A community worker health card consumption behavior modeling method based on big data analysis

Abstract

This invention discloses a method for modeling community employee health card consumption behavior based on big data analysis, comprising the following steps: collecting health card consumption data and user information data to construct an original dataset; performing field standardization and feature construction to generate a feature vector set; inputting a multi-channel embedding dimensionality reduction model, fusing the embedding representations of frequency, category, and health channel to generate a unified behavior embedding vector; inputting an improved hierarchical density clustering model to output a pairing set; constructing a clustering confidence scoring function based on the pairing set to output a clustering confidence scoring matrix; generating a health consumption behavior profile label set containing service preferences, frequency levels, and risk classifications; and employing an adaptive update mechanism to continuously optimize the clustering structure and dynamically update the profile results. This invention can achieve accurate identification and dynamic modeling of community employee health consumption behavior, and has strong universality and scalability.

Description

Technical Field

[0001] This invention relates to the field of health management information processing and intelligent profile generation technology, and in particular to a method for modeling the consumption behavior of community employees' health cards based on big data analysis. Background Technology

[0002] With the continuous improvement of community health management systems and the widespread application of employee health record information, service behavior data based on health cards has become an important data foundation for exploring employee health consumption habits and optimizing the allocation of public health resources. Current technologies for analyzing community employee health card consumption behavior mostly employ statistical summarization or static classification methods, lacking in-depth modeling and structural summarization of the multidimensional attributes of behavior, making it difficult to support refined management needs such as personalized service recommendations and health risk warnings.

[0003] In practical applications, existing modeling methods generally suffer from the following problems: First, they fail to effectively integrate multi-channel features such as consumption frequency, service category, and health attributes, resulting in limited behavioral representation capabilities and difficulty in accurately depicting the full picture of employee health consumption; Second, clustering methods mostly use traditional algorithms with fixed parameters, which cannot dynamically adapt to the problem of blurred consumption behavior boundaries or structural drift, making it difficult to quantify the reliability of clustering results and resulting in poor stability; Third, they lack an effective model adaptive update mechanism, making it difficult to achieve dynamic evolution of the profile structure when faced with continuously increasing consumption data, resulting in insufficient timeliness and continuity of modeling results.

[0004] Therefore, how to provide a modeling method for community employee health card consumption behavior based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a modeling method for community employee health card consumption behavior based on big data analysis. This invention fully integrates multi-channel embedding dimensionality reduction, improved hierarchical density clustering, and adaptive update mechanism to systematically realize structured modeling and profile tag generation of health card consumption behavior. It has the advantages of high behavior recognition accuracy, strong model evolution capability, and high credibility of profile generation results.

[0006] A method for modeling community employee health card consumption behavior based on big data analysis according to an embodiment of the present invention includes the following steps: Collect health card consumption data and user information data of community employees to construct a raw dataset; Perform field standardization and feature construction on the original dataset, and output a set of feature vectors; The feature vector set is input into a multi-channel embedding dimensionality reduction model, and the embedding representation is extracted through the frequency channel, category channel and health channel respectively. The embedding vectors are then fused to generate a unified behavior embedding vector. Input the unified behavior embedding vector into the improved hierarchical density clustering model, perform cluster analysis, and output the set of pairs of embedding vectors and corresponding cluster labels; Based on the pairing set, a clustering confidence scoring function is constructed, and the clustering confidence scoring matrix is ​​output. Based on each embedding vector and corresponding cluster label in the pairing set, a health consumption behavior profile label set including service preferences, frequency levels and risk classifications is generated. An adaptive update mechanism is used to continuously optimize and evolve the clustering structure of health card consumption behavior, resulting in a dynamically updated set of health consumption behavior profile tags.

[0007] Optionally, the collection of health card consumption data and user information data of community employees to construct the original dataset specifically includes: The collected health card consumption data includes fields such as consumption time, service type, consumption amount, and consumption location, which are linked to the unique identity of community employees. The collected user information data includes gender, age, job category, previous diagnosis information, and physical examination indicators. Based on a unified user identifier, health card consumption data and user information data are fused across sources and time-stamped to construct a structured original dataset.

[0008] Optionally, the field standardization and feature construction of the original dataset specifically includes: The numerical fields in the original dataset are normalized using the Z-score standardization method, and the categorical fields are transformed using one-hot encoding. A set of consumer behavior feature vectors is constructed, which includes statistical features of consumption frequency over the past three months, usage ratio features of various service types, monthly average consumption amount and volatility features, health risk factor score features mapped from the most recent physical examination indicators, and structured coding features of user job category and age group.

