A data analysis method and system integrating knowledge graphs and deep learning

Abstract

This invention relates to the field of data processing technology, and more particularly to a data analysis method integrating knowledge graphs and deep learning. The method includes: acquiring multi-modal data and historical market anomaly fluctuation information; constructing a dynamic pharmaceutical knowledge graph; identifying anomaly nodes and causal links; generating joint feature vectors; capturing the evolutionary trends of market indicators to obtain candidate market anomalies; determining implicit market signals; adjusting detection thresholds; and generating a market analysis report. This invention combines multi-modal data with a knowledge graph reflecting the relationships and fluctuations of pharmaceutical market entities, enabling the temporal changes in market dynamics to be fully expressed in a deep learning model. This effectively solves the problems of incomplete feature representation due to single data modalities, difficulty in adapting to complex market environments caused by dynamic market evolution, resulting in low data processing accuracy and slow response speed.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a data analysis method and system that integrates knowledge graphs and deep learning. Background Technology

[0002] With the continuous expansion of the pharmaceutical industry market and the increasing complexity of the market environment, the number of entities in the pharmaceutical industry chain is enormous and their relationships are intricate. The market situation is highly dynamic and changing. Traditional analysis methods based on a single dimension or limited data sources are no longer sufficient to meet the needs of accurate perception of market trends and risk warning. Therefore, how to deeply integrate multi-source market data, establish logical relationship models between various market entities, and achieve early market signal identification and accurate judgment has become a key challenge to improve the market competitiveness and risk management capabilities of pharmaceutical companies.

[0003] Chinese Patent Application Publication No. CN117196694A discloses a method and system for pharmaceutical market data analysis based on big data. The method includes: using data sources from medical institutions, pharmaceutical research centers, and patient self-reports, employing data crawling and cleaning algorithms to collect and clean data, generating a purified raw market dataset; using deep time series analysis to perform historical market data feature analysis, generating a historical market data feature report; using a Bayesian deep learning model combined with probabilistic programming to perform market trend prediction, generating a market trend prediction report; using real-time data stream analysis technology and IoT data integration to perform real-time market monitoring and inventory optimization, generating real-time market monitoring and regional inventory optimization strategies; using multi-level network analysis to study drug interaction networks, generating a drug combination therapy report; and using a hybrid model collaborative filtering algorithm to perform personalized drug recommendations, generating personalized drug recommendation schemes.

[0004] Therefore, the big data-based pharmaceutical market data analysis method has the following problems: the data dimensions processed by this method are limited, making it difficult to fully integrate multi-source heterogeneous information; the analysis model of this method lacks an adaptive adjustment mechanism, and its response to dynamic changes in the market is not sensitive enough; the single data modality of this method leads to incomplete feature representation and difficulty in adapting to the complex market environment caused by dynamic market evolution, resulting in low data processing accuracy and slow response speed. Summary of the Invention

[0005] To address this, the present invention provides a data analysis method and system that integrates knowledge graphs and deep learning. This method overcomes the problems of incomplete feature representation and low data processing accuracy and slow response speed in the prior art due to the single data modality, which makes it difficult to adapt to the complex market environment caused by dynamic market evolution. This is achieved through multi-modal data fusion, dynamic medical knowledge graph construction, and adaptive judgment.

[0006] To achieve the above objectives, on the one hand, the present invention provides a data analysis method that integrates knowledge graphs and deep learning, comprising:

[0007] Real-time acquisition of multi-modal datasets and historical abnormal fluctuation information of the pharmaceutical market, which are obtained by fusing basic drug attribute data, end-user behavior data, macro market indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market.

[0008] A first knowledge graph is constructed based on the logical relationship between the source data and the preset historical market pattern.

[0009] Based on the historical abnormal fluctuation information, several abnormal nodes and causal links of the abnormal nodes in the first knowledge graph are determined, and confidence weights and timestamps are assigned to each relation chain in the causal links to form a dynamic medical knowledge graph with temporal semantics.

[0010] The joint feature vector is determined based on the multi-modal dataset, the dynamic medical knowledge graph, and the preset graph neural network.

[0011] The evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within the preset monitoring period is captured by the preset time sequence neural hybrid model, and the relationship between the source data is constructed according to the evolution trend to obtain a candidate anomaly list;

[0012] The anomaly type of the latent anomaly is determined based on the occurrence probability of latent anomalies in the candidate anomaly list, the preset detection threshold, and the historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph, so as to obtain the latent judgment result.

[0013] The detection threshold is adjusted according to the temporal distribution characteristics of the implicit determination results within the preset correction time.

[0014] An analysis report is generated based on the dynamic medical knowledge graph and the implicit judgment result obtained again after adjusting the detection threshold.

[0015] Furthermore, the process of determining the joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph, and the preset graph neural network includes:

[0016] The multimodal dataset is subjected to feature extraction according to a preset feature extraction algorithm to generate temporal feature representations for each modality;

[0017] The dynamic medical knowledge graph is input into a preset graph neural network to obtain a structured graph embedding representation;

[0018] The temporal feature representation and the graph embedding representation are fused to obtain the joint feature vector.

[0019] Furthermore, the process of constructing the first knowledge graph based on the logical relationship structure of each source data and the preset historical market pattern includes:

[0020] Construct market entity nodes and their logical relationship edges based on the types of pharmaceutical market entities, the business relationships between entities, and the market influence relationships;

[0021] Based on the preset historical market model, determine the feature nodes related to market events and their corresponding market impact relationship edges;

[0022] The logical relationship edges and the market influence relationship edges are combined to form the first knowledge graph used to describe the characteristics of market entity associations and event influence.

[0023] Furthermore, the process of fusing basic drug attribute data, end-user behavior data, macroeconomic market indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market to obtain a multi-modal dataset includes:

[0024] The basic attribute data of the drug, the terminal usage behavior data, the market macro indicator data, the medical academic literature data and the public opinion data are collected in real time from each of the source data, and the collection results are processed by time synchronization, normalization and outlier removal to obtain preprocessed data.

[0025] The preprocessed data is then subjected to feature alignment and modality fusion according to preset rules to obtain the multimodal dataset.

[0026] Furthermore, the process of determining several abnormal nodes and their causal links in the first knowledge graph based on the historical abnormal fluctuation information includes:

[0027] Based on the historical abnormal fluctuation information, abnormal features related to abnormal market events are identified, and each abnormal feature is constructed as an abnormal node.

[0028] Based on the sequential relationship, logical dependency relationship, and market fluctuation evolution law among the abnormal features in the historical abnormal fluctuation information, the causal links between the abnormal nodes are determined.

[0029] The causal links are associated with the first knowledge graph as relational chains between the abnormal nodes, thereby determining a number of abnormal nodes and their causal links in the first knowledge graph.

[0030] Furthermore, the process of assigning confidence weights and timestamps to each relation chain in the causal chain to construct a dynamic medical knowledge graph with temporal semantics includes:

[0031] Based on the frequency of occurrence, co-occurrence relationship, and consistency of fault evolution of each market anomaly in the historical abnormal fluctuation information, the confidence weight corresponding to each relationship chain is determined.

[0032] The timestamps of each relationship chain are determined based on the occurrence time of each market anomaly, the triggering order of the anomaly characteristics, and the duration of the causal link of the anomaly node.

[0033] The confidence weight and the timestamp are respectively assigned to the corresponding relationship chain to form a dynamic medical knowledge graph with temporal semantics.

[0034] Furthermore, the process of capturing the evolution trend of the drug's basic attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period using a preset time-series neural hybrid model, and constructing the relationship influence between the source data based on the evolution trend to obtain a candidate anomaly list includes:

[0035] Based on the temporal changes of the joint feature vector within the preset monitoring period, extract the time-series features corresponding to the basic drug attribute data and the terminal usage behavior data;

[0036] The time series features are input into the time series feature processing unit of the preset time series neural hybrid model to obtain the evolution trend of the basic attribute data of the drug and the terminal usage behavior data over time.

[0037] The evolutionary trend is input into the graph structure processing unit of the preset time-series neural hybrid model to combine the logical relationships between the entities in the pharmaceutical market, construct the relationship influence between the source data, and obtain the candidate anomaly list.

[0038] Furthermore, the process of determining the type of latent anomaly based on the occurrence probability of latent anomalies in the candidate anomaly list, a preset detection threshold, and historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph, in order to obtain the latent determination result, includes:

[0039] The probability of occurrence of latent anomalies in the candidate anomaly list is compared with the preset detection threshold to filter out anomaly candidates corresponding to suspected latent anomalies.

