Intelligent decision-making system for enterprise management consultation based on data analysis
By constructing a causal basis graph and a context state machine, and combining a causal modulator and a closed-loop checker, the problems of insufficient causal relationship revelation and dynamic adaptability in existing systems are solved, realizing the self-evolution and accuracy improvement of enterprise management decision-making systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-03
AI Technical Summary
Existing enterprise management decision-making systems lack the ability to deeply reveal causal relationships, cannot dynamically adapt to changes in the external environment, and lack self-learning and correction mechanisms, leading to a decline in decision-making accuracy.
A causal basis graph is constructed, and a context state machine and a causal modulator module are combined to dynamically adjust and self-correct causal relationships through decision inference and closed-loop verifier, thereby establishing a self-evolution mechanism.
It enables dynamic adaptation to changes in the external environment, deeply reveals the causal logic between decision-making behavior and business results, and improves the accuracy and self-learning ability of the decision-making system.
Smart Images
Figure CN121787533A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and data processing technology, specifically to an intelligent decision-making system for enterprise management consulting based on data analysis. Background Technology
[0002] In the field of enterprise management and operational decision-making, data-driven decision support systems have been widely adopted. Currently, enterprises commonly utilize business intelligence tools, data warehouses, and various data analytics platforms to collect and analyze massive amounts of historical data generated during operations, aiming to extract valuable insights and provide objective basis for decisions in key areas such as marketing, supply chain management, and product pricing. These systems typically employ statistical models or machine learning algorithms to learn patterns and correlations in historical data, and on this basis, generate descriptive reports, diagnostic analyses, and predictions of future trends.
[0003] However, existing decision support systems of this type still suffer from several inherent technical limitations in practical applications. A significant limitation is that these systems rely heavily on mining statistical correlations between data, rather than delving into the causal relationships within the data. Models can identify synchronous trends in the changes of two or more variables in historical data, but they often cannot effectively distinguish whether this association stems from a direct causal driver or merely from a common, unobserved confounding factor. Therefore, when decision-makers attempt to use such systems for "counterfactual analysis" (i.e., assessing the potential effects of a decision intervention that has never been implemented), correlation-based predictions may be severely biased, misleading management decisions and leading to ineffective or even negative resource allocation.
[0004] Furthermore, current mainstream analytical models are often static. Once the model's structure and parameters are trained using historical data, they are typically kept fixed after deployment, or only updated periodically in batches. However, the market environment, macroeconomic background, and industry competitive landscape in which a company operates are constantly and dynamically changing. These structural changes in the external "context" often lead to alterations in the causal logic between internal operational elements of the company. For example, consumers' price sensitivity increases significantly during economic downturns. Existing systems generally lack mechanisms for real-time perception and modeling of such external environmental contexts, thus failing to dynamically adjust their internal analytical logic to adapt to environmental changes. Consequently, the predictive accuracy and decision-making guidance value of the model gradually decline after environmental changes occur.
[0005] Finally, existing decision support systems generally lack an effective, automated closed-loop learning and self-correction mechanism. When decisions based on model predictions are implemented in the real world, deviations inevitably occur between the actual results and the model's predictions. Current technologies typically rely on manual review and model adjustments to address these deviations, a process that is slow and inefficient. The system itself cannot automatically correlate real-world feedback with its internal knowledge assumptions, nor can it pinpoint and attribute erroneous assumptions that lead to deviations, and thus fail to evolve its knowledge base. This makes it difficult for the system to learn from each decision-making practice, and the accuracy of its models cannot continuously and automatically improve over time and with the accumulation of experience. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides an intelligent decision-making system for enterprise management consulting based on data analysis. This system solves the problems that existing enterprise management decision-making systems typically employ static analytical models, lack the ability to reveal true causal relationships, and lack a closed-loop mechanism for continuous learning and self-evolution from the consequences of decisions.
[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent decision-making system for enterprise management consulting based on data analysis, comprising:
[0008] The system includes a causal basis graph construction module, a context state machine module, a causal modulator module, a decision inference module, a closed-loop verifier module, and a knowledge base update module.
[0009] The causal basis graph construction module is used to construct a causal basis graph consisting of nodes and causal edges. The nodes represent quantifiable entities in enterprise operations, and the causal edges represent direct causal relationship assumptions between nodes. Each causal edge is assigned edge attributes, which include at least the strength of the causal effect and time lag. In a preferred embodiment, the causal basis graph construction module establishes the initial structure of the causal basis graph through expert knowledge injection, and supplements and verifies the initial structure through a data-driven causal discovery algorithm.
