AI-based customer intelligent label extraction method and system

Abstract

This application relates to the fields of computer science and artificial intelligence, specifically to an AI-based method and system for extracting intelligent customer tags. The method comprises the following steps: collecting and preprocessing customer data, including behavior logs, transaction records, and text reviews; extracting features from the preprocessed customer data, outputting a feature set, and constructing behavioral pattern features; learning the behavioral pattern features based on a preset machine learning model to predict customer tags; and sorting and storing the predicted customer tags based on customer personalized weights and a preset sorting algorithm. This application is effective in accurately extracting customer tags.

Description

Technical Field

[0001] This application relates to the fields of computer science and artificial intelligence, and specifically to an AI-based customer intelligent label extraction method and system. Background Art

[0002] With the development of big data and artificial intelligence technologies, companies are increasingly focusing on the in-depth analysis and utilization of customer data. Intelligent customer tag extraction is a key component of customer relationship management (CRM). By mining customer behavior, attributes, preferences, and other data, it provides companies with customer profiles, thereby supporting marketing and customer service decisions.

[0003] Existing customer intelligent label extraction technology usually relies on simple statistical analysis methods, which has the following problems: simple statistical analysis is difficult to handle large-scale and complex data, and may not be able to accurately capture the deep characteristics of customers, resulting in low accuracy and relevance of labels.

[0004] Therefore, improvements are needed. Summary of the Invention

[0005] In order to solve the above technical problems, this application provides an AI-based customer intelligent label extraction method and system.

[0006] The first object of the invention of this application is achieved through the following technical solutions:

[0007] An AI-based customer intelligent tag extraction method includes the following steps:

[0008] Collect customer data and pre-process it;

[0009] Extract features from pre-processed customer data, output feature sets, and construct behavioral pattern features;

[0010] Perform feature selection on pre-processed customer data;

[0011] Based on preset multi-level feature screening rules, match data fields related to customer behavior from the preset feature library;

[0012] Analyze the correlation between each data field and customer behavior based on preset statistical methods;

[0013] When the correlation fluctuation coefficient exceeds the preset safety threshold, the manual review mechanism is triggered;

[0014] Select features that have passed the review and whose correlation is higher than the preset threshold as input features;

[0015] Based on the feature selection results, perform feature engineering on the selected features;

[0016] SA1: Determine the type of input features, which include numerical features, categorical features, and text features;

[0017] SA2: When the feature is numerical, calculate the statistics of the sliding window of the input feature, including mean, variance, and skewness, and output the time series derived features;

[0018] SA3: When it is a categorical feature, determine the categorical feature category, where the categorical feature category includes continuous features and discontinuous features;

[0019] SA4: When it is a non-continuous feature, hash the input feature, convert it into a numerical feature, and return to SA1;

[0020] SA5: When it is a continuous feature, the input feature is dynamically binned based on the information gain maximization criterion. When the number of bins is less than or equal to the preset bin threshold, it is converted to a non-continuous feature and returned to SA4;

[0021] SA6: When it is a text feature, perform the following processing: perform semantic vectorization based on the pre-trained word vector model and output the semantic vector;

[0022] SA7: Extract text sentiment polarity scores based on SnowNLP sentiment analysis algorithm;

[0023] SA8: Calculate the keyword TF-IDF weight based on the preset business keyword library;

[0024] SA9: Based on SA1-SA8, output a feature set, which includes time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights;

[0025] Construct behavioral pattern features based on the feature set processed by feature engineering;

[0026] Based on the customer ID and a preset time-decay aggregation rule, the feature set is aggregated to generate a customer-feature matrix, wherein the matrix element value is a standardized feature vector;

[0027] Based on the customer-feature matrix, construct a customer behavior graph network;

[0028] The customer behavior graph network includes nodes and edge weights, wherein the nodes are customer IDs, and the edge weights are similarities of the feature vectors calculated based on a preset improved cosine similarity algorithm;

[0029] Based on the preset NetworkX graph calculation library, the node betweenness centrality is calculated. When the node betweenness centrality exceeds the preset industry benchmark value, the node is identified as a key node;

[0030] Based on the customer-feature matrix, matching and extracting the feature vectors of the key nodes;

[0031] Based on the preset multi-head attention mechanism and the customer-feature matrix, the feature vector of the key node is fused with the feature vector of the current customer to output an attention-enhanced feature;

[0032] Outputting a behavioral pattern feature set, the behavioral pattern feature set including a customer-feature matrix, graph network structural features, and attention enhancement feature nodes, the graph network structural features including nodes, edge weights, and node betweenness centrality;

[0033] Learn behavioral pattern features based on a preset machine learning model and output label prediction results;

