Medical market data classification and analysis method and system based on artificial intelligence

Abstract

The invention relates to the technical field of medical big data analysis, in particular to a medical market data classification and analysis method and system based on artificial intelligence, and the method comprises the steps: firstly obtaining multi-source data, generating a standard event flow, and constructing an event traceability knowledge graph and a sliding window event intensity sequence; secondly, text representation, graph representation and structured representation are fused, and hierarchical market classification and executable vectors are output under time logic constraints; a world model and a return model are established again, and a strategy value and a risk index are obtained in combination with off-line evaluation; and finally, performing robust optimization in the feasible region, and generating a final executable strategy and forming a deployment record according to the condition risk value constraint. According to the method, explainable, audible and robust market insight and strategy decision can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical big data analysis, in particular to a medical market data classification and analysis method and system based on artificial intelligence. BACKGROUND

[0002] The data sources of the medical market are scattered, the policies and procurement cycles change frequently, the channels and inventories are significantly affected by supply and cash flow, and enterprises need to quickly respond within the compliance boundaries in product layout, price strategy and resource allocation. Traditional reports and static analysis cannot express the cross-source causal chain and time constraints, rule scoring lacks quantitative control of the risk tail, and strategy online lacks traceable evidence and gray mechanism, resulting in inconsistent classification, uncontrollable execution and high audit cost. The industry urgently needs an interpretable, auditable and robust decision-making process that can maintain stability under distribution drift, connecting data insights and execution strategies. SUMMARY

[0003] In view of the many problems existing in the prior art, the present application provides a medical market data classification and analysis method and system based on artificial intelligence. The present application uses event traceability knowledge graph and sliding window event intensity sequence as unified representation, fuses text representation, graph representation and structured representation, and outputs hierarchical market classification and executable vector under time logic constraints; combines offline evaluation of world model and return model, and obtains the final executable strategy under the constraints of robust optimization and conditional value at risk, improving accuracy, compliance and robustness.

[0004] A medical market data classification and analysis method based on artificial intelligence, comprising the following steps: Obtain multi-source medical market data and perform entity resolution and encoding, convert the data into atomic events according to event type, generate a standardized event stream after time alignment and deduplication, and construct a timestamped event traceability knowledge graph based on the standardized event stream and generate a sliding window event intensity sequence; Fuse text representation, graph representation and structured representation based on the event traceability knowledge graph and the sliding window event intensity sequence, and output hierarchical market classification and executable vector under the time logic constraints of event calculus; Based on the fused representation, hierarchical market classification and executable vector, establish a world model and a return model and calculate the strategy value and risk indicators using offline evaluation; Construct a robust optimization model within the feasible region determined by time logic and business constraints, obtain the final executable strategy under the conditional value at risk constraint, and generate deployment records corresponding to the final executable strategy.

[0005] A medical market data classification and analysis system based on artificial intelligence, for implementing the medical market data classification and analysis method based on artificial intelligence, the system comprising: an event modeling module configured to acquire multi-source medical market data, perform entity resolution and coding, convert the data into atomic events according to event types, generate a standard event stream through time alignment and deduplication, construct a timestamped event provenance knowledge graph based on the standard event stream, and generate a sliding window event intensity sequence; a fusion determination module configured to perform text representation, graph representation and structured representation fusion based on the event provenance knowledge graph and the sliding window event intensity sequence, and output a hierarchical market classification and an executable vector under the time logic constraint of event calculus; a simulation evaluation module configured to establish a world model and a return model based on the fusion representation, the hierarchical market classification and the executable vector, and calculate a strategy value and a risk indicator by offline evaluation; a robust optimization module configured to construct a robust optimization model within a feasible region determined by time logic and business constraints, obtain a final executable strategy under a conditional value at risk constraint, and generate a deployment record corresponding to the final executable strategy.

[0006] Compared with the prior art, the application has the following advantages and beneficial effects: Through unified modeling of the event provenance knowledge graph and the standard event stream, consistent alignment and causal chain traceability of cross-source data are achieved; Through the sliding window event intensity sequence and the time logic constraint, accurate constraints on policy effectiveness and termination and classification output within the compliance boundary are achieved; Through adaptive fusion of text representation, graph representation and structured representation, robust classification and executable determination under the multi-factor coupling of policy, channel and inventory are achieved; Through offline evaluation of the world model and the return model, value estimation and evidence review of the candidate strategy are achieved without interfering with the online business; Through robust optimization based on the probability distribution distance and conditional value at risk constraint, quantitative control of tail risk and stable income in extreme scenarios are achieved; Through feasible region projection and trust region difference constraint, smooth connection between the strategy and historical behavior is achieved, and online fluctuation and compliance risk are reduced; Through the deployment record and evidence binding mechanism, full-link auditability and replayability are achieved, and the audit and review cycle is shortened; Through hierarchical market classification-based category grouping and gray deployment, controllable volume release and rapid rollback are achieved, and the strategy landing efficiency and safety are improved. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 is a flowchart of the method of the application; Figure 2 is a structural block diagram of the system of the application. DETAILED DESCRIPTION

[0008] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It should be understood, however, that the description is merely exemplary and is not intended to limit the scope of the present disclosure. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure.

[0009] As Figure 1 shown, an artificial intelligence-based medical market data classification and analysis method includes the following steps: Obtain multi-source medical market data and perform entity resolution and encoding, convert data into atomic events according to event type, generate a standardized event stream through time alignment and deduplication, build a timestamped event provenance knowledge graph based on the standardized event stream and generate a sliding window event intensity sequence; Preferably, obtaining multi-source medical market data includes sales data, order data, inventory data, price data, rebate data, budget data, repayment data, policy text, recruitment text, academic summary, meeting minutes, channel correspondence text, customer service text, social public opinion and main data, performing entity resolution and encoding, converting data into atomic events according to event type, generating a standardized event stream through time alignment and cross-source deduplication, building a timestamped event provenance knowledge graph based on the standardized event stream, and generating a sliding window event intensity sequence.

[0010] In the embodiment, the system is composed of a data access layer, an analysis and standardization layer, an event generation layer, a time processing and deduplication layer, a knowledge graph layer, and a time series analysis layer. The data sources include sales data, order data, inventory data, price data, rebate data, budget data, repayment data, policy text, recruitment text, academic summary, meeting minutes, channel correspondence text, customer service text, social public opinion and main data. The access layer establishes a connection according to the source, unifies the character set and time format, adds source identification and ingestion timestamp to all data records, and forms an original data set.

[0011] The entity resolution and coding adopts a two-stage strategy. The first stage performs blocking grouping, and a candidate set is established with the normalized writing of the enterprise name, product registration certificate number, medical institution unified social credit code, regional administrative code, and channel number as the key. The second stage performs pairwise comparison. The absolute difference threshold is used for numerical field comparison, the weighted sum of edit distance and pinyin similarity is used for name field comparison, and the consistency check of the coded standard address library is used for address field comparison. When the comprehensive similarity exceeds the preset threshold, they are merged into the same entity, and the enterprise code, product code, medical institution code, region code, and channel code are output. This process generates an entity mapping table, which is used for unified reference in all subsequent steps to avoid data splicing errors caused by name ambiguity. The implementation effect is that multi-source records are aligned with a unified primary key, and subsequent graph and time series analysis can be stably positioned to the same business object.