[0010] The context state machine module is used to determine the current context state of the system based on an external data signal source. In a preferred embodiment, the context state machine module determines the current context state through a state transition function, which performs state transitions based on the current state and the received external data signal source.
[0011] The causal modulator module is configured to modify the causal base graph according to the current context state and a preset modulation rule to generate a contextualized causal graph. In a preferred embodiment, the causal modulator module modifies the causal base graph by: modifying the attribute of at least one causal edge in the causal base graph; or activating or disabling at least one causal path in the causal base graph.
[0012] The decision inference module is used to perform decision intervention simulation on the contextualized causal graph and to perform causal propagation calculations to generate predicted values for target nodes. In a preferred embodiment, the decision inference module performs the decision intervention simulation by applying a do-operator to set intervention variables on the contextualized causal graph.
[0013] The closed-loop validator module is used to track the observed values of the target node after the decision is executed, and compare the observed values with the predicted values to calculate the deviation. When the deviation triggers a preset condition, the closed-loop validator module performs attribution analysis on the causal path leading to the deviation and identifies suspicious causal paths. In a preferred embodiment, the deviation triggering preset condition is that the cumulative value of the deviation exceeds a dynamic deviation threshold set for the target node.
[0014] The deviation is the cumulative deviation D(T), which is calculated as follows:
[0015]
[0016] in:
[0017] D(T): Cumulative deviation within the observation period T;
[0018] T: Total duration of the observation period;
[0019] t ′ : Represents the integral variable at a time point within the observation period;
[0020] W(t ′ ): Time weighting function, used to assign weights to deviations at different time points;
[0021] The target node at time point t ′ Observed values;
[0022] For: the target node at time point t ′ The predicted value.
[0023] The closed-loop validator module performs the attribution analysis by calculating the suspense score H for each causal edge in the suspicious causal path.ij The calculation method is as follows:
[0024]
[0025] in:
[0026] H ij From node v i to node v j The suspense score of the causal side;
[0027] λ′ ij The causal effect strength of the causal edge in the current context state;
[0028] α ij The contribution factor of the causal edge in the total causal effect from the intervention node to the target node;
[0029] ρ ij The prior confidence level of the causal edge;
[0030] ∈: A smoothing term used to prevent the denominator from being zero.
[0031] The knowledge base update module is used to update the causal base graph or the modulation rules based on the analysis results of the suspicious causal paths. In a preferred embodiment, the knowledge base update module is specifically used to: update the modulation rules in the causal modulator module when the generation of the suspicious causal path is confirmed to be specific to a certain context state; or update the edge attributes in the causal base graph when the generation of the suspicious causal path is confirmed to be related to the cognition of basic causal relationships.
[0032] The second aspect of this invention provides an intelligent decision-making method for enterprise management consulting based on data analysis.
[0033] The method includes the following steps: constructing a causal basis graph containing nodes, causal edges, and edge attributes, wherein the edge attributes include at least causal effect strength and time lag; determining the current context state of the system based on external data signal sources using a context state machine; modifying the causal basis graph according to the current context state and a preset modulation rule to generate a contextualized causal graph; performing a decision intervention simulation on the contextualized causal graph and performing causal propagation calculations to generate predicted values for target nodes; tracking the observed values of the target nodes after decision execution using a closed-loop validator and calculating the deviation between the observed values and the predicted values; when the deviation triggers a preset condition, performing attribution analysis on the causal path causing the deviation to identify suspicious causal paths; and updating the causal basis graph or the modulation rule based on the analysis results of the suspicious causal paths.
[0034] This invention provides an intelligent decision-making system for enterprise management consulting based on data analysis. It has the following beneficial effects:
[0035] 1. This invention sets up a context state machine module and a causal modulator module. First, the context state machine module determines the current context state of the external environment in real time. Then, the causal modulator module generates a contextualized causal graph that matches the real environment based on the state. This allows the entire decision analysis to be carried out on a dynamically adaptable model, thereby improving the adaptability of the decision system to changes in the external environment.