[0034] The behavioral pattern feature set is input into a preset machine learning model to extract the customer behavior pattern representation based on the following multimodal processing flow:

[0035] Generate topological correlation features based on graph neural network, customer-feature matrix, and graph network structure features;

[0036] Generate semantic features based on fully connected networks and attention-enhanced features;

[0037] Concatenate the topological association features and semantic features into vectors to generate a fused feature vector;

[0038] Map the fused feature vector to the label category space and output the customer label category;

[0039] Based on the temperature-scaled Softmax function, the customer label categories are calibrated, the standardized probability distribution is generated, and the probability values ​​are compared;

[0040] When the probability value is the largest, the label prediction result is output, and the label prediction result includes the customer ID and the customer label category;

[0041] Based on the customer's personalized weight and a preset sorting algorithm, the tag prediction results are sorted and stored.

[0042] In a preferred embodiment, the steps of collecting customer data and pre-processing include:

[0043] The customer data includes static data, dynamic data, and text data;

[0044] The static data includes basic customer information, which includes customer ID and registration time;

[0045] The dynamic data includes customer operation behavior data, which includes page visit frequency, function click times and stay time;

[0046] The text data includes customer work order text and social media comment text;

[0047] Based on the multidimensional time series of the customer operation behavior data, the standard deviation of the customer operation behavior data within a preset sliding time window is calculated.

[0048] In a preferred embodiment, the step of sorting and storing the tag prediction results based on customer personalized weights and a preset sorting algorithm includes:

[0049] Based on the behavioral pattern characteristics, assigning personalized weights to each tag prediction result;

[0050] Compare the personalized weight with the preset weight threshold to determine whether the corresponding tag prediction result can participate in the ranking;

[0051] When the personalized weight is lower than the preset weight threshold, the tag prediction result cannot be included in the ranking;

[0052] When the personalized weight is higher than the preset weight threshold, the tag prediction result can participate in the ranking;

[0053] Based on the order of personalized weights, the tag prediction results that can participate in the sorting are sorted and stored from high to low.

[0054] The second objective of this application is achieved through the following technical solutions:

[0055] An AI-based customer intelligent tag extraction, including:

[0056] Module 1: Collect customer data and pre-process;

[0057] The second module: extracts features from pre-processed customer data, outputs feature sets, and constructs behavioral pattern features;

[0058] Feature selection module: performs feature selection on pre-processed customer data;

[0059] The second matching module: matches data fields related to customer behavior from the preset feature library based on preset multi-level feature screening rules;

[0060] The first output module: Based on a preset statistical method, it analyzes the correlation between each data field and customer behavior and outputs the correlation fluctuation coefficient;

[0061] The first comparison module: when the correlation fluctuation coefficient exceeds the preset safety threshold, the manual review mechanism is triggered;

[0062] The first selection module: selects features that have passed the review and whose correlation is higher than the preset threshold as input features;

[0063] Feature engineering module: Based on the feature selection results, feature engineering is performed on the selected features;

[0064] The fifth module includes: determining the type of input features, which include numerical features, categorical features, and text features;

[0065] The sixth module includes: when it is a numerical feature, it calculates the statistics in the sliding window of the input feature, the statistics include mean, variance, skewness, and outputs the time series derived features;

[0066] The seventh module includes: when it is a categorical feature, determining the category of the categorical feature, the categorical feature category includes continuous features and discontinuous features;

[0067] First return module: When it is a non-continuous feature, the input feature is hashed and converted into a numerical feature, and then returned to SA1;

[0068] Second return module: When it is a continuous feature, the input feature is dynamically binned based on the information gain maximization criterion. When the number of bins is less than or equal to the preset bin threshold, it is converted into a non-continuous feature and returned to SA4;

[0069] The second output module: When it is a text feature, it performs the following processing: based on the pre-trained word vector model, it performs semantic vectorization and outputs the semantic vector;

[0070] The first extraction module: extracts the text sentiment polarity score based on the SnowNLP sentiment analysis algorithm;

[0071] The third calculation module: calculates the keyword TF-IDF weight based on the preset business keyword library;

[0072] The eighth module includes: based on SA1-SA8, outputting a feature set, wherein the feature set includes time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights;

[0073] Pattern feature module: constructs behavioral pattern features based on the feature set processed by feature engineering;

[0074] A first generation module: based on the customer ID and a preset time decay aggregation rule, aggregating the feature set to generate a customer-feature matrix, wherein the matrix element value is a standardized feature vector;

[0075] The first building module: constructing a customer behavior graph network based on the customer-feature matrix;