[0040] The abnormal candidates are matched with historical abnormal feature templates in the dynamic medical knowledge graph to obtain template matching results for each abnormal candidate.

[0041] Based on the path association relationship of the abnormal candidates in the causal path rules, the causal consistency of each abnormal candidate is determined;

[0042] Based on the template matching results and the causal consistency, the anomaly type of the corresponding latent anomaly is determined to obtain the latent judgment result.

[0043] Furthermore, the process of adjusting the detection threshold based on the temporal distribution characteristics of the implicit determination results within a preset correction period includes:

[0044] Within the preset correction time, the occurrence time, frequency, and duration of the implicit judgment result are statistically analyzed to obtain the temporal distribution characteristics of the implicit judgment result.

[0045] When the deviation between the time distribution feature and the preset detection threshold is greater than the preset deviation threshold, the preset detection threshold is adjusted according to the degree of deviation.

[0046] On the other hand, the present invention also provides a data analysis system that integrates knowledge graphs and deep learning, comprising:

[0047] The acquisition module is used to acquire in real time a multi-modal dataset and historical abnormal fluctuation information of the pharmaceutical market, which is obtained by fusing basic drug attribute data, end-user behavior data, macro market indicator data, medical academic literature data and public opinion data from various sources of pharmaceutical market data.

[0048] A construction module, which is connected to the acquisition module, is used to construct a first knowledge graph based on the logical association structure of each source data and a preset historical market pattern;

[0049] The constitutive module is connected to the construction module and the acquisition module respectively, and is used to determine a number of abnormal nodes and causal links of abnormal nodes in the first knowledge graph based on the historical abnormal fluctuation information, and to assign confidence weights and timestamps to each relation chain in the causal links to form a dynamic medical knowledge graph with temporal semantics.

[0050] A vector determination module, which is connected to the composition module and the acquisition module respectively, is used to determine a joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph and the preset graph neural network.

[0051] A candidate anomaly determination module, connected to the vector determination module, is used to capture the evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period according to a preset time-series neural hybrid model, and to construct the relationship influence between the source data according to the evolution trend to obtain a candidate anomaly list.

[0052] The latent determination module is connected to the candidate anomaly determination module and the constituent module respectively, and is used to determine the anomaly type of the latent anomaly based on the occurrence probability of the latent anomaly in the candidate anomaly list, the preset detection threshold, and the historical anomaly feature template and causal path rules in the dynamic medical knowledge graph, so as to obtain the latent determination result.

[0053] An adjustment module, which is connected to the latent determination module, is used to adjust the detection threshold according to the time distribution characteristics of the latent determination results within a preset correction time.

[0054] A generation module, which is connected to the constituent module and the implicit determination module respectively, is used to generate an analysis report based on the dynamic medical knowledge graph and the implicit determination result obtained again after adjusting the detection threshold.

[0055] Compared with existing technologies, the beneficial effects of this invention lie in combining multi-modal data, consisting of basic drug attribute data, end-user behavior data, macro-market indicator data, medical academic literature data, and public opinion data, with a knowledge graph reflecting the relationships between pharmaceutical market entities and the evolution of market fluctuations. This allows the temporal changes in dynamic market behavior to be fully expressed in a deep learning model, and the causal relationships between abnormal market events to be clearly characterized in the knowledge graph. Therefore, on the one hand, it can improve the sensitivity of market anomaly identification by utilizing the correlation patterns between multi-source data, enabling information from different dimensions such as sales fluctuations, price changes, market share adjustments, academic attention, and public opinion heat to form mutually corroborating feature structures in a unified feature space, thereby enhancing the ability to detect implicit market signals. On the other hand, by introducing causal links including confidence weights and timestamps, the chronological order, persistence, and influence relationships between market entities of abnormal market features can be quantitatively expressed, thereby improving the interpretability of market anomaly type determination. Furthermore, by analyzing the distribution of implicit determination results over a time period, this invention achieves adaptive adjustment of the detection threshold, allowing the determination strategy to be automatically corrected according to the long-term trend of the market environment, significantly reducing the risk of false positives and false negatives. Ultimately, it can accurately identify and stably output hidden market anomalies in complex market environments, effectively solving the problems of low data processing accuracy and slow response speed caused by incomplete feature representation due to the single data modality and difficulty in adapting to the dynamic evolution of the market in complex market environments.

[0056] Furthermore, by combining multi-modal data—comprising basic drug attribute data, end-user behavior data, macroeconomic market indicator data, medical academic literature data, and public opinion data—with a knowledge graph reflecting the relationships and fluctuations in the pharmaceutical market, the temporal characteristics of market dynamics are fully expressed in the deep learning model, and the causal relationships between market anomalies are clearly characterized in the knowledge graph. Thus, on the one hand, the sensitivity of market anomaly identification can be improved by utilizing the correlation patterns between multi-source data, allowing information from different dimensions such as sales trend changes, price fluctuations, market share adjustments, evolution of academic attention, and migration of public opinion heat to form mutually corroborating feature structures in a unified feature space, thereby enhancing the ability to detect latent market anomalies. On the other hand, by introducing causal links including confidence weights and timestamps, the chronological order, persistence, and transmission relationships between market anomaly features can be quantitatively expressed, thereby improving the interpretability of market anomaly type determination. In addition, this invention analyzes the distribution of latent determination results over a time period to achieve adaptive adjustment of the detection threshold, enabling the determination strategy to automatically correct itself according to the long-term trend of the market environment, significantly reducing the risk of false positives and false negatives. Ultimately, it can accurately identify and stably output hidden market anomalies in complex market environments, effectively solving the problems of low data processing accuracy and slow response speed caused by incomplete feature representation due to the single data modality and difficulty in adapting to the dynamic evolution of the market in complex market environments.

[0057] Furthermore, by mapping the types of pharmaceutical market entities, their business relationships, and market influence relationships to logical nodes and edges, and combining this with anomaly feature nodes generated from historical market fluctuation patterns and their market influence relationship edges, a first knowledge graph is formed. This graph comprehensively describes the relationships between various market entities at the business level and in terms of market influence. At the same time, historical market fluctuation patterns are integrated into the graph structure, enabling systematic modeling of the mutual influence between market entities, fluctuation transmission paths, and market anomaly triggering conditions. This provides a scientific and quantifiable foundation for subsequent joint feature vector construction and implicit market anomaly identification, enabling more accurate analysis and prediction of market dynamics.

[0058] Furthermore, by collecting various types of data from different sources in the pharmaceutical market in real time, including basic drug attribute data, end-user behavior data, macroeconomic market indicator data, medical academic literature data, and public opinion data, and performing time synchronization, normalization, and outlier removal on these data, the system achieves the unification and cleaning of data from different modalities. Based on this, feature alignment and modality fusion are used to form a multi-modal dataset, which fully preserves and quantifies the changing trends and mutual influences among various data types. This dataset can intuitively reflect market dynamics and abnormal fluctuation paths, thus providing rich and highly correlated information for subsequent graph neural network processing and implicit market signal identification, improving the accuracy and reliability of market anomaly detection.

[0059] Furthermore, by analyzing historical abnormal fluctuations in the pharmaceutical market, anomalous features closely related to market anomalies can be identified and constructed as anomalous nodes. By further combining the temporal sequence and logical dependencies of these anomalous features with the evolutionary patterns of market fluctuations, causal links between these anomalous nodes are established and integrated into a first knowledge graph. This method reveals the intrinsic connections and evolutionary paths between various data features and events during market fluctuations, clearly representing anomalous nodes and their interactions within the graph. This provides a reliable structured foundation for subsequent joint feature vector generation, analysis of market anomaly evolution trends, and identification of implicit market signals, enhancing a comprehensive understanding and accurate prediction of market dynamics.

[0060] Furthermore, by assigning confidence weights and timestamps to each relationship chain, the influence intensity and evolutionary order between different abnormal nodes can be intuitively reflected. This allows the dynamic pharmaceutical knowledge graph to not only present the business relationships between market entities but also dynamically depict the development trend of abnormal market events over time. Confidence weights comprehensively consider the frequency of occurrence, co-occurrence relationships, and evolutionary consistency of historical abnormal market events, accurately distinguishing between high-risk and low-risk associations. Timestamps record the triggering order and duration of abnormalities, revealing the sequential relationship and potential transmission path of events within the market cycle. This enables subsequent graph neural network-based analysis to simultaneously utilize structural and temporal evolution information, achieving more accurate judgment and prediction of market dynamics and implicit market signals.

