Machine learning based customer profiling system

Abstract

The present application relates to the field of customer portrait construction, in particular to a customer portrait construction system based on machine learning, comprising a multi-source data acquisition module, a multi-modal feature extraction module, a customer portrait representation learning module and a graph portrait output module, the multi-source data acquisition module integrates data after hash desensitization and graph anomaly detection processing; the multi-modal feature extraction module generates the highest fitness combination features through genetic programming algorithm, forms the portrait feature optimization set, and obtains low-dimensional sparse features by using SAE dimension reduction; the customer portrait representation learning module generates a shared representation vector, on which each customer portrait task tower is set; the graph portrait output module generates a labeled portrait through dynamic optimization of reinforcement learning, and constructs a graph portrait by combining knowledge graph to complete entity relationship, the system accurately depicts the characteristics of hospital customers such as equipment failure risk, maintenance renewal willingness, spare parts demand and training demand, and provides data support for medical equipment service.

Description

Technical Field

[0001] This invention relates to the field of customer profiling, specifically to a customer profiling system based on machine learning. Background Technology

[0002] In the medical device service sector, accurate customer profiling is crucial for improving service quality and optimizing resource allocation. Traditional customer profiling methods have the following limitations:

[0003] Insufficient feature extraction capabilities: Features are often generated through manual screening or simple statistical methods, making it difficult to capture the complex relationships between multimodal features in medical data. Feature redundancy or missing features are prominent, affecting the accuracy of profiling. One-sided and weakly correlated profiling dimensions: Focusing on single-dimensional descriptions, there is a lack of systematic integration of multi-dimensional features such as equipment failure risk, maintenance renewal intention, spare parts demand, and training needs. Furthermore, the intrinsic relationships between features are not revealed, making it difficult to support refined service decisions. Poor model adaptability and interpretability: Traditional statistical models or single algorithms are difficult to handle the nonlinear and high-dimensional characteristics of medical data. Moreover, the profiling results are mostly numerical outputs, lacking intuitive labeling or graphical presentation, making it difficult for business personnel to understand and apply. Therefore, a customer profiling system based on machine learning is needed to solve the above problems. Summary of the Invention

[0004] To address the technical problems raised in the background, this invention provides a customer profiling system based on machine learning.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This invention provides a customer profile building system based on machine learning, including a multi-source data acquisition module, a multimodal feature extraction module, a customer profile representation learning module, and a graph profile output module.

[0007] The multi-source data acquisition module collects data from hospitals and manufacturers' cloud-based IoT devices and clinical texts. It processes this data using hash desensitization and graph anomaly detection algorithms, and then integrates the processed data into a customer profile information repository. The specific process is as follows:

[0008] Real-time IoT operation data of each device is collected by various types of sensors installed in each medical testing device.

[0009] The HL7 gateway connects to the hospital's communication terminal to identify and extract the required clinical text data, including surgical reports and quality control records.

[0010] The maintenance records in the manufacturer's cloud are extracted to include maintenance time, maintenance content, and processing results. The corresponding fault codes are entered to form complete maintenance log data.

[0011] Connect to hospital websites and other platforms to obtain bidding documents and data in real time;

[0012] Real-time acquisition of external data from various hospitals, including hospital level information and financial allocation data;

[0013] Identify fields requiring anonymization from data streams including device IoT operation data, clinical text data, maintenance log data, bidding and tendering text data, and external data. The fields that need to be anonymized include patient ID, doctor's name, contact information, biometric information, and hospital number. A random string is generated for each hospital and updated periodically to serve as the hospital's salt value. The original sensitive value in each record is concatenated with the hospital's salt value as a string, and the SHA-512 algorithm is called to calculate and generate a hash string, which is then overwritten with the original field to obtain the desensitized data.

[0014] The graph anomaly detection algorithm loads the anonymized data into a heterogeneous graph. heterogeneous graph node set Includes equipment, hospitals, fault codes, and spare parts entities, edge collection This includes relationships between entities such as the hospital to which the equipment belongs, the equipment malfunction, and the corresponding spare parts; all nodes are converted into node embedding vectors, representing the node's features and location information, using the following conversion formula: ,in For nodes Embedded vector, For embedded functions, This refers to the node number; the anomaly score value for each node is obtained using the anomaly scoring function. Its calculation logic is as follows: ,in represents the connection weight between a node and its neighboring nodes, reflecting the strength of their association; j is the neighboring node number. For nodes The set of neighboring nodes, Let it be a distance function. for Obtain the anomaly score; get the current anomaly threshold. Its calculation logic is as follows: ,in The average of the anomaly scores for all nodes. To adjust the coefficient, The standard deviation of the anomaly scores for all nodes; if the anomaly score of any node is ≥ If the condition is met, the node is marked with an abnormal label; otherwise, the node is considered normal.

[0015] Align normal nodes and abnormal labels to the timeline to generate a standardized Parquet table. Using hospitals and equipment as the granularity, integrate IoT statistics, clinical text keywords, maintenance records, bidding documents, and external data to form structured profile entries. Map these profile entries to the customer profile information hub in real time.

