Enterprise consultation service system based on big data

Abstract

The invention belongs to the technical field of computers, particularly relates to an enterprise consultation service system based on big data, and aims to solve the problems of data islands, decision lagging and insufficient predictive ability in traditional enterprise consultation service. The system comprises a global data access module, a knowledge graph construction module, a dynamic simulation engine, an intelligent consultation generation module and an interactive feedback optimization module, a time sequence evolution knowledge network is constructed by fusing internal and external multi-source heterogeneous data, multi-scene parallel simulation and intelligent suggestion generation are realized, and closed-loop feedback learning is supported. According to the scheme, the accuracy, foresight and response efficiency of enterprise strategic decision can be improved, and large-scale concurrent service and real-time dynamic optimization are supported.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to an enterprise consulting service system based on big data. BACKGROUND

[0002] In the field of modern enterprise management and consulting services, data-driven decision support systems are gradually becoming the core tools for improving the operational efficiency and strategic planning capabilities of enterprises. With the rapid development of information technology, especially the maturity of big data collection, storage and analysis technology, the demand for management consulting services has shifted from traditional experience-oriented to data-intelligent. The key technologies in this field include multi-source data fusion, business intelligence analysis, knowledge graph construction and predictive modeling, aiming to provide accurate, interpretable and forward-looking consulting recommendations through deep mining of massive structured and unstructured data.

[0003] Among them, the enterprise consulting service system based on big data focuses on the integrated analysis of internal and external operational data of enterprises (such as financial data, market behavior, human resources, supply chain information) and industry macro data, and then generates customized strategic, organizational, process or digital transformation recommendations. The basic goal of such systems is to break through the problems of information lag, sample limitations and subjective judgment biases in traditional consulting services, realize the standardization, replicability and intelligence of the consulting process, and thus improve service efficiency and decision-making quality.

[0004] The existing technology still faces a series of key challenges in achieving the above goals: although some consulting platforms have introduced data analysis modules, they generally lack effective cleaning and semantic alignment mechanisms for multi-source heterogeneous enterprise data, resulting in low analysis result reliability; most systems only provide static reports or simple trend prediction, and fail to build dynamic evolving knowledge models to support causal inference and strategy simulation in complex business scenarios; at the same time, existing architectures often ignore the interactive logic of human-machine collaboration in the consulting process, making it difficult to effectively embed expert experience into algorithm models and realize bidirectional iterative optimization; in addition, the real-time response capability of the system is insufficient, which cannot meet the immediate decision-making needs of enterprises in rapidly changing market environments. These problems are particularly prominent in large-scale, cross-industry enterprise consulting service scenarios, severely restricting the deep application and value release of big data technology in the field of high-end consulting services. Therefore, there is an urgent need for a new type of enterprise consulting service system that can realize data deep fusion, has dynamic knowledge evolution capability and supports human-machine collaborative decision-making. SUMMARY

[0005] The present application aims to provide a big data-based enterprise consulting service system to solve the systematic defects of existing enterprise consulting services in data integration capability, decision support accuracy and service response real-time. Current enterprise consulting services mainly rely on the personal experience of consultants and limited market research data of samples, resulting in strong subjectivity and hysteresis of analysis conclusions, which is difficult to adapt to complex and changing business environment. With the exponential growth of enterprise operation data, the traditional mode cannot effectively tap the potential value of unstructured data, and lacks the modeling ability of dynamic coupling relationship between macroeconomic variables, industry competition pattern and internal operation indicators. In addition, the existing system generally lacks cross-cycle trend prediction and multi-scenario deduction mechanism, resulting in lack of forward-looking and robustness of strategic recommendations. In order to break through the above technical bottleneck, it is urgent to build a technical architecture that can realize global data fusion, deep knowledge extraction and intelligent reasoning collaboration.

[0006] The technical scheme of the present application comprises a global data access module, a knowledge graph construction module, a dynamic simulation engine, an intelligent consultation generation module and an interactive feedback optimization module. The global data access module is used to continuously collect heterogeneous data streams from enterprise internal business systems, public market databases, social media platforms, patent literature libraries and macroeconomic release agencies; the data types cover financial statements, customer behavior logs, supply chain transaction records, public opinion texts, policy documents and industry statistical data; the module has a built-in adaptive protocol parser that can automatically match communication protocols according to the data source interface characteristics and perform preliminary data cleaning and format normalization processing. The knowledge graph construction module receives preprocessed multi-source data, extracts core concept nodes such as enterprises, products, technologies, regions and policies through named entity recognition algorithms; uses a relationship extraction model to determine the association attributes between nodes, including supply chain upstream and downstream, market competition, technology substitution and policy influence path; further, a graph neural network is used to perform semantic completion on the initial topology structure, infer the implicit indirect associations, and assign a dynamic weight to each edge to form a multi-dimensional knowledge network with time evolution characteristics. The dynamic simulation engine is based on the completed knowledge graph, and establishes a complex dynamic model containing external impact factors, internal adjustment variables and state transition functions; the external impact factors include uncontrollable variables such as raw material price fluctuations, regulatory policy changes and major public health events; the internal adjustment variables correspond to controllable parameters such as enterprise pricing strategies, production capacity adjustments and research and development investment intensity; the state transition function is obtained by training a long short-term memory network, which is used to describe the time series response law of key performance indicators of enterprises under different input combinations; the simulation process supports setting multiple independently running scenario branches, each branch can configure differentiated initial conditions and intervention strategies, the system performs multi-path reasoning in parallel, and outputs the probability distribution results under each scenario. The intelligent consultation generation module decodes the received simulation output, combines a pre-set industry logic rule base, and converts quantitative prediction into structured consultation suggestions; the rule base is divided into categories such as manufacturing industry, service industry and high-tech industry, each category contains not less than 200 causal inference logics refined based on empirical research; when a certain indicator deviates from the normal fluctuation interval, the module automatically triggers a root cause tracing process, traces back to the upstream driving factors along the knowledge graph, and generates a three-in-one diagnostic report including problem positioning, impact assessment and coping strategies. The interactive feedback optimization module is deployed on the user terminal side, provides a visual operation interface, allows enterprise managers to manually label the generated consultation suggestions, mark their adoption, implementation effect and supplementary opinions; all feedback information is returned to the system backend through an encrypted channel, used to update the trust score of related nodes in the knowledge graph, and online fine-tune some hyperparameters in the dynamic simulation engine, thereby realizing closed-loop learning.