[0036] 2. This invention constructs a causal basis graph containing nodes, directed causal edges, and quantified attributes. The decision inference module then performs intervention simulation and causal propagation calculations based on the do-operator on this graph. This transforms the basis of decision analysis from traditional statistical correlation to causal relationship, thereby revealing the inherent logic between decision behavior and business results more profoundly and providing technical support for counterfactual inference.
[0037] 3. This invention establishes a complete self-evolution mechanism by setting up a closed-loop validator module and a knowledge base update module. After the decision is executed, the closed-loop validator module verifies the validity and performs attribution analysis on the causal assumptions of the system by comparing the deviation between the prediction and the reality. The knowledge base update module performs targeted correction on the core knowledge base of the system based on the analysis results. This enables the system to continuously learn from practice and improve its accuracy. Attached Figure Description
[0038] Figure 1 This is a system architecture diagram of the present invention;
[0039] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] Example:
[0042] Please see the appendix Figure 1 This invention provides an intelligent decision-making system for enterprise management consulting based on data analysis, comprising:
[0043] The causal basis graph construction module is used to construct a causal basis graph that includes nodes, causal edges, and edge attributes. The edge attributes include at least the causal effect strength and time lag.
[0044] In this embodiment, the core function of the causal basis graph construction module is to build a formalized knowledge base that represents the universal causal relationships between various elements of an enterprise's internal operations. The output of this module, the causal basis graph, provides structured and quantifiable input for subsequent processing steps such as context modulation, decision deduction, and closed-loop verification.
[0045] The causal basis graph is defined as a directed acyclic graph G. CBG = (V, E). Where the node set V consists of a series of nodes v representing quantifiable entities or key performance indicators in the company's operations. i The components include, for example, marketing advertising investment, product prices, new customer registrations, and supply chain costs. The edge set E consists of directed edges ei representing the assumption of direct causal relationships between nodes. ij Composition, a line from node v i Pointing to node v j The directed edges of v represent v i It is v j One direct cause of the change. This directional structural design aims to fundamentally distinguish between causal relationships and statistical correlations.
[0046] To provide a refined description of causal relationships, a module is constructed for each causal edge e. ij Configure a set of quantized edge attributes. Preferably, the edge attributes include: causal effect strength λ. ij Used to quantify the cause node v, assuming other conditions remain unchanged. i The unit change affects the result node v j The expected magnitude of the impact; time lag τ ij Used to characterize the cause node v i After the change, its impact is on the result node v. j The time period required for the above to begin to appear; and the prior confidence level ρ ij This is a normalized value used to measure the degree of certainty of the causal relationship hypothesis. This confidence level will serve as an important basis for the subsequent attribution analysis by the closed-loop verification module.
[0047] To ensure the structural completeness and logical accuracy of the causal basis graph, the causal basis graph construction module adopts a hybrid-driven construction process. This process combines expert knowledge injection with data-driven causal discovery, aiming to integrate the prior knowledge of domain experts with the objective laws of historical data.
[0048] In the expert knowledge injection phase, the module provides interfaces or tools for structuring the experiential knowledge of domain experts and management consultants. This knowledge, existing in unstructured or semi-structured form, improves after-sales service quality and typically significantly reduces churn rates for high-end customers after about a quarter. It is transformed into specific nodes, causal edges, and corresponding initial edge attributes in the graph. This step constructs the initial framework of the graph and establishes the core causal paths with high confidence and consistent business logic.
[0049] In the data-driven causal discovery phase, the module utilizes the historical time-series dataset D accumulated by the enterprise. hist This stage supplements, validates, and quantifies the initial framework built in the expert knowledge injection phase. The goal of this stage is to learn an optimal graph structure E from the data, while satisfying the strong constraints of expert knowledge. * and its parameter set θ * This process can be formalized as an optimization problem, namely, finding a graph structure and parameters that maximize the following objective function:
[0050]
[0051] in,
[0052] G(E,θ) is a set of edges E and parameter set θ (containing λ). ij With τ ij The graphical model defined by (etc.);
[0053] logP(D hist |G(E,θ)) is the log-likelihood term, used to measure the degree to which the current graphical model interprets historical observation data, i.e., the model fit.
[0054] Ψ(G) is a complexity penalty term used to penalize the number of edges in the graph or the overall complexity of the model, in order to avoid spurious causal relationships that lack generalization ability due to model overfitting.