[0016] The multimodal feature extraction module generates the most fitness-optimized combination features through a genetic programming algorithm, integrates them into a portrait feature optimization set, and uses a sparse autoencoder (SAE) for feature dimensionality reduction. The specific process is as follows:

[0017] Several basic profile features are randomly extracted from the customer profile information repository. Any two basic profile features are combined with mathematical operators to construct several profile feature combinations (XR). The current multi-level target variable (YG) is obtained, corresponding to the hospital's maintenance demand level, which includes low, medium, and high demand levels. The system is then grouped by level, and the feature mean of each group is calculated. and sample size ,in Given the grouping numbers, calculate the cumulative probability of each group, and the cumulative probability of the low-demand group. The formula is: NT represents the total sample size of the three groups, and the cumulative probability of the low demand level group and the high demand level group is obtained sequentially. and Look up the standard normal distribution table to get the corresponding cumulative probability. Normality score, obtained through the formula for calculating correlation coefficients of multiple columns, yields the correlation between the combined profile feature value and multi-level target variables. Its calculation logic is as follows: ,in The combined feature mean of all samples. The combined characteristic of all samples, the total standard deviation, when The closer the value is to 1, the stronger the correlation.

[0018] Input the combined values ​​of the portrait features into a preset linear regression model. Its expression is ,in The intercept term represents the baseline value of the target variable when the eigenvalue is 0. For regression coefficients, For the random error term, the absolute value of the regression coefficient is divided by the maximum coefficient value of all combinations of features to be evaluated to obtain the relative weight score;

[0019] A domain knowledge rule base is introduced, such as the requirement that the temperature-related features of the X-ray tube must be monotonically correlated with the rated withstand temperature of the equipment, and the product feature of the number of exposures must meet the radiation safety threshold constraints. The combined features are scored for compliance, with a compliance score of 0-10.

[0020] The fitness score for each combination of profile features is obtained by weighting and summing the relevance score, relative weight score, and compliance score.

[0021] Image feature combinations with fitness scores greater than the preset score are retained and recorded as candidate sets. The retained candidate sets are sorted in descending order of fitness scores. The top 30% of the candidate sets are selected as parent sets for crossover operations, and the rest are used as child sets. The basic features of the parent and child sets are randomly changed for updates. The top ten candidate sets with the highest fitness scores are selected as the preferred image feature set YD.

[0022] Sparse Autoencoder (SAE) Dimensionality Reduction: The optimal set of profile features is mapped to the low-dimensional space of the SAE, and the ReLU activation function is used to introduce nonlinearity to obtain the dimensionality-reduced low-dimensional sparse features. Its formula is expressed as: ,in and The encoder weights and biases are given. It should be noted that the sparse autoencoder can compress high-dimensional medical multimodal features to 20 dimensions, which significantly reduces data complexity while retaining important information for tasks such as equipment failure prediction and maintenance requirement classification.

[0023] The customer profiling representation learning module uses the cross-modal Med-Crossformer as a shared underlying layer, sets up various customer profiling task stacks, and selects corresponding algorithms to obtain accurate customer profiles. The specific process is as follows:

[0024] Med-Crossformer is a Transformer-based cross-modal shared encoder that consists of four Transformer encoder layers. Each layer includes a multi-head self-attention mechanism and a feedforward network, embedding low-dimensional sparse features into the input sequence of the Transformer. After passing through the four encoder layers, a shared representation vector is generated, which serves as the shared bottom layer.

[0025] On the shared underlying layer, each customer profile task tower is set up sequentially. Each customer profile task tower includes equipment failure probability tower, maintenance renewal willingness tower, spare parts demand prediction tower, and training demand level tower.

[0026] The equipment failure probability tower is used to predict the probability of hospital medical equipment failure within the next 7 days. A gradient boosting tree algorithm is introduced, where shared representation vectors are input into the equipment failure probability tower. The contribution of the shared representation vectors to failure prediction is calculated using the Gini index. Shared representation vectors with contributions below a preset threshold are removed, and the remaining representation vectors are labeled as failure feature vectors (GL). A base prediction value of 0.5 is set. Each decision tree (out of a total of 1000 trees) is traversed, and each tree makes a prediction based on the failure feature vector, outputting a corresponding residual value. This residual value is then multiplied by the learning rate. To control the influence of individual trees, the corrected residual values ​​are added to the current prediction score. After several iterations, a linear score is obtained by combining the prediction results of all trees. This linear score is then input into the Sigmoid function to convert it into a prediction probability value. ,like A value greater than 0.5 indicates a high probability of failure. A value less than 0.5 indicates a low probability of failure. The calculation logic for its predicted probability value is as follows: Where m is the index of the decision tree, bs is the global bias term, and exp is the natural exponential function. Let be the predicted value of the fault feature vector by the m-th decision tree;

[0027] A causal forest algorithm is introduced into the maintenance renewal intention tower to assess hospitals' willingness to renew maintenance services. A shared representation vector is input into the maintenance renewal intention tower to identify the intention variable GQ, which has a potential causal relationship with maintenance renewal intention. Logistic regression is then used to predict the hospital's propensity score to accept customized maintenance plans. Its calculation logic is as follows: ,in For the treatment variable, T=1 represents acceptance, T=0 represents non-acceptance, and PO is a probability function; 500 decision trees are generated, each growing on a randomly selected subset of features, and the average treatment effect is calculated for each leaf node. Its calculation logic is as follows: ,in For expectation operator, and To handle the variable's values ​​for accepting and not accepting, the intention variable, propensity score, and average treatment effect are concatenated and then passed through a two-layer fully connected network and an output layer to obtain the renewal intention score.