[0076] The ninth module includes: the customer behavior graph network includes nodes and edge weights, the nodes are customer IDs, and the edge weights are similarities of the feature vectors calculated based on a preset improved cosine similarity algorithm;

[0077] The second comparison module: calculates the node betweenness centrality based on the preset NetworkX graph calculation library, and identifies the node as a key node when the node betweenness centrality exceeds the preset industry benchmark value;

[0078] A third matching module: matching and extracting the feature vectors of the key nodes based on the customer-feature matrix;

[0079] The third output module: based on the preset multi-head attention mechanism and the customer-feature matrix, fuses the feature vector of the key node with the feature vector of the current customer, and outputs the attention enhancement feature;

[0080] The fourth output module outputs a behavioral pattern feature set, wherein the behavioral pattern feature set includes a customer-feature matrix, graph network structure features, and attention enhancement feature nodes. The graph network structure features include nodes, edge weights, and node betweenness centrality.

[0081] The third module: learns behavioral pattern features based on a preset machine learning model and outputs label prediction results;

[0082] The first input module inputs the behavioral pattern feature set into a preset machine learning model and extracts the customer behavior pattern representation based on the following multimodal processing flow:

[0083] The second generation module generates topological correlation features based on the graph neural network, customer-feature matrix, and graph network structure features;

[0084] The third generation module: Generates semantic features based on the fully connected network and attention-enhanced features;

[0085] The fourth generation module: concatenates the topological association features and semantic features to generate a fused feature vector;

[0086] The fifth output module maps the fused feature vector to the label category space and outputs the customer label category;

[0087] The fifth generation module: Based on the temperature-scaled Softmax function, calibrate the customer label categories, generate standardized probability distributions, and compare probability values;

[0088] The sixth output module: when the probability value is the largest, outputs the label prediction result, which includes the customer ID and the customer label category;

[0089] The fourth module: sorts and stores the tag prediction results based on customer personalized weights and a preset sorting algorithm.

[0090] The third objective of this application is achieved through the following technical solutions:

[0091] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned AI-based customer intelligent label extraction method are implemented.

[0092] The fourth objective of this application is achieved through the following technical solutions:

[0093] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned AI-based customer smart label extraction method.

[0094] In summary, this application has at least one of the following implementation effects:

[0095] First, customer data is carefully categorized into static data, dynamic data, and text data to ensure the targetedness and accuracy of subsequent processing. Static data covers basic customer information, such as customer ID and registration time, which provides a basis for understanding the customer's background. Dynamic data focuses on customer operational behavior, such as page visit frequency, function click count, and dwell time. These data reflect customer activity and interests. Text data includes customer ticket text and social media comment text, providing customer feedback and emotional tendencies. By calculating the standard deviation of customer operational behavior data within a preset sliding time window, this method can capture fluctuations and trends in customer behavior, providing a basis for constructing behavioral pattern features. This series of preprocessing measures not only improves data quality but also lays a solid foundation for subsequent feature extraction and label prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 This is a flowchart of an implementation of an AI-based customer intelligent tag extraction method embodiment of the present application;

[0097] Figure 2 This is a flowchart for implementing step S20 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0098] Figure 3 This is a flowchart for implementing step S201 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0099] Figure 4This is a flowchart for implementing step S202 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0100] Figure 5 This is a flowchart for implementing step S203 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0101] Figure 6 This is a flowchart for implementing step S30 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0102] Figure 7 This is a flowchart of an implementation of step S10 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0103] Figure 8 This is a flowchart for implementing step S40 in an embodiment of an AI-based customer intelligent tag extraction method of the present application;

[0104] Figure 9 This is a principle block diagram of a computer device of the present application. DETAILED DESCRIPTION

[0105] The following is combined with Figure 1-9 This application is described in further detail.

[0106] In one embodiment, if Figure 1-6 As shown, this application discloses an AI-based customer intelligent tag extraction method, which specifically includes the following steps:

[0107] S10: Collect customer data and pre-process it;

[0108] S20: Extract features from the pre-processed customer data, output a feature set, and construct behavioral pattern features;

[0109] S201: Perform feature selection on the pre-processed customer data;

[0110] SF1: Based on preset multi-level feature screening rules, match data fields related to customer behavior from the preset feature library;

[0111] SF2: Analyze the correlation between each data field and customer behavior based on preset statistical methods;

[0112] SF3: When the correlation fluctuation coefficient exceeds the preset safety threshold, the manual review mechanism is triggered;

[0113] SF4: Select features that have passed the review and whose correlation is higher than the preset threshold as input features;

[0114] S202: Based on the feature selection results, perform feature engineering on the selected features;