[0061] Furthermore, by extracting and modeling the temporal changes of basic drug attribute data and end-user behavior data in the joint feature vector within a preset monitoring period, it is possible to capture the evolution of the status of each market entity over time. By combining the business relationships between market entities to construct the influence of relationships, potential market anomaly patterns can be accurately identified, achieving high-precision screening of candidate market anomalies. It fully considers the dynamic changes, interactions, and evolution trends between different data dimensions, enabling the market anomaly judgment results to reflect the inherent correlation and causal logic of the pharmaceutical market operation, and improving the timeliness and reliability of market risk prediction.

[0062] Furthermore, by comparing the probability of occurrence of latent market signals in the candidate anomaly list with thresholds, matching historical anomaly feature templates, and analyzing the consistency of causal paths, potential market anomaly types can be accurately identified, enabling precise judgment of market anomaly states. This allows for the organic integration of dynamic characteristics of market data with historical market fluctuation patterns and causal relationships between market entities, revealing potential anomaly transmission paths and mutual influence patterns among various market entities. Consequently, the reliability and accuracy of latent market signal identification are improved in complex market environments, providing a scientific basis for subsequent threshold adjustments and risk warnings.

[0063] Furthermore, by statistically analyzing the temporal distribution characteristics of implicit judgment results, such as their occurrence time, frequency, and persistence, and using deviation as a trigger condition, the detection threshold is adaptively adjusted. This ensures that the threshold adjustment is based on long-term trends rather than instantaneous fluctuations, thereby distinguishing between short-term market fluctuations and persistent anomalies. This strategy utilizes the business logic and statistical correlations between market data. For example, a continuous decline in market share accompanied by negative policy sentiment is more likely to point to real market risks, while isolated short-term promotional activities leading to sales spikes are mostly normal fluctuations. This approach amplifies or suppresses risk at the threshold level. The result is improved sensitivity to persistent and transmissive implicit market anomalies, while reducing false positive rates and judgment instability caused by frequent threshold fluctuations. Short-term verification before and after threshold updates also suppresses over-adjustment, achieving a balance between robustness and response speed. This improves the detection accuracy of implicit market signals and provides more reliable and interpretable early warning evidence for market strategy decisions.

[0064] Furthermore, through real-time acquisition and processing of multi-modal data, the system organically integrates basic drug attribute data, end-user behavior data, market macro-indicator data, medical academic literature data, and public opinion data. It also constructs a knowledge graph by combining historical abnormal fluctuation information with market entity relationships, achieving a global perception of market dynamics. By assigning confidence weights and time stamps to the causal links of abnormal nodes in the knowledge graph, and combining graph neural networks and time-series graph neural hybrid models to analyze joint feature vectors, the system can capture the evolutionary trends of basic drug attribute data and end-user behavior data over time, thereby revealing the mutual influence relationships between various market entities. Historical abnormal templates and causal path rules are used to implicitly determine candidate anomalies, and the detection threshold is dynamically adjusted based on time distribution characteristics, making market anomaly identification more accurate and robust. Overall, the system can achieve early identification and accurate classification of market anomalies through the linkage relationships of multi-dimensional data such as sales trends, price fluctuations, market share, academic attention, and public opinion heat, providing a scientific and reliable basis for pharmaceutical companies' strategic decisions. At the same time, it fully reflects the business connections and dynamic evolution patterns between various data dimensions, improving the accuracy and response speed of market anomaly detection. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the data analysis method that integrates knowledge graphs and deep learning in this embodiment;

[0066] Figure 2 A flowchart for determining the joint feature vector in this embodiment;

[0067] Figure 3 A flowchart for constructing the first knowledge graph in this embodiment;

[0068] Figure 4 This is a schematic diagram of the data analysis system that integrates knowledge graphs and deep learning in this embodiment. Detailed Implementation

[0069] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0070] Please see Figure 1 The diagram shows a flowchart of the data analysis method integrating knowledge graphs and deep learning in this embodiment. On one hand, this embodiment provides a data analysis method integrating knowledge graphs and deep learning, including: real-time acquisition of a multi-modal dataset and historical abnormal fluctuation information of the pharmaceutical market, obtained by fusing basic drug attribute data, terminal usage behavior data, market macroeconomic indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market; constructing a first knowledge graph based on the logical association structure of each source data and a preset historical market pattern; determining several abnormal nodes and causal links of the abnormal nodes in the first knowledge graph based on the historical abnormal fluctuation information, and assigning confidence weights and timestamps to each relation chain in the causal links to form a dynamic pharmaceutical knowledge graph with temporal semantics; and based on the multi-modal dataset... The system uses a dataset, the dynamic medical knowledge graph, and a preset graph neural network to determine a joint feature vector. A preset temporal graph neural hybrid model is used to capture the evolution trend of the basic drug attribute data and the terminal usage behavior data within the joint feature vector over a preset monitoring period. Based on this evolution trend, the system constructs the relationship between the various source data to obtain a candidate anomaly list. The system determines the anomaly type of the latent anomaly based on the occurrence probability of latent anomalies in the candidate anomaly list, a preset detection threshold, and historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph, thus obtaining a latent judgment result. The detection threshold is adjusted based on the temporal distribution characteristics of the latent judgment result within a preset correction period. An analysis report is generated based on the dynamic medical knowledge graph and the latent judgment result obtained after adjusting the detection threshold.

[0071] In this embodiment, the method is applied to a pharmaceutical market monitoring and analysis scenario. The pharmaceutical market entities include pharmaceuticals, medical institutions, sales channels, patient groups, competing companies, and other pharmaceutical industry entities undertaking R&D, production, distribution, or consumption functions. The basic attribute data of pharmaceuticals includes attribute data reflecting the basic characteristics of pharmaceuticals, such as generic name, brand name, dosage form and specifications, indications, pharmacological classification, and inclusion in the medical insurance catalog. The terminal behavior indicator data includes behavioral data obtained through the terminal monitoring system, such as prescription volume, dosage, purchase frequency, inventory level, and market share. The macroeconomic market indicator data includes market size, growth rate, policy impact index, and industry prosperity, such as market environment characteristics obtained from macroeconomic monitoring devices. The medical academic literature data includes research characteristics obtained through academic databases, such as clinical trial results, number of academic papers published, expert consensus, and guideline updates. The public opinion data includes public perception characteristics obtained through the public opinion monitoring system, such as media coverage intensity, social media discussion volume, patient evaluation sentiment, and policy interpretation tendency. The historical abnormal fluctuation information consists of the occurrence time, scope of impact, triggering factors, market reaction, and countermeasures of major events recorded in the market monitoring system, used to characterize the types and evolution processes of market anomalies. The aforementioned data were acquired in real time by accessing various pharmaceutical databases, sales monitoring systems, policy release platforms, academic resource databases, and public opinion monitoring platforms. After time synchronization processing, the data was used to form the multi-modal dataset and historical abnormal fluctuation information for subsequent analysis.

[0072] In this embodiment, the preset historical market pattern is a combination of market anomaly features pre-organized based on typical fluctuations, triggering conditions, and evolutionary patterns of various entities in the pharmaceutical market during long-term operation. This combination is used to characterize the common manifestations and internal relationships of different types of market anomalies. The preset historical market pattern includes anomaly features extracted from historical market data, industry reports, records of major events, market research results, and expert judgments. Examples include: rapid decline in market share, price adjustments, and competitor encroachment corresponding to the patent cliff; changes in sales structure, channel restructuring, and price restructuring corresponding to the impact of volume-based procurement policies; and reshaping of the market landscape, price system adjustments, and changes in physician prescribing behavior corresponding to the impact of new product launches. Each preset historical market pattern includes pre-trigger features, dominant features, and accompanying features, as well as the sequence or logical dependency between these features, used to map structured pattern information of market anomalies when constructing a knowledge graph.

[0073] In this embodiment, the preset graph neural network specifically adopts a graph attention network (GAT) structure. This network aggregates the attribute features and adjacent relationships of each node in the dynamic medical knowledge graph by introducing a neighborhood feature weighting method based on an attention mechanism. The graph attention network includes several graph attention layers. Each graph attention layer calculates attention weights based on the relationship features between a node and its neighboring nodes, and uses these weights to perform a weighted summation of neighborhood features to obtain the updated feature representation of the node. Through multi-layer feature propagation, graph embedding vectors reflecting market entity associations, abnormal pattern propagation, and the influence range of key nodes can be generated. In this embodiment, the dynamic pharmaceutical knowledge graph, after being trained on a preset training set, is processed by a graph attention network to output high-dimensional embedding features of each node, which are then used for subsequent fusion with multi-modal data time-series features. The preset time-series graph neural hybrid model is specifically composed of a Long Short-Term Memory (LSTM) network and a Graph Attention Network (GAT). LSTM is used to model the time-series features of basic drug attribute data and terminal usage behavior data within a preset monitoring period. By inputting multi-moment data of parameters and capturing their short-term fluctuations and long-term trends in state units, it generates a time-series feature representation that reflects the evolution of market parameters. GAT is used to perform graph propagation processing on the time-series features combined with market relationships in the dynamic pharmaceutical knowledge graph, and uses an attention mechanism to weightedly fuse the business impact between market entities. Through the cascaded processing of LSTM and GAT, the hybrid model can simultaneously express the time-varying trends of market parameters and the interrelationships between market entities, thereby generating a hybrid representation that comprehensively reflects dynamic market behavior and business relationships, which is used to output the candidate anomaly list.

