Insurance marketing activity multi-label classification method based on ECC algorithm

Abstract

The invention discloses an insurance marketing activity multi-label classification method based on an ECC (Error Correction Code) algorithm, and the method can effectively model the dependency relationship between labels, improve the classification accuracy under a small data set, reduce the model overfitting risk, reduce the label dimension through rule pre-classification, and improve the feature generalization ability and the stability of a recommendation system.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-label classification method for insurance marketing activities based on the ECC algorithm. Background Technology

[0002] In the insurance industry, marketers reach customers by inviting them to offline events or pushing promotions through online apps, conveying product information and driving sales conversions. Recommending themed events based on customer preferences and historical behavior can improve customer satisfaction and engagement. A customer may have participated in multiple events; predicting their likelihood of participating in marketing campaigns falls under the multi-tag classification domain. Events should be designed to meet the interests and needs of different customer groups, and these interests and needs often overlap, meaning there is a certain correlation between event categories. An algorithm that can capture these correlations should be chosen. Existing marketing campaign matching methods include rule-based recommendations based on human experience and user characteristics, or marketing matching recommendations using machine learning and deep learning methods.

[0003] Existing patents propose a multi-label benefit recommendation method that integrates prior information. First, basic information and card usage information of the target user are obtained; transaction descriptions in the information are categorized into consumption categories; based on the obtained consumption categories, individual consumption preferences and the salient preferences of group goals are analyzed and used as features input into a trained recommendation model, outputting the probability distribution of the target user across different benefits; combining the probability distribution of the target user across different benefits, their lifecycle stage, and benefit reach, a multi-label benefit recommendation strategy is formulated. This method introduces a multi-attention mechanism into the recommendation model, making the model pay more attention to benefits related to salient preferences in individual consumption preferences and group goals, thereby recommending benefits more accurately.

[0004] Existing technologies have the following problems: the label space may be relatively large, increasing the complexity of deep learning models as they process a large number of labels, which also leads to a decrease in model performance. Furthermore, there is a long-tail distribution problem, where many labels may only be associated with a small number of samples, making it difficult for the model to capture them and affecting the overall classification results. Summary of the Invention

[0005] As is known, ECC, as an ensemble learning method, can explicitly utilize dependencies between labels to improve performance when dealing with such dependencies. In some cases, deep learning models require a large amount of labeled data to achieve maximum performance, while ECC can demonstrate good performance even on smaller datasets. The main objective of this invention is to provide a multi-label classification method for insurance marketing activities based on the ECC algorithm. It first pre-classifies the activities according to rules, then employs the ECC ensemble learning method to fully utilize the correlation between category labels, thereby improving the accuracy of multi-label classification of insurance marketing activities. This solves the problems of low model generalization ability, high overfitting risk, and poor performance on small datasets in traditional solutions.

[0006] Another objective of this invention is to propose a multi-label classification device for insurance marketing activities based on the ECC algorithm.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a multi-label classification method for insurance marketing activities based on the ECC algorithm, comprising: S1. Obtain the original insurance marketing campaign data, perform word segmentation and keyword extraction on the campaign name based on the preset campaign theme classification rules, and map the original campaign tags to 5-10 theme categories to form a dimensionality-reduced tag theme space. S2, construct a combined classifier chain model, using a decision tree as the base classifier, and train multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature for chain-like transmission when training the current classifier. S3, perform SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the characteristics of the label topic distribution in order to optimize the model training effect. S4 integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

[0008] In one embodiment of the present invention, the step of performing word segmentation and keyword extraction on the activity name based on a preset activity theme classification rule, mapping the original activity tags to 5-10 theme categories to form a dimensionality-reduced tag theme space, further includes: S11, use the TF-IDF algorithm to calculate the weight of each keyword in the activity name, and filter keywords with weights higher than a preset threshold. Keywords are used as thematic features; S12, Construct a rule optimization model based on the historical activity classification results, and calculate the classification accuracy. With rule coverage The combined metrics are used to dynamically adjust keyword matching rules.

[0009] In one embodiment of the present invention, the step of training multiple classifier chains according to a preset dependency relationship between tag topics, wherein each classifier chain uses the prediction result of the previous classifier as input features for chain-like passing during the training of the current classifier, further includes: S21, based on the Pearson correlation coefficient between labeled topics The ranking strategy for constructing the classifier chain is to prioritize classes with a correlation coefficient absolute value greater than 1. The topics are arranged in descending order; S22, the predicted probability from the previous classifier Compared with the original feature vector The features are concatenated to form a new feature matrix. .