[0007] Preferably, the global data access module sets the data freshness threshold to 15 minutes, automatically starts the reconnection mechanism and sends an alarm signal for the data source that is not updated beyond the time limit; the data cleaning link adopts an outlier detection algorithm based on isolated forest to eliminate outliers deviating from the mean value by more than 3 standard deviations; in the format normalization process, the currency unit is uniformly converted into US dollars, and the timestamp is standardized to UTC+8 time zone. Preferably, the named entity recognition model in the knowledge graph construction module adopts a mixed architecture of BERT-BiLSTM-CRF, and the F1 value on the self-labeled corpus is not less than 92%; the relationship extraction task uses a graph convolution network based on an attention mechanism, which supports the identification of 12 pre-defined relationship types at the same time; the training period of the graph neural network is from 2am to 4am every day, and the node embedding vector is updated using the data increment added in the past 24 hours. Preferably, the dynamic simulation engine configures the minimum time step to be 1 hour, and the maximum simulation span can reach 36 months in the future; the upper limit of the number of scenario branches is set to 16, and the system dynamically allocates GPU computing power according to available computing resources; the training data set of the state transition function contains not less than 100,000 cross-industry enterprise historical operation trajectories, and the sampling frequency is daily granularity; the model verification adopts a rolling cross-validation method to ensure that the prediction error in the future time window is controlled within ±7% of the true value. Preferably, the intelligent consulting generation module has a natural language generation template library built-in, covering 8 major consulting topics such as strategic planning, organizational change and cost optimization; each diagnostic report contains not more than 5 action recommendations sorted by priority, and the priority is calculated by weighting the impact amplitude and implementation urgency; the system adds a confidence label to each recommendation, with a value range of 60% to 99%, and the value is evaluated according to the integrity and reliability of the supporting evidence chain. Preferably, the visual interface of the interactive feedback optimization module provides a three-dimensional heat map display function, with the horizontal axis representing the time sequence, the vertical axis representing different business departments, and the color depth reflecting the risk level distribution; users can retrieve the corresponding simulation path details by clicking any area; the feedback information upload frequency is limited to a maximum of 1 time per hour to prevent high-frequency noise from interfering with model convergence. Preferably, the system is deployed on a distributed cloud computing platform, and containerized orchestration technology is used to realize the elastic scaling of each functional module; inter-module communication follows the RESTful API specification, and the data transmission delay requirement is less than 200 milliseconds; the system processes more than 200 million data records per day, and supports concurrent services for not less than 500 enterprise customers.

[0008] Compared with the prior art, the application has the advantages and positive effects that: This solution fundamentally solves the problem of information fragmentation caused by data silos in traditional consulting services by constructing a data collection system covering multiple internal and external heterogeneous sources. The automated protocol adaptation and real-time cleaning mechanism of the full-domain data access module ensures the quality and timeliness of input data, providing a reliable foundation for subsequent analysis. Compared to traditional methods relying solely on sampling surveys, this system processes data on a scale three orders of magnitude larger, significantly enhancing the representativeness and statistical power of insights. This solution uses knowledge graphs as the core knowledge carrier, achieving a leap from "data" to "cognition." By supplementing implicit connections through graph neural networks, it reveals deep business logic chains that are difficult to discover using traditional analytical methods, such as cross-industry technology spillover effects or secondary impact paths of policy transmission. The dynamic weighting mechanism of the knowledge network allows the strength of associations to evolve over time, capturing the gradual changes in market structure and avoiding the cognitive rigidity caused by static models. This solution introduces a multi-scenario parallel simulation framework, overcoming the limitations of a single prediction path. The dynamic simulation engine not only outputs the most probable development trends but also provides a complete probability distribution view, enabling decision-makers to clearly assess the likelihood and potential losses of various extreme scenarios. This risk probability distribution-based decision support model significantly improves the scientific rigor and resilience of strategic planning compared to traditional qualitative judgment. This solution establishes a complete closed loop from machine output to human feedback. The interactive feedback optimization module enables the system to continuously evolve. Each human intervention is transformed into training signals for model optimization, continuously improving the accuracy of the knowledge graph and the adaptability of the simulation model. With long-term operation, the system will accumulate personalized cognitive models for specific industries or enterprises, providing increasingly customized services that meet actual needs. This solution achieves the standardization and intelligent restructuring of the consulting process, compressing analysis work that originally required weeks to a response time of hours. The structured reports output by the intelligent consulting generation module maintain logical consistency and standardized terminology, eliminating style differences between individual consultants. The system supports 24 / 7 uninterrupted operation and can respond to emergency consulting needs for sudden business events at any time, greatly improving service supply efficiency and response agility. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall technical architecture of the enterprise consulting service system based on big data proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the synergistic effect between the dynamic simulation engine and the knowledge graph in this invention. Detailed Implementation