[0074] In this embodiment, the generated analysis report involves, after adaptively adjusting the detection threshold, performing correlation analysis between the newly obtained implicit judgment results under the updated threshold and the market relationships, historical market patterns, and causal links in the dynamic pharmaceutical knowledge graph. This analysis aims to generate a diagnostic result that reflects the current market state and potential risks. Specifically, the system generates analysis content including the source of the anomaly, key influencing entities, possible direction of fluctuation transmission, and risk level, based on the anomaly type corresponding to the implicit judgment result, the temporal characteristics of the anomaly's occurrence, and the related causal path. It also combines market dependencies in the knowledge graph to mark the potential impact range of the anomaly on adjacent entities or key market nodes. Furthermore, the report includes changes in anomaly-related feature values ​​in multi-modal data and their explanations to assist decision-makers in understanding the causes of the anomaly. Finally, the system automatically generates an analysis report containing anomaly type determination, risk explanation, correlation path description, and suggested countermeasures for use in market strategy formulation and risk management.

[0075] The preset monitoring period is a fixed time length used to statistically analyze the time-series changes of market parameters. It depends on the typical periodicity of market parameter fluctuations and the speed of environmental changes, and is usually set between 1 week and 1 month. In this embodiment, it is set to 2 weeks to ensure the capture of stable trends in market parameter changes and reduce short-term noise interference. The preset detection threshold is a probability threshold used to determine whether the probability of the occurrence of hidden market anomalies meets the anomaly judgment conditions. It depends on the statistical characteristics of parameter distribution during normal market operation and the allowable false alarm rate, and is usually set between 0.6 and 0.9. In this embodiment, it is set to 0.75 to effectively reduce the risk of misjudgment while ensuring sensitivity. The preset correction duration is the time interval used to statistically analyze the time distribution of hidden judgment results and trigger adaptive threshold adjustment. It depends on the stability of the market environment and the response speed of anomaly evolution, and is usually set between 1 month and 1 quarter. In this embodiment, it is set to 6 weeks to ensure that the threshold adjustment is based on sufficient trend information rather than immediate fluctuations.

[0076] By combining multi-modal data—comprising basic drug attribute data, end-user behavior data, macro-market indicator data, medical academic literature data, and public opinion data—with a knowledge graph reflecting the relationships between pharmaceutical market entities and the evolution of market fluctuations, the temporal changes in dynamic market behavior are fully expressed in the deep learning model, and the causal relationships between market anomalies are clearly characterized in the knowledge graph. This allows for two main advantages: firstly, it enhances the sensitivity of market anomaly identification by leveraging the correlation patterns between multi-source data, enabling information from different dimensions such as sales fluctuations, price changes, market share adjustments, academic attention, and public opinion heat to form mutually corroborating feature structures within a unified feature space, thereby strengthening the detection capability of implicit market signals; secondly, by introducing causal links including confidence weights and timestamps, the chronological order, persistence, and influence relationships between market entities of anomaly characteristics are quantified, thus improving the interpretability of market anomaly type determination. Furthermore, by analyzing the distribution of implicit determination results over a time period, this invention achieves adaptive adjustment of the detection threshold, allowing the determination strategy to automatically correct itself according to the long-term trend of the market environment, significantly reducing the risk of false positives and false negatives. Ultimately, it can accurately identify and stably output hidden market anomalies in complex market environments, effectively solving the problems of low data processing accuracy and slow response speed caused by incomplete feature representation due to the single data modality and difficulty in adapting to the dynamic evolution of the market in complex market environments.

[0077] Please see Figure 2The flowchart shown is for determining the joint feature vector in this embodiment. In this embodiment, the process of determining the joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph, and the preset graph neural network includes: extracting features from the multi-modal dataset according to a preset feature extraction algorithm to generate temporal feature representations for each modality; inputting the dynamic medical knowledge graph into the preset graph neural network to obtain a structured graph embedding representation; and fusing the temporal feature representation and the graph embedding representation to obtain the joint feature vector.

[0078] In this embodiment, the preset feature extraction algorithm specifically includes: extracting time-series features from basic drug attribute data and end-user behavior data using wavelet transform and autoregressive models; obtaining trend and correlation features from macroeconomic market indicator data using principal component analysis and Granger causality tests; extracting research hotspots and academic attention features from medical academic literature data using topic modeling and sentiment analysis; extracting public opinion trends and dissemination influence features from public opinion data using sentiment analysis and propagation dynamics models; and performing feature mapping and multi-layer convolutional pooling operations on visual content in policy and public opinion data using convolutional neural networks to obtain a unified time-series feature representation for subsequent joint feature vector construction.

[0079] By combining multi-modal data—comprising basic drug attribute data, end-user behavior data, macro-market indicator data, medical academic literature data, and public opinion data—with a knowledge graph reflecting the relationships and fluctuations in the pharmaceutical market, the temporal characteristics of market dynamics are fully expressed in the deep learning model, and the causal relationships between market anomalies are clearly characterized in the knowledge graph. This allows for two main advantages: firstly, it enhances the sensitivity of market anomaly identification by leveraging the correlation patterns between multi-source data, enabling information from different dimensions such as sales trend changes, price fluctuations, market share adjustments, evolution of academic attention, and shifts in public opinion heat to form mutually corroborating feature structures within a unified feature space, thereby strengthening the detection capability of latent market anomalies; secondly, by introducing causal links including confidence weights and timestamps, it quantifies the chronological order, persistence, and transmission relationships between market entities related to market anomaly features, thus improving the interpretability of market anomaly type determination. Furthermore, by analyzing the distribution of latent determination results over a time period, this invention achieves adaptive adjustment of the detection threshold, allowing the determination strategy to automatically correct itself according to long-term market trends, significantly reducing the risk of false positives and false negatives. Ultimately, it can accurately identify and stably output hidden market anomalies in complex market environments, effectively solving the problems of low data processing accuracy and slow response speed caused by incomplete feature representation due to the single data modality and difficulty in adapting to the dynamic evolution of the market in complex market environments.

[0080] Please see Figure 3The diagram shows a flowchart of constructing the first knowledge graph in this embodiment. In this embodiment, the process of constructing the first knowledge graph based on the logical association structure of each source data and the preset historical market pattern includes: constructing market entity nodes and their logical relationship edges based on the type of the pharmaceutical market entity, the business relationship between each entity, and the market influence relationship; determining feature nodes related to market events and their corresponding market influence relationship edges based on the preset historical market pattern; and combining the logical relationship edges and the market influence relationship edges to form the first knowledge graph used to describe the characteristics of market entity association and event influence.

[0081] In this embodiment, each market entity is first mapped to a node based on its business type, including pharmaceuticals, manufacturers, distributors, medical institutions, and patient groups. The business relationships between these entities are then identified, such as the supply relationship between pharmaceuticals and manufacturers, the treatment relationship between pharmaceuticals and their indications, the competitive relationship between different pharmaceuticals, and the influence dependency relationship between market entities. For example, a price adjustment of a pharmaceutical product might affect the market share of competing products, or the release of a medical insurance policy might affect the sales structure of related pharmaceuticals. This information is used to construct logical relationship edges between nodes in the graph, fully representing the business connections and influence dependency network between market entities. Then, based on a preset historical market pattern, various abnormal fluctuations that have occurred in the market's historical operation are analyzed, such as patent cliffs, the impact of volume-based procurement, the impact of new product launches, and the fermentation of negative public opinion. These abnormally related market features are then constructed as abnormal nodes. Simultaneously, based on market fluctuation patterns, the sequential triggering relationships and causal connections between abnormal events are analyzed. Market influence relationship edges are established between these abnormal nodes to describe the transmission path and potential chain reactions of abnormal fluctuations. Finally, the logical relationship edges between market entities and the market influence relationship edges are combined into the same graph structure, forming a first knowledge graph that includes both business-level information of market entities and historical market anomaly correlation characteristics. In this graph, each market entity node carries its business attributes and market state information, and is also connected to abnormal nodes through causal links, forming a complete knowledge representation with business hierarchy and causal connections.