[0009] Optionally, the step of inputting the feature vector set into a multi-channel embedding dimensionality reduction model, extracting embedding representations through the frequency channel, category channel, and health channel respectively, and fusing them to generate a unified behavior embedding vector specifically includes: The frequency channel receives consumption frequency statistics, usage ratio of various service types, monthly average consumption amount and volatility of the past three months, transforms them using a fully connected perceptron structure, and extracts frequency embedding representations by combining batch normalization and ReLU activation function. The category channel receives category features, performs dense mapping using an embedding coding layer, and then inputs them into a one-dimensional convolutional network and max pooling combination structure to extract category preference embedding representations. The health channel receives the health risk factor score features mapped from the most recent physical examination indicators, performs normalization processing, and then inputs them into a long short-term memory network containing time step control to extract the dynamic health status embedding representation. The frequency embedding representation, category preference embedding representation, and dynamic health status embedding representation are sequentially input into the attention-weighted fusion module in the multi-channel embedding dimensionality reduction model. The fusion coefficients are adaptively allocated according to the correlation weights between features to generate a unified behavior embedding vector.

[0010] Optionally, the improved hierarchical density clustering model specifically includes: A Gaussian kernel function is used to estimate the kernel density of the unified behavior embedding vector set. The local density value of each unified behavior embedding vector in the neighborhood is calculated, and the corresponding density-aware core distance is calculated to obtain the local density value sequence and the core distance set. Based on the core distance set, determine whether there is a density reachability relationship between any two unified behavior embedding vectors, construct a density reachability graph structure, and obtain a connected path graph to represent the density connectivity relationship between sample points; A recursive aggregation operation is performed on the connected path graph to generate a hierarchical clustering tree structure. The hierarchical clustering tree structure is then pruned and optimized using a minimum stability scoring function, and the preliminary clustering results and corresponding initial clustering labels are output. For each initial cluster, the perturbation magnitude, density variation rate and neighborhood label switching rate of the boundary embedding vector are calculated to construct a perturbation sensitivity index and identify boundary embedding vectors with label stability below a preset threshold to obtain a set of suspected drift samples. For a set of suspected drifting samples, the consistency ratio of cluster labels in their respective density neighborhoods is calculated, a neighborhood label confidence index is constructed, and a weighted evaluation mechanism is used to re-evaluate the cluster labels in combination with a perturbation sensitivity index, and the corrected final cluster label set is output. The unified behavior embedding vector set is paired one-to-one with the final cluster label set to generate a structured embedding vector and cluster label pairing set.

[0011] Optionally, the step of constructing a clustering confidence scoring function based on the pairing set and outputting a clustering confidence scoring matrix specifically includes: For each embedding vector, a fixed number of nearest neighbor samples are selected in the unified behavior embedding space. The consistency score of the label is calculated by counting whether these neighbor samples and the target sample belong to the same cluster. The sliding window density estimation method is used to calculate the density distribution of the embedding vectors within the clusters, and obtain the average density value and density variance of each cluster as an index of cluster structure stability. By locating density abrupt change points in the edge region of clusters, the dispersion of samples within the boundary region is calculated, thus forming a boundary clarity index. After normalizing the label consistency score, cluster structure stability index and boundary clarity index respectively, they are weighted and superimposed with fixed weight coefficients to construct a cluster confidence scoring function. The scoring function is used to score each of the embedding vectors, outputting the clustering confidence score. All embedding vectors and scores are then arranged in order to output the clustering confidence score matrix.

[0012] Optionally, the step of generating a health consumption behavior profile tag set, including service preferences, frequency levels, and risk classifications, based on each embedding vector in the pairing set and its corresponding cluster label, specifically includes: The frequency channel features in each embedded vector of the pairing set are divided into intervals to determine the consumption frequency interval and generate frequency level labels. The activity levels of each service category in the category channel are sorted and statistically analyzed to extract the dominant consumption categories and generate service preference tags; Analyze the characteristic dimensions related to health consumption in the health channel, identify abnormally high-frequency services or high-amount expenditure behaviors, and generate risk classification labels; The frequency level label, service preference label and risk classification label are combined to form a multi-dimensional label structure, which serves as a healthy consumption behavior profile with embedded vectors. The output contains health consumption behavior profile tags corresponding to the embedded vectors in all pair sets, forming a health consumption behavior profile tag set.