[0028] A label power set learning algorithm is introduced into the spare parts demand prediction tower. Based on the shared representation vector, the demand for various spare parts of hospitals or departments in the future period is predicted, and the output includes demand labels for various spare parts. Specifically, the shared representation vector is input into the spare parts demand prediction tower, and the demand probability of various spare parts is mapped to several effective combinations. The probability distribution of the corresponding combinations is output through the LightGBM classifier. The top three combinations with the highest probability are selected, and the frequency of each spare part in these three combinations is calculated. When the spare part frequency is ≥0.6, it is marked as needed and the label is 1. Otherwise, it is marked as not needed and its label is 0. The predicted spare parts demand label vector is statistically analyzed.

[0029] The contribution coefficient of each feature in the shared vector to the training demand level is calculated using the Spearman correlation coefficient. The training demand levels include Level 1, Level 2, and Level 3 training demand, with Level 1 training demand < Level 2 < Level 3. Features with contribution coefficients greater than a preset threshold are selected and marked as the training feature set. The training feature set is input into the training demand hierarchy, which is equipped with a fuzzy logic inference algorithm to define a fuzzy set for each feature in the training feature set. Fuzzy sets include low, medium, and high membership functions. The calculation logic for determining the degree to which an input value belongs to a corresponding fuzzy set is as follows: Where b, c, and d are boundary parameters of the fuzzy set; a fuzzy rule base is established, containing several rule combinations; the activation strength of each rule is calculated, and the minimum value of the input membership degree is taken. The calculation logic is as follows: Where g is the rule number, G is the total number of rules, and activation intensity represents the weight of the rule's influence on the output; the centroid method is used to convert the fuzzy output into a clear level, and the current training demand level is mapped through the clear level. If the clear level < 1.5, it is a level 1 training demand; if 1.5 ≤ clear level ≤ 2.5, it is a level 2 training demand; and if 2.5 ≥ 2.5, it is a level 3 training demand.

[0030] The graph profiling module dynamically optimizes and generates labeled profiles using reinforcement learning algorithms, completes entity relationships in the medical knowledge graph, and outputs graph profiles. The specific process is as follows:

[0031] The state is set to integrate the outputs of four task towers, the action is the generated tag combination, the reward is the business feedback after the tag is generated, and the strategy is the rule of selecting the optimal tag combination h based on the current state s.

[0032] A set of label policies is randomly generated. Through reinforcement learning, the policy is continuously tilted towards generating labels with higher rewards. The calculation logic is as follows: ,in Given state s, the value score of the selected label combination h is determined by the score; a higher score indicates a better result for that label. To enhance the learning rate, The immediate reward for the current tag combination. As a discount factor, To determine the maximum future value obtainable under the new state s' after labeling, the optimal equipment failure risk label, optimal maintenance renewal willingness label, optimal spare parts demand list label, and optimal training demand level label are output. The above labels are then summarized to form a structured text.

[0033] Logic for generating graphical representations: ;

[0034] in To create a complete image of the atlas. The core entity is TA, the feature type entity is VA, the feature value entity is TE, and the related entity is TE. , and The corresponding relationships are defined as follows: the core entity is set as "customer," the feature type entities are "fault risk," "renewal intention," "spare parts requirement," and "training level," and the feature value entities are "high," "medium," "low," "CT tube," and "detector," etc. Entities and relationships are extracted from the outputs of the four task towers and transformed into graph triples. All extracted triples are summarized to form the basic data of the graph. Then, a medical knowledge graph is constructed using Neo4j storage. Graph profiles are obtained through the outputs of the real-time task towers. The structured text and graph profiles are matched to a preset template to obtain a medical customer profile template, which is then sent to the user port.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] Comprehensive profile with high prediction accuracy: Using the cross-modal Med-Crossformer as a shared underlying layer, it achieves accurate prediction of multi-dimensional features through four task towers: Equipment failure probability tower (gradient boosting tree) improves the accuracy of failure risk prediction; maintenance renewal willingness tower (causal forest) explores the potential causal relationship of renewal behavior; spare parts demand prediction tower (label power set learning) accurately locates the spare parts that need to be stocked; training demand level tower (fuzzy logic reasoning) realizes the hierarchical quantification of training needs. Multi-dimensional integration makes the customer profile more three-dimensional.

