Operation index multi-dimensional analysis optimization method based on dynamic weight adjustment of AI large model

Abstract

The invention relates to the technical field of operation index analysis, in particular to an operation index multi-dimensional analysis optimization method based on AI large model dynamic weight adjustment, which comprises the following steps: acquiring a market fluctuation index, an operation efficiency index, a financial health index and a supply chain elasticity index of an enterprise; four-dimensional index data of enterprise operation are input into a hypergraph tensor decomposer, the hypergraph tensor decomposer maps four-dimensional indexes into mutually orthogonal hypergraph nodes through orthogonally constrained Tucker decomposition, a decoupling hypergraph is generated, the decoupling hypergraph is injected into a dynamic weight modeling model, and an optimization strategy chain is generated based on a real-time event stream; and the orthogonal constraint condition of the decoupling hypergraph is corrected through tensor gradient back propagation after execution of the optimization strategy, and closed-loop optimization is formed. According to the method, the optimal response action with business semantics and timeliness can be generated according to unexpected situations such as market abnormal fluctuation, supply chain breakage or financial pressure.

Description

Technical Field

[0001] The present invention relates to the technical field of business indicator analysis, and particularly to an optimized method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of an AI large model. Background Art

[0002] In the context of the continuous evolution of the digital business environment, enterprises' demand for the analysis and optimization of multi-dimensional business indicators is increasing day by day. Typical indicator systems cover multiple dimensions such as market volatility, operational efficiency, financial health, and supply chain resilience, and there are complex correlations and dynamic change relationships among different dimensions. Most existing business analysis methods use static weight configuration or linear models to uniformly evaluate various indicators, making it difficult to capture the deep coupling characteristics among high-dimensional indicators and also difficult to adapt to the non-linear linkage effects caused by external event shocks.

[0003] In traditional solutions, methods such as principal component analysis are generally relied on to extract features of multi-dimensional indicators. However, these methods themselves lack the ability to control dimension independence and cannot achieve effective decoupling while maintaining the physical meanings of each indicator. At the same time, the strategy generation module mostly adopts a static decision-making method based on a rule library and lacks the ability to adaptively adjust to changes in the business environment, resulting in a lack of pertinence in strategy execution, easy resource misallocation or response delay.

[0004] In addition, existing systems generally lack a complete strategy optimization mechanism, that is, the results of strategy execution cannot be timely fed back to the indicator modeling process, and the structural biases existing in the analysis model cannot be dynamically corrected, restricting the continuous effectiveness of the optimization strategy and the long-term evolution ability of the model. Summary of the Invention

[0005] The present invention provides an optimized method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of an AI large model, realizing a new analysis and optimization method for orthogonal decoupling of indicators, dynamic evolution of weights, and closed-loop correction of the structure, so as to support accurate decision-making and self-driven update of strategies by enterprises in a complex and uncertain environment.

[0006] An optimized method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of an AI large model includes the following steps: S1: Obtain the market volatility indicator, operational efficiency indicator, financial health indicator, and supply chain resilience indicator of the enterprise, input the four-dimensional indicator data of the enterprise operation into a hypergraph tensor decomposer, and the hypergraph tensor decomposer maps the four-dimensional indicators into mutually orthogonal hypergraph nodes through Tucker decomposition with orthogonal constraints to generate a decoupled hypergraph. Among them, the adjacency entropy values of the market volatility node, operational efficiency node, financial health node, and supply chain resilience node satisfy the independence threshold, taking the value of 0.1; S2: Inject the decoupled hypergraph into the dynamic weight evolution model to generate an optimized policy chain based on the real-time event stream; S3: Form a closed-loop optimization by correcting the orthogonal constraint conditions of the decoupled hypergraph through the tensor gradient backpropagation after the execution of the optimized policy.

[0007] Optionally, in S1: The market volatility indicator includes the industry index volatility; The operation efficiency indicator includes the inventory turnover rate; The financial health indicator includes the current ratio; The supply chain resilience indicator includes the response delay of alternative suppliers.

[0008] Optionally, the Tucker decomposition of the orthogonal constraint includes constructing the four-dimensional index data into a fourth-order tensor according to the time series, and decomposing the fourth-order tensor into a core tensor and four groups of factor matrices, corresponding to the four dimensions of market volatility, operation efficiency, financial health, and supply chain resilience respectively.

[0009] Optionally, after the Tucker decomposition, orthogonality constraints are imposed on each factor matrix respectively. If linear correlation is detected among the column vectors of any group of factor matrices, the re-orthogonalization process is automatically triggered to correct its independence.

