might encode a complex pattern: "accumulation of negative customer sentiment over the past week → decreased production line efficiency → inventory backlog."

[0072] The graph embedding vector is input into the tensor rank constraint compression module, and redundant features are eliminated through non-negative matrix factorization and low-rank projection, and finally a spatiotemporally consistent multimodal joint embedding tensor is output.

[0073] This method uses low-rank decomposition technology to remove redundant information in features (such as highly correlated sales indicators), retain the most discriminative core features, improve the simplicity and generalization ability of feature representation, and reduce the complexity of subsequent calculations.

[0074] The goal of tensor rank-constrained compression is to remove noise and redundant information from the embedding vector. First, the graph embedding vector is reorganized into a three-dimensional tensor (time × node × feature) by time step and fed into a non-negative matrix factorization (NMF) module. The NMF basis matrix is set to (time × 50) and the coefficient matrix is (50 × node × feature). It is iteratively optimized using a multiplicative update rule, enforcing non-negativity constraints to enhance physical interpretability. For example, after decomposition, 50 basis modes may be obtained, where basis mode 3 corresponds to the "customer sentiment-sales positive feedback loop during promotional periods."

[0075] To further reduce dimensionality, a low-rank projection (LRP) technique was employed. Singular value decomposition (SVD) was used to extract the first 30 principal components of the tensor (retaining 95% of the variance), resulting in a projection matrix of (256×30) dimensions. For example, the feature representing "increased defective rate due to equipment aging" in a high-dimensional embedding vector was compressed into three principal components in the low-rank space, corresponding to "vibration frequency offset," "temperature baseline drift," and "increased production cycle time," respectively.

[0076] The final multimodal joint embedding tensor output has dimensions of (time × node × 30) and exhibits spatial and temporal consistency. For example, at time slice t = 2023-10-05, the embedding vector for the node "East China Warehouse" might reflect the chain reaction of "logistics delays → surge in customer complaints → urgent inventory allocation," with a smooth transition to the embeddings of previous and subsequent time steps. This tensor serves as a unified input for downstream tasks (such as causal reasoning and decision optimization), supporting the entire process of enterprise-level intelligent decision-making.

[0077] S202, inputting the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, constructing a manifold projection space for high-dimensional sparse data based on a transfer learning framework, and performing adversarial correction on the feature distribution through an orthogonal constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced class separability, wherein the orthogonal constraint enforces orthogonalization of the weight matrices of the generator and the discriminator through Frobenius norm regularization to suppress mode collapse;

[0078] This method maps high-dimensional sparse data to a low-dimensional manifold space through a generative adversarial network (GAN), and introduces orthogonal constraints to prevent the model from falling into mode collapse. This makes the generated features compact and can effectively distinguish different categories, improving the robustness and interpretability of feature representation, avoiding overfitting problems caused by data sparsity, and enhancing classification and clustering effects in different business scenarios.

[0079] Specifically, we can construct a manifold projection space based on transfer learning based on the multimodal joint embedding tensor, and use the adversarial domain adaptation algorithm to align the feature distributions of the source domain and the target domain to generate the initial manifold projection vector.

[0080] This method uses adversarial learning technology to map data from different business scenarios (such as different branches or time periods) to a unified feature space, eliminating the bias caused by distribution differences, enhancing the model's generalization ability in cross-scenario applications, and avoiding performance degradation due to changes in data distribution.

[0081] The multimodal joint embedding tensor (example size: Batch Size × Time Steps × Features = 128 × 30 × 512) is first constructed using a transfer learning framework to construct a manifold projection space. The core goal of this step is to address the feature distribution differences between the source domain (e.g., historical sales data) and the target domain (e.g., real-time market data). This is implemented in three stages:

[0082] Inter-domain feature alignment: Using the Adversarial Domain Adaptation (ADA) algorithm, source and target domain data are fed into a weighted feature extractor (e.g., a ResNet-18 variant) to generate an initial feature vector (dimension 256). The discriminator network (consisting of three fully connected layers with a hidden layer dimension of 128) drives feature distribution alignment through a binary classification task (distinguishing between the source and target domains). For example, in a retail scenario, the source domain might be sales logs from last year's "Double 11" (Singles' Day) shopping festival, while the target domain might be real-time transaction streams from this year's "618" (June 18th) shopping festival. ADA enables the model to identify common consumption patterns across time periods.

[0083] Manifold Projection Optimization: Maximum Mean Discrepancy (MMD) is used as the inter-domain difference metric, combined with radial basis functions (RBF kernel bandwidth set to 0.5) to calculate distribution distance. In each training iteration, the optimizer (Adam, learning rate 0.001) simultaneously minimizes the classification loss (cross entropy) and the MMD loss (weighted by 0.3), forcing the projected features to overlap in the manifold space. For example, when processing cross-modal alignment of customer review text and sales figures, MMD can effectively eliminate the nonlinear distribution shift between the sentiment polarity of the text (e.g., positive / negative) and the numerical sales figures.

[0084] Dynamic temperature scaling: To address category imbalance (e.g., the volume of data for popular products far outweighs that for long-tail products), a temperature scaling parameter (initial value 1.0) is introduced to dynamically adjust the density of the feature space based on category frequency. The temperature corresponding to low-frequency categories is reduced to 0.7, resulting in a more compact distribution of their feature vectors in the manifold space. This parameter is optimized using a grid search (step size 0.1) using the F1-score of the validation set.

[0085] The final generated initial manifold projection vector (dimension 128) has cross-domain invariance. For example, it can map the "user purchasing power" features of different quarters to the same clustering area in the manifold space, laying the foundation for subsequent adversarial training.

[0086] Inputting the initial manifold projection vector into an orthogonal constrained generative adversarial network, forcing the weight matrices of the generator and the discriminator to be orthogonalized through Frobenius norm regularization, and outputting an adversarial feature distribution after mode collapse suppression;

[0087] This method prevents the generative adversarial network from falling into mode collapse (such as generating a single feature) through orthogonal constraints, ensures feature diversity, improves the richness and discrimination of feature representation, and avoids model degradation.

[0088] The structural design of the Orthogonal GAN (O-GAN) focuses on solving the mode collapse problem of traditional GANs. The specific implementation details are as follows:

[0089] Network architecture configuration:

[0090] Generator: A 4-layer transposed convolutional network (channel count: 512 → 256 → 128 → 64), with spectral normalization (SN) and LeakyReLU activation (negative slope 0.2) following each layer. The input noise vector has a dimension of 100, and the output feature map size is aligned with the manifold projection vector.