[0037] The profile output is intuitive and highly adaptable to business needs: By dynamically optimizing the tag profile through reinforcement learning, and combining it with the medical knowledge graph to build the graph profile, it is finally integrated into a structured template. The tag profile facilitates a quick understanding of the customer's core characteristics, while the graph profile reveals the deep relationships between characteristics, such as the relationship between "high failure risk" and "urgent maintenance needs". Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0039] Figure 1 This is a block diagram showing the module connection of the present invention. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0041] Please refer to Figure 1 As shown, the present invention provides a customer profile construction system based on machine learning, including a multi-source data acquisition module, a multi-modal feature extraction module, a customer profile representation learning module, and a graph profile output module;

[0042] The multi-source data acquisition module collects data from hospitals and manufacturers' cloud-based IoT devices and clinical texts. It processes this data using hash desensitization and graph anomaly detection algorithms, and then integrates the processed data into a customer profile information repository. The specific process is as follows:

[0043] Real-time IoT operation data of each device is collected by various types of sensors installed in each medical testing device.

[0044] The HL7 gateway connects to the hospital's communication terminal to identify and extract the required clinical text data, including surgical reports and quality control records.

[0045] The maintenance records in the manufacturer's cloud are extracted to include maintenance time, maintenance content, and processing results. The corresponding fault codes are entered to form complete maintenance log data.

[0046] Connect to hospital websites and other platforms to obtain bidding documents and data in real time;

[0047] We acquire external data from various hospitals in real time. This external data includes hospital grading information and financial allocation data. It should be noted that the hospital grading information is extracted from the publicly available hospital grading assessment results from authoritative health institutions. These results are released through official channels and cover the grading status of different hospitals. They are an important indicator reflecting the comprehensive strength of hospitals. Financial allocation data is obtained by querying the publicly available fiscal budget information released by relevant local government departments. Some data can be provided by hospitals with authorized allocation information.

[0048] Identify fields requiring anonymization from data streams including device IoT operation data, clinical text data, maintenance log data, bidding and tendering text data, and external data. The fields that need to be anonymized include patient ID, doctor's name, contact information, biometric information, and hospital number. A random string is generated for each hospital and updated periodically to serve as the hospital's salt value. The original sensitive value in each record is concatenated with the hospital's salt value as a string. The SHA-512 algorithm is then used to calculate the hash string, which overwrites the original field, resulting in the de-identified data. The generation logic is as follows: All downstream modules use only hashed pseudonyms to ensure no real personal information is leaked. It should be noted that the original sensitive value refers to the original data value in medical data that has not undergone de-identification processing and can directly or indirectly identify a specific patient, or involves patient privacy and meets the definition of sensitive information in regulations.

[0049] The graph anomaly detection algorithm loads the anonymized data into a heterogeneous graph. heterogeneous graph node set Includes entities such as equipment, hospitals, fault codes, and spare parts, and is used as an edge collection. This includes relationships between entities such as the hospital to which the equipment belongs, the equipment malfunction, and the corresponding spare parts; all nodes are converted into node embedding vectors, representing the node's features and location information, using the following conversion formula: ,in For nodes Embedded vector, For embedded functions, This refers to the node number; the anomaly score value for each node is obtained using the anomaly scoring function. Its calculation logic is as follows: ,in represents the connection weight between a node and its neighboring nodes, reflecting the strength of their association; j is the neighboring node number. For nodes The set of neighboring nodes, Let it be a distance function. for The anomaly score is used to determine the likelihood of a node being abnormal; a higher score indicates a more likely abnormal node. The current anomaly threshold is then obtained. Its calculation logic is as follows: ,in The average of the anomaly scores for all nodes. To adjust the coefficient, The standard deviation of the anomaly scores for all nodes; if the anomaly score of any node is ≥ If the condition is met, the node is marked with an abnormal label; otherwise, the node is considered normal.

[0050] Normal nodes and abnormal labels are aligned to the timeline to generate a standardized Parquet table. Using hospitals and equipment as the granularity, IoT statistics, clinical text keywords, maintenance records, bidding documents, and external data are integrated to form structured profile entries. These profile entries are then mapped in real time to the customer profile information hub for direct use by downstream feature extraction and profile modeling.

[0051] The multimodal feature extraction module generates the most fitness-optimized combination features through a genetic programming algorithm, integrates them into a portrait feature optimization set, and uses a sparse autoencoder (SAE) for feature dimensionality reduction. The specific process is as follows:

[0052] Several basic profile features are randomly extracted from the customer profile information repository, such as X-ray tube temperature and exposure count in IoT devices, surgical duration in clinical texts, and fault frequency in maintenance logs. Any two basic profile features are combined with mathematical operators to construct several profile feature combinations (XR), such as X-ray tube temperature × exposure count, fault frequency ÷ equipment service life, etc. The current multi-level target variable YG is obtained, corresponding to the hospital's maintenance demand level, which includes low, medium, and high demand levels. The groups are then grouped by level, and the feature mean of each group is calculated. and sample size ,in Given the grouping numbers, calculate the cumulative probability of each group, and the cumulative probability of the low-demand group. The formula is: NT represents the total sample size of the three groups, and the cumulative probability of the low demand level group and the high demand level group is obtained sequentially. and Look up the standard normal distribution table to get the corresponding cumulative probability. Normality score, obtained through the formula for calculating correlation coefficients of multiple columns, yields the correlation between the combined profile feature value and multi-level target variables. Its calculation logic is as follows: ,in The combined feature mean of all samples. The combined characteristic of all samples, the total standard deviation, when The closer the value is to 1, the stronger the correlation.