[0055] By solving the above optimization problem, the module can uncover potential causal links that experts have failed to mention, and make more accurate quantitative estimates of the properties of existing causal edges.
[0056] Finally, the causal basis graph construction module integrates the structure generated by expert knowledge injection with the structure of data-driven discovery to form the final causal basis graph G. CBG This graph not only carries the prior wisdom of domain experts but also incorporates statistical evidence from historical data, thus providing a more robust and comprehensive causal knowledge foundation for the entire intelligent decision-making system. The structure and properties of this graph serve as the benchmark for subsequent contextual adjustments by the causal modulator module and the basis for causal propagation calculations by the decision inference module.
[0057] The context state machine module is used to determine the current context state of the system based on external data signal sources.
[0058] In this embodiment, the context state machine module's main function is to formally model and track the dynamically changing external environment that influences enterprise business decisions in real time. This module aims to address the technical problem that traditional decision analysis models typically treat the external environment as a static background or ordinary input variables, leading to insufficient adaptability when the environment undergoes structural changes. By abstracting the external environment into a series of discrete states, this module provides a crucial, dynamic context benchmark for the entire system.
[0059] In a preferred embodiment, the context state machine module is constructed as a finite state machine, which can be formally defined as M. CSM = (S, s0, Σ, δ).
[0060] Wherein, the state set S = {s1, s2, ..., s} k A state s is a finite, discrete set of states in the external environment in which an enterprise exists. k Each of these represents a clearly defined macroeconomic, industry, or competitive environment that has a significant impact on a company's operations. For example, states could include: a period of economic prosperity, an economic recession, a period of technological innovation in the industry, or a price war initiated by major competitors. This discretized definition of states allows for the structured handling of complex external environments.
[0061] The initial state s0∈S is the default environment state when the context state machine module starts up or when there is no explicit external signal input.
[0062] Input symbol set Σ={σ1,σ2,...,σ m A data source is a collection of key events or signals derived from and processed by external data sources. These external data sources may include, but are not limited to: macroeconomic data APIs released by national statistical agencies, news data streams from industry information monitoring platforms, or social media sentiment analysis results for specific market events. The module monitors and analyzes this raw data in real time. When an event that meets preset conditions is detected, a corresponding input symbol σ is generated. m .
[0063] The state transition function δ: S×Σ→S is the core logic driving the state transitions of the context state machine module. This function contains a series of predefined rules to explicitly specify the state transitions in any current state s. i Below, when a specific input symbol σ is received... j When the system transitions to the next state s, lIts functional relationship is s. l =δ(s) i ,σ j For example, a transition rule can be defined as follows: when the system is currently in a period of economic stability and receives an input symbol representing two consecutive quarters of negative GDP growth, the state will transition to a period of economic recession.
[0064] During system operation, the context state machine module continuously monitors the designated external data signal sources. Its internal signal processing unit converts continuous or unstructured external data into discrete input symbols σ. Once a new input symbol is generated, the module immediately calls the state transition function δ, and calculates and updates the system state based on the current context state maintained internally.
[0065] The final output of the context state machine module is its real-time determined current context state s. current This output is crucial for connecting this module to the causal modulator module. Current context state s current This information is passed to the causal modulator module, serving as the sole basis for its selection and application of the corresponding modulation rules. In this way, this module enables the entire intelligent decision-making system to perceive changes in the external environment and ensures that subsequent decision-making analysis is always conducted on a contextualized causal graph that matches the real-world environment, thus providing a foundation for the accuracy and robustness of the decisions.
[0066] The causal modulator module is used to modify the causal basis graph according to the current context state and according to the preset modulation rules in order to generate a contextualized causal graph.
[0067] In this embodiment, the causal modulator module is the core functional unit connecting the causal basis graph and the context state machine module. Its purpose is to dynamically adapt static, universal causal knowledge to specific, changing external environmental contexts, thereby solving the technical problem of traditional analysis models lacking adaptability in the face of different environments.
[0068] The causal modulator module is essentially a rule processing engine, internally storing a series of preset modulation rules. Preferably, these modulation rules correspond to the state set S in the context state machine module. That is, for each context state s in the state set S... k There exists one or a set of corresponding modulation functions F for each. k These modulation functions explicitly define how the underlying causal relationships within a business should change when the business is operating in a specific external environment.