[0013] Optionally, the step of continuously optimizing and evolving the clustering structure of health card consumption behavior using an adaptive update mechanism to obtain a dynamically updated health consumption behavior profile tag set specifically includes: Periodically collect newly added health card consumption data, perform field standardization and feature construction, and generate corresponding feature vectors; The feature vector is input into the multi-channel embedding dimensionality reduction model to generate the corresponding unified behavior embedding vector. For each newly added embedded vector, calculate the similarity with the cluster centers in the existing pairing set. When the preset label consistency condition is met, include it in the current cluster, correct the cluster center position through incremental fusion, and update the pairing set. When multiple samples in the newly added embedding vector overlap with the cluster boundary or label drift occurs, the local structure correction process is triggered to redefine the boundary and adjust the internal structure of the abnormal cluster, and output the updated pairing set. A self-supervised discriminant for structural stability is constructed to monitor the changes in cluster center drift rate and cluster confidence score matrix in pairing sets. If the fluctuation exceeds the dynamic threshold, the model update frequency is adjusted. Based on the updated pairing set, a new set of health consumption behavior labels, including service preferences, frequency levels, and risk classifications, is generated to form a dynamically updated modeling result.

[0014] The beneficial effects of this invention are: (1) This invention achieves multi-dimensional embedding modeling and semantic understanding of health card consumption behavior, improving the refinement and accuracy of user behavior profiles: By constructing a multi-channel embedding dimensionality reduction model, this invention independently models and fuses the consumption frequency, service category, and health status, significantly enhancing the ability to model the potential semantics of consumption behavior. Compared with the traditional single statistical feature modeling method, it can more comprehensively capture the behavioral characteristics of employees under different service categories and health stages, and based on a unified behavior embedding vector, it provides a more stable and highly discriminative feature foundation for subsequent clustering and label generation.

[0015] (2) An improved hierarchical density clustering and clustering confidence scoring mechanism are constructed to effectively enhance the robustness of label generation and the interpretability of the model: The improved hierarchical density clustering model proposed in this invention combines perturbation sensitivity and neighborhood label consistency strategies to re-score and correct boundary samples, reducing mis-clustering caused by sample density differences or label drift. At the same time, by designing a clustering confidence scoring function, three dimensions—label consistency, cluster structure stability, and boundary clarity—are introduced to quantitatively score and visualize the clustering results, making the health consumption behavior profile more interpretable and providing a stronger basis for evaluation.

[0016] (3) An adaptive update mechanism and a structural stability discriminator are introduced to ensure the model has continuous evolution and dynamic adaptability: By periodically collecting new consumer behavior data and combining dynamic cluster center updates, local structure corrections, and model update frequency adjustment mechanisms, this invention constructs a complete adaptive evolution process for consumer behavior modeling. It can effectively cope with dynamic fluctuations in real-world scenarios such as changes in community employee behavior and adjustments to service content, continuously outputting a high-quality healthy consumer behavior profile tag set, and significantly improving the stability, practicality, and service support capabilities of the modeling system. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for modeling the consumption behavior of community employee health cards based on big data analysis, as proposed in this invention. Figure 2 This is a schematic diagram of a method for modeling community employee health card consumption behavior based on big data analysis proposed in this invention; Figure 3 This is a framework diagram of the improved hierarchical density clustering model in the community employee health card consumption behavior modeling method based on big data analysis proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 A method for modeling community employee health card consumption behavior based on big data analysis includes the following steps: Step 1: Collect health card consumption data and user information data of community employees to construct the original dataset; Step 2: Standardize the fields and construct features in the original dataset, and output a set of feature vectors; Step 3: Input the feature vector set into the multi-channel embedding dimensionality reduction model, extract the embedding representation through the frequency channel, category channel and health channel respectively, and fuse them to generate a unified behavior embedding vector; Step 4: Input the unified behavior embedding vector into the improved hierarchical density clustering model, perform cluster analysis, and output the set of pairs of embedding vectors and corresponding cluster labels; Step 5: Based on the pairing set, construct the clustering confidence scoring function and output the clustering confidence scoring matrix; Step 6: Generate a health consumption behavior profile tag set that includes service preferences, frequency levels, and risk classifications based on each embedding vector in the pairing set and the corresponding cluster label; Step 7: Use an adaptive update mechanism to continuously optimize and evolve the clustering structure of health card consumption behavior to obtain a dynamically updated health consumption behavior profile tag set.

[0020] In this embodiment, the step of collecting health card consumption data and user information data of community employees to construct an original dataset specifically includes: The collected health card consumption data includes fields such as consumption time, service type, consumption amount, and consumption location, which are linked to the unique identity of community employees. The collected user information data includes gender, age, job category, previous diagnosis information, and physical examination indicators. Based on a unified user identifier, health card consumption data and user information data are fused across sources and time-stamped to construct a structured original dataset.