[0010] Optionally, the column vectors in the factor matrix are all mapped to the nodes in the decoupled hypergraph, and the nodes are weighted according to the eigenvector norms of the column vectors. The adjacency entropy value of the decoupled hypergraph is calculated based on the cross-dimensional connection relationship between the nodes, which is used to evaluate the coupling degree between the characteristics of each dimension. If the adjacency entropy value of a certain node exceeds the preset independence threshold , the gradient clipping operation will be executed to weaken the coupling effect, or the corresponding local tensor fragment will be reconstructed to restore the orthogonality between the dimensions.

[0011] Optionally, the decoupled hypergraph is constructed based on the factor matrix with orthogonalization constraints. The structure of the decoupled hypergraph needs to meet the limit that the proportion of cross-dimensional connection edges does not exceed the specified value of the overall connection number, and the node degree distribution within each dimension conforms to the power-law characteristic. At the same time, the purity index and time-sensitive index of each dimension are calculated to describe the stability of the overall structure and the response ability to changes during the tensor evolution process, providing basic attribute support for subsequent weight adjustment and policy linkage.

[0012] Optionally, S2 includes converting the generated decoupled hypergraph into a graph structure. In the graph structure, each node includes two types of representation features, corresponding to the dimension decoupling degree and time-sensitive characteristics respectively. The connection edges between the nodes are weighted according to the directional relationship between the node features to reflect their relative deviation degree in the feature space.

[0013] Optionally, after the decoupled hypergraph is converted into a graph structure, a graph neural network model including a time and structure information fusion mechanism is constructed. In the graph neural network model, dilated causal convolution is used to extract the evolution features of the multi-scale event stream in the time dimension; in the structure dimension, the attention mechanism is used to calculate the dynamic weight distribution between nodes in the graph, reflecting the importance relationship between nodes in the current context. The attention mechanism performs weighted calculation based on the combination of node features and the training weight matrix. According to the dynamic weight distribution result and the evolution features, a policy output sequence with time attributes is generated to form an optimized policy chain. The optimized policy chain includes multiple action instructions parsed by the model, and each action is equipped with an optimal execution time, which is applicable to subsequent actual business deployment.

[0014] Optionally, after the optimized policy chain is actually executed, business response data is collected, the change amount of business metrics caused by the policy execution is calculated, and compared with the predicted value to construct a policy effect evaluation index. The policy effect evaluation index includes a business target deviation term and an intervention sensitivity term, which are used to comprehensively evaluate the fitness between the current policy and the model structure.

[0015] Optionally, based on the policy effect evaluation index, the basic representation structure of the decoupled hypergraph is corrected through the backpropagation mechanism, and the orthogonal constraint situation of the tensor factor matrix is adjusted. For each group of factor matrices, the orthogonal offset degree between its internal column vectors is calculated and used as an auxiliary loss term to jointly participate in the model update process with the policy deviation loss term, ensuring that the optimized factor matrix maintains the structural characteristics of dimension decoupling.

[0016] After the update is completed, check the local coupling degree of the decoupled hypergraph structure. If it exceeds the coupling threshold, the incremental update mechanism of the core tensor will be triggered. The incremental update mechanism adjusts according to the pseudo-inverse factor matrix mapped by the newly added input tensor on the basis of keeping the original tensor backbone unchanged, improving the adaptability of the model to sudden business changes and ensuring structural continuity and prediction reliability.

[0017] Advantages of the present invention: In the present invention, by introducing the tensor decomposition method with orthogonal constraints and combining the incremental orthogonalization mechanism, the core indicators of an enterprise in dimensions such as market volatility, operational efficiency, financial health, and supply chain elasticity can be efficiently mapped into independent structured feature representations, reducing the redundant coupling and aliasing problems between multiple indicators in traditional methods. The constructed dynamic decoupled hypergraph not only has the ability to distinguish dimension purity but also meets the requirements of cross-dimension connection control and power-law distribution characteristic verification, providing a model support with clear structure and definite boundaries for subsequent policy simulation and index linkage analysis, and avoiding misjudgment and policy deviation caused by dimension interference.

[0018] The present invention utilizes a neural network structure that integrates time and space, performs joint encoding on the decoupled business indicator hypergraph and real-time event stream, dynamically extracts the attention weight distribution between nodes, forms an executable optimization strategy chain, models the long-term dependencies of external events through dilated causal convolution, and combines the graph attention mechanism guided by changes in node attributes. It can generate optimal response actions with business semantics and timeliness according to sudden situations such as market abnormal fluctuations, supply chain disruptions, or financial pressures, covering key links such as inventory compression, production line adjustment, credit extension, and emergency resource scheduling, comprehensively improving the timeliness and pertinence of strategy formulation.