[0012] Event generation converts standardized data rows into atomic events. Structured sources are mapped to a fixed set of event types according to data table types and field values, subject participants are filled with unified codes from the entity mapping table, event start times are determined by record time or policy effective time, and necessary additional attributes are written to the payload field. Text sources are first divided by paragraph, and medical field named entity recognition models are used to extract elements such as regulation name, applicable area, category restriction, amount threshold, and validity period start and end, and then aligned with the entity mapping table to form policy and procurement events. Each event generates an event identifier, which is obtained by sequentially connecting the event type, participant code, start time, and payload summary and then performing hash calculation to ensure idempotent writing across batches. The implementation effect is that heterogeneous sources are uniformly projected into the same event model, and the same interface can be used for subsequent consumption.

[0013] Time alignment and deduplication perform unified time axis processing on full-quantity events. Relative time is converted to absolute time, and all timestamps are saved in the same time zone. Two sets of rules are executed in parallel. Rule one is key consistent deduplication, which merges and retains the latest payload when the event type, participant code, and start time are completely consistent. Rule two is text approximate deduplication, which identifies and merges repeated text events using a fixed-length fragment set and similarity threshold. When merging, the source credibility and record freshness are used as the dominant choice to retain fields. The implementation effect is to eliminate duplicate records and cross-source conflicts, ensuring that the same fact appears only once and reducing subsequent calculation bias.

[0014] The event-based knowledge graph is constructed using enterprises, products, medical institutions, regions, channels, and policies as node sets, with events serving as the basis for generating time-stamped edges. When an event describes a product's inventory arrival at a medical institution, an edge is generated from the product to the medical institution, with edge attributes including start time, end time, and load summary. Policy effectiveness events generate edges from the policy to the region and product category. To support temporal retrieval, the graph maintains a snapshot index based on time slices, allowing access to valid node relationships on any given date. The graph storage retains event identifiers and source fingerprints, ensuring that any graph relationship can be traced back to the original record, meeting auditing requirements. The implementation provides a structured representation of multi-entity relationships along the time dimension, facilitating subsequent graph message passing and rule inference.

[0015] The sliding window event intensity sequence is used to characterize the activity of different event types over time. For a specified combination of product, region, and medical institution, for each event type, the occurrence frequency or weighted sum of events is statistically analyzed by rolling the window with a fixed width and step size, forming a fixed-length time series. The core calculation expression is as follows: ;

[0016] Event type In window width Deadline The intensity of the event, For the set of events of event type k, For the first The start time of the event, The weight of the first event is given by 1, where 1 represents the indicator function. The event weight expresses a combined value of source confidence and extraction confidence. The weight can be obtained through linear weighting, and the weight coefficients are configured in the system parameter table. This sequence is written into the time-series analysis layer as the time-series input basis for the classification and decision modules. The implementation effect is that discrete events from different sources are stably converted into continuous metrics, which is beneficial for subsequent models to identify periodic, abrupt, and persistent effects.

[0017] To ensure the consistency of each layer, event identification, entity encoding, timestamp, and source fingerprint are one-to-one corresponding in all tables. The knowledge graph layer and the time series analysis layer pass edge attributes and time slice information through primary key connection. Any downstream calculation can be traced back to specific events and original records. When the system is deployed, the data access layer and the parsing layer run in batch processing mode at a fixed frequency. Event generation and time processing are continuously written in the streaming queue. The knowledge graph layer inserts or updates edges according to the incremental update strategy. The time series analysis layer recalculates the sliding window event intensity sequence for the latest period at a fixed time every day. The technical effect of this embodiment is that multi-source data for medical market business is reliably standardized into events and graph structures, forming a unified input that can be directly used for subsequent classification, judgment, simulation, and optimization, reducing misjudgment and repeated calculation caused by data inconsistency across systems, and improving the explainability and auditability of subsequent modeling and decision-making.

[0018] Fusion of text representation, graph representation, and structured representation based on event trace knowledge graph and sliding window event intensity sequence, output hierarchical market classification and executable vector under the time logic constraint of event calculus; The system input is event trace knowledge graph and sliding window event intensity sequence, as well as text fragments and table features corresponding to the sub-business unit. Text fragments come from policy texts, recruitment texts, academic summaries, meeting minutes, channel correspondence texts, customer service texts, and social public opinions. Table features come from sales data, order data, inventory data, price data, rebate data, budget data, and repayment data. The event trace knowledge graph provides entities such as enterprises, products, medical institutions, regions, channels, and policies, as well as time-labeled relationship edges.

[0019] Text representation generation uses a pre-trained language model in the Chinese medical field to perform word segmentation and truncation on each text fragment, outputting a fixed-length vector as the text representation. Graph representation generation takes the time slice of the event trace knowledge graph as input and uses a message passing network containing time encoding to consistently generate vectors on nodes as graph representation. Structured representation generation performs missing indication and standardization on table features and inputs a multilayer perceptron to obtain a vector as the structured representation. The sliding window event intensity sequence is used to calculate the frequency or weighted frequency of each event type on a given product and region and medical institution combination within a fixed window, which is used for subsequent fusion weight inference and time logic evaluation.

[0020] To adaptively fuse the three types of representation, a gated weight generator based on the sliding window event intensity sequence is introduced. The core calculation is: ; ; Where, is the text representation, is the graph representation, is a structured representation, is a shape vector obtained by one-dimensional convolution and pooling of the sliding window event intensity sequence, and is a linear mapping parameter, , , is a three-way fusion weight, is a fusion representation. This design automatically increases the weight of the corresponding mode during the policy-intensive period or supply disturbance period, reducing the misjudgment caused by a single mode.

[0021] The temporal logic constraints of event calculus are used to depict the validity and termination of clauses, and are input as soft constraints during classification and judgment. The core calculation is: ; wherein, is the validity truth value of the th clause at time , is the event count or weighted count, is the termination event count or weighted count, , , is a learnable or rule-set coefficient, is a sigmoid function. The temporal logic truth value and the fusion representation are jointly input to the downstream two-class output head.

[0022] The hierarchical market classification output is organized in a parent-child consistent hierarchy, first judging the upper-level category, and then subdividing to the lower-level category within the restricted space. The executability vector output includes five components of compliance dimension, qualification dimension, channel dimension, inventory dimension, and budget dimension, representing the executable probability under the corresponding constraints. During training, the consistency of hierarchical market classification is constrained by cross-entropy loss, the multi-label output of the executability vector is constrained by binary cross-entropy loss, and a violation penalty of the temporal logic constraint is added. Specifically, when the temporal logic truth value is not valid, the high-confidence output of the related category or dimension is suppressed. During reasoning, the temporal logic truth value is used as a gating signal for post-processing of the executability vector, ensuring that the situation violating the clause is not judged as executable.

[0023] To ensure feasibility, the fusion weight generator and the three-class representation encoder are optimized end-to-end in the same training process; the output dimensions of the text representation and the graph representation are fixed and consistent in the design phase, and the structured representation is aligned to the same dimension through a linear layer; the window width and step size of the sliding window event intensity sequence are uniformly configured in the system parameter table; the event count of the temporal logic directly comes from the time slice query of the event provenance knowledge graph, ensuring traceability and consistency.

[0024] In engineering deployment, the text representation generation, graph representation generation, structured representation generation, fusion weight generation and two types of output heads are packaged into the same inference service, the input is the business unit identifier, the corresponding time range and the text batch to be evaluated, the output is the hierarchical market classification and the executable vector, and the most influential text segment identifier and graph path identifier are attached for audit. Through the above implementation, the system realizes robust classification and executable judgment under complex time constraints and multi-source heterogeneous data, and can maintain output consistency and interpretability in the face of policy changes, supply fluctuations and channel adjustments.