[0021] This implementation method collects health card consumption data and user information data containing the unique identification of community employees, and performs cross-source fusion and timestamp synchronization operations to ensure data consistency and integrity. Based on the fusion of consumption behavior and user health background, a structured original dataset is constructed, providing a multi-dimensional and high-quality input foundation for subsequent behavior modeling, thereby significantly improving the comprehensiveness of health consumption behavior profiles and the ability to identify individual differences.

[0022] In this embodiment, the field standardization and feature construction of the original dataset specifically includes: The numerical fields in the original dataset are normalized using the Z-score standardization method, and the categorical fields are transformed using one-hot encoding. A set of consumer behavior feature vectors is constructed, which includes statistical features of consumption frequency over the past three months, usage ratio features of various service types, monthly average consumption amount and volatility features, health risk factor score features mapped from the most recent physical examination indicators, and structured coding features of user job category and age group.

[0023] This implementation enhances the comparability of features with different dimensions by normalizing numerical fields in the original dataset using Z-score normalization. Categorical fields are transformed using one-hot encoding, effectively preserving the information structure of discrete attributes. Based on this, a set of consumer behavior feature vectors covering multiple dimensions such as behavior frequency, service preference, spending power, health status, and user background is constructed. This provides high-quality input feature representations for subsequent embedding modeling, significantly improving the model's ability to represent behavioral differences and the accuracy of health risk identification.

[0024] In this embodiment, the step of inputting the feature vector set into a multi-channel embedding dimensionality reduction model, extracting embedding representations through the frequency channel, category channel, and health channel respectively, and fusing them to generate a unified behavior embedding vector specifically includes: The frequency channel receives consumption frequency statistics, usage ratio of various service types, monthly average consumption amount and volatility of the past three months, transforms them using a fully connected perceptron structure, and extracts frequency embedding representations by combining batch normalization and ReLU activation function. The category channel receives category features, performs dense mapping using an embedding coding layer, and then inputs them into a one-dimensional convolutional network and max pooling combination structure to extract category preference embedding representations. The health channel receives the health risk factor score features mapped from the most recent physical examination indicators, performs normalization processing, and then inputs them into a long short-term memory network containing time step control to extract the dynamic health status embedding representation. The frequency embedding representation, category preference embedding representation, and dynamic health status embedding representation are sequentially input into the attention-weighted fusion module in the multi-channel embedding dimensionality reduction model. The fusion coefficients are adaptively allocated according to the correlation weights between features to generate a unified behavior embedding vector.

[0025] This implementation constructs a three-channel embedding representation structure to extract embedding features for frequency, category, and health status, effectively achieving a structured expression of multi-dimensional information. The frequency channel employs a combination of a fully connected perceptron, batch normalization, and ReLU function to enhance the nonlinear modeling capability for consumption fluctuations; the category channel introduces convolution and pooling structures to improve the extraction accuracy of local preferences for category features; and the health channel uses a long short-term memory network with time-step control to strengthen the ability to capture trends in health status changes. Finally, a unified behavioral embedding vector is generated through an attention-weighted fusion mechanism, significantly improving the expressive completeness and recognition sensitivity of downstream clustering modeling.

[0026] In this embodiment, the improved hierarchical density clustering model specifically includes: A Gaussian kernel function is used to estimate the kernel density of the unified behavior embedding vector set. The local density value of each unified behavior embedding vector in the neighborhood is calculated, and the corresponding density-aware core distance is calculated to obtain the local density value sequence and the core distance set. Based on the core distance set, determine whether there is a density reachability relationship between any two unified behavior embedding vectors, construct a density reachability graph structure, and obtain a connected path graph to represent the density connectivity relationship between sample points; A recursive aggregation operation is performed on the connected path graph to generate a hierarchical clustering tree structure. The hierarchical clustering tree structure is then pruned and optimized using a minimum stability scoring function, and the preliminary clustering results and corresponding initial clustering labels are output. For each initial cluster, the perturbation magnitude, density variation rate and neighborhood label switching rate of the boundary embedding vector are calculated to construct a perturbation sensitivity index and identify boundary embedding vectors with label stability below a preset threshold to obtain a set of suspected drift samples. For a set of suspected drifting samples, the consistency ratio of cluster labels in their respective density neighborhoods is calculated, a neighborhood label confidence index is constructed, and a weighted evaluation mechanism is used to re-evaluate the cluster labels in combination with a perturbation sensitivity index, and the corrected final cluster label set is output. The unified behavior embedding vector set is paired one-to-one with the final cluster label set to generate a structured embedding vector and cluster label pairing set.