[0115] SA1: Determine the type of input features, which include numerical features, categorical features, and text features;

[0116] SA2: When the feature is numerical, calculate the statistics of the sliding window of the input feature, including mean, variance, and skewness, and output the time series derived features;

[0117] SA3: When it is a categorical feature, determine the categorical feature category, where the categorical feature category includes continuous features and discontinuous features;

[0118] SA4: When it is a non-continuous feature, hash the input feature, convert it into a numerical feature, and return to SA1;

[0119] SA5: When it is a continuous feature, the input feature is dynamically binned based on the information gain maximization criterion. When the number of bins is less than or equal to the preset bin threshold, it is converted to a non-continuous feature and returned to SA4;

[0120] SA6: When it is a text feature, perform the following processing: perform semantic vectorization based on the pre-trained word vector model and output the semantic vector;

[0121] SA7: Extract text sentiment polarity scores based on SnowNLP sentiment analysis algorithm;

[0122] SA8: Calculate the keyword TF-IDF weight based on the preset business keyword library;

[0123] SA9: Based on SA1-SA8, output a feature set, which includes time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights;

[0124] S203: Constructing behavioral pattern features based on the feature set processed by feature engineering;

[0125] SB1: Based on the customer ID and a preset time-decay aggregation rule, aggregate the feature set to generate a customer-feature matrix, where the matrix element values ​​are normalized feature vectors;

[0126] SB2: Based on the customer-feature matrix, construct a customer behavior graph network;

[0127] SB3: The customer behavior graph network includes nodes and edge weights, wherein the nodes are customer IDs, and the edge weights are similarities of the feature vectors calculated based on a preset improved cosine similarity algorithm;

[0128] SB4: Based on the preset NetworkX graph calculation library, calculate the node betweenness centrality. When the node betweenness centrality exceeds the preset industry benchmark value, identify the node as a key node;

[0129] SB5: Based on the customer-feature matrix, matching and extracting the feature vectors of the key nodes;

[0130] SB6: Based on the preset multi-head attention mechanism and the customer-feature matrix, the feature vector of the key node is fused with the feature vector of the current customer to output the attention enhancement feature;

[0131] SB7: Output a behavioral pattern feature set, wherein the behavioral pattern feature set includes a customer-feature matrix, a graph network structure feature, and an attention-enhancing feature node. The graph network structure feature includes nodes, edge weights, and node betweenness centrality.

[0132] S30: Learning the behavior pattern features based on a preset machine learning model and outputting a label prediction result;

[0133] S301: Input the behavioral pattern feature set into a preset machine learning model, and extract the customer behavior pattern representation based on the following multimodal processing flow:

[0134] S302: Generate topological correlation features based on the graph neural network, the customer-feature matrix, and the graph network structure features;

[0135] S303: Generate semantic features based on the fully connected network and attention-enhanced features;

[0136] S304: performing vector concatenation of topological association features and semantic features to generate a fused feature vector;

[0137] S305: Map the fused feature vector to the label category space and output the customer label category;

[0138] S306: Calibrate the customer label categories based on the temperature-scaled Softmax function, generate a standardized probability distribution, and compare the probability values;

[0139] S307: When the probability value is the largest, output the label prediction result, which includes the customer ID and the customer label category;

[0140] S40: Sort and store the tag prediction results based on the customer's personalized weight and a preset sorting algorithm.

[0141] In this embodiment, the method first collects and preprocesses customer data to lay the foundation for subsequent analysis. Next, features are extracted from the preprocessed data to construct behavioral pattern features that can effectively reflect the customer's behavioral habits and preferences. Then, the behavioral pattern features are learned using a preset machine learning model to predict the corresponding labels. Finally, the predicted labels are sorted and stored based on the behavioral pattern features and a preset sorting algorithm for quick retrieval and utilization. This method uses AI technology to automatically extract and sort customer smart labels, improves the efficiency and accuracy of customer data processing, and provides strong support for precision marketing and customer relationship management.

[0142] In step S20, the method first performs feature selection on the preprocessed customer data. This process aims to filter out the most valuable and relevant features for label prediction from a large amount of data, thereby reducing the data dimension and improving the efficiency and accuracy of model training. Then, based on the results of feature selection, feature engineering is performed on the selected features, including operations such as feature transformation and feature encoding, to further optimize the expression form of the features and enhance the model's ability to capture data features. Finally, the feature set processed by feature engineering is used to construct behavioral pattern features. These features can deeply reflect the customer's behavioral patterns and preferences, providing effective input for subsequent machine learning models. Through this series of steps, the method ensures that the data input to the model has high information content and low redundancy, laying a solid foundation for accurately predicting customer labels.