[0010] Please refer to Figure 1 and Figure 2The application provides an enterprise consulting service system based on big data, comprising a global data access module, a knowledge graph construction module, a dynamic simulation engine, an intelligent consulting generation module and an interactive feedback optimization module. The overall technical architecture of the system starts from the real-time convergence of multi-source heterogeneous data, enters the decision simulation stage based on dynamic deduction through structured knowledge extraction and semantic modeling, finally outputs executable structured consulting suggestions, and realizes the closed-loop iterative update of model parameters through user feedback. The system is deployed on a distributed cloud computing platform supporting elastic scaling, and containerized orchestration technology is used for resource scheduling and life cycle management of each functional module to ensure service stability and response efficiency in high concurrency scenarios.

[0011] The global data access module serves as the information input port of the system and is responsible for continuously collecting raw data from multiple external and internal sources. The data collection range covers the financial accounting subsystem, customer relationship management subsystem, supply chain management system and human resource information system in the enterprise internal business system; meanwhile, it accesses public market databases such as the National Bureau of Statistics industry database, Wanda financial terminal, Tianyancha business information library; social media platforms including Weibo, Zhihu, Douyin and other mainstream content communities; patent literature library covering the full-text retrieval interface of China National Intellectual Property Administration and World Intellectual Property Organization (WIPO); the data sources of macroeconomic release agencies include interest rates, inflation rates, purchasing manager indexes and other key indicators released by the International Monetary Fund (IMF), World Bank and central banks of various countries. The above data streams have significant heterogeneous characteristics: financial statements exist in the form of structured tables, fields including operating income, net profit, asset-liability ratio, cash flow and other core indicators; customer behavior logs are unstructured clickstream records, containing user ID, access timestamp, page path, dwell time, conversion events and other fine-grained operation traces; supply chain transaction records are composed of order number, supplier name, commodity code, quantity, unit price, delivery period, default flag and other information; public opinion texts come from news reports, social comments and forum posts, covering topics such as brand reputation, product evaluation and policy interpretation; policy documents are government-issued regulations, guidelines or industry white papers, usually stored in PDF or HTML format; industrial statistical data is released monthly or quarterly, including production, price index, import and export volume, capacity utilization and other macro dimensions.

[0012] To cope with the above complex and diverse data forms and transmission protocols, the global data access module has a built-in adaptive protocol parser. This parser has automatic recognition capability, which can determine the communication protocol type of the target data source according to its network address characteristics, response header information and data packet structure, and dynamically load the corresponding decoding strategy. For data sources using RESTful API interface, the module uses HTTPS protocol to initiate GET request, carries OAuth 2.0 authentication token to complete identity authentication, and obtains JSON format response body; for services providing WebSocket long connection, the module establishes a persistent session channel to receive incremental push message frames in real time; for FTP or SFTP hosted historical data sets, the module configures a timing task to log in to the server and download the latest version of the compressed archive file within the preset time window; some old systems only support direct database access, and the module establishes a connection pool through JDBC driver to execute precompiled SQL statements to extract specified table change records. All received raw data first enters the buffer queue, and then triggers the cleaning process.

[0013] The data cleaning link adopts an outlier detection mechanism based on the Isolation Forest algorithm. This algorithm regards each data record as a point in a p-dimensional space, forms an integrated model by constructing multiple random split trees, calculates the average path length of each sample, and then obtains the anomaly score. When the net profit of a financial record deviates from the industry average by more than 3 standard deviations, or the time interval between two client logins is shorter than the normal human reaction limit (e.g. less than 0.1 seconds), the system determines it as an outlier and removes it. The cleaning process also includes missing value filling strategies: for continuous variables such as sales, linear interpolation is used to complete the trend based on the previous and subsequent periods; for categorical variables such as customer level, the mode filling method is used to fill in the most common category in the region. All processed data is uniformly entered into the format normalization stage.