[0053] Input the combined values ​​of the portrait features into a preset linear regression model. Its expression is ,in The intercept term represents the baseline value of the target variable when the eigenvalue is 0. For regression coefficients, For the random error term, the absolute value of the regression coefficient is divided by the maximum coefficient value of all combinations of features to be evaluated to obtain the relative weight score;

[0054] A domain knowledge rule base is introduced, such as the requirement that the temperature-related features of the X-ray tube must be monotonically correlated with the rated withstand temperature of the equipment, and the product feature of the number of exposures must meet the radiation safety threshold constraints. The combined features are scored for compliance, with a compliance score of 0-10.

[0055] The fitness score for each combination of profile features is obtained by weighting and summing the relevance score, relative weight score, and compliance score.

[0056] Image feature combinations with fitness scores greater than the preset score are retained and recorded as candidate sets. The retained candidate sets are sorted in descending order of fitness scores. The top 30% of the candidate sets are selected as parent sets for crossover operations, and the rest are used as child sets. The basic features of the parent and child sets are randomly changed for updates. The top ten candidate sets with the highest fitness scores are selected as the preferred image feature set YD.

[0057] Sparse Autoencoder (SAE) Dimensionality Reduction: The optimal set of profile features is mapped to the low-dimensional space of the SAE, and the ReLU activation function is used to introduce nonlinearity to obtain the dimensionality-reduced low-dimensional sparse features. Its formula is expressed as: ,in and The encoder weights and biases are given. It should be noted that the sparse autoencoder can compress high-dimensional medical multimodal features to 20 dimensions, which significantly reduces data complexity while retaining important information for tasks such as equipment failure prediction and maintenance requirement classification.

[0058] The customer profiling representation learning module uses the cross-modal Med-Crossformer as a shared underlying layer, sets up various customer profiling task stacks, and selects corresponding algorithms to obtain accurate customer profiles. The specific process is as follows:

[0059] Med-Crossformer is a Transformer-based cross-modal shared encoder that consists of four Transformer encoder layers. Each layer includes a multi-head self-attention mechanism and a feedforward network, embedding low-dimensional sparse features into the input sequence of the Transformer. After passing through the four encoder layers, a shared representation vector is generated, which serves as the shared bottom layer.

[0060] On the shared underlying layer, each customer profile task tower is set up sequentially. Each customer profile task tower includes equipment failure probability tower, maintenance renewal willingness tower, spare parts demand prediction tower, and training demand level tower.

[0061] The equipment failure probability tower is used to predict the probability of hospital medical equipment failing within the next 7 days. The output is a value between 0 and 1, with values ​​closer to 1 indicating a higher probability of failure. Specifically, a gradient boosting tree algorithm is introduced. Shared representation vectors are input into the equipment failure probability tower, and the contribution of these vectors to failure prediction is calculated using the Gini index. Shared representation vectors with a contribution below a preset threshold are removed, and the remaining vectors are labeled as failure feature vectors (GL). A base prediction value of 0.5 is set. Each decision tree (out of a total of 1000 trees) is traversed, and each tree makes a prediction based on the failure feature vector, outputting a corresponding residual value. This residual value is then multiplied by the learning rate. To control the influence of individual trees, the corrected residual values ​​are added to the current prediction score. After several iterations, a linear score is obtained by combining the prediction results of all trees. This linear score is then input into the Sigmoid function to convert it into a prediction probability value. ,like A value greater than 0.5 indicates a high probability of failure. A value less than 0.5 indicates a low probability of failure. The calculation logic for its predicted probability value is as follows: Where m is the index of the decision tree, bs is the global bias term, and exp is the natural exponential function. Let be the predicted value of the fault feature vector by the m-th decision tree;

[0062] A causal forest algorithm is introduced into the maintenance renewal intention tower to assess hospitals' willingness to renew maintenance services. A shared representation vector is input into the maintenance renewal intention tower to identify the intention variable GQ, which has a potential causal relationship with maintenance renewal intention. Logistic regression is then used to predict the hospital's propensity score to accept customized maintenance plans. Its calculation logic is as follows: ,in For the treatment variable, T=1 represents acceptance, T=0 represents non-acceptance, and PO is a probability function; 500 decision trees are generated, each growing on a randomly selected subset of features, and the average treatment effect is calculated for each leaf node. Its calculation logic is as follows: ,in For expectation operator, and To handle the variable's values ​​for accepting and not accepting, the intention variable, propensity score, and average treatment effect are concatenated and then passed through a two-layer fully connected network and an output layer to obtain the renewal intention score.

[0063] A label power set learning algorithm is introduced into the spare parts demand prediction tower. Based on the shared representation vector, the demand for various spare parts of hospitals or departments in the future period is predicted, and the output includes demand labels for various spare parts. Specifically, the shared representation vector is input into the spare parts demand prediction tower, and the demand probability of various spare parts is mapped to several effective combinations. The probability distribution of the corresponding combinations is output through the LightGBM classifier. The top three combinations with the highest probability are selected, and the frequency of each spare part in these three combinations is calculated. When the spare part frequency is ≥0.6, it is marked as needed and the label is 1. Otherwise, it is marked as not needed and its label is 0. The predicted spare parts demand label vector is statistically analyzed.