[0091] Discriminator: 5-layer convolutional network (channel number 64 → 128 → 256 → 512 → 1), using gradient penalty (GP coefficient 10) to stabilize the training process.

[0092] Orthogonality Constraint Implementation: Frobenius norm regularization is introduced in the fully connected layers of the discriminator (for example, the last layer with dimension 512→1). Specifically, for the weight matrix W (size 512×1), its orthogonality loss is calculated: Loss_orth = ||W^TW - I||_F^2. Here, I is the identity matrix, and the regularization coefficient is set to 0.01. This constraint forces the weight vector to approach an orthogonal basis, thereby enhancing the discriminator's feature decoupling capability. For example, when processing cross-modal features, orthogonalization ensures that abstract concepts such as "price sensitivity" and "brand loyalty" are independently modeled in the discriminator's decision space.

[0093] Adversarial training strategy: Utilizing the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework, the generator and discriminator training ratio is set at 1:5. Each batch is fed with 128 samples, and through alternating optimization, the adversarial examples generated by the generator (such as simulated user churn characteristics) gradually approach the real-world data distribution. For example, when generating customer churn warning signals, O-GAN can simultaneously generate both high-risk and low-risk samples, avoiding the simplistic output often associated with traditional GANs due to pattern collapse.

[0094] After training, the discriminator's intermediate layer features (the output of the third convolutional layer, dimension 256) are extracted as an adversarial feature distribution after mode collapse suppression. This distribution exhibits improved class separability, for example increasing the distance between feature clusters of "high-value customers" and "normal customers" by 2.3 times (verified by t-SNE visualization).

[0095] Dynamically correct the adversarial feature distribution, use the Wasserstein distance to measure the difference in feature separability, optimize the discriminator decision boundary through the gradient penalty mechanism, and generate an intermediate feature vector with enhanced class separability;

[0096] This method evaluates the quality of feature distribution through Wasserstein distance and optimizes the discriminator to generate features that are easier to classify, enhance the separability of features, and improve the accuracy of downstream tasks (such as classification or clustering).

[0097] The dynamic correction stage aims to further sharpen category boundaries and eliminate interference from abnormal features. The specific process includes:

[0098] Wasserstein distance optimization: Calculate the Wasserstein-1 distance (Earth Mover's Distance) between the true data distribution P_r and the generated distribution P_g as a quantitative indicator of feature separability. Lipschitz continuity constraints (gradient penalty coefficient λ = 10) are implemented in the critic network (structural identity discriminator) to ensure the stability of the distance estimate. For example, in customer segmentation scenarios, this distance can effectively reflect the degree of separation between "corporate customers" and "individual customers."

[0099] Decision boundary sharpening:

[0100] Difficult Sample Mining: We select the 10% of samples closest to the decision boundary from the adversarial feature distribution (sorted by the absolute value of the critic output) and enhance their features. For example, for "potential churn" samples near the boundary, we generate synthetic samples using MixUp data augmentation (mixing ratio α = 0.2) to increase the data density in the decision boundary area.

[0101] Gradient-guided correction: During backpropagation, Jacobian regularization (with a coefficient of 0.1) is applied to the feature vectors to constrain the smoothness of the feature space in local regions. This prevents excessive curvature of the decision boundary, for example, preventing multiple purchases by the same user from being misclassified as different categories.

[0102] Dynamic temperature scheduling: A cosine annealing strategy is introduced to adjust the learning rate (initial 0.0002, cycle 50 epochs), combined with an early stopping mechanism (training is terminated when the validation set accuracy does not improve for five consecutive epochs) to prevent overfitting. Simultaneously, the feature projection temperature is linearly decayed from 1.0 to 0.5, gradually enhancing the clustering compactness of the feature space.

[0103] After dynamic correction, the intermediate feature vector (dimension 256) exhibits improved class separability. For example, in a product recommendation scenario, the center distance between feature clusters of different categories (such as 3C digital products vs. beauty and skincare) increased by 40%, while the intra-class variance decreased by 25%.

[0104] The intermediate feature vector is input into the manifold compactification module, and a spectral clustering-driven dimensionality reduction algorithm is used to compress the high-dimensional sparse features, and finally a low-dimensional compact semantic embedding vector is output.

[0105] This method compresses high-dimensional features into low-dimensional space through spectral clustering technology while retaining core semantic information, improving computational efficiency and enhancing the interpretability of features, making it easier for business personnel to understand and use them.

[0106] The goal of the manifold compaction module is to compress the 256-dimensional intermediate features into a more interpretable low-dimensional space (such as 32 dimensions) while maintaining the integrity of the semantic information. Specific implementation technologies include:

[0107] Spectral clustering pre-segmentation:

[0108] Constructing a feature similarity matrix: Using an adaptive Gaussian kernel (with a bandwidth σ determined by the median of the k-nearest-neighbor distances, k=15) to calculate the similarity between samples. For example, in a customer segmentation scenario, the similarity matrix can capture the implicit associations between "high-spending, low-frequency" and "low-spending, high-frequency" groups.

[0109] Laplacian matrix construction: The similarity matrix is symmetrically normalized to generate the graph Laplacian matrix L=D⁻¹ / ²(DW) D⁻¹ / ², where D is the degree matrix and W is the similarity matrix.

[0110] Eigendecomposition: Select the eigenvectors corresponding to the first 10 smallest eigenvalues of L to form the initial low-dimensional embedding (dimension 10). This step projects the high-dimensional features into spectral space, facilitating subsequent nonlinear dimensionality reduction.

[0111] Multi-layer autoencoder optimization:

[0112] Encoder structure: 4-layer fully connected network (256→128→64→32), each layer is followed by BatchNorm and Swish activation function.

[0113] Decoder structure: symmetrical 4-layer inverse encoding network (32→64→128→256).

[0114] Loss function: Combining the reconstruction loss (MSE) with the spectral clustering loss (KL divergence, weighted 0.5) ensures that the reduced-dimensional embedding preserves the original information while conforming to the structural characteristics of the spectral space. For example, in supply chain forecasting tasks, this design ensures that key semantic concepts such as "shipping delay risk" are linearly separable in the low-dimensional space.

[0115] Manifold smoothing post-processing:

[0116] Locally Linear Embedding (LLE) fine-tuning: Perform local neighborhood reconstruction (number of neighbors k = 20) on the reduced 32-dimensional vector to eliminate distortions that may be introduced by the autoencoder.

[0117] Density Peak Clustering: This method calculates the local density (radius ε = 0.5) and relative distance of each sample to identify the cluster center. For example, in financial fraud detection, this method can highlight the center of anomalous transaction clusters.