[0027] This implementation method achieves stable and highly discriminative partitioning of a unified behavior embedding vector set by designing an improved hierarchical density clustering method. A hierarchical clustering tree structure is generated through recursive aggregation, and a minimum stability scoring function is introduced for pruning, improving the robustness of the initial clustering partitioning. A perturbation sensitivity index is constructed by fusing perturbation amplitude, density variation rate, and label switching rate, and label correction is performed by combining neighborhood label confidence, improving the accuracy and consistency of boundary determination and providing a stable foundation for downstream health behavior profiling and service adaptation.

[0028] In this embodiment, the step of constructing a clustering confidence scoring function based on the pairing set and outputting a clustering confidence scoring matrix specifically includes: For each embedding vector, a fixed number of nearest neighbor samples are selected in the unified behavior embedding space. The consistency score of the label is calculated by counting whether these neighbor samples and the target sample belong to the same cluster. The sliding window density estimation method is used to calculate the density distribution of the embedding vectors within the clusters, and obtain the average density value and density variance of each cluster as an index of cluster structure stability. By locating density abrupt change points in the edge region of clusters, the dispersion of samples within the boundary region is calculated, thus forming a boundary clarity index. After normalizing the label consistency score, cluster structure stability index and boundary clarity index respectively, they are weighted and superimposed with fixed weight coefficients to construct a cluster confidence scoring function. The scoring function is used to score each of the embedding vectors, outputting the clustering confidence score. All embedding vectors and scores are then arranged in order to output the clustering confidence score matrix.

[0029] This implementation constructs a clustering confidence scoring function that integrates label consistency scores, cluster structure stability indicators, and boundary clarity indicators. This allows for multi-dimensional measurement of the credibility of health consumption behavior clusters, enhancing the credibility assessment capability of clustering results. By combining sliding window density estimation with boundary density abrupt change detection, it effectively identifies loosely structured or unstable label distribution boundary regions, improving the perception accuracy of clustering noise samples and potential drift behaviors. After outputting the clustering confidence scoring matrix, it can rank the clustering stability of samples in the unified behavior embedding vector set, facilitating the selection of high-confidence samples for subsequent behavior profiling and risk identification tasks, significantly enhancing the accuracy of profiling labels and the overall robustness of the system.

[0030] In this embodiment, the step of generating a health consumption behavior profile tag set, including service preferences, frequency levels, and risk classifications, based on each embedding vector in the pairing set and its corresponding cluster label, specifically includes: The frequency channel features in each embedded vector of the pairing set are divided into intervals to determine the consumption frequency interval and generate frequency level labels. The activity levels of each service category in the category channel are sorted and statistically analyzed to extract the dominant consumption categories and generate service preference tags; Analyze the characteristic dimensions related to health consumption in the health channel, identify abnormally high-frequency services or high-amount expenditure behaviors, and generate risk classification labels; The frequency level label, service preference label and risk classification label are combined to form a multi-dimensional label structure, which serves as a healthy consumption behavior profile with embedded vectors. The output contains health consumption behavior profile tags corresponding to the embedded vectors in all pair sets, forming a health consumption behavior profile tag set.

[0031] This implementation method divides the frequency channel features in the unified behavior embedding vector into intervals to generate frequency level labels, sorts the category channels by activity to extract the dominant consumption categories to generate service preference labels, and combines the health channel to identify abnormal spending behavior to generate risk classification labels. This constructs a multi-dimensional health consumption behavior profile label structure, which achieves a comprehensive and detailed characterization of user consumption behavior features. While ensuring the accuracy of label generation, it provides stable and interpretable feature support for health risk identification and behavior intervention strategy design.

[0032] In this embodiment, the step of continuously optimizing and evolving the clustering structure of health card consumption behavior using an adaptive update mechanism to obtain a dynamically updated health consumption behavior profile tag set specifically includes: Periodically collect newly added health card consumption data, perform field standardization and feature construction, and generate corresponding feature vectors; The feature vector is input into the multi-channel embedding dimensionality reduction model to generate the corresponding unified behavior embedding vector. For each newly added embedded vector, calculate the similarity with the cluster centers in the existing pairing set. When the preset label consistency condition is met, include it in the current cluster, correct the cluster center position through incremental fusion, and update the pairing set. When multiple samples in the newly added embedding vector overlap with the cluster boundary or label drift occurs, the local structure correction process is triggered to redefine the boundary and adjust the internal structure of the abnormal cluster, and output the updated pairing set. A self-supervised discriminant for structural stability is constructed to monitor the changes in cluster center drift rate and cluster confidence score matrix in pairing sets. If the fluctuation exceeds the dynamic threshold, the model update frequency is adjusted. Based on the updated pairing set, a new set of health consumption behavior labels, including service preferences, frequency levels, and risk classifications, is generated to form a dynamically updated modeling result.