[0010] In one embodiment of the present invention, it further includes: S5, based on feedback data from customers' historical participation in activities, constructs a dynamic adjustment module to calculate the prediction confidence of each classifier chain. When the confidence level is below the threshold When this happens, it triggers parameter updates in the rule classification module and retraining of the classifier chain.

[0011] To achieve the above objectives, a second aspect of the present invention provides a multi-label classification device for insurance marketing activities based on the ECC algorithm, comprising: The data acquisition and preprocessing module is used to acquire raw insurance marketing activity data, perform word segmentation and keyword extraction on the activity name based on the preset activity theme classification rules, and map the original activity tags to 5-10 theme categories to form a dimensionality-reduced tag theme space. The classifier chain building module is used to build a combined classifier chain model. It uses a decision tree as the base classifier and trains multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature when training the current classifier. The data augmentation module is used to perform SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the characteristics of the label topic distribution in order to optimize the model training effect. The results integration and decision-making module integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

[0012] The methods, apparatus, electronic devices, and computer-readable storage media of the present invention can effectively model the dependencies between labels of insurance marketing activities, improve classification accuracy and model robustness on small datasets, and reduce the risk of overfitting. Attached Figure Description

[0013] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a multi-label classification method for insurance marketing activities based on the ECC algorithm, provided for an embodiment of the present invention; Figure 2 A flowchart illustrating another multi-label classification method for insurance marketing activities based on the ECC algorithm provided in this embodiment of the invention; Figure 3 A chain diagram of a combined classifier provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a multi-label classification device for insurance marketing activities based on the ECC algorithm, provided in an embodiment of the present invention. Detailed Implementation

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0016] The following are the technical terms that may be used in this invention: Multi-label classification: Each sample can belong to multiple categories simultaneously; ECC: Ensemble of Classifier Chains, which considers the dependencies between labels by constructing a chain of classifiers, where the output of each classifier is part of the input of the next classifier; Decision tree: A decision tree is a tree structure in which each internal node represents a test on an attribute, each branch represents a test output, and each leaf node represents a category.

[0017] The following description, with reference to the accompanying drawings, describes a multi-label classification method and apparatus for insurance marketing activities based on the ECC algorithm, according to an embodiment of the present invention.

[0018] Example 1 This embodiment provides a multi-label classification method for insurance marketing activities based on the ECC algorithm. For example... Figure 1 As shown, the method includes the following steps: S1. Obtain the original insurance marketing activity data, perform word segmentation and keyword extraction on the activity name based on a preset activity theme classification rule, and map the original activity labels to 5 - 10 theme categories to form a reduced - dimensional label theme space.

[0019] Specifically, in some implementations, this step aims to map the original high - dimensional activity label space to 5 - 10 theme categories through a text - processing method based on a preset activity theme classification rule, so as to form a reduced - dimensional label theme space. This step is a key pre - processing link in the entire multi - label classification system, and its technical implementation is based on the combination of natural language processing (NLP) and a rule engine. [[ID=N]]

[0020] First, perform text cleaning on the original activity name, including removing punctuation marks and stop words (such as "of", "and", "activity", etc.), and use a Chinese word - segmentation tool (such as jieba, HanLP, etc.) to perform word - segmentation on the activity name. The word - segmentation granularity is usually set to a词性过滤模式 (pos - filtering mode), retaining nouns, verbs, and proper nouns to improve the accuracy of keyword extraction. Subsequently, based on the preset activity theme classification rule, perform keyword matching and frequency statistics on the word - segmented results. For example, if the activity name contains keywords such as "Mid - Autumn Festival", "reunion", "festival", etc., it can be classified into the "festival - type" theme; if it contains words such as "health", "physical examination", "protection", etc., it is classified into the "health protection - type".

[0021] The key parameters involved in this step include the granularity setting of the word - segmentation tool, the version of the stop - word library (such as using the "Harbin Institute of Technology Stop - Word List" or the "Baidu Stop - Word List"), the threshold of keyword matching (such as the keyword appears ≥ 2 times or the TF - IDF value ≥ 0.5), and the coverage and priority sorting of the theme classification rule. In addition, the number of theme categories is controlled between 5 - 10 to ensure the dimensionality reduction effect of the label space while retaining sufficient semantic discrimination.

[0022] This step is applicable to scenarios such as customer behavior prediction and personalized recommendation in insurance marketing activities. For example, in the customer's historical participation records, the original activity labels may be as many as 1000 types, and through the rule mapping of this step, they can be compressed into 5 - 10 theme categories, thereby significantly reducing the input dimension of the model and improving the training efficiency and classification performance of the subsequent ECC algorithm.