[0019] The present invention constructs a complete "execution - feedback - correction" closed-loop optimization link. Through the joint optimization of the policy effect loss and the tensor orthogonality loss, it not only realizes the quantitative evaluation of the policy execution effect but also synchronously corrects the decoupled structure of the tensor factor matrix, thereby dynamically adjusting the structural stability and response ability of the model. When it identifies an abnormal increase in the coupling strength between dimensions, it can automatically trigger a local tensor incremental update, maintaining the continuous adaptation ability of the hypergraph structure to new data, avoiding policy failures or misguidance caused by data structure drift, and providing highly reliable and highly self-consistent intelligent optimization support for enterprises in complex business environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only for the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0021] Figure 1 It is a schematic flowchart of the method according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the graph neural network structure according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The present invention will be described in detail below in conjunction with the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for a more specific description of the embodiments and are not intended to specifically limit the present invention.

[0023] It should be noted that in the specification, terms such as "one embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not necessarily every embodiment includes such specific features, structures, or characteristics. Additionally, when combining an embodiment to describe a specific feature, structure, or characteristic, implementing such a feature, structure, or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge scope of those skilled in the relevant art.

[0024] Generally, terms can be understood at least in part from their use in context. For example, at least in part depending on the context, the term "one or more" used herein can be used to describe any feature, structure, or characteristic in a singular sense, or can be used to describe a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey a set of exclusive factors, but rather, at least in part depending on the context, can allow for the existence of other factors that are not necessarily explicitly described.

[0025] As Figure 1 - Figure 2 shown, a multi-dimensional analysis and optimization method for business indicators based on dynamic weight adjustment of an AI large model includes the following steps: S1: Obtain the market volatility indicator, operation efficiency indicator, financial health indicator, and supply chain resilience indicator of the enterprise, input the four-dimensional indicator data of the enterprise operation into a hypergraph tensor decomposer, and the hypergraph tensor decomposer maps the four-dimensional indicators into mutually orthogonal hypergraph nodes through Tucker decomposition with orthogonal constraints to generate a decoupled hypergraph. Among them, the adjacency entropy values of the market volatility node, operation efficiency node, financial health node, and supply chain resilience node satisfy the independence threshold of taking the value of 0.1; S2: Inject the decoupled hypergraph into a dynamic weight evolution model to generate an optimization policy chain based on the real-time event stream; S3: Correct the orthogonal constraint conditions of the decoupled hypergraph through the tensor gradient backpropagation after the execution of the optimization policy to form a closed-loop optimization.

[0026] S1 specifically includes: S11, multi-dimensional indicator data collection and tensor construction, obtained through the enterprise data interface: Market volatility indicator: Industry index volatility (standard deviation of the industry benchmark price within the calculation period): , where represents the industry index volatility, represents the standard deviation processing, represents the industry benchmark price sequence within the calculation period; Operation efficiency indicator: Inventory turnover rate (ratio of sales to average inventory cost): , where is the inventory turnover rate; Financial health indicator: Current ratio (the ratio of current assets to current liabilities): , where is the current ratio; Supply chain resilience indicator: Alternative supplier response latency (the average time from the interruption of the primary supplier to the restoration of supply by the alternative supplier): , where is the alternative supplier response latency; Construct the four-dimensional index data into a fourth-order tensor according to the time series , where T is the number of time slices, and M / O / F / S correspond to the number of feature quantities in each dimension respectively; S12, Orthogonal constrained Tucker tensor decomposition: Perform Tucker decomposition on the fourth-order tensor , expressed as: ; where is the core tensor, and the core tensor captures the interaction pattern between dimensions, are the factor matrices of each dimension, , representing the independent features of each dimension, represents the factor matrix of the market volatility dimension, represents the factor matrix of the operational efficiency dimension, represents the factor matrix of the financial health dimension, represents the factor matrix of the supply chain resilience dimension.