[0118] The final 32-dimensional semantic embedding vector has the following characteristics: in the e-commerce scenario, the cosine similarity of the embedding vectors of the same user in different sessions exceeds 0.85, while the similarity between different users is less than 0.3, proving that its compactness and separability meet the requirements of downstream tasks.

[0119] S203, performing spatiotemporal causal association mining on the semantic embedding vector, extracting multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjusting causal thresholds through reinforcement learning, and outputting a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search;

[0120] This method uses meta-path analysis technology to mine multi-hop association rules in enterprise data (such as "customer A → product B → market trend C"), and dynamically optimizes the confidence of causal relationships through reinforcement learning, ultimately generating a causal graph that reflects business logic, revealing hidden business laws that are difficult to discover with traditional analysis methods, and assisting enterprise decision makers in understanding complex business relationships and optimizing strategic planning and resource allocation.

[0121] Specifically, we can build a deep probabilistic reasoning model for meta-path optimization based on semantic embedding vectors, extract initial multi-hop association rules through the path walking algorithm, and generate a set of candidate meta-paths.

[0122] This method uses meta-path analysis technology to simulate multi-hop path wandering (such as "customer → product → market trend") based on semantic embedding vectors, mines potential association rules from enterprise data, and forms a preliminary set of candidate meta-paths. It can discover long-chain business logic that is difficult to capture with traditional analysis methods, provide rich candidate association relationships for subsequent causal reasoning, and avoid missing important implicit business laws.

[0123] Semantic embedding vectors are low-dimensional, compact feature representations generated by the previous steps. They contain the fused semantic information of a company's multi-source data (such as transaction records, customer texts, and sensor time series). The core goal of the deep probabilistic reasoning model for meta-path optimization is to mine multi-hop association rules across entities and modalities from these embeddings.

[0124] Meta-path walking algorithm design:

[0125] A restart random walk strategy (Restart Random Walk with Meta-Path Constraints) is used to perform path exploration on the constructed enterprise knowledge graph. Each walk step is limited to three hops (for example, "Customer A → Purchase → Product B → Belongs to → Category C → Association → Promotion D"), and attention-weighted transition probabilities are introduced to control direction. For example, when walking to a product node, the model dynamically adjusts the probability of moving to the "Category" or "Supplier" node based on the customer's purchase frequency (a numerical feature in the embedding vector) and the sentiment polarity of the product description text (a semantic feature in the embedding vector). The weight ratio can be set to 7:3.

[0126] Deep Probabilistic Reasoning Model Construction:

[0127] Using a variational autoencoder (VAE) framework, the sequence generated by path walks is encoded into a latent probability distribution. The encoder employs a bidirectional LSTM (Long Short-Term Memory) network to capture path context, with a hidden layer dimension set to 256. The decoder generates candidate meta-paths through Monte Carlo sampling. For example, for the path "user → click → ad → association → marketing channel," the model calculates the conditional probability of this path appearing in historical data (e.g., 0.85) and then assigns a weighted score based on features such as path length and node type diversity.

[0128] Candidate meta-path generation:

[0129] Through threshold filtering (such as retaining paths with a confidence score ≥ 0.7) and diversity control (retaining at most the top three paths of the same type), a set of candidate meta-paths is finally formed. For example, in a retail scenario, the following typical meta-paths may be generated:

[0130] Path 1: Customer → Purchase → Product → Viewed → Similar Products → Association → Discount (Confidence 0.82);

[0131] Path 2: Supplier → Supply → Product → Complaint → Customer Service Ticket → Association → Logistics Delay (Confidence Level 0.75).

[0132] Each meta-path is accompanied by multi-dimensional attribute labels, including path length, node type combination, time decay coefficient, etc.

[0133] Perform reinforcement learning on the candidate meta-path set, dynamically adjust the causal threshold based on the temporal difference error, select meta-paths with confidence higher than the preset value through the path reward mechanism, and output the optimized causal rule set;

[0134] This method uses reinforcement learning to dynamically evaluate the causal strength of each meta-path, screens high-confidence association rules (such as "customer complaints → product defects → sales decline") through a reward mechanism, eliminates low-correlation paths, optimizes the causal rule set, improves the reliability of causal relationships, and ensures that the final output meta-path map only contains high-confidence business logic, reduces noise interference, and improves the credibility of decision support.

[0135] The reinforcement learning framework aims to dynamically optimize the causal threshold of the meta-path (i.e., the minimum confidence level for determining whether a causal relationship exists) to balance rule coverage and accuracy. The specific process is as follows:

[0136] Definition of state space and action space:

[0137] State: A three-dimensional vector consisting of the current causal threshold (initial value 0.7), the average confidence of candidate meta-paths (e.g., 0.74), and the path diversity index (calculated by Shannon entropy, ranging from 0 to 1).

[0138] Action: Adjust the causal threshold in steps of ±0.05 (e.g. increase from 0.7 to 0.75 or decrease from 0.65).

[0139] Reward function design:

[0140] The reward value is calculated by weighting three parts:

[0141] Rule Accuracy Reward: The percentage of meta-paths triggered within the verification period (e.g., 7 days) that lead to the expected results in actual business (e.g., if the actual purchase conversion rate increases by 15% after path 1 is triggered, the reward is +0.3).

[0142] Coverage penalty: If the threshold is too high and the number of valid paths is less than 20% of the total number, a penalty of -0.2 is applied.

[0143] Diversity bonus: For every 0.1 increase in the path type distribution entropy, the bonus is +0.1.

[0144] Temporal Difference (TD) Learning and Policy Optimization:

[0145] The action-value function is updated using the Q-learning algorithm, with a learning rate of 0.01 and a discount factor of γ = 0.9. For example, when the threshold is adjusted from 0.7 to 0.75, the accuracy of the rule improves but the coverage decreases. The system updates the Q table based on the TD error (the difference between the actual reward and the predicted Q value), gradually converging to the optimal policy.

[0146] Path reward mechanism screening:

[0147] Perform secondary filtering on meta-paths that pass the dynamic threshold using the Path Importance Score (PIS):

[0148] PIS = 0.6 × confidence + 0.3 × causal strength + 0.1 × business weight

[0149] Causal strength was calculated using a Granger causality test (with a lag order of 3), and business weights were assigned by domain experts (e.g., a weight of 0.8 for the supply chain path and 0.6 for the marketing path). Paths with a PIS ≥ 0.65 were retained to form the optimized causal rule set.