[0143] In step S201, the method adopts a multi-level feature screening mechanism. First, based on the preset multi-level feature screening rules, data fields related to customer behavior are matched from the preset feature library. This step ensures the pertinence and comprehensiveness of feature selection. Next, the preset statistical method is used to analyze the correlation between each data field and customer behavior, and the degree of influence of each feature on customer behavior patterns is evaluated through quantitative analysis. In order to ensure the accuracy and reliability of feature selection, when the correlation fluctuation coefficient exceeds the preset safety threshold, the method will trigger a manual review mechanism to eliminate the influence of outliers and noise data through manual intervention. Finally, those features that have passed the review and whose correlation is higher than the preset threshold are selected as input features. These features have been proven to be the most valuable for predicting customer behavior patterns. Through this series of meticulous feature selection processes, the method effectively improves the quality of features and the performance of the prediction model, providing a solid data foundation for subsequent behavior pattern feature construction and label prediction.

[0144] In step S202, the method uses a series of feature engineering techniques to carefully process the input features to enhance the expressive power of the features and the learning effect of the model. First, different processing strategies are adopted according to the type of input features (numerical, categorical, textual). For numerical features, their statistics (such as mean, variance, skewness) within the sliding window are calculated to generate time-series derived features to capture the dynamic changes of the features. For categorical features, continuous and discontinuous features are further distinguished. Discontinuous features are converted into numerical features through hash coding, while continuous features are dynamically binned based on the information gain maximization criterion and converted into discontinuous features for processing when necessary. For textual features, semantic vectorization is performed using a pre-trained word vector model, text sentiment polarity scores are extracted, and keyword TF-IDF weights are calculated, thereby converting text data into structured numerical features. Through these processes, a feature set containing time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights is finally output. This step effectively improves the quality and diversity of features, providing richer and more accurate data input for constructing behavioral pattern features and subsequent label predictions, thereby improving the overall performance and prediction accuracy of the model.

[0145] In step S203, this method further refines and enhances behavioral pattern features by constructing a customer behavior graph network and applying an attention mechanism. First, based on customer IDs and preset time-decay aggregation rules, the feature set is aggregated into a customer-feature matrix. The matrix elements are standardized feature vectors. This step effectively organizes and normalizes the feature data. Next, the customer-feature matrix is ​​used to construct a customer behavior graph network, where nodes represent customer IDs and edge weights are based on feature vector similarities calculated using a modified cosine similarity algorithm. This forms a correlation network of customer behavioral patterns. By calculating node betweenness centrality and identifying key nodes, this method highlights customers with high influence in the network, whose feature vectors are further extracted and utilized. Using a multi-head attention mechanism, the feature vectors of key nodes are fused with the feature vector of the current customer to generate attention-enhanced features. This process effectively improves the importance and discriminability of the features. Ultimately, the output is a behavioral pattern feature set consisting of the customer-feature matrix, graph network structural features, and attention-enhanced feature nodes. The graph network structural features encompass node, edge weight, and node betweenness centrality. This series of steps not only deepens our understanding of customer behavior patterns, but also enhances the expressiveness of features through graph networks and attention mechanisms, providing more accurate and rich input for subsequent label predictions, significantly improving the performance and effectiveness of the prediction model.

[0146] In step S30, the method uses a multimodal processing flow and a preset machine learning model to conduct in-depth analysis and learning of the behavioral pattern feature set to accurately predict customer labels. First, the behavioral pattern feature set is input into the model. A graph neural network is then used to combine the customer-feature matrix and graph network structural features to generate topological association features. These features reflect the structure and relationship information of customers in the behavioral graph network. Simultaneously, attention-enhanced features are processed through a fully connected network to generate semantic features that capture the deeper meaning and importance of customer behavior. Next, the topological association features and semantic features are vector-concatenated to form a fused feature vector. This step integrates multiple feature information and enhances the expressive power of the features. The fused feature vector is mapped to the label category space, and the customer label category is output. The label category is then calibrated using a temperature-scaled Softmax function to generate a normalized probability distribution. Finally, the probability values ​​are compared. When the probability value is the largest, the label prediction result, including the customer ID and customer label category, is output. This series of processing flows fully leverages the advantages of graph neural networks and fully connected networks, achieving the fusion of multimodal features and accurate prediction of label categories, significantly improving the accuracy and efficiency of customer label extraction, and providing strong support for subsequent marketing strategy formulation and customer relationship management.