[0025] Preferably, the text representation is generated by a pre-trained language model, the graph representation is generated by a time-series message passing of a graph neural network, and the structured representation is generated by feature embedding and a multi-layer perception machine. The text representation, graph representation and structured representation are fused through attention mechanism and interaction to obtain a fused representation.

[0026] The system aligns three types of inputs at the same time section: one is the sentence segment from policy text, recruitment text, academic summary, meeting minutes, channel correspondence text, customer service text and social public opinion; two is the slice of the event trace knowledge graph at that time and the neighborhood of the target business object; three is the table features composed of sales data, order data, inventory data, price data, return point data, budget data and repayment data, and a sliding window event intensity sequence is used as a time series clue.

[0027] Text representation generation: Chinese word segmentation and subword encoding are performed on each text sentence, truncated to a fixed length, and sent to a pre-trained language model in the medical field; the pooling vector of the first mark is taken as the text representation, and the sentence segment identifier is reserved for subsequent audit. To enhance the sensitivity to clauses and numerical values, two additional embeddings, "trigger word" and "numerical position", are added to the word embedding layer.

[0028] Graph representation generation: the time slice of the event trace knowledge graph is taken as the input, the two-hop neighborhood of the node where the target business object is located is selected, and the edge features are derived from the type and time difference of the corresponding atomic event. A graph neural network containing time encoding is used for message passing, which first performs linear transformation according to edge type, then performs weight decay according to time difference and aggregates to adjacent nodes, and finally reads the graph representation at the target node. To avoid the dominance of hot nodes, the aggregated vector is normalized with a function of node degree.

[0029] Structured representation generation: standardization is performed on continuous features and a missing indication is output; discrete features are mapped to dense vectors using an embedding dictionary and then concatenated with continuous features; the concatenation result is sent to a multi-layer perception machine to output a structured representation. This path shares the same batch index with the text path and the graph path during the training period, ensuring that the three representations are fully aligned by sample and time.

[0030] The adaptive fusion of three-path representation takes the sliding window event intensity sequence as the gating signal. First, the sliding window event intensity sequence is input into one-dimensional convolution and pooling to obtain a shape vector, which is then spliced with the concise statistics of the three-path representation, the three-path weights are calculated and weighted and interacted. The core calculation is as follows: ; ; wherein, is the text representation vector, is the graph representation vector, is the structured representation vector, is the shape vector extracted from the sliding window event intensity sequence, is the confidence statistic of the text representation, is the confidence statistic of the graph representation, is the confidence statistic of the structured representation, is the weight matrix, is the bias vector, is the text channel weight, is the graph channel weight, is the structured channel weight, is the fusion representation. To capture the interaction between modalities, a low-rank term of element-wise product can be added after the above weighting and to enhance the nonlinearity, but the interface is not changed.

[0031] The training process inputs the fusion representation into two output heads: the hierarchical market classification head and the executability judgment head. The hierarchical market classification outputs the categories in the order from parent class to child class to avoid inconsistent paths; the executability judgment outputs five components of compliance dimension, qualification dimension, channel dimension, inventory dimension and budget dimension. The loss function uses the weighted sum of classification loss and multi-label loss, and introduces a penalty term based on temporal logic to suppress high-confidence outputs that do not match the true value of the clause. During inference, the executability components are gated by the true value of the temporal logic to ensure that situations that do not meet the clause are not judged as executable.

[0032] In engineering implementation, the three-path encoder and the fusion layer are provided in the same inference service, and the input is the time range, the target business object identifier and the text batch identifier. The text path returns the text representation and the adopted sentence segment identifier, the graph path pulls the neighborhood from the event trace knowledge graph by the graph computing engine according to the time slice, and the structured path retrieves data from the feature warehouse and synchronizes with the missing indication. To ensure consistency, all samples use the same primary key and event identifier, the sliding window event intensity sequence and the temporal logic true value are directly derived from the slice query of the event trace knowledge graph, and any fusion result can be located to the corresponding text segment, graph path and table field, which is convenient for auditing and backtracking.

[0033] Through the above implementation, the system automatically increases the text channel weight in the policy-intensive period, increases the graph channel weight in the network structure change period, and increases the structured channel weight in the sales and inventory driven period, to realize adaptive modeling for different business scenarios. The fusion representation provides stable and interpretable input for subsequent counterfactual simulation, offline evaluation and robust optimization, reduces the chain effect of single modal distortion, and improves the reliability and feasibility of classification and executable decision.

[0034] Preferably, the temporal logic constraints of event calculus are output by calculating the temporal truth value and rule violation degree, and jointly optimizing with the supervision loss, to output the hierarchical market classification and executability vector.

[0035] The system input is an event trace knowledge graph and a sliding window event intensity sequence, and a fusion representation aligned to the same time slice. First, a clause rule library is established. The rule library is stored in a table structure, including clause identification, trigger event type, termination event type, applicable object range, associated output target, clause weight and evidence pointer. The association between the clause and the output target is maintained by using a mapping table, for example, mapping "qualification validity" to the qualification dimension in the executability vector, mapping "selection and price limit" to the channel dimension, mapping "budget and repayment constraints" to the budget dimension, and mapping "policy range restriction" to the visible category set of hierarchical market classification. The rule library only records business verifiable conditions, and does not introduce abstract logic that is difficult to implement.

[0036] The temporal truth value calculation depends on the time slice of the event trace knowledge graph and the sliding window event intensity sequence. For each clause, the weighted count of its effective event and termination event is counted at a given business object and time point to obtain the temporal truth value. The core calculation expression is: ; The true value of the first clause at time t is The true value of the first clause at time t is The weighted count of the effective event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the effective event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the effective event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the effective event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the termination event of the first clause at time t is The weighted count of the termination event of the first clause at time t is

[0037] The time-based truth values ​​are transformed into constraint signals that can be perceived by the model optimization. These constraint signals are divided into two categories: one is a hierarchical market classification constraint, which prevents certain categories from being predicted when the clause is invalid; the other is an enforceability vector constraint, which prevents the corresponding dimension from giving a high probability when the clause is invalid. To keep the implementation simple, a rule violation degree is constructed during the training phase, and the output score associated with the clause is monotonically and consistently penalized with the time-based truth value. The core calculation expression is: ; For the degree of violation of the general rules, For the first The weight of each clause In order to be with the first The model output score is associated with each clause. The model output score is derived from the target category probability or the corresponding component of the executability vector in the hierarchical market classification. This expression means that a penalty is imposed when the ground truth value is low but the model gives a high score, thus forcing the model to learn the constraint that "a clause should not be output with high confidence if it is not effective."

[0038] Supervised training employs joint loss, combining conventional supervision with constraint penalties. The core calculation expression is: ; For joint losses, The monitoring loss due to hierarchical market classification, For multi-label supervision loss of executability vectors, These are the weighting coefficients for the degree of rule violation. The labels of the training data come from historical real records or manually verified samples, and the clause weights and coefficients are determined through parameter tuning on the validation set.

[0039] During the inference phase, the time truth value serves as a gating signal in post-processing. For outputs associated with clauses, element-wise lower bound constraints are applied to the prediction results and the time truth value to suppress outputs that violate the clauses. Specifically, when the time truth value is less than a threshold, the visible set of the corresponding category is removed, and the probability of the corresponding dimension is reduced below the threshold. This process does not change the model structure and is only performed on the output side, facilitating auditing and playback.