[0014] In the format normalization process, the system performs a series of standardization conversion rules. All currency units are converted to US dollars, with the exchange rate benchmark taken from the daily middle price of the People's Bank of China, and if the original data does not specify the currency, it is determined by context keyword matching (such as "yuan", "yen", "euro") combined with IP geolocation assistance; time stamps are uniformly adjusted to UTC+8 time zone and stored in ISO 8601 standard format "YYYY-MM-DDTHH:MM:SS"; geographic location information is converted to latitude and longitude coordinates by calling a geocoding service, with a precision control of city level or above; text content is processed by Unicode normalization, removing invisible characters and rich text tags, and retaining the pure text body. The normalized data is timestamped and source identified, and written into the topic partition of the distributed message middleware Kafka, for downstream module subscription and consumption.

[0015] The global data access module sets the data freshness threshold to 15 minutes. The system maintains a heartbeat monitoring table that records the last successful update time of each registered data source. Whenever new data is written to Kafka, the corresponding source's timestamp is refreshed. A background daemon scans this table every 30 seconds, and if it finds that any source's update delay exceeds 15 minutes, it immediately initiates a reconnection mechanism: first, it attempts to reinitialize the connection configuration, release old session resources, and rebuild the network link; if three consecutive retries fail, it sends an alarm signal to the operation and maintenance monitoring center, notifying the technical personnel to intervene in troubleshooting. This mechanism ensures the continuity and real-time nature of data flow, avoiding analysis interruptions caused by single-point failures.

[0016] The knowledge graph construction module is responsible for converting the cleaned and normalized multi-source data into a knowledge network with semantic associations. When the module starts, it subscribes to the data streams of all related topics from Kafka and batch-pulls the records to be processed according to time windows. The first step is named entity recognition, which aims to extract core concept nodes with business significance from unstructured text. The system uses a BERT-BiLSTM-CRF hybrid architecture to achieve this task. The BERT pre-trained language model is responsible for generating context-sensitive word vector representations, capturing the deep semantics of words in sentences; the bidirectional long short-term memory network further extracts sequence features and learns the context dependency; and the conditional random field layer serves as a decoder, integrating global information to output the optimal label sequence. The model is fine-tuned on a self-labeled corpus containing over 500,000 sentences of artificially labeled enterprise annual report summaries, industry research report paragraphs, and policy interpretation articles, with five categories of labeled classes covering enterprise names, product models, technical terms, administrative divisions, and legal regulation names. The evaluation results on the test set show that the F1 value is stable at over 92%, meeting the requirements of industrial-level applications.

[0017] The second step is relation extraction, which is used to determine the semantic relationship between identified entities. The system deploys a graph convolution network model based on attention mechanism, supporting the simultaneous identification of 12 predefined relationship types, including but not limited to "upstream and downstream of the supply chain", "direct competition", "technology cooperation", "equity investment", "patent licensing", "policy support", "geographical agglomeration", "talent flow", "alternative products", "complementary services", "common customers", and "regulatory affiliation". The model input is two entities and the co-occurrence sentence they are in, focusing on key words expressing their relationship (such as "purchasing from", "joint venture with", "restricted by") through multi-head attention mechanism, generating a relationship vector representation. The training data comes from historical contract text mining, enterprise announcement automatic parsing, and expert rule labeling, totaling 1.2 million positive and negative sample pairs. The model output is a 12-dimensional probability distribution, indicating the most likely relationship category for the current entities.

[0018] The third step is the construction of the atlas topology and semantic completion. The system takes all identified entities as nodes and the extracted relationships as edges to preliminarily construct the original graph structure. However, due to the noise and information loss in the original data, many indirect associations that actually exist cannot be directly observed. Therefore, the system introduces a graph neural network for semantic completion. Specifically, a graph attention network (GAT) is used for embedding representation learning of nodes. The initial feature vector of each node is obtained by concatenating its attribute fields, such as the registered capital, establishment time, industry code, and revenue growth rate in the last three years of the enterprise. GAT uses multi-layer aggregation operations to continuously integrate the information of its neighbors, and finally obtains a d-dimensional (d=256) dense embedding vector. On this basis, the cosine similarity between any two nodes is calculated. If it exceeds the preset threshold of 0.85 and there is no direct edge connection, it is inferred that there is a potential association and the corresponding relationship type label is assigned. For example, if A enterprise and B enterprise are registered in the same park, the management team has cross-employment, and the main business is highly consistent, even if there is no public transaction record, the system can still infer that they constitute a "direct competition" relationship.

[0019] The fourth step is dynamic weight assignment. The knowledge graph is not a static structure, and the strength of its edges evolves over time. The system assigns a dynamic weight w(t) to each edge, reflecting the importance of the association at the current time. The weight update frequency is once an hour, and the calculation is based on real-time signals such as recent association events (such as signing contracts, filing lawsuits), social media mention frequency change rate, stock price linkage coefficient, and supply chain on-time delivery rate fluctuations. The weight calculation formula is as follows:

[0020] where, represents the association weight between nodes and at time ; is the decay coefficient, taking a value of 0.6, to reduce the impact of outdated information; is the time decay factor, set to 0.1 / hour, so that the influence of events that have not been updated for more than 24 hours is reduced to 13.5% of the initial value; is the time interval (in hours) from the last valid event; represents the standardized score of the th type of real-time signal ( =1 to , =5), including public opinion intensity, transaction frequency, capital association, technology citation times, and personnel flow frequency; the normalization factor before the summation term ensures balanced contributions. This formula ensures that the graph can sensitively reflect the dynamic changes of business relationships and avoid being trapped in historical inertia.