[0082] By mapping the types of pharmaceutical market entities, their business relationships, and market influence relationships to logical nodes and edges, and combining this with anomaly feature nodes generated from historical market fluctuation patterns and their market influence relationship edges, a first knowledge graph is formed. This graph comprehensively describes the relationships between market entities at the business level and in terms of market influence. At the same time, historical market fluctuation patterns are integrated into the graph structure, enabling systematic modeling of the mutual influence between market entities, fluctuation transmission paths, and market anomaly triggering conditions. This provides a scientific and quantifiable foundation for subsequent joint feature vector construction and implicit market anomaly identification, enabling more accurate analysis and prediction of market dynamics.

[0083] Specifically, the process of fusing basic drug attribute data, end-user behavior data, market macro-indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market to obtain a multi-modal dataset includes: real-time collection of the basic drug attribute data, end-user behavior data, market macro-indicator data, medical academic literature data, and public opinion data from each of the aforementioned source data, and performing time synchronization processing, normalization processing, and outlier removal processing on the collection results to obtain preprocessed data; and performing feature alignment and modality fusion on the preprocessed data according to preset rules to obtain the multi-modal dataset.

[0084] In this embodiment, various data sources in the pharmaceutical market (including but not limited to drug attribute databases, sales and transaction systems, market monitoring platforms, academic literature databases, and public opinion monitoring systems) are collected in real time during the data update process. These data include: basic drug attribute data collection: collecting generic names, brand names, dosage forms, specifications, indications, pharmacological classifications, etc., for each drug, obtained through database queries or API calls using data interfaces; end-user behavior data collection: collecting sales revenue, sales volume, sales channel distribution, price fluctuations, etc., for each drug, collected periodically using a sales data monitoring system; macro-market indicator data collection: collecting market size, growth rate, policy impact index, industry prosperity, etc., obtained through market research institutions and statistical platforms; medical academic literature data collection: collecting clinical trial results, number of academic papers published, expert consensus updates, etc., obtained through retrieval from academic databases and knowledge bases; and public opinion data collection: collecting media coverage popularity, social media discussion volume, patient evaluation sentiment, etc., collected in real time through a public opinion monitoring system. The collected data undergoes initial time synchronization to ensure alignment across different sources at the same time granularity. Normalization is then implemented to eliminate dimensional differences, ensuring all data indicators fall within a uniform numerical range. Simultaneously, outlier removal is performed to eliminate data biases introduced by acquisition errors or system malfunctions, resulting in high-quality preprocessed data. Subsequently, the preprocessed data is processed according to predefined feature alignment rules and modality fusion strategies: key features of drug attributes, sales transactions, market indicators, academic literature, and public opinion data are time-aligned and mapped to corresponding business dimensions, then fused to generate a unified multi-modal dataset. This preserves the correlations and interactions between various data types, forming a joint feature representation that can be directly used as input for subsequent graph neural networks, facilitating market dynamics analysis, anomaly detection, and the identification of implicit market signals.

[0085] By collecting various types of data from different sources in the pharmaceutical market in real time, including basic drug attribute data, end-user behavior data, macroeconomic market indicator data, medical academic literature data, and public opinion data, and performing time synchronization, normalization, and outlier removal on these data, the system achieves the unification and cleaning of data from different modalities. Based on this, feature alignment and modality fusion are used to form a multi-modal dataset, which fully preserves and quantifies the changing trends and mutual influences among various data types. This dataset can intuitively reflect market dynamics and abnormal fluctuation paths, providing rich and highly correlated information for subsequent graph neural network processing and implicit market signal identification, thereby improving the accuracy and reliability of market anomaly detection.

[0086] Specifically, the process of determining several abnormal nodes and causal links of the abnormal nodes in the first knowledge graph based on the historical abnormal fluctuation information includes: identifying abnormal features related to abnormal market events based on the historical abnormal fluctuation information, and constructing each abnormal feature as an abnormal node; determining the causal links between the abnormal nodes based on the sequential relationship, logical dependency relationship, and market fluctuation evolution law among the abnormal features in the historical abnormal fluctuation information; and associating the causal links as relationship chains between the abnormal nodes to the first knowledge graph to determine several abnormal nodes and causal links of the abnormal nodes in the first knowledge graph.

[0087] In this embodiment, abnormal features related to abnormal market events are identified based on historical abnormal fluctuation information in the pharmaceutical market. Specifically, firstly, statistical analysis is performed on historical drug basic attribute data, end-user behavior data, market macroeconomic indicator data, medical academic literature data, and public opinion data from various data sources to calculate their mean, standard deviation, and abnormal threshold. When a data indicator exceeds a preset judgment threshold or displays an abnormal fluctuation pattern, it is marked as an abnormal feature. Each abnormal feature is constructed as an abnormal node in a knowledge graph, containing node attributes: indicator type, indicator value, abnormal timestamp, and historical market impact level. Subsequently, based on the chronological order, logical dependencies, and market fluctuation evolution patterns of historical abnormal market events, the causal relationships between the abnormal nodes are determined. The specific rules are as follows: If node A's anomaly precedes node B's anomaly, and the anomaly value of A has a statistical correlation with the anomaly event of B (correlation coefficient greater than a preset correlation threshold of 0.7), then a causal edge is established between A and B. Each causal edge is assigned a weight, calculated based on historical co-occurrence frequency and event evolution consistency, typically set between 0.1 and 1.0; in this embodiment, it is initially set to 0.8. A timestamp is appended to each causal edge to represent the time interval of the anomaly trigger, reflecting the duration of the anomaly evolution. Finally, the anomaly nodes and their causal links are integrated into the first knowledge graph. Nodes and edges combine to form a graph structure for subsequent graph neural network calculations and joint feature vector generation. This constructed knowledge graph not only retains market entity association information but also clarifies the causal relationships between anomaly features, enabling the prediction and analysis of implicit market risks.

[0088] By analyzing historical abnormal fluctuations in the pharmaceutical market, anomalous features closely related to market anomalies can be identified and constructed as anomalous nodes. Further, by combining the temporal sequence and logical dependencies of these anomalous features with the evolutionary patterns of market fluctuations, causal links between these anomalous nodes are established and integrated into a first knowledge graph. This method reveals the intrinsic connections and evolutionary paths between various data features and events during market fluctuations, clearly representing anomalous nodes and their interactions within the graph. This provides a reliable structured foundation for subsequent joint feature vector generation, analysis of market anomaly evolution trends, and identification of implicit market signals, enhancing a comprehensive understanding and accurate predictive capability of market dynamics.

[0089] Specifically, the process of assigning confidence weights and timestamps to each relation chain in the causal chain to construct a dynamic medical knowledge graph with temporal semantics includes: determining the confidence weights corresponding to each relation chain based on the frequency of occurrence, co-occurrence relationships, and consistency of fault evolution of each market anomaly event in the historical abnormal fluctuation information; determining the timestamps of each relation chain based on the occurrence time of each market anomaly event, the triggering order of the anomaly features, and the duration of the causal chain of the anomaly node; and assigning the confidence weights and timestamps to the corresponding relation chains to form a dynamic medical knowledge graph with temporal semantics.

[0090] In this embodiment, for each causal link, the frequency and co-occurrence relationships of relevant market anomalies in historical abnormal fluctuation information are first statistically analyzed, and the consistency of fault evolution is analyzed to calculate the confidence weight of each relationship chain. This weight reflects the strength and reliability of the association between anomaly nodes. Then, based on the specific occurrence time of each market anomaly event, the order in which anomaly features are triggered, and the duration of the causal link between anomaly nodes, the timestamp of each relationship chain is determined to identify the temporal evolution order of the events. Finally, the confidence weight and timestamp are assigned to the corresponding relationship chain, so that the dynamic medical knowledge graph not only presents the structural association between nodes, but also dynamically reflects the evolution of anomalies over time, providing accurate temporal and relational information for subsequent anomaly analysis and prediction based on graph neural networks.