[0064] The purpose of the training needs hierarchy is to accurately classify the training needs of hospitals or departments by analyzing the differences between their actual capabilities and needs in areas such as the operation and maintenance of medical equipment. Specifically, the contribution coefficient of each feature in the shared vector to the training needs hierarchy is calculated using the Spearman correlation coefficient. The training needs hierarchy includes Level 1, Level 2, and Level 3 training needs, with Level 1 training needs < Level 2 < Level 3. Features with contribution coefficients greater than a preset threshold are selected and marked as the training feature set. The training feature set is input into the training demand hierarchy, which is equipped with a fuzzy logic inference algorithm to define a fuzzy set for each feature in the training feature set. Fuzzy sets include low, medium, and high membership functions. The calculation logic for determining the degree to which an input value belongs to a corresponding fuzzy set is as follows: Where b, c, and d are boundary parameters of the fuzzy set; establish a fuzzy rule base, which contains several rule combinations. For example, if the operation assessment score is low and the number of errors is high, the training demand is high; if the operation assessment score is medium and the frequency of new technologies is medium, the training demand is medium; if the operation assessment score is high and the number of errors is low, the training demand is low; calculate the activation strength of each rule and take the minimum value of the input membership degree. The calculation logic is as follows: Where g is the rule number, G is the total number of rules, and activation intensity represents the weight of the rule's influence on the output; the centroid method is used to convert the fuzzy output into a clear level, and the current training demand level is mapped through the clear level. If the clear level < 1.5, it is a level 1 training demand; if 1.5 ≤ clear level ≤ 2.5, it is a level 2 training demand; and if 2.5 ≥ 2.5, it is a level 3 training demand.

[0065] The graph profiling module dynamically optimizes and generates labeled profiles using reinforcement learning algorithms, completes entity relationships in the medical knowledge graph, and outputs graph profiles. The specific process is as follows:

[0066] The label generation is dynamically optimized through reinforcement learning algorithms. Specifically, the state is set as the output integration of four task towers, the action is the generated label combination, such as high failure risk and high training need, and the reward is the business feedback after the label is generated. For example, if a customer actually renews after generating a label with high renewal potential, the reward is +10 points, and if the customer does not renew, the reward is -5 points. The strategy is to select the optimal label combination h according to the current state s, and the goal is to maximize long-term rewards.

[0067] A set of label policies is randomly generated. Through reinforcement learning, the policy is continuously tilted towards generating labels with higher rewards. The calculation logic is as follows: ,in Given state s, the value score of the selected label combination h is determined by the score; a higher score indicates a better result for that label. To enhance the learning rate, The immediate reward for the current tag combination. As a discount factor, To determine the maximum future value obtainable under the new state s' after labeling, the optimal equipment failure risk label, optimal maintenance renewal intention label, optimal spare parts demand list label, and optimal training demand level label are output. These labels are then summarized to form a structured text, such as the XX Hospital customer profile: High failure risk (maintenance recommended within 72 hours), Medium renewal intention (account manager follow-up required), Spare parts need to be stocked (CT tube - emergency, detector - routine), Training demand level: High (specialized training recommended).

[0068] Logic for generating graphical representations: ;

[0069] in To create a complete image of the atlas. The core entity is TA, the feature type entity is VA, the feature value entity is TE, and the related entity is TE. , and The corresponding relationships are defined as follows: the core entity is set as "customer," the feature type entities are "fault risk," "renewal intention," "spare parts requirement," and "training level," and the feature value entities are "high," "medium," "low," "CT tube," and "detector," etc. Entities and relationships are extracted from the outputs of the four task towers and transformed into graph triples. All extracted triples are summarized to form the basic data of the graph. Then, a medical knowledge graph is constructed using Neo4j storage. Graph profiles are obtained through the outputs of the real-time task towers. The structured text and graph profiles are matched to a preset template to obtain a medical customer profile template, which is then sent to the user port.

[0070] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization (e.g., Min-Max scaling to map medical device IoT data and maintenance log features to the [0,1] interval), which will not be elaborated here. The formulas are obtained by software simulation through the collection of massive amounts of hospital equipment operation data, clinical operation records, and maintenance service feedback, fitting the calculation logic that is closest to the real scenario. The preset parameters in the formulas (such as the number of decision trees in the fault probability tower, the clarity level threshold of the training demand level, etc.) are set by those skilled in the art based on the actual data distribution of the medical device service scenario.

[0071] The above embodiments can be implemented, in whole or in part, through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs that, when loaded or executed on a computer, generate, in whole or in part, the processes or functions described in the embodiments of the present invention (such as multi-source data desensitization processing, multimodal feature extraction, customer profile task tower prediction, and map profile generation, etc.). The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices (such as an embedded processor in a medical device service terminal).