[0021] The training cycle of the graph neural network is set from 2:00 AM to 4:00 AM every day. During this time period, the system loads all the newly added data from the past 24 hours, recalculates the node embedding vectors of the affected subgraphs, and updates the global index. After the training is complete, the new version of the knowledge graph is activated online, and the old version is transferred to the cold standby state. The entire process is carried out in an isolated sandbox environment to prevent interference with online query services.

[0022] The dynamic simulation engine is based on the completed knowledge graph and performs multi-scenario deduction on the future development path of the enterprise. The core of the engine is a composite dynamics model composed of three elements: external impact factors, internal adjustment variables, and state transition functions. External impact factors represent macro variables that the enterprise cannot control but have a significant impact on its operations, including fluctuations in international oil prices, tariff rates imposed by major trading partners, the level of public health emergencies, the duration of key raw material supply disruptions, and the tightening of digital currency regulatory policies. These factors are injected in real-time by the global data access module as input conditions for the simulation.

[0023] Internal adjustment variables correspond to strategic parameters that can be actively adjusted by the management of the enterprise, and are used to simulate the differences in the effects of different operating strategies. The variable set includes product pricing float ratio, advertising budget growth rate, R&D investment as a percentage of revenue, capacity expansion multiple, employee recruitment scale growth rate, and inventory turnover target. Each variable has a legal value range constraint, such as a pricing float that cannot exceed ±30% and an R&D intensity that cannot be lower than 50% of the industry average. Users can set different variable combinations for each scenario branch in the interactive interface to explore the optimal strategy space.

[0024] The state transition function describes the evolution of key performance indicators over time under given input conditions. The system uses a long short-term memory network (LSTM) to train this function. The input layer of the model receives a time series vector of length T, containing observations of internal and external variables over the past T=365 days, with a total dimension of 150 (50 external factors, 30 internal variables, and 70 other auxiliary features). The hidden layer consists of 3 stacked LSTM units, each containing 128 memory cells with forget, input, and output gate structures, capable of capturing long-term dependencies. The output layer predicts a sequence of key indicators for the next H=90 days, including revenue growth rate, gross margin, return on equity, customer satisfaction index, and market share ratio. The training data set contains more than 100,000 historical operating trajectories of cross-industry enterprises, with a daily sampling frequency, covering manufacturing, retail, software services, biomedicine, and other fields, ensuring good generalization ability of the model.

[0025] The training process of the state transition function follows a strict validation mechanism. The system adopts a rolling cross-validation method, sliding the training-test window on the time axis: first, train the model with the data of the previous n years and predict the actual performance of the n+1 year; then move the window forward one year and repeat the process. All prediction results are compared with the true values, and the root mean square error (RMSE) and the mean absolute percentage error (MAPE) are calculated. Only when the MAPE is less than 7% in the last 5 rolling periods, the model is considered qualified and put into production environment. Unqualified models will trigger the hyperparameter tuning process, adjusting learning rate, batch size, regularization strength and other parameters before retraining.

[0026] The simulation process supports setting up a maximum of 16 independent running scenario branches. Each branch can be configured with unique initial conditions (such as current asset-liability ratio of 60% or 40%), external shock paths (such as oil price rising 40% or falling 20% in the next 6 months) and internal adjustment strategies (such as aggressive expansion or conservative contraction). The system uses a GPU cluster to perform the deduction tasks of all branches in parallel, with a minimum time step of 1 hour and a maximum simulation span of up to 36 months in the future. In each iteration, the state transition function takes the system state at the previous time and the current input variables as the basis to predict the KPI values at the next time, and feeds back to the trust score of the relevant nodes in the knowledge graph, forming a closed loop of influence.

[0027] The simulation output not only contains the expected path of KPIs under each scenario, but also provides a complete view of the probability distribution. The system performs kernel density estimation on the results of the 16 branches to draw the probability density curve of the key indicators at different time nodes in the future. For example, it can visually show that the probability of the enterprise's net profit reaching break-even at the 18th month is 68%, and the probability of exceeding 100 million yuan is 12%. This presentation based on risk probability distribution enables decision-makers to comprehensively evaluate the likelihood of various extreme scenarios and potential losses, rather than being limited to a single deterministic prediction.

[0028] The intelligent consulting generation module performs semantic decoding on the received simulation output and converts it into structured consulting advice that enterprise managers can understand. The module maintains a natural language generation template library internally, covering eight consulting themes: strategic planning, organizational change, cost optimization, market expansion, technological innovation, risk management, capital operation, and digital transformation. Each theme has no less than 30 sub-templates, totaling more than 240 preset sentence structures. The templates use a slot filling mechanism, with key variables dynamically injected by simulation results. For example, "it is recommended to lower the price of {product line name} by {percentage} % in the future {time interval}, which is expected to increase market share to {target value} %, but it needs to be noted that gross margin may decrease by {value} percentage points."