[0091] By assigning confidence weights and timestamps to each relationship chain, the influence intensity and evolutionary order between different abnormal nodes can be intuitively reflected. This allows the dynamic pharmaceutical knowledge graph to not only present the business relationships between market entities but also dynamically depict the development trend of abnormal market events over time. Confidence weights comprehensively consider the frequency of occurrence, co-occurrence relationships, and evolutionary consistency of historical market abnormal events, accurately distinguishing between high-risk and low-risk associations. Timestamps record the triggering order and duration of abnormalities, revealing the sequential relationship and potential transmission path of events within the market cycle. This enables subsequent graph neural network-based analysis to simultaneously utilize structural and temporal evolution information, achieving more accurate judgment and prediction of market dynamics and implicit market signals.

[0092] Specifically, the process of capturing the evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period according to a preset time-series graph neural hybrid model, and constructing the relationship influence between the source data based on the evolution trend to obtain a candidate anomaly list includes: extracting time-series features corresponding to the basic drug attribute data and the terminal usage behavior data based on the temporal changes of the joint feature vector within the preset monitoring period; inputting the time-series features into the time-series feature processing unit of the preset time-series graph neural hybrid model to obtain the evolution trend reflecting the changes of the basic drug attribute data and the terminal usage behavior data over time; and inputting the evolution trend into the graph structure processing unit of the preset time-series graph neural hybrid model to combine the logical associations between the pharmaceutical market entities and construct the relationship influence between the source data to obtain the candidate anomaly list.

[0093] In this embodiment, firstly, based on the pre-fused joint feature vector, including basic drug attribute data, end-user behavior data, market macro-indicator data, medical academic literature data, and public opinion data of each market entity during operation, time-series data are acquired according to a preset monitoring period. Secondly, features are extracted from these time-series data, including mean, variance, peak value, frequency domain features, wavelet features, trend coefficients, etc., to form time-series feature representations corresponding to basic drug attribute data and end-user behavior data. The time-series feature representation can characterize the evolution trend of each data indicator over time, including information such as indicator fluctuation amplitude, upward or downward trend, and periodic changes. Then, the time-series features are input into the time-series feature processing unit of a preset time-series graph neural hybrid model. This unit combines recurrent neural networks (such as LSTM and GRU) or temporal convolutional networks (TCN) to learn temporal features, extract dynamic features that evolve over time, and capture the temporal correlation and potential coupling relationship between basic drug attribute data and end-user behavior data, thereby obtaining a trend representation reflecting the evolution of the state of each market entity. Next, the obtained evolutionary trend is input into a graph structure processing unit. This unit, based on a dynamic pharmaceutical knowledge graph, maps each market entity node and its relationships into a graph structure. Through GAT (Graph Attribution-Action Graph), message passing and feature aggregation are performed. Combined with business relationships between market entities (such as supply relationships, competitive relationships, and influence dependencies), a relationship influence matrix between market entities is constructed, revealing possible abnormal influence paths between market entities. For example, a market anomaly of a core drug may cause fluctuations in the market share of related competing products. Finally, a candidate anomaly list is generated based on the graph structure processing results, including potentially problematic market entity nodes and their anomaly types, anomaly intensity indicators, and impact ranges, providing a basis for subsequent implicit market signal determination and threshold adjustment. By using LSTM and GAT for joint modeling, we can simultaneously capture the evolution of market parameters over time and the impact of business interactions between market entities, thereby significantly improving the accuracy and reliability of market anomaly detection.

[0094] By extracting and modeling the temporal changes of basic drug attribute data and end-user behavior data in the joint feature vector within a preset monitoring period, it is possible to capture the evolution of the status of each market entity over time. Combined with the business relationships between market entities, it can construct the influence of relationships, thereby accurately identifying potential market anomaly patterns and achieving high-precision screening of candidate market anomalies. It fully considers the dynamic changes, interactions, and evolution trends between different data dimensions, so that the market anomaly judgment results can reflect the inherent correlation and causal logic of the operation of the pharmaceutical market, improving the timeliness and reliability of market risk prediction.

[0095] Specifically, the process of determining the anomaly type of a latent anomaly based on the occurrence probability of latent anomalies in the candidate anomaly list, a preset detection threshold, and historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph to obtain a latent judgment result includes: comparing the occurrence probability of latent anomalies in the candidate anomaly list with the preset detection threshold to filter out anomaly candidates corresponding to suspected latent anomalies; matching the anomaly candidates with historical anomaly feature templates in the dynamic medical knowledge graph to obtain template matching results for each anomaly candidate; determining the causal consistency of each anomaly candidate based on the path association relationship of the anomaly candidate in the causal path rules; and determining the anomaly type to which the corresponding latent anomaly belongs based on the template matching results and the causal consistency to obtain a latent judgment result.

[0096] In this embodiment, the occurrence probability of each anomaly in the candidate anomaly list is first compared with a preset detection threshold. The preset detection threshold is 0.75. Anomalies with an occurrence probability not lower than this threshold are screened as "suspected hidden market anomaly candidates" and enter the subsequent knowledge verification stage; candidates with an occurrence probability lower than this threshold are temporarily archived for model optimization and long-term trend analysis. For each suspected hidden market anomaly candidate, the system will initiate dual verification using the knowledge graph. The first verification is template matching: the system extracts the multi-dimensional feature vector of the candidate anomaly and calculates the similarity with the historical anomaly feature templates pre-stored in the dynamic pharmaceutical knowledge graph. The calculation uses the cosine similarity algorithm, and the template matching threshold is set to 0.8. Candidates with matching results not lower than this threshold are considered to conform to known market anomaly patterns. The second verification is causal consistency testing: the causal path related to the candidate anomaly is traced in the knowledge graph, and two judgments are made: first, the overall weight of the path is calculated, with a preset threshold of 0.7; second, whether the time sequence of the anomaly events conforms to the historical experience time window (this system presets an adjustable value between 1 day and 1 week). Candidates that simultaneously meet the path weight threshold and time window constraints are deemed to have "established" causal consistency. Finally, the system integrates the template matching and causal consistency results to execute an authoritative judgment: if a candidate anomaly simultaneously meets the requirements of a template matching degree ≥ 0.8 and causal consistency, it is confirmed as a latent market anomaly and its specific anomaly type is marked; if the template matching degree is high but causal consistency is not established, it is marked as an "anomaly to be observed," prompting market analysts to conduct focused investigation; if the template matching degree is low but causal consistency is established, it is marked as an "emerging potential risk," and the system will monitor it and learn new features; if neither condition is met, it is judged as a "false alarm" and excluded. The final latent judgment result will clearly list the confirmed anomalous market entities, anomaly types, judgment confidence levels, and associated causal paths, providing a direct and reliable basis for market decision-making. The template matching threshold is crucial for balancing judgment sensitivity and false alarm risk, and its value is typically set between 0.7 and 0.9. The optimal balance between high reliability in judgment and effective risk warning is achieved by maintaining high accuracy in identifying typical market anomalies while maintaining sufficient sensitivity to high-risk hidden market anomalies that closely match certain characteristics.

[0097] By comparing the probability of occurrence of latent market signals in the candidate anomaly list with thresholds, matching historical anomaly feature templates, and analyzing the consistency of causal paths, potential market anomaly types can be accurately identified, enabling precise judgment of market anomaly states. This allows for the organic integration of dynamic characteristics of market data with historical market fluctuation patterns and causal relationships between market entities, revealing potential anomaly transmission paths and mutual influence patterns among various market entities. Consequently, the reliability and accuracy of latent market signal identification are improved in complex market environments, and a scientific basis is provided for subsequent threshold adjustments and risk warnings.

[0098] Specifically, the process of adjusting the detection threshold based on the temporal distribution characteristics of the latent determination results within a preset correction period includes: statistically analyzing the occurrence time, frequency, and duration of the latent determination results within the preset correction period to obtain the temporal distribution characteristics of the latent determination results; and adjusting the preset detection threshold based on the degree of deviation when the deviation between the temporal distribution characteristics and the preset detection threshold is greater than a preset deviation threshold.

[0099] In this embodiment, the adaptive adjustment mechanism of the detection threshold is as follows: within a preset correction period of 6 weeks, the frequency and persistence of the occurrence of implicit judgment results are statistically analyzed to obtain time distribution characteristics; then, the deviation between this distribution characteristic and the current preset detection threshold (initial value is 0.75) is calculated. When the deviation exceeds the preset threshold by 20%, the detection threshold is dynamically adjusted according to the degree and direction of the deviation. If the deviation is positive and exceeds 20%, the detection threshold is adjusted upward by a certain amount (upper limit is 0.9) to reduce false alarms; if the deviation is negative and exceeds 20%, it is adjusted downward accordingly (lower limit is 0.6) to improve detection sensitivity, thereby ensuring that the system can always maintain optimal judgment performance as the market environment changes.