[0023] It should be noted that the "词性过滤模式" in the original text seems to be a Chinese term that needs to be accurately translated according to the specific meaning in the context. Here, a rough translation is provided for the overall content. If there are official or more accurate translations for relevant terms, they can be adjusted accordingly.This step, through rule-driven label mapping, effectively alleviates the problems of increased model complexity and long-tail distribution caused by an excessively large label space, while enhancing the semantic correlation between labels. Furthermore, this method maintains high classification stability even on small sample datasets, providing a structured and semantic input feature space for subsequent multi-label classification models, thereby improving the overall system's generalization ability and prediction accuracy.

[0024] Furthermore, S1 includes: S11, use the TF-IDF algorithm to calculate the weight of each keyword in the activity name, and filter keywords with weights higher than a preset threshold. Keywords are used as thematic features.

[0025] Specifically, in some implementations, this step uses the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to calculate the weights of keywords in the activity name, thereby filtering out the most representative topic features. The core of this step lies in quantifying the importance of keywords in the document set to extract feature words that effectively represent the activity topic, providing high-quality input features for subsequent multi-label classification models.

[0026] The TF-IDF algorithm measures the importance of a word in a document by calculating the product of its term frequency (TF) in a single document and its inverse document frequency (IDF) across the entire document set. Specifically, for an activity name document set... Each activity name It is considered a document containing several keywords. The formula for calculating TF-IDF is:

[0027] in, Keywords In the document Frequency of occurrence in Indicates the presence of keywords The number of documents, This represents the total number of documents. This formula effectively reduces the weight of common words (such as "event," "invitation," etc.) while increasing the weight of keywords that appear frequently in specific events but are less common in the overall documents.

[0028] Preset threshold Used to filter keywords, its value is usually adjusted according to the actual data distribution; for example, it can be set to [value] in an experiment. or This ensures that only keywords that are highly distinctive to the event theme are retained. Furthermore, during the word segmentation process, Chinese word segmentation tools such as jieba or HanLP can be used, combined with a stop word list (such as the "Harbin Institute of Technology Stop Word List") for filtering, to improve the accuracy of keyword extraction.

[0029] This step is applicable to scenarios where the event name is thematically modeled in insurance marketing campaigns. For example, in historical campaigns, if an event is named "2024 Mid-Autumn Festival Customer Appreciation Event," keywords such as "Mid-Autumn Festival" and "appreciation" are likely to have high weights based on TF-IDF calculations, and thus be selected as thematic features of the event. These features will serve as input to an ECC multi-label classification model to predict customer participation tendencies across various events.

[0030] This step extracts representative keywords using the TF-IDF algorithm, effectively reducing feature dimensionality and improving the model's generalization ability. Simultaneously, by setting reasonable... Thresholds can filter out noisy words and retain keywords that are highly relevant to the activity theme, thereby enhancing the ability of subsequent classification models to capture the dependencies between labels and improving the overall classification accuracy.

[0031] S12, Construct a rule optimization model based on the historical activity classification results, and calculate the classification accuracy. With rule coverage The combined metrics are used to dynamically adjust keyword matching rules.

[0032] Specifically, in some implementations, this step involves constructing a rule optimization model to dynamically adjust the classification results of historical activities, thereby improving the adaptability and accuracy of keyword matching rules in multi-label classification tasks. The core of this model lies in jointly optimizing classification accuracy. With rule coverage ,in , , , These represent the number of true positives, true negatives, false positives, and false negatives, respectively. This indicates the number of activities matching the rule. This represents the total number of activities. By jointly evaluating these two metrics, the system can quantify the performance of the current rules in classification tasks, thus providing a basis for subsequent rule adjustments.

[0033] This step first extracts keywords from activity names based on historical activity data and constructs preliminary classification rules. Then, these rules are applied to the activity dataset to generate preliminary classification results. The system calculates the corresponding classification results by counting the number of matches and the classification correctness of each rule. and Value. Furthermore, a weighted joint indicator is used. ,in This is an adjustable parameter used to balance the priorities of accuracy and coverage. In practical applications, The value can be dynamically set according to business needs. For example, in scenarios that emphasize recommendation accuracy, it can be set to... In scenarios where the breadth of recommendations is emphasized, settings can be configured. .