[0027] Apply an orthogonality constraint to the factor matrices to ensure: ; Orthogonal error threshold , trigger column vector re-orthogonalization, represents the identity matrix, which is used to represent the orthogonal constraint target, represents the orthogonal error threshold, indicating the non-orthogonal degree of the factor matrix; S13: Hypergraph node generation and coupling detection: ; ; ; ; Hypergraph adjacency entropy calculation formula: ; where represents the th column vector in the market volatility dimension, represents the corresponding market volatility hypergraph node, represents the weight of the market volatility node, which is equal to the norm of its eigenvector, represents the th column vector in the operational efficiency dimension, represents the corresponding operational efficiency hypergraph node, represents the weight of the operational efficiency node, represents the th column vector in the financial health dimension, represents the corresponding financial health hypergraph node, represents the weight of the financial health node, represents the th column vector in the supply chain resilience dimension, represents the corresponding supply chain resilience hypergraph node, represents the weight of the supply chain resilience node, represents the node 's adjacency entropy, indicating its connection complexity with nodes in other dimensions, represents the node 's set of adjacent nodes, represents the node and 's cross - dimensional connection weight, represents the sum of all node connection weights.

[0028] If any node , then execute: a. Gradient clipping: , The learning rate or adjustment coefficient of gradient clipping, is the gradient of the adjacency entropy with respect to the current factor vector; b. Trigger local tensor decomposition and update the corresponding slice in.

[0029] S14, Dynamic decoupled hypergraph generation: Generate a decoupled hypergraph from the orthogonalized factor matrix, satisfying the following constraints: The number of cross - dimensional connection edges is reduced to less than the original data; ; The node attribute encoding is as follows: Dimension purity: ; Time sensitivity: 。

[0030] is the purity index of dimension, indicating the degree of orthogonality preservation. The purity index of dimension, indicating the degree of orthogonality preservation. is the factor matrix after orthogonalization of dimension The factor matrix after orthogonalization of dimension is the time sensitivity index of dimension indicating the change response of its core tensor over time.

[0031] S2 includes: S21, encoding the decoupled hypergraph into graph-structured data, where: The node feature vector is expressed as: , where represents the dimension purity, represents the time sensitivity; The edge weight calculation method: , where represents node and node the vector angle between them, is the edge weight between node and node , is the decay factor.

[0032] S22, constructing a spatio-temporal fusion graph neural network structure, including the following: Temporal convolutional layer: Performing dilated causal convolution on the real-time event stream, and using the temporal convolutional layer to extract temporal feature representations ; Graph attention layer: Calculating dynamic weights based on the attention mechanism between nodes: , where , are the attention weight matrices, is the set of adjacent nodes of node , represents the , th node feature vector in the graph.

[0033] Policy decoder: Generating an optimized policy chain with time intervals according to the node weight distribution: where is the preset action set, is the execution timestamp corresponding to the policy action, represents a single policy and its timestamp in the optimized policy chain, is the policy action selected at time point .

[0034] S3 specifically includes: S31. After executing the policy chain, collect the change amount of business metrics , and calculate the policy effect loss: , where represents the predicted business change value of the model; represents the actual business observation value; is the regularization term coefficient, represents the change gradient (intervention sensitivity) of the hypergraph structure for the pre-operation ; S32. Based on the loss function, backpropagate to correct the orthogonal constraint conditions in the hypergraph: The orthogonality loss function of the factor matrix is defined as: ; The joint gradient update rule is: , where is the learning rate; is the loss weight balance factor, taking 0.3.

[0035] If the current maximum adjacency entropy value satisfies: ; then trigger the incremental update of the core tensor : , where is the tensor learning rate, is the pseudo-inverse of the dimensional factor matrix, respectively represent the core tensors at the th and th moments, is the new input data tensor.

[0036] The present invention covers any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without these detailed descriptions. Additionally, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0037] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A multi-dimensional analysis optimization method for business indicators based on dynamic weight adjustment of AI large models, characterized in that It includes the following steps: S1: Obtain the market volatility index, operation efficiency index, financial health index, and supply chain resilience index of the enterprise. Input the four-dimensional index data of the enterprise operation into the hypergraph tensor decomposer. The hypergraph tensor decomposer maps the four-dimensional index into mutually orthogonal hypergraph nodes through Tucker decomposition with orthogonal constraints, generating a decoupled hypergraph. Among them, the adjacency entropy values of the market volatility node, operation efficiency node, financial health node, and supply chain resilience node satisfy the independence threshold; S2: Inject the decoupled hypergraph into the dynamic weight evolution model to generate an optimized policy chain based on the real-time event stream; S3: Correct the orthogonal constraint conditions of the decoupled hypergraph through the backpropagation of the tensor gradient after the execution of the optimized policy to form a closed-loop optimization.