[0100] In this embodiment, the specific calculation method for the time distribution characteristics is as follows: Within a preset 6-week correction period, the system uses three quantitative indicators to characterize the distribution of implicit judgment results: anomaly frequency density, average duration percentage, and cluster index. Anomaly frequency density F is the average number of anomaly judgments per unit time (week), calculated as F = N / 6, where N is the total number of anomalies within the correction period. Average duration percentage D is the percentage of the total duration of all anomaly events to the correction period, calculated as D = (Σ(Te-Ts)) / 6 × 100%, where Te and Ts are the end and start timestamps of a single anomaly, respectively. Cluster index C measures the degree of anomaly clustering over time; its value is the ratio of the number of anomalies contained in the longest continuous anomaly period to the total number of anomalies N. Finally, the system calculates the degree of deviation using the weighted formula: Deviation δ = α × ΔF + β × ΔD + γ × ΔC (where α, β, and γ are preset adjustment weight coefficients, and ΔF, ΔD, and ΔC are the differences between the indicators of the current period and the baseline period, respectively). This objectively and quantitatively assesses the matching degree between the system state and the current detection threshold. The preset value for the anomaly frequency density benchmark is 1.5 times / week, depending on the system's historical false alarm level under normal market conditions, and is typically set between 1.0 and 2.0 times / week. In this embodiment, it is set to 1.5 times / week, which effectively distinguishes between normal market fluctuations and frequent anomalies. The preset value for the average duration ratio benchmark is 3%, depending on the stability requirements of the market during normal operation, and is typically set between 2% and 5%. In this embodiment, it is set to 3%, which accurately identifies significant changes in the duration of anomalies. The preset value for the cluster index benchmark is 0.25, depending on the system's sensitivity requirements to anomaly clustering characteristics, and is typically set between 0.2 and 0.4. In this embodiment, the value is set to 0.25, which can reliably detect the cluster distribution pattern of abnormal events. The preset values ​​for the weighting coefficients are α=0.5, β=0.3, and γ=0.2, depending on the relative importance of each feature indicator in the threshold adjustment. Generally, the weight allocation should ensure that frequency indicators are dominant (α≥0.4), persistence indicators are secondary (β≈0.3), and cluster indicators are auxiliary (γ≤0.3). The weight combination set in this embodiment can ensure that the threshold adjustment decision not only focuses on the core frequency changes, but also takes into account the comprehensive impact of persistence and cluster characteristics.

[0101] In this embodiment, the hidden market anomalies refer to those early risk signs that, in the operation of the pharmaceutical market, do not show obvious limits in a single data dimension, but can only be revealed through multi-dimensional data coupling and time-series evolution analysis. Specifically, they are subtle coordinated changes in multi-dimensional data such as behavior, price, market share, academic attention, and public opinion heat, which are slowly transmitted along the market entity correlation and causal links. For example, when a patent expiration event for a core product is detected, although its end-user usage remains stable, the synergistic relationship between the product's market share and the number of academic publications of competing products can be captured by a time-series model and verified by the pre-set "patent event impact" template and causal path in the knowledge graph. Another typical scenario is that after a policy release event is detected, although price indicators do not change significantly, the dynamic correlation between the frequency of medical institution purchases and channel inventory levels shows a specific pattern. The graph neural network hybrid model can capture its statistical significance from multiple periods of data and achieve accurate identification by matching the "policy-related impact" template and verifying the causal link of "change in purchase frequency → fluctuation in inventory level → restructuring of price system". In addition, after a new product launch event is detected, although the weak correlation between doctors' prescribing behavior characteristics and competitors' market share indicators does not reach the conventional detection threshold, the system can identify its implicit market structure change pattern by integrating multi-dimensional time-series features and market competition correlation paths in the knowledge graph. This embodiment achieves early detection and precise positioning of potential risks in the pharmaceutical market through the deep integration of data-driven and knowledge reasoning.

[0102] By statistically analyzing the temporal distribution characteristics of latent judgment results, such as their occurrence time, frequency, and persistence, and using deviation as a trigger condition, the detection threshold is adaptively adjusted. This ensures that threshold adjustments are based on long-term trends rather than instantaneous fluctuations, thereby distinguishing between short-term market fluctuations and persistent anomalies. This strategy leverages the business logic and statistical correlations between market data. For example, a continuous decline in market share accompanied by negative policy sentiment is more likely to point to real market risks, while isolated short-term promotional activities leading to sales spikes are often normal fluctuations. This approach amplifies or suppresses risk at the threshold level. The result is improved sensitivity to persistent and transmissive latent market anomalies, while reducing false positive rates and judgment instability caused by frequent threshold fluctuations. Short-term verification before and after threshold updates also suppresses over-adjustment, achieving a balance between robustness and response speed. This improves the detection accuracy of latent market signals and provides more reliable and interpretable early warning data for market strategy decisions.

[0103] Please see Figure 4 As shown, this is a schematic diagram of the data analysis system integrating knowledge graphs and deep learning in this embodiment. Furthermore, this embodiment also provides a data analysis system integrating knowledge graphs and deep learning, including:

[0104] The acquisition module is used to acquire in real time a multi-modal dataset and historical abnormal fluctuation information of the pharmaceutical market, which is obtained by fusing basic drug attribute data, end-user behavior data, macro market indicator data, medical academic literature data and public opinion data from various sources of pharmaceutical market data.

[0105] A construction module, which is connected to the acquisition module, is used to construct a first knowledge graph based on the logical association structure of each source data and a preset historical market pattern;

[0106] The constitutive module is connected to the construction module and the acquisition module respectively, and is used to determine a number of abnormal nodes and causal links of abnormal nodes in the first knowledge graph based on the historical abnormal fluctuation information, and to assign confidence weights and timestamps to each relation chain in the causal links to form a dynamic medical knowledge graph with temporal semantics.

[0107] A vector determination module, which is connected to the composition module and the acquisition module respectively, is used to determine a joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph and the preset graph neural network.

[0108] A candidate anomaly determination module, connected to the vector determination module, is used to capture the evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period according to a preset time-series neural hybrid model, and to construct the relationship influence between the source data according to the evolution trend to obtain a candidate anomaly list.

[0109] The latent determination module is connected to the candidate anomaly determination module and the constituent module respectively, and is used to determine the anomaly type of the latent anomaly based on the occurrence probability of the latent anomaly in the candidate anomaly list, the preset detection threshold, and the historical anomaly feature template and causal path rules in the dynamic medical knowledge graph, so as to obtain the latent determination result.

[0110] An adjustment module, which is connected to the latent determination module, is used to adjust the detection threshold according to the time distribution characteristics of the latent determination results within a preset correction time.

[0111] A generation module, which is connected to the constituent module and the implicit determination module respectively, is used to generate an analysis report based on the dynamic medical knowledge graph and the implicit determination result obtained again after adjusting the detection threshold.

[0112] By real-time acquisition and processing of multi-modal data, the system organically integrates basic drug attribute data, end-user behavior data, market macro-indicator data, medical academic literature data, and public opinion data. It also constructs a knowledge graph by combining historical abnormal fluctuation information with market entity relationships, achieving a global perception of market dynamics. By assigning confidence weights and time stamps to the causal links of abnormal nodes in the knowledge graph, and combining graph neural networks and time-series graph neural hybrid models to analyze joint feature vectors, the system can capture the evolutionary trends of basic drug attribute data and end-user behavior data over time, thereby revealing the mutual influence relationships between various market entities. Historical abnormal templates and causal path rules are used to implicitly determine candidate anomalies, and the detection threshold is dynamically adjusted based on time distribution characteristics, making market anomaly identification more accurate and robust. Overall, the system can achieve early identification and accurate classification of market anomalies through the linkage relationships of multi-dimensional data such as sales trends, price fluctuations, market share, academic attention, and public opinion heat, providing a scientific and reliable basis for pharmaceutical companies' strategic decisions. It also fully reflects the business connections and dynamic evolution patterns between various data dimensions, improving the accuracy and response speed of market anomaly detection.