[0072] The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one hospital data center, vendor server, or cloud platform to another via wired (e.g., fiber optic, Ethernet) or wireless (e.g., infrared, wireless LAN, microwave) means. The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that contains one or more sets of available media. The available media can be magnetic media (e.g., hard disk, magnetic tape), optical media (e.g., optical disc), or semiconductor media (e.g., solid-state drive), suitable for storing multi-source data, feature vectors, and profiling results during the medical customer profile construction process.

[0073] It should be understood that in the various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic (e.g., multi-source data collection and desensitization need to be completed first, then feature extraction is performed, and finally a customer profile is generated), and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0074] Those skilled in the art will recognize that the units and algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein (such as genetic programming to generate combinatorial features, gradient boosting trees to predict fault probabilities, reinforcement learning to optimize label strategies, etc.) can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution (such as the real-time requirements of medical data, the accuracy requirements of customer profiling, etc.). Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0075] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules (multi-source data acquisition module, multi-modal feature extraction module, etc.) is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or modules shown or discussed may be through some interfaces. The indirect coupling or communication connection between the devices or modules may be electrical, mechanical, or other forms (such as the transmission of IoT data between hospital equipment sensors and data acquisition modules via wired communication).

[0076] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs (e.g., depending on the size of the hospital customer, the spare parts demand prediction tower or training demand level tower may be selectively activated).

[0077] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0078] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention (such as multi-source data processing, feature extraction, image generation, etc.). The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention (such as adjusting the algorithm parameters for feature extraction, optimizing the entity relationship rules for graph profiling, etc.) should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A customer profiling construction system based on machine learning, comprising a multi-source data acquisition module, a multi-modal feature extraction module, a customer profiling representation learning module, and a graph-based profiling output module, characterized in that: The multi-source data acquisition module collects IoT device data and clinical text data from hospitals and manufacturers' cloud platforms. It processes the data using hash desensitization algorithms and graph anomaly detection algorithms, and integrates the processed data into a customer profile information database. The multimodal feature extraction module generates the combination features with the highest fitness through a genetic programming algorithm, integrates them into a portrait feature optimization set, and uses a sparse autoencoder (SAE) for feature dimensionality reduction. The customer profile representation learning module uses the cross-modal Med-Crossformer as a shared underlying layer, sets up equipment failure probability tower, maintenance renewal willingness tower, spare parts demand prediction tower, and training demand level tower, and selects the corresponding algorithm to obtain accurate customer profiles. The graph profiling module dynamically optimizes and generates tagged profiles through reinforcement learning algorithms, while the medical knowledge graph completes the entity relationships and outputs graph profiles. The two are integrated into a medical customer profile template and sent to the user's end.

2. The customer profiling system based on machine learning according to claim 1, characterized in that, The customer profile representation learning module uses the cross-modal Med-Crossformer as a shared underlying layer, sets up various customer profile task stacks, and selects corresponding algorithms to obtain accurate customer profiles. The specific process is as follows: Low-dimensional sparse features are embedded into the input sequence of the Transformer. After passing through 4 layers of encoders, a shared representation vector is generated. The shared representation vector serves as the shared bottom layer. On the shared bottom layer, task towers for each customer profile are set up in sequence. Each customer profile task tower includes a tower for equipment failure probability, a tower for maintenance renewal willingness, a tower for spare parts demand prediction, and a tower for training demand level.

3. The customer profiling system based on machine learning according to claim 2, characterized in that, The steps for the customer profile representation learning module to obtain the predicted probability value and renewal intention score are as follows: A gradient boosting tree algorithm is introduced. Shared representation vectors are input into the device fault probability tower. The contribution of shared representation vectors to fault prediction is calculated using the Gini index. Shared representation vectors with a contribution lower than the preset value are removed, and the remaining representation vectors are marked as fault feature vectors. The base prediction value is set to 0.

5. Each decision tree is traversed, with a total of 1000 trees. Each tree makes a prediction based on the fault feature vector and outputs the corresponding residual value. The residual value is multiplied by the learning rate for correction, and then the corrected residual value is accumulated into the current prediction score. After several iterations, a linear score is obtained by combining the prediction results of all trees. The linear score is input into the Sigmoid function to convert it into a prediction probability value. A causal forest algorithm is introduced into the maintenance and renewal intention tower. The shared representation vector is input into the maintenance and renewal intention tower to identify intention variables that have a potential causal relationship with maintenance and renewal intentions. Logistic regression is used to predict the hospital's propensity score for accepting customized maintenance and renewal plans. 500 decision trees are generated, each growing on a randomly selected subset of features. The average treatment effect is calculated for each leaf node. The intention variables, propensity scores and average treatment effects are concatenated and the renewal intention score is obtained through a two-layer fully connected network and an output layer.