[0069] During system runtime, the causal modulator module receives the current context state s from the context state machine module. current Upon receiving this status, the module retrieves and activates the rule set associated with s from its rule base. current The corresponding modulation function F current Subsequently, the module will use the causal basis map G generated by the causal basis map construction module. CBG As input, and apply the activated modulation function F current The graph is then processed. The output of this processing is a temporary, dynamically adjusted graph, namely, a contextualized causal graph G. sit This process can be formally expressed as:
[0070] G sit =F current (G CBG );
[0071] Modulation function F current For causal basis maps G CBG The modifications are mainly achieved through the following two methods:
[0072] One method is attribute modulation. This approach primarily adjusts the edge attributes of causal edges in a causal basis graph. For G... CBG any causal edge e in ij In context s current New attributes, such as causality
[0073] Effect intensity λ′ ij With time lag τ′ ij It is generated through calculation using a specific modulation function. This calculation process can be expressed as:
[0074] λ′ ij =f λ,current (λ ij );
[0075] τ′ ij =f τ,current (τ ij );
[0076] Where, λ ij and λ ij f is the fundamental attribute value of the causal edge in the causal basis graph; λ,current and f τ,current It is specific to the current context state s current The property modulation function. For example, in a context defined as an industry price war, the effect strength modulation function for the causal side from product pricing to market share may amplify the absolute value of its impact.
[0077] The second approach is structural modulation. This method modifies the topological structure of the causal basis graph itself, representing a more profound adjustment. Structural modulation is achieved by activating or disabling certain causal edges in the graph. In certain contexts, some potential causal relationships may be triggered or revealed, while in other contexts, some known causal paths may be blocked or become inapplicable.
[0078] By performing the above modulation operations, the causal modulator module generates a contextualized causal graph G. sit It is no longer a static knowledge base, but a dynamic model that accurately reflects the interactions between variables under a specific current environment. This module effectively decouples basic causal logic from environmental influencing factors, allowing the maintenance of causal relationships and adaptation to environmental changes to be carried out separately. Finally, the module outputs the generated contextualized causal map to the subsequent decision-making inference module, providing a calibrated, high-fidelity analytical foundation for accurate intervention simulation and consequence prediction under this specific context.
[0079] The decision inference module is used to perform decision intervention simulations and causal propagation calculations on a contextualized causal graph to generate predicted values for target nodes.
[0080] In this embodiment, the decision inference module is the core functional execution unit following the causal modulator module. Its main function is to simulate a specific decision intervention proposed by the user based on a context-adapted causal graph, and quantitatively calculate the potential consequences of this intervention as it propagates along the causal chain.
[0081] The input to the decision inference module is the contextualized causal graph G generated by the causal modulator module, which reflects the current specific environment. sit This module provides an interactive interface for users to receive formalized decision inference requests. In a preferred embodiment, the user's decision problem is transformed into one or more intervention operations based on the do-operator. This do-operator aims to distinguish between active intervention and passive observation, ensuring that the simulation depicts the causal effect resulting from a forced change in a variable's value, rather than merely observing the correlation after the change. This intervention operation can be formally denoted as do(V I :=C), where V I C is a set of one or more nodes in the graph that are being intervened, and C is the target value set for these nodes.
[0082] Upon receiving the intervention command, the decision inference module uses the contextualized causal graph G. sit The causal propagation computation is initiated. This computation process first involves the intervention set v iThe value of the middle node is forcibly updated to its target value CC. Subsequently, based on the topological order of the graph, the module propagates the effect of the change to all its descendant nodes, starting from the affected node and level by level.
[0083] For graph G sit Any non-intervention node v j Its value v at time point t j (t), is computed as the set of all its direct parent nodes Pa(v). j The aggregation result is influenced by the aggregation process. This aggregation calculation is characterized by the following formula:
[0084]
[0085] in:
[0086] v j (t): represents node v j The predicted value at time point t;
[0087] g j (·): This is related to node v j The associated aggregation function, which is used to calculate the combined effect of influences from multiple parent nodes, can be a linear or nonlinear function;
[0088] Pa(v j ): This is in the contextualized causal graph G sit In the middle, node v j The set of all direct parent nodes;
[0089] λ′ ij : is from the parent node v i To child node v j The strength of the causal effect of the causal edge in the current context, this value comes from G. sit ;
[0090] Δv i (t-τ′ ij ): Represents the parent node v i The change in the value of relative to its baseline value, taking into account the time lag τ′. ij Then, at the past time point (t-τ′) ij (occurred)
[0091] τ′ ij : is from parent node v i To child node v j The causal edge has a time lag in the current context, and this value comes from G. sit ;
[0092] ∈ j (t): is related to node vj The relevant random perturbation term is used to characterize the influence of other random factors not included in the model.