[0150] Performing a Monte Carlo tree search on the causal rule set, evaluating the causal strength by simulating counterfactual scenarios, and correcting the path weights in combination with Bayesian posterior probabilities to generate a meta-path probability map of spatiotemporal causal associations;

[0151] This method uses Monte Carlo tree search to simulate causal relationships in different business scenarios (such as "If the product is not defective, will sales rebound?"), and combines Bayesian probability to dynamically correct path weights to generate a probability map that reflects the true causal strength, enhance the robustness of causal reasoning, avoid overfitting or false associations, enable the map to adapt to changes in different business environments, and improve the adaptability of decision-making.

[0152] Monte Carlo Tree Search (MCTS) is used to evaluate the causal effects of meta-paths in counterfactual reasoning. It is divided into four stages:

[0153] Selection:

[0154] Starting from the root node (initial causal rule set), child nodes are selected using the UCB1 (Upper Confidence Bound) formula. For example, the UCB value of a path is calculated as:

[0155] ;

[0156] Where Q is the cumulative reward of the path, N is the number of visits, T is the total number of visits, and the exploration coefficient c is set to 1.414. Paths with high UCB values are preferentially selected to enter the next layer.

[0157] Expansion:

[0158] When encountering an incompletely explored node, expand it to a new child node. For example, for the path "customer → complaint → work order → association → logistics delay," generate a counterfactual scenario: assuming a 20% reduction in logistics response time, simulate a change in customer complaint rates. This expansion requires injecting random perturbations (e.g., ±5% fluctuations in logistics parameters).

[0159] Simulation:

[0160] We ran the counterfactual scenario 1,000 times in the virtual environment and recorded the causal strength metric. For example, if the simulation showed that the complaint rate dropped by 12% after logistics acceleration, the causal strength would be 0.12 (normalized to the interval [0, 1]).

[0161] Backpropagation:

[0162] The simulation results are used to reversely update the Q value and visit count of the path nodes. At the same time, the Bayesian posterior probability is used to modify the path weight:

[0163] Prior distribution: Assume that the path weights follow a Beta distribution (α=2, β=2 indicates a neutral prior).

[0164] Posterior update: Based on the number of successes (e.g., 800 successful) and failures (200 invalid) obtained from the simulation, update to Beta(802, 202).

[0165] Finally, the path weight takes the posterior mean (802 / (802+202)≈0.799).

[0166] Through multiple rounds of iteration (e.g., 10,000 simulations), a meta-path probability graph is generated. The weight of each edge in the graph represents the probability strength of the causal relationship, for example:

[0167] Edge A (supply chain delay → production disruption): probability 0.92;

[0168] Side B (promotional activity → short-term sales): probability 0.85 but with a confidence interval of [0.78, 0.89].

[0169] The meta-path probability graph is input into the graph pruning module, and low-significance edges are removed using the information entropy threshold, and finally a spatiotemporal causal meta-path graph containing implicit business logic is output.

[0170] This method is based on information entropy analysis, automatically pruning low-significance edges (such as weakly correlated or paths with no actual business significance), retaining key causal chains, and forming a streamlined and high-value meta-path map. It improves the interpretability and practicality of the map, allowing decision makers to focus on the most influential business logic, avoid information overload, and improve decision-making efficiency.

[0171] The goal of graph pruning is to remove noisy edges and redundant connections to improve the interpretability and computational efficiency of the graph. The specific process is as follows:

[0172] Information entropy calculation:

[0173] The information entropy of each node's outgoing edge set is calculated to measure the degree of connection disorder. For example, a supplier node has 3 outgoing edges:

[0174] Edge 1: Supply delay → Production problem (weight 0.9);

[0175] Edge 2: Supply delay → inventory overstock (weight 0.7);

[0176] Edge 3: Delivery delay → Customer service complaint (weight 0.4).

[0177] The probability distribution is normalized to [0.45, 0.35, 0.20], and the information entropy is:

[0178] H=−(0.45ln0.45+0.35ln0.35+0.20ln0.20)≈1.03

[0179] Set an entropy threshold (such as 0.7), and nodes above this value need to be pruned.

[0180] Significance test:

[0181] Fisher's exact test is used to evaluate the statistical significance of edges. For example, if an edge appears 200 times in the historical data and 180 of them lead to the expected result, the p-value is calculated to be 1.2×10⁻ 6 , which is lower than the significance level (α=0.01), so it is retained.

[0182] Spectral clustering to remove redundancy:

[0183] Perform spectral clustering on the retained edges, setting the number of clusters K to 5 and the feature vector dimension to 50. Within a cluster, only the edges with a similarity greater than 0.8 are retained. For example, if a cluster contains three similar paths: "Customer churn → revenue decline," "Customer churn → market share loss," and "Customer churn → stock price decline," only the first one (weight 0.92) is retained.

[0184] Strengthening of space-time constraints:

[0185] For edges with time-decay properties (such as the short-term effects of promotional activities), we impose time window constraints. For example, if the causal effect of a path decays to less than 30% of its initial value after 30 days, we add a time-effect label to it and display it as a dotted line in the graph.

[0186] The resulting spatiotemporal causal metapath graph can be directly used for decision support. For example, in a supply chain optimization scenario, if the graph shows that the path from "raw material price increase → supplier switching delay → production plan adjustment" has a high causal strength (0.89), the system will recommend establishing a database of alternative suppliers in advance and trigger automatic adjustments to procurement strategies.

[0187] S204, input the spatiotemporal causal meta-path graph into the dynamic game adversarial interpretation framework, generate the optimal decision-making strategy based on the attention-driven policy optimization mechanism, balance the rule interpretability and commercial value gain through the adversarial perturbation compensation algorithm constrained by KL divergence, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. In particular, the framework jointly optimizes the stability of the interpretation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

[0188] This method combines game theory and adversarial learning techniques to maximize business value while ensuring the explainability of decisions, and verifies the robustness of decisions through counterfactual analysis. It ultimately generates an intelligent decision-making map that can guide actual business, providing explainable and high-value decision support, avoiding the uncontrollable risks of black-box AI models, and enhancing the model's resistance to abnormal data and adversarial attacks.

[0189] Specifically, we can build an attention-driven policy optimization model based on the spatiotemporal causal metapath graph, dynamically assign priority weights to rule interpretations through a multi-head self-attention mechanism, and generate an initial decision strategy vector.

[0190] This method uses a multi-head self-attention mechanism to analyze the key paths in the causal graph (such as "supply chain delay → cost increase → profit decrease"), dynamically assign priorities to different rules, and form a preliminary decision-making strategy to ensure that the decision-making strategy focuses on the most influential causal relationships, avoids treating all rules equally, and improves the targetedness and effectiveness of the strategy.