[0029] When a key metric deviates from its normal fluctuation range, the module automatically triggers a root cause analysis process. Assuming simulations show a sharp increase in customer churn rate to 18% in the next six months (historical average is 8%), the system immediately initiates a diagnostic procedure. First, it locates the corresponding node for the metric in the knowledge graph, then recursively searches upstream driving factors along the reverse edges. The search algorithm prioritizes traversing strongly correlated paths with weights higher than 0.7, checking for the following: competitors launching alternative products, key technical personnel leaving, negative social media sentiment, decreased logistics delivery timeliness, and delayed after-sales service response. For each potential cause found, the system assesses its contribution—that is, the explanatory power of the factor's change on the target metric—calculated by multiplying the Granger causality test statistic by the path weight. Finally, a comprehensive diagnostic report is generated, encompassing problem identification, impact assessment, and response strategies.

[0030] The problem identification section clearly states: "The main reason for the increased customer churn rate is that competitor X's new model Y, released three months ago, offers better value for money, leading to a decrease in the relative attractiveness of our product Z." The impact assessment section quantifies the consequences: "Without intervention, it is estimated that the cumulative customer churn will reach 23,000 users in the next six months, resulting in a direct revenue loss of approximately 140 million yuan." The response strategy section proposes an action plan: "It is recommended to immediately launch a product upgrade project, focusing on enhancing the features of model Z in terms of function A and user experience B. R&D investment is expected to increase by 35%, and the iteration can be completed within 8 months. After launch, it is expected to win back 70% of churned customers."

[0031] Each diagnostic report contains no more than five priority-ranked action recommendations. Priority is calculated by weighting two dimensions: impact magnitude and urgency. Impact magnitude refers to the expected benefits or risk mitigation value that the measure would bring if successfully implemented, expressed in RMB 10,000. Urgency is determined by the system based on the derivative of the rate of deterioration of relevant indicators; a higher value indicates a greater need for immediate response. Priority Score The calculation formula is:

[0032] in, For the final priority score, normalize to the 0-1 range; This represents the original value of the impact magnitude of the current recommendation; and These represent the minimum and maximum impact values ​​among all candidate suggestions, respectively. To implement the original urgency score; and This serves as its extreme boundary. Weighting coefficients of 0.6 and 0.4 reflect system design preferences: prioritizing the overall value of measures over mere time sensitivity. After ranking, the highest-scoring recommendations should be prominently displayed on the first page of the report.

[0033] The system attaches a confidence label to each suggestion, ranging from 60% to 99%. The confidence is a comprehensive assessment of the integrity of the supporting evidence chain and the reliability of the source. The evidence chain integrity measures the number of logical jumps from the suggestion to the core hypothesis, deducting 8% for each level of indirect reasoning; the source reliability scores according to the authority level of the data source: government statistical bulletins are 95 points, listed company annual reports are 90 points, third-party research reports are 75 points, social media rumors are 50 points, and the final weighted average is the total score and is mapped to the confidence interval. For example, a production capacity warning suggestion based on the cross-validation of national statistical bureau data and enterprise financial reports has a confidence of 97%; while a supply chain risk warning based on network rumors has a confidence of only 62%.

[0034] The interactive feedback optimization module is deployed on the user terminal side, providing a visual operation interface that allows enterprise managers to manually label the generated consulting suggestions. The interface uses responsive web design, compatible with desktop browsers and mobile devices. The core functional area is the three-dimensional heat map display panel: the horizontal axis represents the time series of the next 24 months, the vertical axis lists the five functional departments of the enterprise's sales department, production department, R&D department, finance department, and human resources department, and the color depth reflects the comprehensive risk level distribution under the combination of each dimension. The color spectrum transitions from green (low risk) to yellow (medium) to red (high risk). Users can view specific values by hovering the mouse, and click on any cell to retrieve the corresponding simulation path details, including original data sources, model parameter configurations, intermediate calculation results, and alternative scenario comparisons.

[0035] Users can perform four labeling operations on each consulting suggestion on the interface: fully adopt, partially adopt, temporarily not adopt, and supplementary opinions. If you choose "fully adopt", you need to fill in the expected implementation start and end dates and the responsible person information; if you choose "partially adopt", you need to check the specific accepted sub-items and explain the modification reason; if you choose "temporarily not adopt", you need to choose the rejection reason category (such as insufficient resources, inconsistent strategy, inappropriate timing, etc.); if you choose "supplementary opinions", you can input free text to describe the missing factors or local constraints. All feedback information is encrypted by AES-256 and returned to the system backend through HTTPS protocol.

[0036] The uploading frequency of feedback information is limited to a maximum of 1 submission per hour. This throttling mechanism prevents individual users from submitting test data or malicious noise at a high frequency, which can interfere with model convergence. Upon receiving feedback, the system first performs a legality check: it checks whether the digital signature is valid, whether the field format is compliant, and whether the semantic content contains offensive language. Data that passes the verification enters two processing paths: first, the trust score of the relevant nodes in the knowledge graph is updated. For example, if a warning about "raw material price increase" is repeatedly confirmed to be accurate, the trust score of the corresponding knowledge node will gradually increase from the initial 0.7 to 0.92. Conversely, if there are three consecutive prediction errors, the trust score will be reduced to below 0.5, triggering a retraining signal. Second, some hyperparameters in the dynamic simulation engine are fine-tuned online. Specifically, using the small-batch gradient descent method, the user adoption rate is used as a reward signal to adjust the output layer weights of the LSTM model related to a specific industry, so that the model is closer to the actual decision-making preference in future predictions.