[0093] By iteratively executing the above causal propagation calculations until the influence propagates to all terminal nodes without child nodes in the graph, or to the target nodes specified by the user, the decision inference module ultimately generates a prediction sequence for the future values of one or more target nodes.
[0094] The predicted sequence This is the final output of the decision-making simulation module. This output provides decision-makers with a quantitative prediction of the series of consequences that a specific intervention might trigger in the current environment. Subsequently, this predicted sequence will be passed to the closed-loop validator module as a benchmark for comparison with future real-world observations, thus providing the necessary input for the self-verification and evolution of the entire system.
[0095] The closed-loop validator module is used to track the observed values of the target node after the decision is executed and compare them with the predicted values to calculate the deviation. When the deviation triggers a preset condition, the causal path that caused the deviation is analyzed to identify suspicious causal paths.
[0096] In this embodiment, the closed-loop verifier module is the core component for realizing the self-evolution and knowledge correction functions of the system of the present invention. This module establishes a path from real-world observation results to the system's internal knowledge base, aiming to resolve potential discrepancies between the causal model and the dynamically changing real world, and to attribute these discrepancies, thereby driving the continuous iteration of the model.
[0097] The workflow of the closed-loop validator module involves generating predicted values for the target node in the decision inference module. It is then activated after the corresponding decision is implemented. Its inputs include: the target node prediction sequence output by the decision inference module, and the real-world observation sequence corresponding to the target node obtained through the data acquisition interface.
[0098] In a preferred embodiment, the module first performs deviation quantification and tracking. It continuously compares the predicted values with the observed values and calculates the cumulative deviation D(T) between them. This cumulative deviation considers not only the instantaneous magnitude of the deviation but also its persistence over a period of time. This calculation process is characterized by the following formula:
[0099]
[0100] in:
[0101] D(T): represents the cumulative bias over an observation period of total duration T;
[0102] T: The total duration of the observation period, whether preset or dynamically adjusted;
[0103] t ′ : represents the integral variable at time points within the observation period;
[0104] W(t ′ ): This is a time weighting function, preferably an exponential decay function, used to assign a higher weight to recent deviations compared to long-term deviations;
[0105] For the target node at time point t ′ Real-world observations;
[0106] For: the target node at time point t ′ The predicted values generated by the decision inference module.
[0107] Subsequently, the closed-loop verifier module incorporates a verification trigger mechanism. This mechanism does not respond to all deviations; instead, it initiates further in-depth analysis only when the cumulative deviation reaches a statistically significant level. The trigger condition is: when the calculated cumulative deviation D(T) exceeds a preset dynamic deviation threshold θ for the target node. target At that time, the attribution analysis process is activated. This design aims to filter out errors within the normal range caused by factors such as random market fluctuations.
[0108] Once the verification is triggered, the module's core function—attribution of suspected causal paths—begins. This step aims to pinpoint the most likely causal assumption errors leading to significant prediction bias. The module first traces back the main causal paths relied upon from the intervention node to the target node during the decision-making process. Then, for each causal edge e on that path... ij The module calculates its suspense score H. ij This quantifies the likelihood that the edge is a source of bias. The suspense score is calculated as follows:
[0109]
[0110] in:
[0111] H ij : For node v i to node v j The suspense score of the causal edge indicates that the higher the score, the greater the likelihood that it is a source of bias;
[0112] |λ′ ij |: The absolute value of the causal effect strength of the causal edge in the current context state, reflecting the potential influence of the edge itself;
[0113] α ij : This is the contribution factor of the causal edge to the total causal effect from the intervention node to the target node, used to measure the importance of the edge in this specific inference;
[0114] ρ ij : This represents the prior confidence of the causal edge. This value comes from the causal basis graph. The lower the confidence of the causal hypothesis, the more suspicious it is when bias occurs.
[0115] ∈: is a constant smoothing term whose value is close to zero, used to prevent the denominator from being zero.