[0034] This step has significant application value in insurance marketing activities. For example, in customer activity recommendation systems, dynamically adjusting keyword matching rules can effectively address issues such as changes in activity categories over time and shifts in customer interests. In practical implementation, the system can periodically re-evaluate the rule set. The system calculates the value of each rule and adjusts its weights by adding, deleting, merging, or modifying the rules based on the evaluation results, thereby achieving adaptive optimization of the rules. This method not only improves the generalization ability of the classification model but also reduces the cost of manually maintaining the rules, enhancing the scalability and stability of the system.

[0035] S2. Construct a combined classifier chain model, using a decision tree as the base classifier, and train multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature when training the current classifier.

[0036] Specifically, in this application, constructing an Ensemble of Classifier Chains (ECC) model is one of the core steps in achieving multi-label classification for insurance marketing campaigns. This step uses a decision tree as the base classifier and explicitly models the dependencies between labels through a chain structure, thereby improving the accuracy and generalization ability of the classification.

[0037] The ECC model consists of multiple classifier chains, with each classifier in the chain trained sequentially according to a predefined label topic order. Specifically, the classifier in each classifier chain... During training, each classifier uses not only the original input features (such as customer age, gender, historical policy information, activity preferences, etc.), but also the preceding data. The prediction results of each classifier are used as input for new features, forming a chain-like propagation mechanism. This structure allows the correlation between labels to be effectively utilized. For example, the prediction results of "festival" activities can be used as input features for the classification of "family" activities, thereby enhancing the model's ability to model the potential relationships between labels.

[0038] Each decision tree classifier can be trained using the CART (Classification and Regression Tree) algorithm, with the splitting criterion being Gini impurity, and the tree depth can be set to... To balance model complexity and computational efficiency, the order of the classifier chain can be ordered using the label correlation matrix. Correlation can be calculated using the Pearson correlation coefficient or mutual information (MI), with a threshold set as follows: This ensures that the labels in the chain have significant dependencies. Furthermore, the number of classifier chains in the ECC model can be set to... This is to enhance the robustness of the model.

[0039] This step is applicable to customer activity recommendation scenarios in insurance marketing systems. For example, after a customer profile is built, the system inputs the customer feature vector into the ECC model to predict the probability of their participation in various activity themes. Finally, the prediction results of all chains are integrated through a voting mechanism (such as majority voting or weighted voting) to output multi-label prediction results for each activity category, which are then used for subsequent personalized activity pushes.

[0040] By employing a chain-like structure and ensemble learning mechanism, the problems of large label space and imbalanced label distribution are effectively alleviated, improving the model's classification performance on small sample data. Meanwhile, the decision tree, as a base classifier, offers good interpretability, facilitating subsequent model tuning and business decision support.

[0041] Furthermore, S2 includes: S21, based on the Pearson correlation coefficient between labeled topics The ranking strategy for constructing the classifier chain is to prioritize classes with a correlation coefficient absolute value greater than 1. The topics are arranged in descending order.

[0042] Specifically, in some implementations, this application proposes to calculate the Pearson correlation coefficient between labeled topics. This involves constructing a ranking strategy for the classifier chain to optimize the performance of the combined classifier chain (ECC) in multi-label classification tasks. The core of this step lies in leveraging the statistical correlation between label topics to guide the order of labels in the classifier chain, thereby enhancing the model's ability to model label dependencies.

[0043] First, rule-based topic classification is performed on the original activity labels, mapping high-dimensional activity names to low-dimensional topic categories (such as holidays, health, finance, etc.), thereby reducing the number of labels and improving the model's generalization ability. Then, for each topic category, the Pearson correlation coefficient between it and other topic categories is calculated. ,in Indicates the first Does the sample participate in the first...? One topic, Indicate whether to participate in the One topic, and These represent the average participation rates for the two topics. This formula quantifies the linear correlation between tagged topics, providing a basis for constructing subsequent chain structures.

[0044] Set a correlation coefficient threshold The typical value range is This threshold is used to filter label-topic pairs that are significantly related. In practical applications, this threshold can be adjusted based on the label distribution characteristics of the dataset to balance model complexity with the accuracy of label dependency modeling.

[0045] This step is applicable to predicting customer participation across multiple theme categories in insurance marketing campaigns. For example, in historical customer behavior data, if the absolute value of the correlation coefficient between "holiday" and "health" activities is... Therefore, when constructing the classifier chain, the "festival class" is placed before the "health class" so that when predicting the "health class" label, the prediction results of the "festival class" can be used as input features to enhance the model's ability to capture the dependencies between labels.

[0046] By introducing statistical correlations between labeled topics and optimizing the label order in the classifier chain, the prediction accuracy and stability of the model in multi-label classification tasks are improved. Compared with traditional rule-based or deep learning methods, this approach maintains good generalization ability on small sample datasets and effectively alleviates the problems of long-tailed label distribution and model overfitting.