[0033] This implementation method achieves continuous access and incremental adaptation of user behavior data by periodically collecting new health card consumption data and performing unified feature construction and embedding modeling. It introduces a cluster center drift detection and label consistency evaluation mechanism to dynamically adjust and optimize the cluster structure, effectively suppressing modeling errors caused by label drift. Combined with a self-supervised discriminator for structural stability, it monitors cluster credibility and adjusts the model update frequency to ensure the timeliness and stability of the health consumption behavior profile label set, significantly improving the model's sensitivity and adaptability to changes in new data.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a health card consumption behavior data modeling scenario at a local urban community health service center. Service record data from approximately 4,000 registered users across six subdistricts were selected, including fields such as item code, service type, consumption time, service location, amount, and basic employee information. The data underwent structured preprocessing on a monthly basis, with a quarterly behavior observation period. A time-sliding window mechanism was constructed to embed each user's service records into a fixed-length consumption behavior sequence for modeling. The goal is to generate a structured consumption behavior profile tag set, including service preferences, consumption frequency levels, and health risk classifications, using the modeling method proposed in this invention based on deep consumption embedding and improved clustering structure, providing decision-making references for subdistrict health departments.

[0035] First, the original consumption records are standardized by unifying the format, completing missing data, and aligning the time. Records without service item numbers or with abnormal amounts are removed, resulting in a stable structured consumption dataset covering fields such as item ID, category code, consumption period, and service tag. Then, a multi-granularity embedding strategy is designed for consumption behavior. First, a one-hot vector is constructed using service categories to represent basic preference information. Then, behavioral statistical features such as amount, frequency, and service continuity within a period are introduced. These are fused into a unified consumption behavior embedding vector using a three-layer perceptron embedding structure, with a unified 128-dimensional dimension to facilitate subsequent high-dimensional clustering. This embedding strategy balances the density of behavior with the distinguishability of service preference patterns, avoiding the sensitivity of traditional statistical clustering methods to category sparsity and amount fluctuations.

[0036] This invention employs an improved hierarchical density clustering model to perform a six-stage structural processing on the consumer embedding vector set. First, a high-dimensional feature space adjacency graph is constructed, and a candidate edge set is generated based on a distance threshold function and a label consistency strategy, forming a dense subgraph structure. Then, a composite density estimation method based on distance and service labels is defined to extract the local density distribution trend and output a preliminary density gradient map. Subsequently, a hierarchical expansion structure constrained by the gradient map is used to construct multi-scale cluster candidates. Next, based on a neighborhood confidence adjustment mechanism, label consistency correction is performed on the clusters to which boundary samples belong, enhancing cluster stability. Then, a clustering confidence scoring mechanism is introduced, constructing a confidence function based on the paired embedding vector set to output a scoring matrix. Finally, cluster labels are reconstructed based on the confidence matrix to generate a high-confidence consumer behavior clustering structure, providing reliable input for subsequent profile label generation.

[0037] A clustering confidence scoring mechanism is used to confirm labels and adjust clustering results for all embedded vectors, avoiding the problems of ambiguous boundaries or uncertain number of clusters in traditional clustering. A label pairing set construction method is introduced. Based on each set of embedded vectors and its clustering labels, statistical indicators such as service preference distribution, consumption amount variance, and service time coverage are used to construct a clustering confidence scoring function and output a clustering confidence scoring matrix. This scoring matrix is ​​then used to evaluate clustering quality and guide subsequent structural label generation. By traversing the embedded samples in each pairing set, their label frequency, service type distribution, and density level are integrated to generate a three-level health consumption behavior profile label set containing service preferences (such as traditional Chinese medicine health care, rehabilitation, psychological counseling, etc.), frequency levels (such as high-frequency, medium-frequency, low-frequency), and health risk classifications (such as high, medium, low), forming a structured output format. Table 1 below shows a partial example of structured data output based on the system model: Table 1. Example Data Table of Community Employee Health Card Consumption Behavior Profile

[0038] Analysis of Table 1 reveals that users A001 and A003 exhibit typical low-risk profile characteristics such as high frequency of service use, clear core service preferences, and low risk factors. The system consistently categorizes them into the same cluster and assigns them high label confidence scores, indicating good stability of their health behaviors. In contrast, user A004 exhibits low and loosely distributed service frequency, with main services concentrated on non-core rehabilitation items such as massage. Combined with multiple abnormal indicators in their health record (such as excessive BMI and abnormal dietary records), the system classifies them as a high-risk user. The clustering structure and profile label results highly match the offline case situation, demonstrating the accuracy and practical applicability of the system modeling.