[0040] In terms of engineering implementation, the rule base is maintained in the form of a configurable table, with fields including clause identification, clause type, trigger event type, termination event type, applicable object key, associated output key, and evidence pointer. The event provenance knowledge graph provides a time query interface, returning a list of events and weights related to a given object and clause. The time truth value calculation module performs batch processing, caching the results in a key-value store with keys composed of object, clause, and date to ensure high concurrency matching with model reasoning. Violation degree and joint loss are implemented as custom operators in the training framework, with parameters and logs written to a monitoring system to support online playback and offline review.

[0041] The effects of this embodiment are reflected in three points. First, classification and determination can converge stably when regulations and recruitment boundaries change, reducing false positives such as "policy not yet in effect but already determined to be executable." Second, the constraint signal comes from traceable events and graph relationships, and any determination result can be traced back to specific clauses and event evidence, meeting audit requirements. Third, time truth value and rule violation degree only rely on standard queries of events and graphs, and do not require rewriting the model when clauses are modified or new clauses are added. Only new mappings and parameters need to be added to the rule base, reducing operational costs. Through the above design, the model output is guided by executable event calculus and temporal logic constraints, ensuring that hierarchical market classification and executable vectors meet data rules and business constraints, and that conclusions are actionable, reviewable, and maintainable.

[0042] Preferably, hierarchical market classification uses a hierarchical classifier that maintains parent-child consistency, and the executable vector includes multi-label probabilities for compliance, qualification, channel, inventory, and budget dimensions.

[0043] The system input is a fusion representation, time slice query results of the event provenance knowledge graph, and summary statistics of sliding window event intensity sequences. The training data consists of historical business records and manually reviewed samples, with annotations including two parts: path labels for hierarchical market classification, organized in a four-level category tree, with recommendations for upper-level categories to lower-level categories in the order of indication group, region, medical institution level, and channel status; and five dimension labels for the executable vector, corresponding to compliance, qualification, channel, inventory, and budget, with binary labels.

[0044] The hierarchical market classification head uses a series of four-level classifiers. Each level receives the fusion representation and the prediction distribution of the previous level to generate the probability of this level, while using a mask to normalize only the sub-classes allowed by the previous level, avoiding contradictions where the parent class is not true but the child class is true. To explicitly maintain the parent-child consistency, a consistency penalty is introduced: ; For the parent-child consistency penalty, is the child class probability, To correspond to the parent class probability, sum up all parent-child pairs. This penalty is combined with cross-entropy training at each level to stabilize the output that meets the hierarchical constraints.

[0045] The executable vector head is composed of five independent binary output units, which linearly transform and nonlinearly activate the fusion representation to obtain five probability components. Supervision uses multilabel binary cross-entropy: For multilabel loss, For any dimension of the five dimensions, For the true label of this dimension, For the predicted probability of this dimension, Indicates compliance, Indicates qualification, Indicates channel, Indicates inventory, Indicates budget. Five components share the fusion representation but do not share parameters with each other, facilitating independent calibration by dimension. Two types of output are jointly optimized, and the overall goal is: For joint loss, For the sum of cross-entropy of four-level classifiers, For the parent-child consistency penalty weight, For the multilabel loss weight. The weights are determined by grid search on the validation set. When training, the training set and validation set are divided according to time before and after to avoid information leakage; class imbalance is handled by sampling or loss weighting.

[0046] In the inference process, the system first outputs the four-level class probability, then masks and normalizes the lower-level classes according to the parent class probability to obtain consistent path predictions; then it outputs the five-dimensional executable vector. To adapt to business use scenarios, thresholding and calibration are introduced: temperature scaling is used for probability calibration on four-level classes, and then rounding is performed according to the maximum probability path; for the five dimensions, thresholds are selected on the validation set to achieve a pre-set balance based on recall and precision. The thresholds are stored as configuration items for quick switching in different customer scenarios.

[0047] The combination of event traceability knowledge graph and sliding window event intensity sequence is reflected in two points: one is at the data entry, limiting the visible class set with time slicing, for example, the regional partition and channel state change with policy and recruitment changes; two is in the explainable output, the system provides supporting evidence for each prediction, including the adopted text fragment identifier, the participated graph path identifier, and the triggered event key, which facilitates manual review.

[0048] ​​In engineering implementation, the hierarchical market classification head and the executability vector head are deployed in a shared service process, with input being fusion representation and temporal context identification, and output being path probability and five-dimensional probability and evidence index. Model parameters and thresholds are managed in a versioned manner, and inference logs record prediction results, evidence index and primary key, supporting playback and audit. Offline batch processing is retrained or fine-tuned once a day to adapt to policy and channel changes; online services are released in gray scale when version switching, monitoring parent-child consistency violation rate and five-dimensional stability.

[0049] The effect of this embodiment is that, on the one hand, through the explicit constraint of parent-child consistency, the unreasonable output of "parent class negation and child class affirmation" is eliminated, and the business credibility of hierarchical market classification is improved; on the other hand, through five-dimensional independent modeling and dimension-by-dimension calibration, it can be directly used as input for subsequent simulation evaluation and robust optimization, reducing rule conflict and manual correction cost, and realizing smooth connection and auditable landing from classification to executability.

[0050] Based on fusion representation, hierarchical market classification and executability vector, world model and return model are established, and strategy value and risk indicators are calculated by offline evaluation; The system first constructs training trajectories. Taking the combination of product, region and medical institution as the smallest business unit, the fusion representation, hierarchical market classification, executability vector, sliding window event intensity sequence and historical actions and results of each period are aggregated into sequence samples in chronological order. The fusion representation comes directly from the upstream fusion module; the hierarchical market classification is used to limit the subsequent visible category space; the executability vector provides executability prior in five dimensions of compliance, qualification, channel, inventory and budget. Historical actions are extracted from marketing logs, price and repayment logs, and mapped into four types of continuous or discrete decisions: frequency of investment, channel allocation ratio, academic activity intensity, price and repayment parameters. Result data includes conversion quantity, gross profit, compliance passing quantity, inventory change.

[0051] The world model is used to approximate the state transition and short-term feedback of the business environment. The world model receives the current state and action and predicts the next period state, while giving a short-term estimate of the result. The core relationship is: ; ; The next period state prediction is The world model is The state vector at time is The action vector at time is The current return prediction is For the reward model, the state vector is obtained by concatenating the probability vector and executability vector of the fused representation and hierarchical market classification, as well as the statistics of the sliding window event intensity sequence. The action vector consists of parameterized representations of four types of decisions. During training, time is split before and after, missing values ​​are filled in forwards and missing indicators are added, and the loss is the sum of the state prediction error and the reward prediction error. The world model can be implemented using a gated recurrent unit or a lightweight temporal encoder, and the reward model can be implemented using a multilayer perceptron.

[0052] The business definition of returns uses a configurable weighted sum, which both covers business objectives and facilitates auditing. A typical approach is to treat conversion and gross profit as positive terms, and compliance failures and overstocking as penalty terms, with weights set by a parameter table and adjusted according to industry strategies. To reduce bias from a single perspective, the return model is scaled during the validation phase to ensure comparability of return definitions across different business units.