[0047] S22, the predicted probability from the previous classifier Compared with the original feature vector The features are concatenated to form a new feature matrix. .

[0048] Specifically, in some implementations, the predicted probability of the previous classifier is used... Compared with the original feature vector The features are concatenated to form a new feature matrix. This step is one of the key steps in the Combined Classifier Chain (ECC) algorithm proposed in this application to achieve label dependency modeling. This step introduces the output of the preceding classifier as the input feature of the current classifier, thereby explicitly modeling the correlation between labels in the chain structure and improving the overall performance of multi-label classification.

[0049] This concatenation operation is typically performed during the feature engineering phase. Original feature vectors Include There are 10 samples, each sample has 100 samples. Dimensional features, such as customer age, gender, historical activity records, and policy type. The predicted probability from the previous classifier. It is the front The prediction results for each label, each element Indicates the first The nth sample pair The probability of each tag participating. The concatenation operation uses column-wise concatenation, that is... As a newly added feature column, with The features are spliced ​​together to form a new feature matrix. This operation can be achieved using the `concatenate` or `torch.cat` functions in mainstream machine learning frameworks such as NumPy or PyTorch.

[0050] Dimensions of the concatenated feature matrix This directly impacts the training efficiency of subsequent base classifiers (such as decision trees) and the model performance. Typically, The value ranges from 10 to 50, while The granularity of classification can be set to 5 to 10 depending on the actual activity theme. The concatenated feature vectors need to be standardized, for example, using Min-Max normalization or Z-score normalization, to ensure that the dimensions of different features are consistent and to improve the convergence speed and stability of the model.

[0051] This step is widely used in insurance marketing campaigns to predict customer participation across multiple activity themes. For example, after customer profiles are built, the ECC algorithm is used to sequentially predict whether customers will participate in activity themes such as "holidays," "health," and "financial planning." The prediction results at each step serve as input features for the next classifier, thus forming a progressive label prediction chain. This method is particularly suitable for scenarios where there are strong dependencies between labels, such as customers who participate in "health" activities are more likely to participate in sub-categories like "health lectures" or "medical checkup discounts."

[0052] This step effectively enhances the model's ability to perceive dependencies between labels. By using preceding predictions as input to the current classifier, the model can leverage existing label information to assist in predicting subsequent labels, thereby improving overall classification accuracy and recall. Experiments show that, on the same dataset, this method achieves an F1-score improvement of over 5% compared to traditional multi-label classification models (such as binary correlation models), and is particularly robust in scenarios with sparse label distribution or class imbalance.

[0053] S3 performs SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the label topic distribution characteristics in order to optimize the model training effect.

[0054] Specifically, in this application, the step of "performing SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the label topic distribution characteristics to optimize model training effect" is a key step in improving the performance of multi-label classification models. This step mainly addresses the problem of imbalanced label category distribution in insurance marketing activities. It uses the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to synthesize and expand minority class samples, thereby improving the model's classification ability on long-tail labels.

[0055] The SMOTE algorithm generates new synthetic samples by performing linear interpolation between minority class samples. Specifically, for each minority class sample... From its Randomly select a sample from the nearest neighbors. And generate synthetic samples according to the following formula. :

[0056] in, A random number between 0 and 1, used to control the synthesis of samples. and The interpolation position between them. In some implementations, The value of is usually set to 5 or 10 to ensure the diversity and representativeness of the synthesized samples. In addition, to adapt to multi-label classification tasks, variants of SMOTE such as ML-SMOTE (Multi-Label SMOTE) can be adopted, which introduces label correlation in the label space to ensure that the synthesized samples are not only reasonable in the feature space, but also conform to the actual label co-occurrence pattern in the label space.

[0057] This step is applicable to scenarios where there is a significant category imbalance in insurance marketing campaigns, such as campaigns on certain themes (e.g., "health protection" or "holiday promotions") that only attract a small number of customers. SMOTE enhancement can effectively improve the model's ability to identify these tags, thereby increasing overall recommendation accuracy and coverage.

[0058] By synthesizing samples that conform to the characteristics of the label topic distribution, the model bias problem caused by class imbalance is alleviated, the classifier's generalization ability to long-tail labels is enhanced, and thus the performance of the ECC ensemble model in multi-label classification tasks is improved.

