Policy generation method, apparatus, device, medium, and program product
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BANK OF COMM CO LTD ANHUI BRANCH
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-21
Smart Images
Figure CN122434570A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence, and more particularly to a strategy generation method, apparatus, device, medium, and program product. Background Technology
[0002] With the rapid development of big data and artificial intelligence technologies, customer profiling can help banks gain a deeper understanding of customer characteristics, including behavioral preferences, financial status, and risk levels, through in-depth analysis of large amounts of customer transaction and behavioral data, thereby supporting the development of personalized services and marketing strategies.
[0003] Existing technologies mainly rely on statistical analysis and feature engineering methods to evaluate the predictive ability of features through binning, WOE (Weight of Evidence) encoding, and IV (Information Value) calculation, and generate detailed reports containing information such as feature distribution and sample proportion.
[0004] However, reports generated by existing technologies often contain a large number of complex statistical indicators, with multiple data layers and high information density. Business personnel find it difficult to quickly extract core features and determine business recommendations, which reduces the accuracy and efficiency of marketing decisions. Summary of the Invention
[0005] This application provides strategy generation methods, apparatus, devices, media, and program products to improve the accuracy and efficiency of marketing decisions.
[0006] In a first aspect, embodiments of this application provide a strategy generation method, including:
[0007] Obtain multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data;
[0008] Multiple original feature data are vectorized separately to obtain the feature vector corresponding to each original feature data. The similarity between each feature vector is calculated, and redundant features are removed based on the similarity to obtain the filtered feature vector.
[0009] For the filtered feature vectors, calculate the Pearson correlation coefficient between each feature vector and the customer classification result. Select the first quantitative feature as the significance feature according to the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significance features as the core feature.
[0010] Customers are segmented based on salient features to generate customer segmentation results. The core features and customer segmentation results are stored in a structured knowledge base. In response to business queries, the core features are retrieved from the structured knowledge base and input into a large language model to generate strategy suggestions for business queries.
[0011] Optionally, customers are segmented based on salient features to generate customer segmentation results, specifically including:
[0012] The significant feature data of each customer is used as input, and the binary K-means algorithm is used for clustering.
[0013] The input salient feature data is regarded as an initial cluster. The total variance of the initial cluster is calculated. Using the K-means++ initialization method, the feature vector with the smallest total variance of all feature vectors in the initial cluster is selected as the cluster center point for binary splitting. New cluster centers are generated and customers are reassigned. The center point selection and splitting steps are recursively repeated until the preset number of clusters is reached.
[0014] Assign a hierarchical label describing the characteristics of the corresponding group to each cluster generated in order to produce customer stratification results.
[0015] Optionally, the core features and customer segmentation results can be stored in a structured knowledge base, specifically including:
[0016] The core features, salient features, and customer segmentation results are stored in a structured knowledge base in a standardized format.
[0017] When a new customer is acquired, feature calculation, feature extraction, and clustering stratification steps are performed on the new customer, and the updated results are stored in a structured knowledge base; when an existing customer is queried, the corresponding data is retrieved from the structured knowledge base.
[0018] Optionally, the retrieved core features are input into a large language model to generate strategy suggestions for business queries, specifically including:
[0019] Determine the corresponding prompt word template based on the retrieved core features;
[0020] The core features and prompt word templates are input into the large language model so that the large language model can output policy suggestions. The prompt word templates are used to guide the large language model to generate policy suggestions based on structured data.
[0021] Verify whether the strategy suggestions output by the large language model meet the requirements of text accuracy and logical consistency; if not, regenerate the model.
[0022] Optionally, the raw feature data includes structured behavioral data and unstructured text data;
[0023] Vectorizing any original feature data yields a feature vector corresponding to the original feature data, specifically including:
[0024] If the original feature data is structured behavioral data, then the structured behavioral data is binned, and a feature vector corresponding to the structured behavioral data is constructed based on the evidence weight value, information value and positive and negative sample ratio of each bin.
[0025] If the original feature data is unstructured text data, then text feature indicators are extracted through natural language processing, the text feature indicators are binned, and feature vectors corresponding to the unstructured text data are constructed based on the evidence weight value, information value and positive and negative sample ratio of each bin.
[0026] Optionally, the structured knowledge base is updated regularly, including adding new feature data, removing outdated feature data, and merging duplicate feature data.
[0027] Secondly, embodiments of this application provide a strategy generation apparatus, comprising:
[0028] The acquisition module is used to acquire multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data.
[0029] The processing module is used to vectorize multiple original feature data separately to obtain the feature vector corresponding to each original feature data, calculate the similarity between each feature vector, remove redundant features based on the similarity, and obtain the filtered feature vector.
[0030] The processing module is also used to calculate the Pearson correlation coefficient between each feature vector and the customer classification result for the filtered feature vectors, select the first quantitative feature as the significance feature according to the sorting of the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significance features as the core feature.
[0031] The processing module is also used to segment customers based on salient features and generate customer segmentation results; store the core features and customer segmentation results in a structured knowledge base; retrieve the core features from the structured knowledge base in response to business queries; input the retrieved core features into a large language model to generate strategy suggestions for business queries.