[0037] The overall system is deployed on a distributed cloud computing platform, and the Kubernetes container orchestration framework is used to achieve elastic scaling of each functional module. Each module is encapsulated as an independent Docker image, with resource requests (CPU 2 cores, memory 4 GB) and limits (CPU 4 cores, memory 8 GB). The Prometheus monitoring component collects real-time CPU usage, memory usage, network IO, and other indicators of each container. When the load of a module exceeds 80% for 5 minutes, the horizontal expansion strategy is automatically triggered, and new replica instances are created to join the service cluster. When the load falls below 30% and lasts for 10 minutes, the excess replicas are safely terminated to save resources. Inter-module communication follows the RESTful API specification, with a request-response delay requirement of less than 200 milliseconds. The API gateway implements rate limiting, identity authentication, and request logging to ensure system security and traceability.

[0038] The system processes more than 200 million data records per day, with an average throughput of 2400 records per second. It supports concurrent services for at least 500 enterprise customers. Each customer has an independent data space and model instance, logically isolated by tenant ID. The database uses a sharding strategy, dividing physical storage nodes by industry and geographic region to avoid hotspots. The cache layer uses Redis clusters to store frequently accessed knowledge graph fragments and simulation intermediate results, with a hit rate of 89%. Batch task scheduling relies on the Airflow workflow engine to ensure that background jobs such as ETL processes, model training, and graph updates are executed in order according to dependencies.

[0039] The embodiment builds a complete intelligent consulting closed loop from data perception to cognitive generation to action feedback through the cooperative work of the above modules. The global data access module breaks the limitation of traditional consulting relying on limited samples, realizes panoramic scanning of the internal and external environment of the enterprise; the knowledge graph construction module sublimates fragmented information into structured cognition, and reveals the business logic network hidden under the surface; the dynamic simulation engine breaks through the single prediction paradigm, provides multi-path and probabilistic future deduction capability; the intelligent consulting generation module converts complex model output into readable and highly operational decision suggestions; the interactive feedback optimization module gives the system the ability to continuously evolve, so that it becomes more accurate and meets customer needs as the service time extends. The entire system compresses the depth analysis that originally takes several weeks to complete to an hour-level response, supports 7x24-hour uninterrupted operation, and can respond to emergency consulting needs of sudden business crises at any time, fundamentally restructuring the technical paradigm and delivery mode of enterprise consulting services.

[0040] Compared with the prior art, the present scheme realizes three fundamental transitions. First, the data foundation goes from sampling to global. Traditional consulting is limited by manpower and cost, and often draws conclusions based on hundreds of questionnaires or dozens of enterprise interviews, which is limited in representation. The total amount of data accessed by the system reaches PB level, covering the real operation traces of thousands of enterprises, with a statistical efficiency improved by 3 orders of magnitude, significantly reducing the risk of conclusion bias. Second, the analysis paradigm goes from static description to dynamic deduction. Traditional reports are mostly status listing and trend extrapolation, lacking description of nonlinear effects and critical points; the system can reveal the transmission path and chain reaction of black swan events such as policy mutation, technology subversion or market collapse through dynamic modeling and multi-scenario simulation, helping customers to lay down resilience defense lines in advance. Third, the service mode goes from one-time delivery to continuous evolution. Traditional consulting projects end with the termination of the project, and the knowledge is deposited in the consultant's brain and is difficult to reuse; the system accumulates organizational memory through feedback loops, forming personalized cognitive models that can be applied to similar customers, realizing the scale replication and value-added of experience assets.

[0041] The system shows significant technical advantages in practical applications. A large manufacturing enterprise faced the background of escalating international trade frictions and used the system to simulate three coping strategies: maintaining the status quo, transferring production capacity, and developing alternative markets. The simulation results show that under the status quo scheme, the probability of profit decline is 76% in the next 18 months, while the alternative market scheme, although with a large upfront investment, has a 63% chance of achieving a profit turnaround within three years. The enterprise adjusted its strategic focus accordingly and successfully avoided significant business risks. In another case, an Internet company found that the root cause of its slowing user growth was not the intensification of market competition, but the imbalance between new and old users caused by the internal recommendation algorithm. After the system suggested restructuring the incentive mechanism, the retention rate rebounded by 14 percentage points the following month. These cases demonstrate that the system not only identifies problems but also accurately locates the root causes and provides verifiable solutions.

[0042] In summary, the technical solution disclosed in this embodiment builds a full-chain intelligent consulting system covering data collection, knowledge modeling, dynamic simulation, intelligent generation, and feedback optimization, systematically addressing the inherent defects of traditional enterprise consulting services in data breadth, analysis depth, and time accuracy. The system not only improves the quality and efficiency of single consulting services but also establishes a sustainable cognitive infrastructure, providing strong technical support for enterprises to make scientific decisions in complex and uncertain environments.