[0059] S4 integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

[0060] Specifically, in this application proposal, the step of "integrating the prediction results of multiple classifier chains, determining the final classification label of the target customer under each topic category through a voting mechanism, and outputting the multi-label classification decision for insurance marketing activities" is a key step in achieving multi-label classification prediction. This step is based on the ensemble learning framework of Ensemble of Classifier Chains (ECC), which uses multiple chain-structured classifiers to predict the likelihood of customer participation in different activity topic categories, and uses a voting mechanism to fuse the prediction results, thereby outputting the final multi-label classification decision.

[0061] Each classifier chain consists of multiple binary classification models, each responsible for predicting whether a customer participates in activities within a specific topic category. In this chained structure, the prediction result of the preceding classifier serves as one of the input features for the next classifier, thus explicitly modeling the dependencies between labels. In this scheme, the base classifier employs a decision tree model, whose input features include basic customer attributes (such as age, gender, and occupation), historical behavioral features (such as the frequency of past activity participation and preferred topics), and the predicted labels output by the preceding classifiers. By constructing multiple chains and randomly arranging the label order, ECC can capture the correlations between labels from different perspectives, enhancing the model's robustness and generalization ability.

[0062] Voting mechanisms typically employ majority voting or weighted voting strategies. In majority voting, if a label is predicted as "participate" by more than half of the chains, the final decision is "participate." Weighted voting assigns different weights to chains based on their prediction performance (such as accuracy and F1 score), and calculates a weighted average to determine the label's state. In this scheme, a weighted voting strategy may be optionally adopted, where the chain weights can be dynamically adjusted based on cross-validation results, for example, by setting a weight threshold. ,and This is to ensure the stability and accuracy of the prediction results.

[0063] This step is widely used in personalized customer recommendation scenarios in insurance marketing campaigns. For example, after a customer profile is built, the system inputs customer characteristics into the ECC model. Through voting on prediction results from multiple chains, it outputs multi-label classification results for customers under themes such as "Holidays," "Health," and "Financial Management." Marketers can then use these results to accurately push activities that match customer interests, improving customer engagement and conversion rates.

[0064] By integrating the prediction results of multiple chain classifiers, the limitations of a single model in label dependency modeling are effectively mitigated, while classification performance is improved under small sample sizes and imbalanced class data. The addition of a voting mechanism further enhances the stability and consistency of the prediction results, thereby improving the recommendation accuracy and customer satisfaction of insurance marketing campaigns.

[0065] Furthermore, it also includes: S5, based on feedback data from customers' historical participation in activities, constructs a dynamic adjustment module to calculate the prediction confidence of each classifier chain. When the confidence level is below the threshold When this happens, it triggers parameter updates in the rule classification module and retraining of the classifier chain.

[0066] Specifically, in some implementations, the dynamic adjustment module in this application proposes to construct a prediction confidence evaluation mechanism for the classifier chain by monitoring customer feedback data from historical participation activities in real time, thereby achieving adaptive optimization of the model. The core technical principle of this module is based on a nonlinear mapping of the prediction accuracy of the classifier chain using a logistic function to quantify the reliability of its predictions. Specifically, the prediction confidence... The calculation formula is:

[0067] in, Indicates the first The accuracy of each classifier chain on the current test set. and These are adjustable parameters that control the steepness of the function and the baseline confidence threshold, respectively. In practical applications, It is usually set to 10. Setting it to 0.7 ensures that the confidence level remains high even when the accuracy is slightly below the ideal value, thereby avoiding frequent triggering of the model update mechanism and reducing system overhead.

[0068] when The value is lower than the preset confidence threshold. At this time, the system will trigger parameter updates in the rule classification module and retraining of the classifier chain. This threshold... The parameter can be set according to business needs. For example, setting it to 0.85 indicates that when the prediction confidence of the classifier chain is below 85%, its predictive ability is considered to have declined, and model optimization is required. Parameter updates include adjustments to the keyword matching rules in the rule classification module, such as adding or deleting identification keywords for specific topics, and optimizing word frequency thresholds. The retraining of the classifier chain is based on the latest customer engagement data and updated label topic classification results. SMOTE technology is used to handle the class imbalance problem to improve the model's generalization ability on long-tail labels.

[0069] This step has significant technical value in insurance marketing campaigns. Through a dynamic adjustment mechanism, the system can automatically optimize the model structure and parameters based on actual classification results, thereby improving the accuracy and stability of multi-label classification. In practical deployments, this module can be integrated into customer profiling systems or recommendation engines to respond in real time to changes in customer behavior and achieve continuous optimization of personalized activity recommendations.