[0191] The spatiotemporal causal metapath graph stores implicit business logic in a graph structure, with nodes representing business entities (e.g., customers, products, and supply chain nodes) and edges representing causal relationships (e.g., "Increased customer complaint rate → increased product return rate"). To construct an attention-driven policy optimization model, the nodes and metapaths in the graph are first encoded as high-dimensional embedding vectors. The embedding dimension of each node is set to 512, and the GraphSAGE algorithm is used to aggregate features within its three-hop neighborhood. For example, the embedding of a customer node might incorporate sentiment analysis of their purchase history, complaint text, and quality indicators of associated products.

[0192] The Multi-Head Self-Attention (MHSA) mechanism is used at this stage to dynamically evaluate the explanatory priority of different meta-paths. The model employs eight independent attention heads, each focusing on causal patterns along different dimensions. For example, head 1 might focus on lagged effects in the temporal dimension (e.g., the two-month delay in the impact of a promotion on quarterly sales), while head 2 might focus on regional diffusion effects in the spatial dimension (e.g., the impact of a regional supply chain disruption on the national distribution network). The computational process for each head is as follows:

[0193] Query-key-value generation: The node embedding vector is linearly transformed to generate query, key, and value matrices, all with dimensions of 64×8 (i.e., 8 total heads and 64 dimensions per head).

[0194] Attention weight calculation: The query and key are dot-producted, scaled (by a factor of √64), and then normalized using Softmax to obtain the weight distribution. For example, the attention weight for the meta-path "Raw material price increase → Production cost increase → Pricing strategy adjustment" might be as high as 0.8, while the weight for "Employee satisfaction → Production efficiency" is only 0.2, reflecting the former's greater importance in the current business scenario.

[0195] Feature fusion: The outputs of each head are concatenated and fused through a fully connected layer to generate a comprehensive attention weight matrix.

[0196] Ultimately, the initial decision strategy vector is formed by the weighted sum of each node's embedding and the corresponding attention weight. For example, for the "Pricing Strategy Adjustment" node, its strategy vector might include features such as cost sensitivity (0.75) and market competition intensity (0.63), with weights determined by its pivotal role in multiple meta-paths. This vector has a dimension of 1024, encompassing the key drivers of business decisions.

[0197] Input the initial decision policy vector into the KL divergence constrained adversarial perturbation compensation module, optimize the geometric distribution of the interpretation boundary through the implicit policy gradient algorithm, and output an intermediate decision policy with enhanced perturbation resistance;

[0198] This method optimizes the stability of decision-making strategies through adversarial learning technology, so that they can still maintain a reasonable interpretation boundary when facing data noise or adversarial attacks, avoid strategy failure due to minor disturbances, enhance the robustness of decision-making strategies, ensure reliable execution in complex or adversarial business environments, and reduce decision-making risks.

[0199] The goal of the adversarial perturbation compensation module is to enhance the robustness of the decision-making strategy to input noise while maintaining the interpretability of the rules. The specific implementation is divided into three stages:

[0200] Adversarial perturbation generation: The Fast Gradient Signed Method (FGSM) is used to add perturbations to the initial policy vector. The perturbation amplitude ε is set to 0.1 (corresponding to a 10% maximum change in the eigenvalue), and the direction is determined by the gradient of the policy loss function. For example, if the gradient of the "cost sensitivity coefficient" dimension of a policy vector is positive, the perturbation will intentionally increase this coefficient to test the stability of the decision.

[0201] KL Divergence Constraint: This constraint uses the Kullback-Leibler Divergence (KL Divergence) to limit the deviation of the perturbed policy distribution from the original distribution. The threshold is set to 0.05 (i.e., the distribution difference does not exceed 5%). The policy vector is treated as a Gaussian distribution, and the KL divergence of its mean μ and covariance Σ is calculated. For example, if the original policy vector has μ = [0.5, 0.3, ...] and the perturbed policy vector has μ' = [0.52, 0.29, ...], then if the KL divergence exceeds the threshold, the perturbation is reduced.

[0202] Implicit Policy Gradient Optimization: The Proximal Policy Optimization (PPO) algorithm is used to update policy parameters. In each iteration, 100 strategies are sampled from the perturbed policy distribution and evaluated for their commercial value gain (e.g., percentage increase in profit margin) and interpretability (e.g., how well the decision rule matches the meta-path graph). Gradient update step size is adjusted through importance sampling, and the learning rate is set to 0.001 to ensure the stability of policy updates.

[0203] The resulting intermediate decision strategy is more resilient to interference. For example, in testing, when 5% noise (simulating data collection errors) was injected into the input data, the fluctuation in the business value of the perturbed strategy decreased from ±15% to ±3%, while the explanatory weights of key decision rules (such as "prioritizing the supply of high-profit products") remained stable.

[0204] Perform dynamic game equilibrium solution on the intermediate decision strategy, use Nash equilibrium iterative algorithm to balance rule interpretability and commercial value gain, and generate a decision strategy map for counterfactual robustness verification;

[0205] This method combines game theory to find the optimal balance between explainability (such as clear decision-making rules) and business value (such as profit maximization), and verifies the robustness of the strategy through counterfactual analysis, providing decision-making solutions that are both easy to understand and can bring actual business benefits, avoiding sacrificing business results due to excessive pursuit of explainability, or vice versa.

[0206] The dynamic game model models rule interpretability (scored by business experts) and commercial value gains (quantified by financial indicators) as the interests of the two players. The Nash equilibrium solution process is as follows:

[0207] Payment matrix construction:

[0208] The strategy space of the interpretability side is to adjust the matching threshold between the decision rule and the meta-path graph (0.6~0.9, step size 0.05).

[0209] The strategic space for commercial value providers is to adjust the resource allocation weight (e.g., the market budget accounts for 20% to 50%).

[0210] The payment value is calculated through Monte Carlo simulation. For example, when the matching threshold is 0.8 and the market budget accounts for 40%, the explainability score is 85 points and the profit margin increases by 12%.

[0211] Iterative solution:

[0212] The initial strategy pair is set to (matching threshold 0.7, market budget 35%).

[0213] In each round of iteration, both sides choose the best response based on the opponent's strategy:

[0214] The explainability party fixes the market budget and searches for a matching threshold that maximizes its own payment (e.g., from 0.7 to 0.75).

[0215] The commercial value side fixes the matching threshold and adjusts the budget allocation (e.g., from 35% to 38%).

[0216] The convergence condition is set as the policy change is less than 1% for 10 consecutive rounds or the maximum number of iterations is 1000.