[0147] Figure 7 , step S10 includes:

[0148] S101: The customer data includes static data, dynamic data, and text data;

[0149] S102: The static data includes basic customer information, which includes customer ID and registration time;

[0150] S103: The dynamic data includes customer operation behavior data, which includes page visit frequency, function click count, and stay time;

[0151] S104: The text data includes customer work order text and social media comment text;

[0152] S105: Based on the multidimensional time series of the customer operation behavior data, calculate the standard deviation of the customer operation behavior data within a preset sliding time window.

[0153] In this embodiment, in step S10, the method first carefully classifies the customer data, including static data, dynamic data and text data, to ensure the pertinence and accuracy of subsequent processing. Static data covers the basic information of customers, such as customer ID and registration time. This information provides a basis for understanding the customer's background. Dynamic data focuses on the customer's operational behavior, such as page visit frequency, function click count and length of stay. These data reflect the customer's activity and interests. Text data includes customer work order text and social media comment text, which provides customer feedback and emotional tendencies. By calculating the standard deviation of customer operational behavior data within a preset sliding time window, the method can capture the fluctuations and trends of customer behavior and provide a basis for the construction of behavioral pattern features. This series of preprocessing measures not only improves the quality of the data, but also lays a solid foundation for subsequent feature extraction and label prediction.

[0154] Figure 8 , step S40 includes:

[0155] S401: Assigning a personalized weight to each tag prediction result based on the behavior pattern characteristics;

[0156] S402: Compare the personalized weight with the preset weight threshold to determine whether the corresponding tag prediction result can participate in the ranking;

[0157] S403: When the personalized weight is lower than the preset weight threshold, the tag prediction result cannot be included in the ranking;

[0158] S404: When the personalized weight is higher than a preset weight threshold, the tag prediction result can be included in the ranking;

[0159] S405: Based on the order of personalized weights, the tag prediction results that can participate in the sorting are sorted from high to low and stored.

[0160] In this embodiment, in step S40, the method assigns personalized weights to each tag prediction result and sorts them based on these weights to ensure the proper prioritization and relevance of the tag prediction results. First, a personalized weight is calculated for each tag prediction result based on behavioral pattern characteristics. These weights reflect the importance and relevance of different tags to the customer's behavioral pattern. Next, the personalized weights are compared with a preset weight threshold to determine which tag prediction results are eligible for sorting. This step ensures that only tags with a certain level of importance are considered, thereby improving the effectiveness of the sorting. Results with personalized weights below the preset threshold are excluded from the sorting process, preventing unimportant or low-relevance tags from interfering with the final results. Results with personalized weights above the preset threshold are sorted and stored in descending order of weight, ensuring that the most relevant and important tags are prioritized and used. Through this sorting mechanism, the method not only improves the organization and accessibility of tag prediction results, but also enhances their practical application value, providing more accurate and organized tag data for subsequent decision support, customer segmentation, and personalized services.

[0161] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0162] In one embodiment, an AI-based customer smart tag extraction system is provided. The AI-based customer smart tag extraction system corresponds to the AI-based customer smart tag extraction method in the above embodiment. The AI-based customer smart tag extraction system includes:

[0163] Module 1: Collect customer data and pre-process;

[0164] The second module: extracts features from pre-processed customer data, outputs feature sets, and constructs behavioral pattern features;

[0165] Feature selection module: performs feature selection on pre-processed customer data;

[0166] The second matching module: matches data fields related to customer behavior from the preset feature library based on preset multi-level feature screening rules;

[0167] The first output module: Based on a preset statistical method, it analyzes the correlation between each data field and customer behavior and outputs the correlation fluctuation coefficient;

[0168] The first comparison module: when the correlation fluctuation coefficient exceeds the preset safety threshold, the manual review mechanism is triggered;

[0169] The first selection module: selects features that have passed the review and whose correlation is higher than the preset threshold as input features;

[0170] Feature engineering module: Based on the feature selection results, feature engineering is performed on the selected features;

[0171] The fifth module includes: determining the type of input features, which include numerical features, categorical features, and text features;

[0172] The sixth module includes: when it is a numerical feature, it calculates the statistics in the sliding window of the input feature, the statistics include mean, variance, skewness, and outputs the time series derived features;

[0173] The seventh module includes: when it is a categorical feature, determining the category of the categorical feature, the categorical feature category includes continuous features and discontinuous features;

[0174] First return module: When it is a non-continuous feature, the input feature is hashed and converted into a numerical feature, and then returned to SA1;

[0175] Second return module: When it is a continuous feature, the input feature is dynamically binned based on the information gain maximization criterion. When the number of bins is less than or equal to the preset bin threshold, it is converted into a non-continuous feature and returned to SA4;

[0176] The second output module: When it is a text feature, it performs the following processing: based on the pre-trained word vector model, it performs semantic vectorization and outputs the semantic vector;