[0116] By calculating and sorting the suspense scores of all causal edges along a path, the closed-loop validator module ultimately generates an analysis result containing one or more suspicious causal paths. This analysis result clearly identifies the causal relationships most likely to be misjudged or invalid in the current context. This result, as the final output of this module, will be passed to the knowledge base update module, providing a concrete and actionable basis for subsequent targeted corrections to the knowledge base.
[0117] The knowledge base update module is used to update the causal base map or modulation rules based on the analysis results of suspicious causal paths.
[0118] In this embodiment, the knowledge base update module is the final execution unit for the closed-loop learning and self-evolution capabilities of the intelligent decision-making system of this invention. The function of this module is to transform the attribution analysis results generated by the closed-loop verifier module into permanent and targeted corrections to the system's core knowledge base, thereby completing a full feedback loop from "prediction-observation-bias-attribution-correction".
[0119] The knowledge base update module is triggered when the closed-loop validator module successfully identifies one or more suspicious causal paths. Its input consists of the analysis results, which include these suspicious causal paths and their corresponding suspense scores.
[0120] In a preferred embodiment, to ensure the accuracy and reliability of knowledge updates, the knowledge base update module includes a human-computer interaction confirmation step. The module first automatically converts the attribution analysis results from the closed-loop validator module into structured, human-readable corrective hypotheses. This generates a hypothesis in the form of detecting a state s in context. k Below, there is a persistent bias in the prediction from the intervention node to the target node, where the causal path v i →v j The suspense score is the highest. It is suggested that the causal effect strength λ′ of this path in this context be considered. ij The written recommendations for reassessment should be submitted to domain experts or management consultants for final review.
[0121] After receiving external confirmation, the knowledge base update module performs precise and targeted knowledge base updates according to a pre-defined update logic, rather than retraining the model globally. The core innovation of this module lies in its ability to distinguish the nature of the bias and adopt different update strategies for different types of bias.
[0122] Specifically, the module update strategy includes:
[0123] If the bias is identified as being specific to a particular context state—that is, the causal relationship is accurate in most other environments, but only in the specific context s that currently triggers the bias—the bias is not necessarily true. current The following behavior is abnormal. In this case, the module will update the modulation rule in the target causal modulator module. It will only modify the state s current Associated modulation function F current For example, adjusting the function f used to calculate the strength of the causal effect in this situation. λ,current The parameters. In this way, the system only corrects causal cognition in specific contexts, while maintaining the universality and stability of the causal basis map.
[0124] If the deviation is confirmed to reflect a generalized cognitive error regarding the underlying causal relationship—meaning the causal assumption may be problematic in any context—then the module will update the target to the causal basis map G maintained by the causal basis map construction module. CBG Update operations may include: directly adjusting the causal edge e. ij The basic causal effect strength λ ij Or time lag τ ij Or, more importantly, reduce its prior confidence ρ ij This reflects a decrease in the reliability of the causal hypothesis. Lowering the confidence level will make the edge more likely to be flagged as suspicious in future attribution analyses.
[0125] Furthermore, when the existing causal graph structure cannot reasonably explain the observed persistent biases—that is, when adjusting the properties of existing causal edges fails to eliminate the bias—the knowledge base update module triggers a structural relearning mechanism. This module marks the relevant node regions causing the biases and sends a request to the causal basis graph construction module, suggesting a new, more targeted data-driven causal discovery process on the subset of data involved in those regions. This aims to explore potentially previously undiscovered confounding variables or hidden causal links within those regions.
[0126] By performing one or more of the above update operations, the knowledge base update module will embed the new insights learned from real-world biases into a permanent component of the system's knowledge base. The operation of this module enables the entire intelligent decision-making system to continuously adapt to environmental changes, revise its internal assumptions, and improve its own knowledge system, thereby achieving a dynamic and constantly evolving decision support capability.
[0127] Please see the appendix Figure 2 A data-driven intelligent decision-making method for enterprise management consulting, comprising:
[0128] S1: Construct a causal basis graph that includes nodes, causal edges, and edge attributes. The edge attributes include at least the causal effect strength and time lag.
[0129] S2: Determine the current context state of the system based on external data signal sources using the context state machine;
[0130] S3: Based on the current context state and according to the preset modulation rules, modify the causal basis graph to generate a contextualized causal graph;
[0131] S4: On the contextualized causal graph, perform decision intervention simulation and causal propagation calculation to generate predicted values for target nodes;
[0132] S5: After the decision is executed, the observed value of the target node is tracked by the closed-loop validator, and the deviation between the observed value and the predicted value is calculated. When the deviation triggers the preset condition, the causal path that caused the deviation is analyzed to identify the suspicious causal path.