[0113] For the few technical parameters not explicitly specified in this embodiment, such as the specific hyperparameters of the Graph Attention Network (GAT) and Long Short-Term Memory Network (LSTM) (e.g., number of network layers, hidden unit dimension, number of attention heads, etc.) and the specific fusion method of the joint feature vector (e.g., concatenation or summation), all fall within the scope of conventional engineering optimization selection in this technical field. Those skilled in the art will understand that these parameters can be adjusted on a preset training set using standard hyperparameter tuning methods (e.g., grid search, random search, or Bayesian optimization), or set according to specific computational resources and accuracy requirements, based on common industry practices. The determination of these parameters is a routine experimental step in the implementation process and does not rely on any non-public know-how that requires creative effort to acquire; therefore, it will not pose any substantial obstacle to those skilled in the art reproducing and implementing this invention.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data analysis method integrating knowledge graphs and deep learning, characterized in that, include: Real-time acquisition of multi-modal datasets and historical abnormal fluctuation information of the pharmaceutical market, which are obtained by fusing basic drug attribute data, end-user behavior data, macro market indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market. A first knowledge graph is constructed based on the logical relationship between the source data and the preset historical market pattern. Based on the historical abnormal fluctuation information, several abnormal nodes and causal links of the abnormal nodes in the first knowledge graph are determined, and confidence weights and timestamps are assigned to each relation chain in the causal links to form a dynamic medical knowledge graph with temporal semantics. The joint feature vector is determined based on the multi-modal dataset, the dynamic medical knowledge graph, and the preset graph neural network. The evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within the preset monitoring period is captured by the preset time sequence neural hybrid model, and the relationship between the source data is constructed according to the evolution trend to obtain a candidate anomaly list; The anomaly type of the latent anomaly is determined based on the occurrence probability of latent anomalies in the candidate anomaly list, the preset detection threshold, and the historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph, so as to obtain the latent judgment result. The detection threshold is adjusted according to the temporal distribution characteristics of the implicit determination results within the preset correction time. An analysis report is generated based on the dynamic medical knowledge graph and the implicit judgment result obtained again after adjusting the detection threshold.

2. The data analysis method integrating knowledge graphs and deep learning according to claim 1, characterized in that, The process of determining the joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph, and the preset graph neural network includes: The multimodal dataset is subjected to feature extraction according to a preset feature extraction algorithm to generate temporal feature representations for each modality; The dynamic medical knowledge graph is input into a preset graph neural network to obtain a structured graph embedding representation; The temporal feature representation and the graph embedding representation are fused to obtain the joint feature vector.

3. The data analysis method integrating knowledge graphs and deep learning according to claim 2, characterized in that, The process of constructing the first knowledge graph based on the logical relationships and preset historical market patterns of the source data includes: Construct market entity nodes and their logical relationship edges based on the types of pharmaceutical market entities, the business relationships between entities, and the market influence relationships; Based on the preset historical market model, determine the feature nodes related to market events and their corresponding market impact relationship edges; The logical relationship edges and the market influence relationship edges are combined to form the first knowledge graph used to describe the characteristics of market entity associations and event influence.

4. The data analysis method integrating knowledge graphs and deep learning according to claim 3, characterized in that, The process of fusing basic drug attribute data, end-user behavior data, macroeconomic market indicator data, medical academic literature data, and public opinion data from various sources in the pharmaceutical market to obtain a multi-modal dataset includes: The basic attribute data of the drug, the terminal usage behavior data, the market macro indicator data, the medical academic literature data and the public opinion data are collected in real time from each of the source data, and the collection results are processed by time synchronization, normalization and outlier removal to obtain preprocessed data. The preprocessed data is then subjected to feature alignment and modality fusion according to preset rules to obtain the multimodal dataset.

5. The data analysis method integrating knowledge graphs and deep learning according to claim 4, characterized in that, The process of determining several abnormal nodes and their causal links in the first knowledge graph based on the historical abnormal fluctuation information includes: Based on the historical abnormal fluctuation information, abnormal features related to abnormal market events are identified, and each abnormal feature is constructed as an abnormal node. Based on the sequential relationship, logical dependency relationship, and market fluctuation evolution law among the abnormal features in the historical abnormal fluctuation information, the causal links between the abnormal nodes are determined. The causal links are associated with the first knowledge graph as relational chains between the abnormal nodes, thereby determining a number of abnormal nodes and their causal links in the first knowledge graph.

6. The data analysis method integrating knowledge graphs and deep learning according to claim 5, characterized in that, The process of assigning confidence weights and timestamps to each relation chain in a causal chain to construct a dynamic medical knowledge graph with temporal semantics includes: Based on the frequency of occurrence, co-occurrence relationship, and consistency of fault evolution of each market anomaly in the historical abnormal fluctuation information, the confidence weight corresponding to each relationship chain is determined. The timestamps of each relationship chain are determined based on the occurrence time of each market anomaly, the triggering order of the anomaly characteristics, and the duration of the causal link of the anomaly node. The confidence weight and the timestamp are respectively assigned to the corresponding relationship chain to form a dynamic medical knowledge graph with temporal semantics.

7. The data analysis method integrating knowledge graphs and deep learning according to claim 6, characterized in that, The process of capturing the evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period using a preset time-series neural hybrid model, and constructing the relationship influence between the source data based on the evolution trend to obtain a candidate anomaly list includes: Based on the temporal changes of the joint feature vector within the preset monitoring period, extract the time-series features corresponding to the basic drug attribute data and the terminal usage behavior data; The time series features are input into the time series feature processing unit of the preset time series neural hybrid model to obtain the evolution trend of the basic attribute data of the drug and the terminal usage behavior data over time. The evolutionary trend is input into the graph structure processing unit of the preset time-series neural hybrid model to combine the logical relationships between the entities in the pharmaceutical market and construct the relationship influence between the source data to obtain the candidate anomaly list.

8. The data analysis method integrating knowledge graphs and deep learning according to claim 7, characterized in that, The process of determining the type of latent anomaly based on the occurrence probability of latent anomalies in the candidate anomaly list, a preset detection threshold, and historical anomaly feature templates and causal path rules in the dynamic medical knowledge graph to obtain the latent judgment result includes: The probability of occurrence of latent anomalies in the candidate anomaly list is compared with the preset detection threshold to filter out anomaly candidates corresponding to suspected latent anomalies. The abnormal candidates are matched with historical abnormal feature templates in the dynamic medical knowledge graph to obtain template matching results for each abnormal candidate. Based on the path association relationship of the abnormal candidates in the causal path rules, the causal consistency of each abnormal candidate is determined; Based on the template matching results and the causal consistency, the anomaly type to which the corresponding latent anomaly belongs is determined to obtain the latent judgment result.

9. The data analysis method integrating knowledge graphs and deep learning according to claim 8, characterized in that, The process of adjusting the detection threshold based on the temporal distribution characteristics of the implicit determination results within a preset correction period includes: Within the preset correction time, the occurrence time, frequency, and duration of the implicit judgment result are statistically analyzed to obtain the temporal distribution characteristics of the implicit judgment result. When the deviation between the time distribution feature and the preset detection threshold is greater than the preset deviation threshold, the preset detection threshold is adjusted according to the degree of deviation.

10. A data analysis system integrating knowledge graphs and deep learning, based on the data analysis method integrating knowledge graphs and deep learning as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to acquire in real time a multi-modal dataset and historical abnormal fluctuation information of the pharmaceutical market, which is obtained by fusing basic drug attribute data, end-user behavior data, macro market indicator data, medical academic literature data and public opinion data from various sources of pharmaceutical market data. A construction module, which is connected to the acquisition module, is used to construct a first knowledge graph based on the logical association structure of each source data and a preset historical market pattern; The constitutive module is connected to the construction module and the acquisition module respectively, and is used to determine a number of abnormal nodes and causal links of abnormal nodes in the first knowledge graph based on the historical abnormal fluctuation information, and to assign confidence weights and timestamps to each relation chain in the causal links to form a dynamic medical knowledge graph with temporal semantics. A vector determination module, which is connected to the composition module and the acquisition module respectively, is used to determine a joint feature vector based on the multi-modal dataset, the dynamic medical knowledge graph and the preset graph neural network. A candidate anomaly determination module, connected to the vector determination module, is used to capture the evolution trend of the basic drug attribute data and the terminal usage behavior data in the joint feature vector within a preset monitoring period according to a preset time-series neural hybrid model, and to construct the relationship influence between the source data according to the evolution trend to obtain a candidate anomaly list. The latent determination module is connected to the candidate anomaly determination module and the constituent module respectively, and is used to determine the anomaly type of the latent anomaly based on the occurrence probability of the latent anomaly in the candidate anomaly list, the preset detection threshold, and the historical anomaly feature template and causal path rules in the dynamic medical knowledge graph, so as to obtain the latent determination result. An adjustment module, which is connected to the latent determination module, is used to adjust the detection threshold according to the time distribution characteristics of the latent determination results within a preset correction time. A generation module, which is connected to the constituent module and the implicit determination module respectively, is used to generate an analysis report based on the dynamic medical knowledge graph and the implicit determination result obtained again after adjusting the detection threshold.

Images