[0053] Offline evaluation is used to compare the value and risk of candidate strategies when online trials are not possible. The system first models historical behavioral strategies, employing multi-class logistic regression for discrete decisions and a conditional Gaussian model for continuous decisions, and performs probability calibration using a calibration set. Subsequently, it combines the world model and the reward model to obtain a value function estimate, which is then combined with importance weights to calculate a double-robust estimate. The core calculations are as follows: ; For the value estimation of candidate strategies, To assess the sample size, For the first The state vector of each evaluation sample at the sampling time. For the first The historical action vectors of each evaluation sample at the same moment For value function estimation, Let be the action probability of the candidate policy in the state. For historical behavior strategy probabilities, For historical returns, in implementation, the value function can be obtained by rolling the world model several times and accumulating returns. An upper limit is set on the importance weights to reduce variance, and robust handling is performed on samples with outlier weights. The risk indicator uses quantile loss as a conditional value at risk measure and is calculated through a concise auxiliary variable form: ; Confidence level Under the following conditions, the value of risk, For quantile variables, For the sample size, For loss samples based on candidate strategies, The non-negative part is. The loss sample is composed of negative return or insufficient target achievement. In order to obtain interval estimation, the system uses self-sampling to repeatedly evaluate the value and risk, outputs the confidence interval and records the feature distribution of unstable samples for subsequent risk control and audit.

[0054] In order to ensure traceability and maintainability, the system uses a unified primary key throughout the state, action, return, value function and importance weight. All estimation results are accompanied by the time range of training data, model version and parameter hash. Each run of offline evaluation forms a solidified report, including the value point estimation, confidence interval, conditional value at risk, high-risk scenario list triggered by the candidate strategy, and the corresponding evidence reference, which can go back to specific events and text fragments in the event traceability knowledge graph.

[0055] In deployment, the world model and the return model provide state rolling and value function query interfaces as inference services. Offline evaluation is performed as a batch task, and evaluation tasks are automatically generated for new strategies and the results are archived. The effect of this embodiment is to provide stable, interpretable and auditable value and risk evaluation results for candidate strategies without interfering with online business, and the evaluation link seamlessly connects with upstream fusion representation and downstream optimization decision, providing directly usable input for subsequent robust optimization and gray online.

[0056] Preferably, the offline evaluation uses a double-robust offline evaluation method to calculate the strategy value and risk indicators. The historical behavior strategy is estimated by a probability model, the value function is calculated by the world model and the return model, and the weight is reweighted based on data distribution alignment.

[0057] Offline evaluation takes time-ordered business trajectories as input. Each trajectory is composed of a state vector, an action vector, a single-period return and a time label. The state vector is obtained by splicing the fusion representation, the probability vector of hierarchical market classification, the executability vector, the statistics of sliding window event intensity, and the inventory and budget snapshot; The action vector comes from marketing and price and payment logs, which are uniformly mapped into elements such as delivery frequency, channel allocation, academic activity intensity, price parameters and payment parameters. Before the data enters the evaluation module, time segmentation, missing indication and standardization are completed to ensure consistent training and evaluation standards.

[0058] The historical behavior strategy is used to describe the selection probability of historical decisions. For discrete decisions, multi-class logistic regression or gradient boosting tree is used to output category probability; for continuous decisions, conditional Gaussian or mixed density network is used to output conditional density, and continuous components are discretized by interval or directly calculated density value. The probability is calibrated by order preserving or temperature scaling, and the log likelihood and precision are tested on an independent validation set. The calibration parameter and model version are included in the audit list.

[0059] The world model and reward model are used to provide value function estimates. The world model takes a state vector and an action vector as input and predicts the next period state and intermediate quantities; the reward model takes the same input and outputs a one-period reward. Both use a combination of light sequence networks and multi-layer perceptrons, and take the weighted sum of state errors and reward errors as the training objective. To ensure comparability across sub-business units, the reward model is quantile calibrated and group scaled, and the calibration parameters are stored in a parameter table.

[0060] The policy value is estimated using double robustness. The expression is When implemented, the ratio of importance weights is capped and the truncation proportion is recorded to avoid excessive variance.

[0061] To reduce the bias caused by domain migration, data distribution alignment weights are introduced. By calculating the sample similarity between the evaluation set and the target set and decoupling the weights, the alignment weight of each sample is obtained; the similarity can be obtained by weighting the distance of the sliding window event intensity and the distance of the fusion representation, and the alignment weight is normalized and participates in the re-weighted estimation. The core calculation of re-weighting is: ; wherein is the data distribution alignment weight of the i-th sample, and the weight is derived from the distribution matching process and satisfies non-negativity and sum normalization.

[0062] The risk indicator is measured using conditional value at risk. The core calculation is as follows: . The loss sample is constructed from negative returns or target unfulfillment, and the value estimate and risk indicator are given self-help confidence intervals during evaluation.

[0063] To ensure traceability, the evaluation process outputs a versioned report for each batch, which includes the value point estimate, confidence interval, conditional value at risk, alignment weight distribution, proportion of importance weights truncated, and evidence link of representative samples of the candidate strategy. The evidence link can go back to the event key and text segment identifier of the event source knowledge graph, facilitating auditing and review.

[0064] In engineering implementation, offline evaluation runs in batch processing tasks, with input being the evaluation time range and candidate strategy configuration identifier; historical behavior strategy, world model and reward model are pulled by model warehouse; distribution alignment weight is calculated by alignment service and written to disk; estimator module calculates double robustness value and conditional value at risk and generates report. All modules use a unified primary key to pass through state, action, return and evidence, ensuring that the results are repeatable and verifiable.

[0065] ​​Through the above implementation, the application can compare the value and risk of different candidate strategies stably without intervention of online business, and reduce the influence of domain migration through data distribution alignment; the evaluation of link, upstream representation learning and downstream optimization deployment alignment is outputted, which can be directly used for quantitative indicators of robust optimization and gray online.

[0066] Preferably, the world model adopts a recurrent neural network to establish state transition, the reward model adopts a multilayer perceptron to establish a reward function, and stress tests are performed in a scenario library of policy scenarios, supply scenarios, public opinion scenarios and competition scenarios to obtain scenario stability and risk indicators.

[0067] The system first constructs a time-ordered business trajectory. Taking the combination of products, regions and medical institutions as the minimum unit, the probability vector of integrated representation, hierarchical market classification, executable vector, statistical quantity of sliding window event intensity, inventory and budget snapshot, and historical action and reward are collected period by period to form state sequence and action sequence. Missing values are filled forward and missing indicators are generated synchronously, and all numerical fields are standardized according to the minimum and maximum intervals of the training period to ensure consistency of online and offline criteria.

[0068] The world model is used to depict the evolution of state with action, and a recurrent neural network is used to realize it. The input is the state vector and action vector of the current period, and the output is the state prediction and intermediate quantity prediction of the next period. The core relationship is: ; ; is the state prediction of the next period, is the state transition function parameterized by the recurrent neural network, is the current state vector, is the current action vector, is the current period reward prediction, is the reward function parameterized by the multilayer perceptron. The training is time forward and backward, and the goal is to minimize the weighted sum of state prediction error and reward prediction error, and early stop in the validation set to prevent overfitting. The state vector is fixed by the probability vector of integrated representation, hierarchical market classification, executable vector, statistical quantity of sliding window event intensity, inventory and budget snapshot; the action vector is fixed by the encoding of launch frequency, channel allocation, academic activity intensity, price parameter and repayment parameter, discrete decision adopts one-hot encoding, and continuous decision adopts interval normalization.

[0069] The reward model gives single-period business return, which is composed of conversion, gross profit, compliance cost, overstock penalty, etc. The weights are stored in the parameter table for auditing and parameter adjustment. After training, different business units are scaled and calibrated to make the return comparable.