[0133] S6: Update the causal base map or modulation rules based on the analysis results of the suspected causal paths.
[0134] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A data-driven intelligent decision-making system for enterprise management consulting, characterized in that: include: The causal basis graph construction module is used to construct a causal basis graph that includes nodes, causal edges, and edge attributes, wherein the edge attributes include at least the causal effect strength and time lag. The context state machine module is used to determine the current context state of the system based on external data signal sources. The causal modulator module is used to modify the causal basis map according to the current context state and according to the preset modulation rules to generate a contextualized causal map. The decision inference module is used to perform decision intervention simulation and causal propagation calculation on the contextualized causal graph to generate predicted values for target nodes. The closed-loop validator module is used to track the observed values of the target node after the decision is executed, and compare them with the predicted values to calculate the deviation. When the deviation triggers a preset condition, the causal path that caused the deviation is analyzed to identify suspicious causal paths. The knowledge base update module is used to update the causal base map or the modulation rules based on the analysis results of the suspected causal paths.
2. The system according to claim 1, characterized in that, The causal basis map construction module is specifically used for: The initial structure of the causal basis map is established by injecting expert knowledge. The initial structure is supplemented and verified using a data-driven causal discovery algorithm.
3. The system according to claim 1, characterized in that, The causal modulator module modifies the causal basis map in the following ways: Modify the properties of at least one causal edge in the causal basis graph; Activate or disable at least one causal path in the causal basis map.
4. The system according to claim 1, characterized in that, The closed-loop validator module performs the attribution analysis by calculating the suspense score H of each causal edge in the suspicious causal path. ij The calculation method is as follows: in, H ij For node v i to node v j The suspense score of the causal edge; λ′ ij The causal effect strength of the causal edge in the current context state; α ij The contribution factor of the causal edge in the total causal effect from the intervention node to the target node; ρ ij Let be the prior confidence level of the causal edge; ∈ is a smoothing term used to prevent the denominator from being zero.
5. The system according to claim 4, characterized in that, The deviation triggering preset condition is as follows: The cumulative value of the deviation exceeds the dynamic deviation threshold set for the target node.
6. The system according to claim 5, characterized in that, The deviation is the cumulative deviation D(T), which is calculated as follows: Where D(T) is the cumulative deviation within the observation period T; T is the observation period; t ′ For integration, it represents a point in time within the observation period; W(t ′ () is a time weighting function used to assign weights to deviations at different time points; For the target node at time point t ′ Observed values; For the target node at time point t ′ The predicted value.
7. The system according to claim 1, characterized in that, The knowledge base update module is specifically used for: When the generation of the suspected causal path is confirmed to be specific to a certain context state, the modulation rules in the causal modulator module are updated; When the generation of the suspected causal path is confirmed to be related to the understanding of basic causal relationships, the edge attributes in the causal basis graph are updated.
8. The system according to claim 1, characterized in that, The decision inference module performs decision intervention simulation by applying the do-operator to set intervention variables on the contextualized causal graph.
9. The system according to claim 1, characterized in that, The context state machine module determines the current context state through a state transition function, which performs state transitions based on the current state and the received external data signal source.
10. A data-driven intelligent decision-making method for enterprise management consulting, comprising a data-driven intelligent decision-making system for enterprise management consulting according to any one of claims 1-9, characterized in that, Includes the following steps: S1: Construct a causal basis graph containing nodes, causal edges, and edge attributes, wherein the edge attributes include at least the causal effect strength and time lag; S2: Determine the current context state of the system based on external data signal sources using the context state machine; S3: Based on the current context state and according to the preset modulation rules, modify the causal basis graph to generate a contextualized causal graph; S4: On the contextualized causal graph, perform decision intervention simulation and causal propagation calculation to generate predicted values for target nodes; S5: After the decision is executed, the observed value of the target node is tracked by the closed-loop validator, and the deviation between the observed value and the predicted value is calculated. When the deviation triggers a preset condition, the causal path that caused the deviation is analyzed to identify the suspicious causal path. S6: Update the causal base map or the modulation rules based on the analysis results of the suspected causal path.