[0039] Finally, a structural evolution mechanism is introduced to achieve dynamic model correction through a dual-track driven adaptive update approach: on the one hand, new data is periodically collected and its labels are judged to be consistent with the original cluster centers. If they are consistent, the cluster centers are adjusted through incremental fusion; on the other hand, if the label drift trend is obvious or there are multiple sample labels with blurred boundaries, a local structure corrector is triggered to reconstruct the cluster structure. Furthermore, a self-supervised discriminant for structural stability is designed to dynamically control the update frequency by judging the cluster center drift rate and the fluctuation range of the confidence interval, ensuring that the modeling results maintain both stability and sensitivity during long-term use.

[0040] This embodiment verifies the effectiveness of the present invention in processing consumer behavior modeling and health profile tag generation using real community health card consumption data. The proposed consumer behavior embedding method has stronger representation capabilities, the improved hierarchical density clustering structure enhances clustering credibility, the paired tag generation mechanism achieves accurate extraction of multi-dimensional structural tags, and the model has stable evolution and update capabilities. The system can effectively classify and personalize the characteristics of users with different profiles, greatly improving the efficiency and effectiveness of service organizations in resource allocation, health intervention, and precision governance.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for modeling community employee health card consumption behavior based on big data analysis, characterized in that, Includes the following steps: Collect health card consumption data and user information data of community employees to construct a raw dataset; Perform field standardization and feature construction on the original dataset, and output a set of feature vectors; The feature vector set is input into a multi-channel embedding dimensionality reduction model, and the embedding representation is extracted through the frequency channel, category channel and health channel respectively. The embedding vectors are then fused to generate a unified behavior embedding vector. Input the unified behavior embedding vector into the improved hierarchical density clustering model, perform cluster analysis, and output the set of pairs of embedding vectors and corresponding cluster labels; Based on the pairing set, a clustering confidence scoring function is constructed, and the clustering confidence scoring matrix is ​​output. Based on each embedding vector and corresponding cluster label in the pairing set, a health consumption behavior profile label set including service preferences, frequency levels and risk classifications is generated. An adaptive update mechanism is used to continuously optimize and evolve the clustering structure of health card consumption behavior, resulting in a dynamically updated set of health consumption behavior profile tags.

2. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The process of collecting health card consumption data and user information data from community employees to construct the original dataset specifically includes: The collected health card consumption data includes fields such as consumption time, service type, consumption amount, and consumption location, which are linked to the unique identity of community employees. The collected user information data includes gender, age, job category, previous diagnosis information, and physical examination indicators. Based on a unified user identifier, health card consumption data and user information data are fused across sources and time-stamped to construct a structured original dataset.

3. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The specific steps of field standardization and feature construction of the original dataset include: The numerical fields in the original dataset are normalized using the Z-score standardization method, and the categorical fields are transformed using one-hot encoding. A set of consumer behavior feature vectors is constructed, which includes statistical features of consumption frequency over the past three months, usage ratio features of various service types, monthly average consumption amount and volatility features, health risk factor score features mapped from the most recent physical examination indicators, and structured coding features of user job category and age group.

4. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The step of inputting the feature vector set into a multi-channel embedding dimensionality reduction model, extracting embedding representations through the frequency channel, category channel, and health channel respectively, and fusing them to generate a unified behavior embedding vector specifically includes: The frequency channel receives consumption frequency statistics, usage ratio of various service types, monthly average consumption amount and volatility of the past three months, transforms them using a fully connected perceptron structure, and extracts frequency embedding representations by combining batch normalization and ReLU activation function. The category channel receives category features, performs dense mapping using an embedding coding layer, and then inputs them into a one-dimensional convolutional network and max pooling combination structure to extract category preference embedding representations. The health channel receives the health risk factor score features mapped from the most recent physical examination indicators, performs normalization processing, and then inputs them into a long short-term memory network containing time step control to extract the dynamic health status embedding representation. The frequency embedding representation, category preference embedding representation, and dynamic health status embedding representation are sequentially input into the attention-weighted fusion module in the multi-channel embedding dimensionality reduction model. The fusion coefficients are adaptively allocated according to the correlation weights between features to generate a unified behavior embedding vector.

5. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The improved hierarchical density clustering model specifically includes: A Gaussian kernel function is used to estimate the kernel density of the unified behavior embedding vector set. The local density value of each unified behavior embedding vector in the neighborhood is calculated, and the corresponding density-aware core distance is calculated to obtain the local density value sequence and the core distance set. Based on the core distance set, determine whether there is a density reachability relationship between any two unified behavior embedding vectors, construct a density reachability graph structure, and obtain a connected path graph to represent the density connectivity relationship between sample points; A recursive aggregation operation is performed on the connected path graph to generate a hierarchical clustering tree structure. The hierarchical clustering tree structure is then pruned and optimized using a minimum stability scoring function, and the preliminary clustering results and corresponding initial clustering labels are output. For each initial cluster, the perturbation magnitude, density variation rate and neighborhood label switching rate of the boundary embedding vector are calculated to construct a perturbation sensitivity index and identify boundary embedding vectors with label stability below a preset threshold to obtain a set of suspected drift samples. For a set of suspected drifting samples, the consistency ratio of cluster labels in their respective density neighborhoods is calculated, a neighborhood label confidence index is constructed, and a weighted evaluation mechanism is used to re-evaluate the cluster labels in combination with a perturbation sensitivity index, and the corrected final cluster label set is output. The unified behavior embedding vector set is paired one-to-one with the final cluster label set to generate a structured embedding vector and cluster label pairing set.

6. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The process of constructing a clustering confidence scoring function based on the pairing set and outputting a clustering confidence scoring matrix specifically includes: For each embedding vector, a fixed number of nearest neighbor samples are selected in the unified behavior embedding space. The consistency score of the label is calculated by counting whether these neighbor samples and the target sample belong to the same cluster. The sliding window density estimation method is used to calculate the density distribution of the embedding vectors within the clusters, and obtain the average density value and density variance of each cluster as an index of cluster structure stability. By locating density abrupt change points in the edge region of clusters, the dispersion of samples within the boundary region is calculated, thus forming a boundary clarity index. After normalizing the label consistency score, cluster structure stability index and boundary clarity index respectively, they are weighted and superimposed with fixed weight coefficients to construct a cluster confidence scoring function. The scoring function is used to score each of the embedding vectors, outputting the clustering confidence score. All embedding vectors and scores are then arranged in order to output the clustering confidence score matrix.

7. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The step of generating a health consumption behavior profile tag set, including service preferences, frequency levels, and risk classifications, based on each embedding vector in the pairing set and its corresponding cluster label, specifically includes: The frequency channel features in each embedded vector of the pairing set are divided into intervals to determine the consumption frequency interval and generate frequency level labels. The activity levels of each service category in the category channel are sorted and statistically analyzed to extract the dominant consumption categories and generate service preference tags; Analyze the characteristic dimensions related to health consumption in the health channel, identify abnormally high-frequency services or high-amount expenditure behaviors, and generate risk classification labels; The frequency level label, service preference label and risk classification label are combined to form a multi-dimensional label structure, which serves as a healthy consumption behavior profile with embedded vectors. The output contains health consumption behavior profile tags corresponding to the embedded vectors in all pair sets, forming a health consumption behavior profile tag set.

8. The method for modeling community employee health card consumption behavior based on big data analysis according to claim 1, characterized in that, The adaptive update mechanism is used to continuously optimize and evolve the clustering structure of health card consumption behavior, resulting in a dynamically updated health consumption behavior profile tag set, specifically including: Periodically collect newly added health card consumption data, perform field standardization and feature construction, and generate corresponding feature vectors; The feature vector is input into the multi-channel embedding dimensionality reduction model to generate the corresponding unified behavior embedding vector. For each newly added embedded vector, calculate the similarity with the cluster centers in the existing pairing set. When the preset label consistency condition is met, include it in the current cluster, correct the cluster center position through incremental fusion, and update the pairing set. When multiple samples in the newly added embedding vector overlap with the cluster boundary or label drift occurs, the local structure correction process is triggered to redefine the boundary and adjust the internal structure of the abnormal cluster, and output the updated pairing set. A self-supervised discriminant for structural stability is constructed to monitor the changes in cluster center drift rate and cluster confidence score matrix in pairing sets. If the fluctuation exceeds the dynamic threshold, the model update frequency is adjusted. Based on the updated pairing set, a new set of health consumption behavior labels, including service preferences, frequency levels, and risk classifications, is generated to form a dynamically updated modeling result.

Images