[0070] The scenario library is used for stress testing and contains four types of parameterized scenarios. Policy scenarios modify the validity and termination of clauses through event provenance knowledge graph slices, directly affecting the temporal logic truth values and changing the visible category set. Supply scenarios modify inventory-related components and related pathways by injecting events that disrupt or advance the arrival of goods. Public opinion scenarios affect demand-related pathways and the weight of text evidence by superimposing positive and negative peaks on the sliding window event intensity. Competition scenarios affect share-related components and price sensitivity by adding events such as the launch, promotion, or withdrawal of competing products. The above scenarios are parameterized in tables, including the effective time, duration, intensity, and target object, and any changes can be replayed and reproduced.

[0071] In each scenario, a fixed initial state and candidate strategy are used to roll the world model for multiple steps to obtain the predicted trajectory and accumulate returns. Compared with the baseline scenario, the stability and risk indicators are output. Stability is measured by the amplitude of output variation, and the core calculation is: ; The stability of the scenario is The number of rolling steps is The joint output vector at time in the scenario includes the probability vector and the executable vector of the hierarchical market classification, The same vector of the baseline scenario is The dimension of the joint output vector is

[0072] The implementation process is as follows: 1. Load the specified version of the world model and the return model; 2. Generate scenario injection events from the event provenance knowledge graph according to the configuration, and refresh the temporal logic truth value and visible category set in real time; 3. Select actions according to the candidate strategy at each period, and use the world model to roll to obtain the state sequence and return sequence; 4. Calculate the stability and conditional value at risk, and output the scenario report. The report includes stability, conditional value at risk, key time points, triggered clause identification, injection event list, and adopted text and graph path evidence, all with replayable timestamps and primary keys, meeting compliance auditing.

[0073] Through the above implementation, the recurrent neural network and multilayer perceptron give reproducible state evolution and return prediction under the unified state definition and action coding. The scenario library covers key disturbances in policy, supply, public opinion, and competition in a parameterized manner. Stability and conditional value at risk form a consistent quantitative benchmark, allowing candidate strategies to complete interpretable and reliable evaluations under multiple extreme conditions before going live.

[0074] A robust optimization model is constructed within the feasible region determined by temporal logic and business constraints. The final executable strategy is obtained under the conditional value at risk constraint, and the deployment record corresponding to the final executable strategy is generated.

[0075] The system enters the strategy solving stage after completing the fusion representation, hierarchical market classification, executability vector and offline evaluation. Taking the combination of products, regions and medical institutions as the decision unit, the time true value is calculated from the rule base and event trace knowledge graph to generate a hard constraint list for the period, including action boundaries, budget upper limit, channel frequency upper limit, inventory non-exceeding boundary, compliance and qualification prerequisites. The time logic and business constraints are translated into standard constraints by the same rule compiler to form the feasible region. The offline evaluation service provides expected return and loss samples for each candidate action and gives sample grouping labels to cover normal, policy changes, supply disturbances and public opinion fluctuations.

[0076] The core optimization objective and risk constraint of the application are described in a single expression: ; Wherein, is the action vector; is the feasible region defined by the time logic and business constraints; is the expected return estimate given by the offline evaluation; is the loss based on the candidate action; is the conditional value at risk with a confidence level of ; is the risk upper limit.

[0077] For landing solution, first perform constraint compilation and consistency check. The rule compiler translates budget, inventory, channel, compliance and qualification into linear or convex constraints, writes into the constraint list and records the effective time and version number, and generates a feasibility check script for subsequent projection and pre-online check to ensure consistency between training and deployment. Then determine the risk range and sample construction: the loss is obtained by linear combination of negative return and key indicator non-achievement degree, and the weight is stored in the parameter table; the conditional value at risk is calculated using sample approximation and quantile auxiliary variable method in the realization layer; the sample grouping weight is input into the optimizer as a configuration item to ensure balanced performance of the strategy under different business conditions.

[0078] The solving process adopts a two-stage type of "main problem optimization + feasible projection". The optimizer uses the sample average target and risk item to form a differentiable target, and uses sequential quadratic programming or projection gradient method to update the action vector; after each update, the feasibility check script is called to project the action to ensure that it falls back into the feasible region; if the projection distance exceeds the threshold, record it as high feasibility deviation and prompt to tighten the target or supplement evidence. To control the estimation variance, perform truncation on the importance weight and abnormal samples, and write the truncation ratio into the report.

[0079] After obtaining the action vector of each decision unit, the system generates the final executable strategy, and solidifies the version together with the feasible region snapshot, time true value summary, model and parameter hash. To verify that the constraints and risks are not broken, a number of decision units are randomly selected, and a short-term trajectory is played back using the world model to check that the conditional value at risk does not exceed the upper limit, the resources and channels do not exceed the limit, and the inventory does not exceed the limit. Online in a gray way, gradually increase the volume according to the category of hierarchical market classification; Run monitoring to track the constraint violation rate and risk indicators, and once the threshold is approached, automatically restore to the previous version according to the rollback parameters in the deployment record.

[0080] The deployment record is a structured document, including at least the strategy version number, the constraint version number, the confidence level and upper limit of the conditional value at risk, the optimizer and key hyperparameters, the sample grouping and weight, the truncation ratio, the maximum distance of the feasible projection, the playback result summary, and the traceable evidence reference (clause identifier, event key, text segment identifier, and graph path identifier). All fields are time-stamped and fingerprinted for easy review and accountability.

[0081] By compiling the time true value and business constraints into a unified feasible region, and maximizing the expected return and tail risk of constraints in a single expression, executable and auditable strategy solving is achieved; Sample grouping and gray online mechanism makes the strategy controllable in the case of policy mutation, supply interruption and public opinion fluctuation; The deployment record links evidence and parameters, making it easy for customers and auditors to review, ensuring that the strategy can be implemented, interpreted and maintained.

[0082] Preferably, the robust optimization model defines the uncertainty distribution family based on the probability distribution distance of ambiguity sets, and optimizes the expected return under the most unfavorable distribution of the uncertainty distribution family, while setting constraints on the conditional value at risk of loss to limit tail risk.

[0083] The system takes the results of offline evaluation of fusion representation, hierarchical market classification, executable vector, world model and return model as input, and establishes decision units for the combination of products, regions and medical institutions within the same business cycle. First, calculate the time true value according to the event traceability knowledge graph, and list the actions that do not meet the clauses as forbidden items; Then compile the upper limit of the budget, the upper limit of the channel frequency, the non-exceeding of the inventory, the compliance and the necessary items into standard constraints to form a feasible region. Then extract the sample set from the offline evaluation service, which contains state summary, candidate action, expected return, loss value and time label. To cover distribution migration, construct a distance matrix between samples according to the state summary and environment label, and use the distance as the ground truth cost metric to obtain the experience distribution and the ambiguity set radius parameter of its neighborhood. The radius is derived from the distance quantile of the historical window and is solidified in the parameter table.

[0084] Robust optimization describes the objective and risk control in a single core expression: ; where, is the action vector, is the feasible region, is the sample distribution, is the probability distribution distance with radius defined ambiguity set, is the empirical distribution (contained in ), is the expected return estimate given by offline evaluation, , is the loss, is the confidence level, is the upper tail risk limit.