[0177] The first extraction module: extracts the text sentiment polarity score based on the SnowNLP sentiment analysis algorithm;

[0178] The third calculation module: calculates the keyword TF-IDF weight based on the preset business keyword library;

[0179] The eighth module includes: based on SA1-SA8, outputting a feature set, wherein the feature set includes time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights;

[0180] Pattern feature module: constructs behavioral pattern features based on the feature set processed by feature engineering;

[0181] A first generation module: based on the customer ID and a preset time decay aggregation rule, aggregating the feature set to generate a customer-feature matrix, wherein the matrix element value is a standardized feature vector;

[0182] The first building module: constructing a customer behavior graph network based on the customer-feature matrix;

[0183] The ninth module includes: the customer behavior graph network includes nodes and edge weights, the nodes are customer IDs, and the edge weights are similarities of the feature vectors calculated based on a preset improved cosine similarity algorithm;

[0184] The second comparison module: calculates the node betweenness centrality based on the preset NetworkX graph calculation library, and identifies the node as a key node when the node betweenness centrality exceeds the preset industry benchmark value;

[0185] A third matching module: matching and extracting the feature vectors of the key nodes based on the customer-feature matrix;

[0186] The third output module: based on the preset multi-head attention mechanism and the customer-feature matrix, fuses the feature vector of the key node with the feature vector of the current customer, and outputs the attention enhancement feature;

[0187] The fourth output module outputs a behavioral pattern feature set, wherein the behavioral pattern feature set includes a customer-feature matrix, graph network structure features, and attention enhancement feature nodes. The graph network structure features include nodes, edge weights, and node betweenness centrality.

[0188] The third module: learns behavioral pattern features based on a preset machine learning model and outputs label prediction results;

[0189] The first input module inputs the behavioral pattern feature set into a preset machine learning model and extracts the customer behavior pattern representation based on the following multimodal processing flow:

[0190] The second generation module generates topological correlation features based on the graph neural network, customer-feature matrix, and graph network structure features;

[0191] The third generation module: Generates semantic features based on the fully connected network and attention-enhanced features;

[0192] The fourth generation module: concatenates the topological association features and semantic features to generate a fused feature vector;

[0193] The fifth output module maps the fused feature vector to the label category space and outputs the customer label category;

[0194] The fifth generation module: Based on the temperature-scaled Softmax function, calibrate the customer label categories, generate standardized probability distributions, and compare probability values;

[0195] The sixth output module: when the probability value is the largest, outputs the label prediction result, which includes the customer ID and the customer label category;

[0196] The fourth module: sorts and stores the tag prediction results based on customer personalized weights and a preset sorting algorithm.

[0197] Optionally, also include:

[0198] The first module includes: the customer data includes static data, dynamic data, and text data;

[0199] The second module includes: the static data includes customer basic information, and the customer basic information includes customer ID and registration time;

[0200] The third module includes: the dynamic data includes customer operation behavior data, and the customer operation behavior data includes page visit frequency, function click number and stay time;

[0201] The fourth module includes: the text data includes customer work order text and social media comment text;

[0202] The first calculation module is used to calculate the standard deviation of the customer operation behavior data within a preset sliding time window based on the multidimensional time series of the customer operation behavior data.

[0203] Optionally, also include:

[0204] Weight module: assigns personalized weights to each tag prediction result based on the behavioral pattern characteristics;

[0205] The first judgment module compares the personalized weight with the preset weight threshold to determine whether the corresponding label prediction result can participate in the ranking;

[0206] Second judgment module: when the personalized weight is lower than the preset weight threshold, the tag prediction result cannot be included in the ranking;

[0207] The third judgment module: when the personalized weight is higher than the preset weight threshold, the tag prediction result can be included in the ranking;

[0208] Storage module: Based on the order of personalized weights, the label prediction results that can participate in the sorting are sorted and stored from high to low.

[0209] For the specific definition of an AI-based customer intelligent label extraction system, please refer to the definition of an AI-based customer intelligent label extraction method above, which will not be repeated here. The various modules in the above-mentioned AI-based customer intelligent label extraction system can be implemented in whole or in part through software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0210] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store customer tags. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, an AI-based customer intelligent tag extraction method is implemented.

[0211] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, an AI-based customer smart label extraction method is implemented.

[0212] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, an AI-based customer smart label extraction method is provided.