Claims

1. A big data-based enterprise consulting service system, characterized by, The application comprises: a global data access module for continuously collecting raw data streams from multiple heterogeneous data sources and performing adaptive protocol parsing, data cleaning and format normalization on the raw data streams to obtain preprocessed data; a knowledge graph construction module for receiving the preprocessed data, constructing an initial knowledge graph through named entity recognition and relation extraction, and using a graph neural network to perform semantic completion and dynamic weight assignment on the initial knowledge graph to form a multi-dimensional knowledge network with time evolution characteristics; a dynamic simulation engine for constructing a complex dynamics model containing external impact factors, internal adjustment variables and state transition functions based on the multi-dimensional knowledge network, and performing multi-path reasoning in parallel according to multiple configured scenario branches to output the probability distribution results of key performance indicators under each scenario branch; an intelligent consultation generation module for receiving the probability distribution results, combining a pre-set industry logic rule base for semantic decoding, converting quantitative prediction into structured consultation suggestions, and generating a diagnostic report containing problem positioning, impact assessment and coping strategies by tracing back upstream driving factors along the multi-dimensional knowledge network when detecting that a key performance indicator deviates from a pre-set threshold; an interactive feedback optimization module for providing a visual operation interface to receive user manual annotations on the structured consultation suggestions, and returning encrypted annotation information to the system backend to update the trust score of related nodes in the multi-dimensional knowledge network and online fine-tune part of the hyperparameters of the complex dynamics model, thereby realizing closed-loop learning.

2. The big data based business consulting service system according to claim 1, wherein, The global data access module comprises: an adaptive protocol parser for automatically matching communication protocols according to data source interface characteristics and loading corresponding decoding strategies; a data cleaning unit for using outlier detection algorithms to remove outliers and performing missing value imputation; a format normalization unit for uniformly converting data from different sources into a standardized format.

3. The big data based enterprise consulting service system as claimed in claim 2, wherein, The knowledge graph construction module comprises: a named entity recognition unit for extracting core concept nodes of enterprises, products, technologies, regions and policies from unstructured text; a relation extraction unit for determining the association attributes between the core concept nodes, including upstream and downstream of the supply chain, market competition, technology substitution and policy influence paths; a graph topology construction and semantic completion unit for constructing an initial topology by taking the core concept nodes as graph nodes and the association attributes as graph edges, and using a graph neural network to infer implicit indirect associations; a dynamic weight assignment unit for assigning dynamic weights to graph edges that reflect the importance of the association at the current time, and periodically updating the dynamic weights according to real-time signals.

4. The big data based enterprise consulting service system according to claim 3, wherein, The dynamic simulation engine comprises: an external impact factor injection unit for taking uncontrollable macro variables as input conditions for simulation; an internal adjustment variable configuration unit for receiving user-set variable combinations corresponding to strategic parameters adjustable by enterprise management; a state transition function unit for describing the evolution law of enterprise key performance indicators over time under given input conditions; A multi-scenario parallel deduction unit is configured to configure differentiated initial conditions and intervention strategies for each scenario branch, and perform parallel deduction tasks to generate the probability distribution results.

5. The big data based business consultancy service system as claimed in claim 4, wherein, The intelligent consultation generation module comprises: A semantic decoding and rule matching unit is configured to match the probability distribution results with the causal inference logic in the industry logic rule base; A structured report generation unit is configured to fill the matching results into a natural language generation template to form the structured consultation suggestions; A root cause tracing and diagnosis unit is configured to recursively search upstream driving factors along the reverse edges of the multi-dimensional knowledge network under a trigger condition, and evaluate the contribution of each factor to generate the diagnosis report.

6. The big data based business consulting service system as claimed in claim 5, wherein, The interactive feedback optimization module comprises: A visualization interface unit is configured to display the risk level distribution in the form of a three-dimensional heat map, and support the user to call up the details of the corresponding simulation path; An artificial annotation receiving unit is configured to receive the user's annotation operation on the adoption, implementation effect and supplementary opinions of the structured consultation suggestions; A feedback information processing unit is configured to perform legality verification on the annotation information, and distribute it to the knowledge graph construction module and the dynamic simulation engine to perform model updating.

7. The big data based business consulting service system as claimed in claim 6, wherein, The root cause tracing and diagnosis unit is configured to preferentially traverse the strongly correlated paths in the multi-dimensional knowledge network with a weight higher than a preset threshold, to locate the potential causes leading to the deviation of the key performance indicators.

8. The big data based business consulting service system as claimed in claim 7, wherein, The structured report generation unit is further configured to calculate a priority score for each action suggestion, and sort the suggestions according to the priority score, wherein the priority score is calculated by weighting the impact amplitude and implementation urgency. 9.The big data based business consulting service system according to claim 8, wherein, The structured report generation unit is further configured to attach a confidence label to each action suggestion, and the value of the confidence label is evaluated according to the integrity of the supporting evidence chain and the reliability of the data source. 10.The big data based business consulting service system according to claim 9, wherein, The system is deployed on a distributed cloud computing platform, each functional module is containerized and encapsulated, and communicates through interfaces complying with the RESTful API specification, to realize elastic scaling and high-concurrency services.

Images