[0085] Ambiguity set construction and worst-case distribution solving are implemented by engineering "adversarial reweighting": first, according to the sample distance matrix and radius, limit the weight offset of each sample; second, fix the action vector, select the weight vector that minimizes the expected return to get the worst-case distribution; third, fix the worst-case distribution, update the action vector to improve the expected return. This inner-outer double loop runs in batch mode, the inner layer is implemented by weight update iteration with constraints, and the outer layer updates the action by sequential quadratic programming or projected gradient method. To ensure constraint consistency, a feasible projection is performed after each outer layer update to project the action vector back to the feasible region; if the projection distance exceeds the threshold, it is recorded as a feasibility deviation and triggers a parameter review.

[0086] Conditional value at risk constraints are calculated by risk services, using a sample approximation method with auxiliary variables, which can be executed in parallel with adversarial reweighting without model modification. Risk services group loss samples (normal, policy changes, supply disturbances, public opinion fluctuations), calculate conditional value at risk separately, and compare with the upper limit; important weights and extreme losses are truncated and recorded in reports, and the truncation ratio is used as a mandatory check item before the upper limit.

[0087] In terms of parameter and hyperparameter management, ambiguity set radius, risk confidence level, risk upper limit, sample grouping weight, feasible projection threshold, optimization step size, and maximum iteration number are saved in both configuration files and databases, and configuration changes enter the versioned pipeline. All intermediate quantities (weight vector, projection distance, constraint active set) are archived by decision unit for easy reproduction and audit.

[0088] After solving, the system aggregates the action vectors of each decision unit to form the final executable strategy, and generates a deployment record together with the feasible region snapshot, ambiguity set parameters, risk scope, model, and configuration hash. Online uses a gray strategy, gradually increasing the volume in groups classified by hierarchical market classification; the monitoring module continuously tracks the constraint violation rate, conditional value at risk, and return deviation, and once the threshold is approached, it rolls back to the previous version according to the rollback parameters in the deployment record.

[0089] The effect of this implementation is that, without changing the upstream data and model, the ambiguity set constructed by the probability distribution distance covers the uncertainty brought by distribution migration; the adversarial reweighting gives a conservative evaluation under the most unfavorable distribution, and the conditional value at risk constraint suppresses tail risk; the feasible projection and versioned deployment record ensure that the policy is executable, traceable, and maintainable.

[0090] Preferably, the final executable policy is online using trust domain difference constraints to control the change amplitude of the final executable policy and the historical policy, and the gray deployment is performed according to the category grouping of the hierarchical market classification, and the deployment record corresponding to the final executable policy is generated.

[0091] The system generates a candidate policy after completing robust optimization, compares and constrains it with the historical policy, and performs gray online according to the category grouping of the hierarchical market classification. To avoid business fluctuations and compliance risks caused by large jumps in the policy, the trust domain difference constraint is used to control the change amplitude of the candidate policy and the historical policy, and the real-time monitoring and rollback mechanism is used to ensure executability and auditability.

[0092] First, build a baseline before going online: train the historical policy from the marketing and price and payment logs in the near future, output the action distribution under the given state; read the action distribution of the candidate policy under the same state from the offline evaluation results. The state definition is consistent with the upstream, consisting of fusion representation, probability vector of hierarchical market classification, executability vector, statistics of sliding window event intensity, inventory and budget snapshot. For continuous actions, use segmentation or Gaussian approximation to get the distribution, and for discrete actions, use classification distribution representation to ensure that the two types of policies are comparable in the same action space.

[0093] The trust domain measures the difference between the candidate policy and the historical policy in terms of average information difference, and the following formula is used as the core constraint: ; Where, is the number of extracted state samples, is the th state sample, is the action distribution of the candidate policy at this state, is the action distribution of the historical policy at this state, is the relative entropy, is the trust domain threshold. For ease of engineering implementation, state samples come from the near real-time cache of the to-be-online grouping; when the metric exceeds the threshold, the action distribution of the candidate policy is scaled down by a certain proportion and re-verified until the threshold is met and the feasible region projection verification (budget, inventory, channel, compliance and qualification) is passed.

[0094] The gray-scale deployment is in the form of hierarchical market classification category grouping, following the order of "small-scale pilot, batch expansion". Each group is bound to independent monitoring and rollback thresholds, including three types of indicators: first, the compliance and qualification hard constraint violation rate is zero; second, the interval threshold of key operating indicators (such as the relative deviation of conversion and gross profit within the limited range); third, the rolling condition value-at-risk does not exceed the upper limit set before the online. The monitoring window and sampling frequency are configured according to the group size, and only one-tenth of the volume is expanded in the first batch. After two consecutive monitoring windows meet the standards, the volume is expanded; any indicator out of range will trigger an automatic rollback to the historical strategy or the previous version of the strategy.

[0095] During the online process, the online gate service calculates a lightweight online difference metric before requesting to enter the strategy service (sampling actions from the current state, estimating the relative entropy of a single point and averaging it within the window). When the window average approaches the threshold, it automatically reduces the exploration intensity and action amplitude, and records the "threshold approach" event. All online decisions are subject to feasible region projection, ensuring that the budget upper limit is not exceeded, the inventory is not out of bounds, the channel frequency is not exceeded, and the compliance and qualification requirements are met; if the projection distance exceeds the alarm threshold, the volume of the group is temporarily expanded and a diagnostic report is output.

[0096] The deployment record is solidified in a structured document, including: strategy version number, historical strategy version number, trust domain threshold, number of sampling states, monitoring window parameters, online grouping and volume ratio, feasible region snapshot (budget, inventory, channel, compliance and qualification key coefficients), risk range and upper limit, online gate and projection alarm threshold, rollback condition, model and configuration hash, evidence reference (clause identifier, event key, text segment identifier, graph path identifier). All fields are time-stamped and signed with a fingerprint for easy review and accountability.

[0097] In terms of system architecture, the gate service, strategy service, projection service, monitoring service, and audit service are deployed independently and interact through a unified primary key and version number. The gate service maintains a state cache and difference metric; the strategy service provides action distribution for candidate strategies and historical strategies; the projection service implements feasible projection after constraint compilation; the monitoring service continuously calculates operating indicators and risk indicators; the audit service writes deployment records and stores only-increase-not-modification for key fields. The version switching adopts a double-channel hot switching strategy, which verifies the metrics and indicators in the shadow channel before switching, and retains the rollback handle after switching until the two monitoring windows are stable.

[0098] Through the above mechanism, the final executable strategy gradually expands while maintaining a bounded difference with the historical strategy, and can be rolled back at any time; all online decisions can be traced back to specific evidence and parameters, ensuring that the strategy online process is executable, interpretable, auditable, and adaptable to the compliance requirements and business rhythm of the medical market.

[0099] For example, Figure 2As shown, an artificial intelligence-based medical market data classification and analysis system is used to implement the artificial intelligence-based medical market data classification and analysis method. The system comprises: An event modeling module is configured to acquire multi-source medical market data, perform entity resolution and coding, convert the data into atomic events according to event types, generate a standardized event stream through time alignment and deduplication, construct a timestamped event provenance knowledge graph based on the standardized event stream, and generate a sliding window event intensity sequence. The event modeling module is deployed on a general x86 server cluster, and a front-end acquisition gateway is equipped with a gigabit Ethernet interface and a clock synchronization (NTP / PTP). Multi-source data is accessed through a message queue. The computing nodes use multi-core CPUs and large-capacity memories to perform entity resolution and coding, use memory hash structures to complete time alignment and cross-source deduplication, and NVMe SSDs to carry disk logging and intermediate columnar data. Graph construction is performed by storage-intensive nodes, which use memory mapping and incremental writing to generate a timestamped event provenance knowledge graph. The sliding window event intensity sequence is accumulated in real time by a memory ring buffer and periodically written to object storage for downstream consumption. The cluster is interconnected through high-speed Ethernet, with redundant power supplies and RAID to ensure continuous operation.