[0213] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0214] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A customer intelligent label extraction method based on AI, characterized in that: Including steps: Collect customer data and pre-process it; Extract features from pre-processed customer data, output feature sets, and construct behavioral pattern features; Perform feature selection on pre-processed customer data; Based on preset multi-level feature screening rules, match data fields related to customer behavior from the preset feature library; Analyze the correlation between each data field and customer behavior based on preset statistical methods; When the correlation fluctuation coefficient exceeds the preset safety threshold, the manual review mechanism is triggered; Select features that have passed the review and whose correlation is higher than the preset threshold as input features; Based on the feature selection results, perform feature engineering on the selected features; SA1: Determine the type of input features, which include numerical features, categorical features, and text features; SA2: When the feature is numerical, calculate the statistics of the sliding window of the input feature, including mean, variance, and skewness, and output the time series derived features; SA3: When it is a categorical feature, determine the categorical feature category, where the categorical feature category includes continuous features and discontinuous features; SA4: When it is a non-continuous feature, hash the input feature, convert it into a numerical feature, and return to SA1; SA5: When it is a continuous feature, the input feature is dynamically binned based on the information gain maximization criterion. When the number of bins is less than or equal to the preset bin threshold, it is converted to a non-continuous feature and returned to SA4; SA6: When it is a text feature, perform the following processing: perform semantic vectorization based on the pre-trained word vector model and output the semantic vector; SA7: Extract text sentiment polarity scores based on SnowNLP sentiment analysis algorithm; SA8: Calculate the keyword TF-IDF weight based on the preset business keyword library; SA9: Based on SA1-SA8, output a feature set, which includes time-series derived features, semantic vectors, text sentiment polarity scores, and keyword TF-IDF weights; Construct behavioral pattern features based on the feature set processed by feature engineering; Aggregating the feature set based on the customer ID and a preset time-decay aggregation rule to generate a customer-feature matrix, wherein the matrix element values ​​are normalized feature vectors; Based on the customer-feature matrix, construct a customer behavior graph network; The customer behavior graph network includes nodes and edge weights, wherein the nodes are customer IDs, and the edge weights are similarities of the feature vectors calculated based on a preset improved cosine similarity algorithm; Based on the preset NetworkX graph calculation library, the node betweenness centrality is calculated. When the node betweenness centrality exceeds the preset industry benchmark value, the node is identified as a key node; Based on the customer-feature matrix, matching and extracting the feature vectors of the key nodes; Based on the preset multi-head attention mechanism and the customer-feature matrix, the feature vector of the key node is fused with the feature vector of the current customer to output an attention-enhanced feature; Outputting a behavioral pattern feature set, the behavioral pattern feature set including a customer-feature matrix, graph network structural features, and attention enhancement feature nodes, the graph network structural features including nodes, edge weights, and node betweenness centrality; Learn behavioral pattern features based on a preset machine learning model and output label prediction results; The behavioral pattern feature set is input into a preset machine learning model to extract the customer behavior pattern representation based on the following multimodal processing flow: Generate topological correlation features based on graph neural network, customer-feature matrix, and graph network structure features; Generate semantic features based on fully connected networks and attention-enhanced features; Concatenate the topological association features and semantic features into vectors to generate a fused feature vector; Map the fused feature vector to the label category space and output the customer label category; Based on the temperature-scaled Softmax function, the customer label categories are calibrated, the standardized probability distribution is generated, and the probability values ​​are compared; When the probability value is the largest, the label prediction result is output, and the label prediction result includes the customer ID and the customer label category; Based on the customer's personalized weight and a preset sorting algorithm, the tag prediction results are sorted and stored.

2. The AI-based customer intelligent tag extraction method according to claim 1 is characterized in that: The steps of collecting customer data and pre-processing include: The customer data includes static data, dynamic data, and text data; The static data includes basic customer information, which includes customer ID and registration time; The dynamic data includes customer operation behavior data, which includes page visit frequency, function click times and stay time; The text data includes customer work order text and social media comment text; Based on the multidimensional time series of the customer operation behavior data, the standard deviation of the customer operation behavior data within a preset sliding time window is calculated.

3. The AI-based customer intelligent tag extraction method according to claim 1 is characterized in that: The step of sorting and storing the tag prediction results based on the customer personalized weight and the preset sorting algorithm includes: Based on the behavioral pattern characteristics, assigning personalized weights to each tag prediction result; Compare the personalized weight with the preset weight threshold to determine whether the corresponding tag prediction result can participate in the ranking; When the personalized weight is lower than the preset weight threshold, the tag prediction result cannot be included in the ranking; When the personalized weight is higher than the preset weight threshold, the tag prediction result can participate in the ranking; Based on the order of personalized weights, the tag prediction results that can participate in the sorting are sorted and stored from high to low.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the AI-based customer smart label extraction method as described in claims 1-3 are implemented.

5. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the AI-based customer smart label extraction method as described in claims 1-3.

Images