[0217] Counterfactual Verification: Counterfactual testing is performed on equilibrium strategies. For example, assuming a supplier suddenly runs out of stock, the system uses the metapath graph to trace the causal chain (e.g., "supplier rating declines → probability of delivery delays increases") and automatically generates contingency strategies (e.g., activating backup suppliers and adjusting production plans). Verification metrics include supply chain recovery time after strategy execution (target: <48 hours) and explanatory consistency score (requires >80 points).

[0218] In the final output of the decision strategy map, each node is associated with the optimal strategy parameters and counterfactual test results. For example, a node with strategy parameters of 0.78 matching threshold and 42% market budget has an average recovery time of 36 hours under five counterfactual scenarios and an explainability score of 82.

[0219] The decision strategy map is input into the manifold projection adversarial enhancement module, and the abnormal samples are compressed by generating the discriminator feature space of the adversarial network, and finally an enterprise-level intelligent decision map is output.

[0220] This method uses the discriminator of a generative adversarial network (GAN) to identify and filter out abnormal samples (such as extreme market conditions or data anomalies), ensuring that the final decision map is highly reliable in real business scenarios. It further enhances the practicality of the decision map, enabling it to adapt to the uncertainty in the real business environment and provide enterprises with stable and feasible intelligent decision-making support.

[0221] The Manifold Projection Adversarial Enhancement Module uses a Generative Adversarial Network (GAN) to identify and correct abnormal samples in the policy. The GAN structure is as follows:

[0222] Generator: 4-layer fully connected network, input is a 128-dimensional noise vector, output is a 1024-dimensional synthetic policy vector, and the activation function is LeakyReLU (negative slope 0.2).

[0223] Discriminator: 3-layer convolutional network (kernel size 3×3, stride 2), output is the probability of policy authenticity (0~1), and the last layer uses Sigmoid activation.

[0224] The training process introduces two key mechanisms:

[0225] Anomalous Sample Detection: The discriminator's intermediate layer features (layer 2 outputs) are used to construct the manifold space of the policy vector. The Mahalanobis distance (MD) between the input policy and the center of the manifold is calculated, with a threshold of 3σ (σ being the standard deviation of the training set distance). For example, a policy containing contradictory rules (e.g., "increase advertising investment and reduce marketing budget simultaneously") might have a MD of 5.2, making it considered an anomaly.

[0226] Manifold Projection Correction: Anomaly policies are projected onto the normal manifold space through gradient descent optimization. The optimization goal is to minimize the L2 distance (weight 0.7) and the discriminator feature distance (weight 0.3) between the projected policy and the original policy. The maximum number of iterations is 50. For example, after projection, the contradictory rule of an anomaly policy is corrected to "increase advertising investment in the eastern region while reducing budgets in inefficient markets in the west."

[0227] In the final output of the enterprise-level intelligent decision-making graph, each node is accompanied by a disturbance-resistant strategy version and manifold space coordinates. For example, the original strategy vector of a node has manifold coordinates of [0.3, 0.6]. After adversarial enhancement, its outlier score drops from 0.89 to 0.12, and the business value gain stabilizes in the range of 8% to 10%. All strategies are stored on the blockchain, ensuring that the decision-making process is traceable and auditable.

[0228] It can be seen that according to the multi-source heterogeneous data of the enterprise, a multimodal hypergraph neural network is used to perform dynamic modal alignment to generate a spatiotemporal consistent multimodal joint embedding tensor; the multimodal joint embedding tensor is input into the orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced category separability; the spatiotemporal causal association of the semantic embedding vector is mined to output a spatiotemporal causal meta-path graph containing implicit business logic; the spatiotemporal causal meta-path graph is input into the dynamic game adversarial interpretation framework, and finally an enterprise-level intelligent decision-making graph with counterfactual robustness is output, which can efficiently integrate multimodal data and perform intelligent analysis, thereby improving the accuracy and effectiveness of the mining results.

[0229] Another embodiment of the present invention provides an enterprise big data mining system based on artificial intelligence, see Figure 3 , the system may include:

[0230] Fusion module 301 is used to perform dynamic modal alignment based on the enterprise's multi-source heterogeneous data using a multimodal hypergraph neural network. It fuses cross-modal features through a self-attention-driven dynamic hypergraph construction algorithm to generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0231] A correction module 302 is configured to input the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, construct a manifold projection space for high-dimensional sparse data based on a transfer learning framework, perform adversarial correction on the feature distribution through an orthogonal constrained generative adversarial network, and generate a low-dimensional, compact semantic embedding vector with enhanced class separability. The orthogonal constraint enforces orthogonalization of the weight matrices of the generator and the discriminator through Frobenius norm regularization to suppress mode collapse.

[0232] Adjustment module 303, configured to perform spatiotemporal causal association mining on the semantic embedding vector, extract multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjust causal thresholds through reinforcement learning, and output a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of meta-paths through a path reward mechanism and Monte Carlo tree search;

[0233] Output module 304 is used to input the spatiotemporal causal meta-path graph into a dynamic game adversarial interpretation framework, generate an optimal decision-making strategy based on an attention-driven policy optimization mechanism, balance rule interpretability and commercial value gain through an adversarial perturbation compensation algorithm constrained by KL divergence, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. The framework jointly optimizes the stability of the interpretation boundary and the decision confidence through an implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance decision robustness.

[0234] It can be seen that according to the multi-source heterogeneous data of the enterprise, a multimodal hypergraph neural network is used to perform dynamic modal alignment to generate a spatiotemporal consistent multimodal joint embedding tensor; the multimodal joint embedding tensor is input into the orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced category separability; the spatiotemporal causal association of the semantic embedding vector is mined to output a spatiotemporal causal meta-path graph containing implicit business logic; the spatiotemporal causal meta-path graph is input into the dynamic game adversarial interpretation framework, and finally an enterprise-level intelligent decision-making graph with counterfactual robustness is output, which can efficiently integrate multimodal data and perform intelligent analysis, thereby improving the accuracy and effectiveness of the mining results.