[0100] A fusion determination module is configured to perform text representation, graph representation, and structured representation fusion based on the event provenance knowledge graph and the sliding window event intensity sequence, and output hierarchical market classification and executable vectors under the time logic constraints of event calculus. The fusion determination module uses a heterogeneous resource pool of GPU inference servers and general CPU nodes. The GPU is responsible for batch inference of pre-trained language models and graph neural networks, and the video memory resides the model weights and intermediate vectors. The CPU is responsible for word segmentation, graph neighborhood extraction, feature standardization, and result assembly. The nodes use PCIe high-bandwidth buses and NVMe local caches to accelerate weight loading and batch processing; the inference results are returned through shared memory or zero-copy channels to reduce copy overhead. To ensure stable latency, the GPU nodes are configured with low-latency network interfaces and support RDMA; the vectorized features and sliding window features are resident in the memory key-value database, and the gating and fusion are completed on the CPU. Finally, the hierarchical market classification and executable vectors are output.

[0101] The simulation evaluation module is used for establishing a world model and a return model based on the fusion representation, the hierarchical market classification and the executability vector, and calculating strategy value and risk indicators by offline evaluation; the simulation evaluation module adopts a CPU / GPU hybrid computing cluster. Multi-step rolling of the world model and batch evaluation of the return model are executed in parallel on a GPU, which is suitable for large-scale matrix and sequence operations; trajectory splicing, state caching and sample resampling are completed on a large-memory CPU node. Historical behavior strategy probability estimation and importance weight calculation are realized by a CPU parallel library, evaluation jobs are distributed through a task queue, and results are persisted to object storage and columnar data warehouse. To improve throughput, local NVMe is used as an intermediate cache and asynchronous writing is enabled; to ensure reproducibility, model versions and configurations are hashed in a read-only parameter library, and evaluation logs are time-stamped through a time synchronization service, which facilitates auditing and comparison.

[0102] The robust optimization module is used for constructing a robust optimization model in a feasible region determined by time logic and business constraints, obtaining a final executable strategy under a conditional value at risk constraint, and generating a deployment record corresponding to the final executable strategy; the robust optimization module uses a multi-core CPU as the main computing resource to run numerical optimization and feasible region projection, and calls a GPU for acceleration when large-scale gradient updating is required. A feasible region matrix and a sample weight vector are maintained in memory, and optimization iterations are delivered through shared memory queues and lock-free queues; risk calculation and conditional value at risk are completed in parallel by independent services and returned to the optimizer through a high-speed network card. The online side is composed of a load balancer, a gating service and a strategy service: the gating service maintains historical strategy distribution in memory, calculates differences online and controls the release amount; the strategy service loads the final executable strategy from a read-only parameter library; the audit service writes the deployment record to a write-once-read-many storage, and completes signature and verification through a hardware security module to ensure that the record is tamper-proof.

[0103] The above is only an embodiment of the present application and is not used to limit the present application. The present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. An artificial intelligence-based medical market data classification and analysis method, characterized by, The system comprises the following steps: Obtain multi-source medical market data and perform entity resolution and coding, convert data into atomic events according to event types, generate a standardized event stream through time alignment and deduplication, build a timestamped event provenance knowledge graph based on the standardized event stream, and generate a sliding window event intensity sequence; Based on the event provenance knowledge graph and the sliding window event intensity sequence, perform text representation, graph representation, and structured representation fusion, and output hierarchical market classification and executable vectors under the time logic constraints of event calculus; Based on the fused representation, hierarchical market classification, and executable vector, establish a world model and a return model, and calculate the strategy value and risk indicators using offline evaluation; Within the feasible region determined by time logic and business constraints, build a robust optimization model, and obtain the final executable strategy under the conditional value at risk constraint and generate the deployment record corresponding to the final executable strategy.

2. The method of claim 1, wherein, Obtain multi-source medical market data including sales data, order data, inventory data, price data, rebate data, budget data, cashback data, policy texts, recruitment texts, academic minutes, meeting minutes, channel correspondence texts, customer service texts, social public opinion, and main data, perform entity resolution and coding, convert data into atomic events according to event types, generate a standardized event stream through time alignment and cross-source deduplication, build a timestamped event provenance knowledge graph based on the standardized event stream, and generate a sliding window event intensity sequence.

3. The method of claim 1, wherein, Text representation is generated by a pre-trained language model, graph representation is generated by a time-series message passing of a graph neural network, and structured representation is generated by feature embedding and a multilayer perceptron. The text representation, graph representation, and structured representation are fused through attention mechanisms and interactions to obtain a fused representation.

4. The method of claim 1, wherein, The time logic constraints of event calculus are obtained by calculating the time true value and the rule violation degree, and are jointly optimized with the supervised loss to output the hierarchical market classification and the executable vector.

5. The method of claim 1, wherein, The hierarchical market classification uses a hierarchical classifier that maintains the parent-child consistent relationship, and the executable vector includes multi-label probabilities of compliance, qualification, channel, inventory, and budget dimensions.

6. The method of claim 1, wherein, The offline evaluation uses a double-robust offline evaluation method to calculate the strategy value and risk indicators, the historical behavior strategy is estimated by a probability model, the value function is calculated by the world model and the return model, and the data distribution alignment weight is used for re-weighted evaluation.

7. The method of claim 1, wherein, The world model uses a recurrent neural network to establish state transition, the return model uses a multilayer perceptron to establish a return function, and stress tests are performed in the scenario library of policy scenarios, supply scenarios, public opinion scenarios, and competition scenarios to obtain scenario stability and risk indicators.

8. The method of claim 1, wherein, The robust optimization model defines an uncertain distribution family based on the probability distribution distance of ambiguity sets, and optimizes the expected return under the most unfavorable distribution of the uncertain distribution family, while setting constraints on the conditional value at risk of loss to limit tail risk.

9. The method of claim 1, wherein, The final executable strategy is controlled by a trust domain difference constraint to limit the change amplitude of the final executable strategy and the historical strategy, and is deployed in gray according to the categories of hierarchical market classification, and deployment records corresponding to the final executable strategy are generated.

10. An artificial intelligence-based medical market data classification and analysis system for implementing the artificial intelligence-based medical market data classification and analysis method according to any one of claims 1 to 9, characterized in that, The system comprises: An event modeling module is configured to acquire multi-source medical market data, perform entity resolution and coding, convert the data into atomic events according to event types, generate a standard event stream through time alignment and deduplication, construct a timestamped event provenance knowledge graph based on the standard event stream, and generate a sliding window event intensity sequence; A fusion determination module is configured to perform text representation, graph representation and structured representation fusion based on the event provenance knowledge graph and the sliding window event intensity sequence, and output a hierarchical market classification and an executable vector under the time logic constraint of event calculus; A simulation evaluation module is configured to establish a world model and a return model based on the fusion representation, the hierarchical market classification and the executable vector, and calculate a strategy value and a risk indicator by offline evaluation; A robust optimization module is configured to construct a robust optimization model in a feasible region determined by time logic and business constraints, obtain a final executable strategy under a conditional value at risk constraint, and generate a deployment record corresponding to the final executable strategy.

Images