4. The customer profiling system based on machine learning according to claim 3, characterized in that, The specific process of the customer profile representation learning module obtaining spare parts demand label vectors and training demand levels is as follows: A label power set learning algorithm is introduced into the spare parts demand prediction tower. The shared representation vector is input into the spare parts demand prediction tower, and the demand probability of various spare parts is mapped to several effective combinations. The probability distribution of the corresponding combinations is output by the LightGBM classifier. The top three combinations with the highest probability are selected, and the frequency of each spare part in these three combinations is calculated. When the spare part frequency is ≥0.6, it is marked as needed and the label is 1. Otherwise, it is marked as not needed and its label is 0. The predicted spare parts demand label vector is statistically analyzed. The contribution coefficient of each feature of the shared vector to the training demand level is calculated using the Spearman correlation coefficient. Features with a contribution coefficient greater than the preset value are selected and marked as the training feature set. The training feature set is input into the training demand level tower, which is equipped with a fuzzy logic inference algorithm. Each feature of the training feature set is defined as a fuzzy set, and the degree to which the input value belongs to the corresponding fuzzy set is calculated using the membership function. A fuzzy rule base is established, which contains several rule combinations. The activation strength of each rule is calculated, and the minimum value of the input membership degree is taken. The center of gravity method is then used to convert the fuzzy output into a clear level. The clear level is then mapped to the current training requirement level. If the clear level is less than 1.5, it is a level 1 training requirement. If 1.5 ≤ clear level ≤ 2.5, it is a level 2 training requirement. If 2.5 ≥ 2.5, it is a level 3 training requirement.

5. The customer profiling system based on machine learning according to claim 1, characterized in that, The image mapping module dynamically optimizes and generates labeled images through reinforcement learning algorithms. The specific process is as follows: The state is set to integrate the outputs of four task towers, the action is the generated tag combination, the reward is the business feedback after the tag is generated, and the strategy is the rule of selecting the optimal tag combination h based on the current state s. A set of label policies is randomly generated, and the policy is biased towards high-reward labels using a reinforcement learning formula. The calculation logic is as follows: ,in Given state s, the value score of the selected label combination h is determined by the score; a higher score indicates a better result for that label. To enhance the learning rate, The immediate reward for the current tag combination. As a discount factor, To determine the maximum future value obtainable under the new state s' after labeling, the optimal equipment failure risk label, optimal maintenance renewal willingness label, optimal spare parts demand list label, and optimal training demand level label are output. The above labels are then summarized to form a structured text.

6. The customer profiling system based on machine learning according to claim 5, characterized in that, The graph profiling module completes entity relationships using a medical knowledge graph to output graph profiling, and integrates structured text to obtain medical customer profile templates, specifically as follows: Logic for generating graphical representations: ; in To create a complete image of the atlas. The core entity is TA, the feature type entity is VA, the feature value entity is TE, and the related entity is TE. , and The corresponding relationships are defined as follows: the core entity is set as "customer," the feature type entities are "fault risk," "renewal intention," "spare parts requirement," and "training level," and the feature value entities are "high," "medium," "low," "CT tube," and "detector." Entities and relationships are extracted from the outputs of the four task towers and transformed into graph triples. All extracted triples are aggregated to form the basic data of the graph. Then, a medical knowledge graph is constructed using Neo4j storage. Graph profiles are obtained through the outputs of the real-time task towers. The structured text and graph profiles are matched to a preset template to obtain a medical customer profile template, which is then sent to the user port.

7. The customer profiling system based on machine learning according to claim 1, characterized in that, The multimodal feature extraction module generates the combination features with the highest fitness through a genetic programming algorithm, and integrates them into a profile feature optimization set. The specific process is as follows: Several basic profile features are randomly extracted from the customer profile information collection. Any two basic profile features and mathematical operators are used to construct several profile feature combinations to obtain the current multi-level target variable, which corresponds to the maintenance demand level of the hospital. The maintenance demand level includes low demand, medium demand and high demand levels. The profile feature combination value and the multi-level target variable are obtained by grouping by level and calculating the feature mean and sample size of each group. The cumulative probability of each group is calculated in turn. By looking up the standard normal distribution table, the normality score of the corresponding cumulative probability is obtained. The correlation between the profile feature combination value and the multi-level target variable is obtained by using the formula for calculating the correlation coefficient of multiple columns. Input the combined values ​​of the portrait features into a preset linear regression model. Its expression is ,in For the intercept term, For regression coefficients, For the random error term, the absolute value of the regression coefficient is divided by the maximum coefficient value of all combinations of features to be evaluated to obtain the relative weight score; a domain knowledge rule base is introduced to score the compliance of the combined features. The fitness score for each combination of profile features is obtained by weighting and summing the relevance score, relative weight score, and compliance score. Image feature combinations with fitness scores greater than the preset score are retained and recorded as candidate sets. The retained candidate sets are sorted in descending order of fitness scores. The top 30% of the candidate sets are selected as parent sets for crossover operations, and the rest are used as child sets. The basic features of the parent and child sets are randomly changed for updates. The top ten candidate sets with the highest fitness scores are selected as the preferred image feature set YD.

8. The customer profiling system based on machine learning according to claim 7, characterized in that, The multimodal feature extraction module uses dimensionality reduction via a sparse autoencoder (SAE), specifically as follows: The optimized feature set of the image is mapped to the low-dimensional space of SAE, and the ReLU activation function is used to introduce nonlinearity to obtain the dimensionality-reduced low-dimensional sparse features. Its formula is expressed as: ,in and For encoder weights and biases.

Images