[0235] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0236] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0237] S201: Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used to perform dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0238] S202, inputting the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, constructing a manifold projection space for high-dimensional sparse data based on a transfer learning framework, and performing adversarial correction on the feature distribution through an orthogonal constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced class separability, wherein the orthogonal constraint enforces orthogonalization of the weight matrices of the generator and the discriminator through Frobenius norm regularization to suppress mode collapse;

[0239] S203, performing spatiotemporal causal association mining on the semantic embedding vector, extracting multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjusting causal thresholds through reinforcement learning, and outputting a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search;

[0240] S204, input the spatiotemporal causal meta-path graph into the dynamic game adversarial interpretation framework, generate the optimal decision-making strategy based on the attention-driven policy optimization mechanism, balance the rule interpretability and commercial value gain through the adversarial perturbation compensation algorithm constrained by KL divergence, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. In particular, the framework jointly optimizes the stability of the interpretation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

[0241] It can be seen that according to the multi-source heterogeneous data of the enterprise, a multimodal hypergraph neural network is used to perform dynamic modal alignment to generate a spatiotemporal consistent multimodal joint embedding tensor; the multimodal joint embedding tensor is input into the orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced category separability; the spatiotemporal causal association of the semantic embedding vector is mined to output a spatiotemporal causal meta-path graph containing implicit business logic; the spatiotemporal causal meta-path graph is input into the dynamic game adversarial interpretation framework, and finally an enterprise-level intelligent decision-making graph with counterfactual robustness is output, which can efficiently integrate multimodal data and perform intelligent analysis, thereby improving the accuracy and effectiveness of the mining results.

[0242] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0243] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0244] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0245] S201: Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used to perform dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition.

[0246] S202, inputting the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, constructing a manifold projection space for high-dimensional sparse data based on a transfer learning framework, and performing adversarial correction on the feature distribution through an orthogonal constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced class separability, wherein the orthogonal constraint enforces orthogonalization of the weight matrices of the generator and the discriminator through Frobenius norm regularization to suppress mode collapse;

[0247] S203, performing spatiotemporal causal association mining on the semantic embedding vector, extracting multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjusting causal thresholds through reinforcement learning, and outputting a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search;

[0248] S204, input the spatiotemporal causal meta-path graph into the dynamic game adversarial interpretation framework, generate the optimal decision-making strategy based on the attention-driven policy optimization mechanism, balance the rule interpretability and commercial value gain through the adversarial perturbation compensation algorithm constrained by KL divergence, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. In particular, the framework jointly optimizes the stability of the interpretation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

[0249] It can be seen that according to the multi-source heterogeneous data of the enterprise, a multimodal hypergraph neural network is used to perform dynamic modal alignment to generate a spatiotemporal consistent multimodal joint embedding tensor; the multimodal joint embedding tensor is input into the orthogonal adversarial manifold learning module to generate a low-dimensional, compact semantic embedding vector with enhanced category separability; the spatiotemporal causal association of the semantic embedding vector is mined to output a spatiotemporal causal meta-path graph containing implicit business logic; the spatiotemporal causal meta-path graph is input into the dynamic game adversarial interpretation framework, and finally an enterprise-level intelligent decision-making graph with counterfactual robustness is output, which can efficiently integrate multimodal data and perform intelligent analysis, thereby improving the accuracy and effectiveness of the mining results.

[0250] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. An enterprise big data mining method based on artificial intelligence, characterized in that: The method comprises: Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used for dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition. The multimodal joint embedding tensor is input into an orthogonal adversarial manifold learning module. A manifold projection space for high-dimensional sparse data is constructed based on a transfer learning framework. The feature distribution is adversarially corrected through an orthogonal-constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced class separability. The orthogonal constraint enforces orthogonalization of the weight matrices of the generator and discriminator through Frobenius norm regularization, thereby suppressing mode collapse. Performing spatiotemporal causal association mining on the semantic embedding vectors, extracting multi-hop association rules using a deep probabilistic reasoning model with meta-path optimization, dynamically adjusting causal thresholds through reinforcement learning, and outputting a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search; The spatiotemporal causal meta-path graph is input into a dynamic game adversarial interpretation framework, and the optimal decision-making strategy is generated based on the attention-driven policy optimization mechanism. The rule interpretability and commercial value gain are balanced through the KL divergence-constrained adversarial perturbation compensation algorithm, and finally an enterprise-level intelligent decision graph with counterfactual robustness is output. Among them, the framework jointly optimizes the stability of the explanation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

2. The method according to claim 1, characterized in that Based on the enterprise's multi-source heterogeneous data, a multimodal hypergraph neural network is used for dynamic modal alignment. A self-attention-driven dynamic hypergraph construction algorithm is used to fuse cross-modal features and generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm introduces an inter-modal causal reasoning mechanism and eliminates the temporal drift noise of cross-domain data through tensor decomposition, including: Based on the company's multi-source heterogeneous data, including structured transaction records, unstructured customer feedback text, and IoT time series data, the modal alignment tensor decomposition algorithm is used to align timestamps and disambiguate semantics of the multi-source heterogeneous data, generating an initial tensor for cross-modal time series alignment. The initial tensor is input into the self-attention driven dynamic hypergraph construction module, the cross-modal feature association weights are calculated through the inter-modal causal reasoning mechanism, the time drift noise is filtered using the causal mask matrix, and the dynamic hypergraph adjacency matrix is output; Performing spatiotemporal joint embedding learning on the dynamic hypergraph adjacency matrix, using a multi-head graph attention network to aggregate the spatiotemporal dependencies of cross-modal nodes to generate a graph embedding vector with multimodal feature fusion; The graph embedding vector is input into the tensor rank constraint compression module, and redundant features are eliminated through non-negative matrix factorization and low-rank projection, and finally a spatiotemporally consistent multimodal joint embedding tensor is output.

3. The method according to claim 2, characterized in that The multimodal joint embedding tensor is input into an orthogonal adversarial manifold learning module, a manifold projection space of high-dimensional sparse data is constructed based on a transfer learning framework, and adversarial correction is performed on the feature distribution through an orthogonal constrained generative adversarial network to generate a low-dimensional, compact semantic embedding vector with enhanced category separability, wherein the orthogonal constraint forces the weight matrices of the generator and the discriminator to be orthogonalized through Frobenius norm regularization to suppress mode collapse, including: Based on the multimodal joint embedding tensor, a manifold projection space based on transfer learning is constructed. The adversarial domain adaptation algorithm is used to align the feature distributions of the source domain and the target domain to generate the initial manifold projection vector. Inputting the initial manifold projection vector into an orthogonal constrained generative adversarial network, forcing the weight matrices of the generator and the discriminator to be orthogonalized through Frobenius norm regularization, and outputting an adversarial feature distribution after mode collapse suppression; Dynamically correct the adversarial feature distribution, use the Wasserstein distance to measure the difference in feature separability, optimize the discriminator decision boundary through the gradient penalty mechanism, and generate an intermediate feature vector with enhanced class separability; The intermediate feature vector is input into the manifold compactification module, and a spectral clustering-driven dimensionality reduction algorithm is used to compress the high-dimensional sparse features, and finally a low-dimensional compact semantic embedding vector is output.