2. The multi-dimensional analysis and optimization method for business indicators based on dynamic weight adjustment of AI large models according to claim 1, characterized in that In the above S1: The market volatility indicator includes the industry index volatility; The operation efficiency indicator includes the inventory turnover rate; The financial health indicator includes the current ratio; The supply chain resilience indicator includes the response delay of alternative suppliers.

3. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of the AI large model according to claim 2, characterized in that The Tucker decomposition of the orthogonal constraint includes constructing the four-dimensional index data into a fourth-order tensor according to the time series, and decomposing the fourth-order tensor into a core tensor and four groups of factor matrices, corresponding to the four dimensions of market volatility, operation efficiency, financial health, and supply chain resilience respectively.

4. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of AI large models according to claim 3, characterized in that, After the Tucker decomposition, orthogonality constraints are imposed on each factor matrix respectively. If linear correlation is detected among the column vectors of any group of factor matrices, the re-orthogonalization process is automatically triggered to correct its independence.

5. The multi-dimensional analysis and optimization method for business indicators based on dynamic weight adjustment of AI large models according to claim 3, characterized in that, The column vectors in the factor matrix are all mapped to the nodes in the decoupled hypergraph, and the nodes are weighted according to the eigenvector norms of the column vectors. Based on the cross-dimensional connection relationships between the nodes, the adjacency entropy value of the decoupled hypergraph is calculated to evaluate the coupling degree between the features of each dimension. If the adjacency entropy value of a certain node exceeds the preset independence threshold , a gradient clipping operation will be performed to weaken the coupling effect, or the corresponding local tensor fragment will be reconstructed to restore the orthogonality between the dimensions.

6. The multi-dimensional analysis and optimization method for business indicators based on dynamic weight adjustment of AI large model according to claim 3, characterized in that, The decoupled hypergraph is constructed based on the factor matrices with orthogonalization constraints. The structure of the decoupled hypergraph needs to meet the limit that the proportion of cross-dimensional connection edges does not exceed the specified value of the overall connection number, and the node degree distribution within each dimension conforms to the power-law characteristic. At the same time, calculate the purity index and time-sensitive index of each dimension to describe the stability of the overall structure and the response ability to changes during the tensor evolution process, providing basic attribute support for subsequent weight adjustment and policy linkage.

7. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of AI large models according to claim 6, characterized in that, The above S2 includes converting the generated decoupled hypergraph into a graph structure. In the graph structure, each node includes two types of representation features, corresponding to the dimension decoupling degree and time-sensitive characteristics respectively. The connection edges between nodes are weighted according to the directional relationship between node features to reflect their relative deviation degree in the feature space.

8. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of AI large models according to claim 7, characterized in that, After the decoupled hypergraph is converted into a graph structure, a graph neural network model including a time and structure information fusion mechanism is constructed. In the time dimension of the graph neural network model, the evolution features of the multi-scale event stream are extracted through dilated causal convolution; in the structure dimension of the graph neural network model, the attention mechanism is used to calculate the dynamic weight distribution between nodes in the graph, reflecting the importance relationship between nodes in the current context. The attention mechanism is weighted and calculated based on the combination of node features and the training weight matrix. According to the dynamic weight distribution result and the evolution features, generate a policy output sequence with time attributes to form an optimized policy chain. The optimized policy chain includes multiple action instructions parsed by the model, and each action is equipped with an optimal execution time.

9. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of AI large models according to claim 8, characterized in that, After the actual execution of the optimized policy chain, collect the business response data, calculate the change amount of the business indicators caused by the policy execution, and compare it with the predicted value to construct a policy effect evaluation index. The policy effect evaluation index includes a business objective deviation term and an intervention sensitivity term, which are used to comprehensively evaluate the fitness between the current policy and the model structure.

10. An optimization method for multi-dimensional analysis of business indicators based on dynamic weight adjustment of AI large models according to claim 9, characterized in that, Based on the above-mentioned policy effect evaluation metrics, the basic representation structure of the decoupled hypergraph is corrected through the backpropagation mechanism, the orthogonal constraint of the tensor factor matrix is adjusted. For each group of factor matrices, calculate the orthogonal offset degree between its internal column vectors, and use it as an auxiliary loss term to jointly participate in the model update process with the policy deviation loss term, ensuring that the optimized factor matrix maintains the structural characteristics of dimensional decoupling; After the update is completed, check the local coupling degree of the decoupled hypergraph structure. If it exceeds the coupling threshold, the incremental update mechanism of the core tensor will be triggered. On the basis of keeping the original tensor backbone unchanged, the incremental update mechanism is adjusted according to the pseudo-inverse factor matrix mapped by the newly added input tensor.

Images