[0070] The multi-label classification method for insurance marketing activities based on rule-based pre-classification and ECC chain structure in this invention introduces a dynamic adjustment module based on historical customer feedback data to calculate and evaluate the prediction confidence of the classifier chain in real time. When the confidence is lower than a set threshold, parameter updates and model retraining are triggered, thereby further improving the adaptability and stability of multi-label classification in dynamically changing scenarios and significantly enhancing classification performance under conditions of small samples and class imbalance.

[0071] Example 2 like Figure 2 As shown in the embodiment of the present invention, the method for activity theme classification based on rule-based methods is as follows: In the original customer activity participation data, the activity label is the specific activity name, and each activity name is a unique string. There are a large number of activities. These activities have certain characteristics, target audiences, and themes. The rule-based method for theme classification helps the model learn more generalized features without relying on specific activity names. For example, if there are 1000 types of activities in the historical data, theme classification can reduce them to 5-10 categories.

[0072] Data preparation: Collect all activity names and perform preprocessing, such as cleaning, removing stop words, and word segmentation.

[0073] Define rules: Based on the understanding of the activity and prior knowledge, define a set of rules to identify the theme of the activity. These rules are based on keywords and phrases. Activity names follow certain patterns or rules. For example, activities held during specific holidays, revolving around the holiday theme and containing keywords such as "Mid-Autumn Festival" and "Spring Festival," can be classified as "holiday activities."

[0074] Applying rules: Apply the defined rules to each activity name to extract information about the topic. Find specific keywords and calculate the frequency of words related to a specific topic.

[0075] Categorization: Based on the extracted information, assign a theme to each activity name.

[0076] Evaluate the classification results: Manually check the classification results.

[0077] Optimize the rules: If there are problems with the evaluation results, adjust or expand the rules and repeat the above steps.

[0078] Ensemble of Classifier Chains, such as Figure 3 As shown.

[0079] Base classifier selection: Use decision trees as base classifiers to train a binary classification model for each category, i.e., the label is whether the user participated in the activity or not.

[0080] Prepare training data: Obtain historical insurance customer activity data and customer and policy characteristics, and divide the dataset into training and test sets. Preprocess the data, such as feature selection, missing value handling, and label recoding. For imbalanced data, use the SMOTE method to increase the sample size.

[0081] Sequential training of base classifiers: Several chain ordering methods for classifiers are determined, and each base classifier is trained sequentially. When training the current base classifier, the prediction results of the previous base classifier are used as additional features input to the current base classifier.

[0082] Verification and Adjustment: Integrate the centralized chain-based tag classification results and use a voting method to determine whether to invite or push notifications to customers for each activity category.

[0083] The multi-label classification method for insurance marketing activities based on the ECC algorithm in this invention first pre-classifies the activities according to rules, and then uses the ECC ensemble learning method to make full use of the correlation between category labels to improve the accuracy of multi-label classification of insurance marketing activities. This solves the problems of low model generalization ability, high risk of overfitting, and poor performance on small datasets in traditional solutions.

[0084] Example 3 This invention also provides a multi-label classification device 10 for insurance marketing activities based on the ECC algorithm, such as... Figure 4 As shown, the device 10 includes: The data acquisition and preprocessing module 100 is used to acquire raw insurance marketing activity data, perform word segmentation and keyword extraction on the activity name based on the preset activity theme classification rules, and map the original activity tags to 5-10 theme categories to form a dimensionality-reduced tag theme space. The classifier chain construction module 200 is used to construct a combined classifier chain model. It uses a decision tree as the base classifier and trains multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature for chain-like transmission when training the current classifier. The data augmentation module 300 is used to perform SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the label topic distribution characteristics in order to optimize the model training effect. The Results Integration and Decision Module 400 integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

[0085] Furthermore, the data acquisition and preprocessing module is also used for: The TF-IDF algorithm is used to calculate the weight of each keyword in the campaign name, and keywords with weights higher than a preset threshold are selected. Keywords are used as thematic features; A rule optimization model is constructed based on the historical activity classification results, and the classification accuracy is calculated. With rule coverage The combined metrics are used to dynamically adjust keyword matching rules.

[0086] Furthermore, the classifier chain building module is also used for: Based on the Pearson correlation coefficient between tag topics The ranking strategy for constructing the classifier chain is to prioritize classes with a correlation coefficient absolute value greater than 1. The topics are arranged in descending order; The predicted probability of the previous classifier Compared with the original feature vector The features are concatenated to form a new feature matrix. .

[0087] Furthermore, it also includes: The dynamic adjustment module is used to calculate the prediction confidence of each classifier chain based on feedback data from customers' historical participation in activities. When the confidence level is below the threshold When this happens, it triggers parameter updates in the rule classification module and retraining of the classifier chain.