4. The method according to claim 3, characterized in that The spatiotemporal causal association mining is performed on the semantic embedding vector, a deep probabilistic reasoning model with meta-path optimization is used to extract multi-hop association rules, causal thresholds are dynamically adjusted through reinforcement learning, and a spatiotemporal causal meta-path graph containing implicit business logic is output. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of the meta-path through a path reward mechanism and Monte Carlo tree search, including: Based on the semantic embedding vector, a deep probabilistic reasoning model for meta-path optimization is constructed. The initial multi-hop association rules are extracted through the path walking algorithm to generate a set of candidate meta-paths. Perform reinforcement learning on the candidate meta-path set, dynamically adjust the causal threshold based on the temporal difference error, select meta-paths with confidence higher than the preset value through the path reward mechanism, and output the optimized causal rule set; Performing a Monte Carlo tree search on the causal rule set, evaluating the causal strength by simulating counterfactual scenarios, and correcting the path weights in combination with Bayesian posterior probabilities to generate a meta-path probability map of spatiotemporal causal associations; The meta-path probability graph is input into the graph pruning module, and low-significance edges are removed using the information entropy threshold, and finally a spatiotemporal causal meta-path graph containing implicit business logic is output.

5. The method according to claim 4, characterized in that The spatiotemporal causal metapath graph is input into a dynamic game adversarial interpretation framework, an attention-driven policy optimization mechanism is used to generate an optimal decision-making strategy, and a KL divergence-constrained adversarial perturbation compensation algorithm is used to balance rule interpretability and commercial value gain. Ultimately, an enterprise-level intelligent decision graph with counterfactual robustness is output. The framework jointly optimizes the stability of the interpretation boundary and the decision confidence through an implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance decision robustness, including: Based on the spatiotemporal causal metapath graph, an attention-driven policy optimization model is constructed. The priority weights of rule interpretations are dynamically assigned through a multi-head self-attention mechanism to generate an initial decision strategy vector. Input the initial decision policy vector into the KL divergence constrained adversarial perturbation compensation module, optimize the geometric distribution of the interpretation boundary through the implicit policy gradient algorithm, and output an intermediate decision policy with enhanced perturbation resistance; Perform dynamic game equilibrium solution on the intermediate decision strategy, use Nash equilibrium iterative algorithm to balance rule interpretability and commercial value gain, and generate a decision strategy map for counterfactual robustness verification; The decision strategy map is input into the manifold projection adversarial enhancement module, and the abnormal samples are compressed by generating the discriminator feature space of the adversarial network, and finally an enterprise-level intelligent decision map is output.

6. An enterprise big data mining system based on artificial intelligence, characterized by: The system comprises: A fusion module is used to perform dynamic modal alignment based on the enterprise's multi-source heterogeneous data using a multimodal hypergraph neural network. This module fuses cross-modal features using a self-attention-driven dynamic hypergraph construction algorithm to generate a spatiotemporally consistent multimodal joint embedding tensor. The dynamic hypergraph construction algorithm incorporates an inter-modal causal reasoning mechanism and eliminates temporal drift noise in cross-domain data through tensor decomposition. A correction module is configured to input the multimodal joint embedding tensor into an orthogonal adversarial manifold learning module, construct a manifold projection space for high-dimensional sparse data based on a transfer learning framework, perform adversarial correction on the feature distribution through an orthogonal-constrained generative adversarial network, and generate a low-dimensional, compact semantic embedding vector with enhanced class separability. The orthogonal constraint enforces orthogonalization of the weight matrices of the generator and discriminator through Frobenius norm regularization, thereby suppressing mode collapse. An adjustment module is configured to perform spatiotemporal causal association mining on the semantic embedding vector, extract multi-hop association rules using a deep probabilistic reasoning model optimized for meta-paths, dynamically adjust causal thresholds through reinforcement learning, and output a spatiotemporal causal meta-path graph containing implicit business logic. The deep probabilistic reasoning model jointly optimizes the confidence and causal strength of meta-paths through a path reward mechanism and Monte Carlo tree search; The output module is used to input the spatiotemporal causal meta-path graph into a dynamic game adversarial interpretation framework, generate the optimal decision-making strategy based on the attention-driven policy optimization mechanism, balance the rule interpretability and commercial value gain through the KL divergence-constrained adversarial perturbation compensation algorithm, and finally output an enterprise-level intelligent decision graph with counterfactual robustness. The framework jointly optimizes the stability of the explanation boundary and the decision confidence through the implicit policy gradient algorithm, and introduces the manifold projection technology of adversarial samples to enhance the decision robustness.

7. The system according to claim 6, characterized in that The fusion module is specifically used to: Based on the company's multi-source heterogeneous data, including structured transaction records, unstructured customer feedback text, and IoT time series data, the modal alignment tensor decomposition algorithm is used to align timestamps and disambiguate semantics of the multi-source heterogeneous data, generating an initial tensor for cross-modal time series alignment. The initial tensor is input into the self-attention driven dynamic hypergraph construction module, the cross-modal feature association weights are calculated through the inter-modal causal reasoning mechanism, the time drift noise is filtered using the causal mask matrix, and the dynamic hypergraph adjacency matrix is output; Performing spatiotemporal joint embedding learning on the dynamic hypergraph adjacency matrix, using a multi-head graph attention network to aggregate the spatiotemporal dependencies of cross-modal nodes to generate a graph embedding vector with multimodal feature fusion; The graph embedding vector is input into the tensor rank constraint compression module, and redundant features are eliminated through non-negative matrix factorization and low-rank projection, and finally a spatiotemporally consistent multimodal joint embedding tensor is output.

8. The system according to claim 7, characterized in that The correction module is specifically used to: Based on the multimodal joint embedding tensor, a manifold projection space based on transfer learning is constructed. The adversarial domain adaptation algorithm is used to align the feature distributions of the source domain and the target domain to generate the initial manifold projection vector. Inputting the initial manifold projection vector into an orthogonal constrained generative adversarial network, forcing the weight matrices of the generator and the discriminator to be orthogonalized through Frobenius norm regularization, and outputting an adversarial feature distribution after mode collapse suppression; Dynamically correct the adversarial feature distribution, use the Wasserstein distance to measure the difference in feature separability, optimize the discriminator decision boundary through the gradient penalty mechanism, and generate an intermediate feature vector with enhanced class separability; The intermediate feature vector is input into the manifold compactification module, and a spectral clustering-driven dimensionality reduction algorithm is used to compress the high-dimensional sparse features, and finally a low-dimensional compact semantic embedding vector is output.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images