[0088] This invention discloses a multi-label classification device for insurance marketing activities based on the ECC algorithm. By introducing a dynamic adjustment module based on historical customer feedback data, the device calculates and evaluates the prediction confidence of the classifier chain in real time. When the confidence is lower than a set threshold, parameter updates and model retraining are triggered, thereby further improving the adaptability and stability of multi-label classification in dynamically changing scenarios and significantly enhancing classification performance under conditions of small samples and class imbalance.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A multi-label classification method for insurance marketing activities based on the ECC algorithm, characterized in that, include: S1. Obtain the original insurance marketing campaign data, perform word segmentation and keyword extraction on the campaign name based on the preset campaign theme classification rules, and map the original campaign tags to 5-10 theme categories to form a dimensionality-reduced tag theme space. S2, construct a combined classifier chain model, using a decision tree as the base classifier, and train multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature for chain-like transmission when training the current classifier. S3, perform SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the characteristics of the label topic distribution in order to optimize the model training effect. S4 integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

2. The method as described in claim 1, characterized in that, The process of segmenting and extracting keywords from activity names based on preset activity theme classification rules, mapping original activity tags to 5-10 theme categories to form a dimensionality-reduced tag theme space, also includes: S11, use the TF-IDF algorithm to calculate the weight of each keyword in the activity name, and filter keywords with weights higher than a preset threshold. Keywords are used as thematic features; S12, Construct a rule optimization model based on the historical activity classification results, and calculate the classification accuracy. With rule coverage The combined metrics are used to dynamically adjust keyword matching rules.

3. The method as described in claim 1, characterized in that, The step of training multiple classifier chains according to a preset dependency relationship between tag topics, wherein each classifier chain uses the prediction result of the previous classifier as input features for chain-like passing during the training of the current classifier, further includes: S21, based on the Pearson correlation coefficient between labeled topics The ranking strategy for constructing the classifier chain is to prioritize classes with a correlation coefficient absolute value greater than 1. The topics are arranged in descending order; S22, the predicted probability from the previous classifier Compared with the original feature vector The features are concatenated to form a new feature matrix. .

4. The method as described in claim 1, characterized in that, Also includes: S5, based on feedback data from customers' historical participation in activities, constructs a dynamic adjustment module to calculate the prediction confidence of each classifier chain. When the confidence level is below the threshold When this happens, it triggers parameter updates in the rule classification module and retraining of the classifier chain.

5. A multi-label classification device for insurance marketing activities based on the ECC algorithm, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw insurance marketing activity data, perform word segmentation and keyword extraction on the activity name based on the preset activity theme classification rules, and map the original activity tags to 5-10 theme categories to form a dimensionality-reduced tag theme space. The classifier chain building module is used to build a combined classifier chain model. It uses a decision tree as the base classifier and trains multiple classifier chains according to the preset dependency relationship between the label topics. Each classifier chain uses the prediction result of the previous classifier as the input feature when training the current classifier. The data augmentation module is used to perform SMOTE data augmentation on imbalanced samples in the training data to generate synthetic samples that conform to the characteristics of the label topic distribution in order to optimize the model training effect. The results integration and decision-making module integrates the prediction results of multiple classifier chains, determines the final classification label of the target customer under each topic category through a voting mechanism, and outputs the multi-label classification decision for insurance marketing activities.

6. The apparatus as claimed in claim 5, characterized in that, The data acquisition and preprocessing module is also used for: The TF-IDF algorithm is used to calculate the weight of each keyword in the campaign name, and keywords with weights higher than a preset threshold are selected. Keywords are used as thematic features; A rule optimization model is constructed based on the historical activity classification results, and the classification accuracy is calculated. With rule coverage The combined metrics are used to dynamically adjust keyword matching rules.

7. The apparatus as claimed in claim 5, characterized in that, The classifier chain construction module is also used for: Based on the Pearson correlation coefficient between tag topics The ranking strategy for constructing the classifier chain is to prioritize classes with a correlation coefficient absolute value greater than 1. The topics are arranged in descending order; The predicted probability of the previous classifier Compared with the original feature vector The features are concatenated to form a new feature matrix. .

8. The apparatus as claimed in claim 5, characterized in that, Also includes: The dynamic adjustment module is used to calculate the prediction confidence of each classifier chain based on feedback data from customers' historical participation in activities. When the confidence level is below the threshold When this happens, it triggers parameter updates in the rule classification module and retraining of the classifier chain.

Images