[0032] Optionally, the processing module is also used to take the salient feature data of each customer as input and perform clustering using the binary K-means algorithm;
[0033] The input salient feature data is regarded as an initial cluster. The total variance of the initial cluster is calculated. Using the K-means++ initialization method, the feature vector with the smallest total variance of all feature vectors in the initial cluster is selected as the cluster center point for binary splitting. New cluster centers are generated and customers are reassigned. The center point selection and splitting steps are recursively repeated until the preset number of clusters is reached.
[0034] Assign a hierarchical label describing the characteristics of the corresponding group to each cluster generated in order to produce customer stratification results.
[0035] Optionally, the processing module is also used to store the core features, salient features, and customer segmentation results into a structured knowledge base in a standardized format;
[0036] When a new customer is acquired, feature calculation, feature extraction, and clustering stratification steps are performed on the new customer, and the updated results are stored in a structured knowledge base; when an existing customer is queried, the corresponding data is retrieved from the structured knowledge base.
[0037] Optionally, the processing module is also used to determine the corresponding prompt word template based on the retrieved core features;
[0038] The core features and prompt word templates are input into the large language model so that the large language model can output policy suggestions. The prompt word templates are used to guide the large language model to generate policy suggestions based on structured data.
[0039] Verify whether the strategy suggestions output by the large language model meet the requirements of text accuracy and logical consistency; if not, regenerate the model.
[0040] Optionally, the multiple original feature data include structured behavioral data and unstructured text data; the processing module is also used to bin the structured behavioral data when the original feature data is structured behavioral data, and to construct the feature vector corresponding to the structured behavioral data based on the evidence weight value, information value and positive and negative sample ratio of each bin;
[0041] If the original feature data is unstructured text data, then text feature indicators are extracted through natural language processing, the text feature indicators are binned, and feature vectors corresponding to the unstructured text data are constructed based on the evidence weight value, information value and positive and negative sample ratio of each bin.
[0042] Optionally, the processing module is also used to periodically update the structured knowledge base, including adding new feature data, removing outdated feature data, and merging duplicate feature data.
[0043] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;
[0044] The memory stores instructions that the computer executes;
[0045] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.
[0046] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.
[0047] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.
[0048] The strategy generation method, apparatus, device, medium, and program product provided in this application embodiment acquires multiple original feature data of customers and calculates the evidence weight value and information value corresponding to each original feature data. After vectorizing the multiple original feature data to obtain feature vectors, the similarity between each feature vector is calculated, and redundant features are removed based on the similarity. For the selected feature vectors, the Pearson correlation coefficient between each feature vector and the customer classification result is calculated, and the first quantitative feature is selected as the salient feature according to the ranking. The second quantitative feature with the highest positive correlation among the salient features is selected as the core feature. Customers are segmented based on the salient features to generate customer stratification results. The core features and customer stratification results are stored in a structured knowledge base. In response to business queries, the core features are retrieved from the structured knowledge base and input into a large language model to generate strategy suggestions for business queries. This solves the problems of low processing efficiency, difficulty in extracting and interpreting complex feature information, and difficulty in transforming complex data into actionable strategy suggestions caused by the reliance on manual analysis in traditional customer profile report interpretation. Therefore, it achieves the effect of improving the efficiency of customer profile interpretation and the accuracy and reliability of marketing strategy generation. Attached Figure Description
[0049] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0050] Figure 1 A flowchart illustrating the strategy generation method provided in this application;
[0051] Figure 2 A schematic diagram of the strategy generation system provided in this application;
[0052] Figure 3 A schematic diagram of the strategy generation device provided in this application;
[0053] Figure 4 A schematic diagram of the structure of the electronic device provided in this application.
[0054] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0055] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0056] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation portals for users to choose to authorize or refuse.
[0057] In the financial sector (such as banking and insurance), customer profiling and targeted marketing are key tools for improving business efficiency and customer satisfaction. Banks and other financial institutions collect and analyze multi-dimensional data such as customer transaction records, behavioral preferences, and financial status to build customer profiles in order to identify potential needs and optimize product recommendation strategies.
[0058] Existing customer profiling technologies mainly rely on statistical analysis and feature engineering methods. They evaluate the predictive ability of features through binning, WOE (Weight of Evidence) coding, and IV (Information Value) calculation, and generate detailed reports containing information such as feature distribution and sample proportion.
[0059] However, reports generated by existing customer profiling technologies typically contain a large number of complex statistical indicators with multiple data layers and high information density, making it difficult for business personnel to quickly extract core features and determine business recommendations. Furthermore, with the explosive growth of data volume, manual report interpretation and analysis are time-consuming and prone to missing important information. The analytical efficiency cannot meet the needs of real-time business decision-making, easily leading to problems such as delayed marketing strategy development and unreasonable resource allocation.
[0060] The strategy generation method provided in this application obtains multiple original feature data of customers and calculates the evidence weight and information value corresponding to each original feature data. After vectorizing the multiple original feature data to obtain feature vectors, the similarity between each feature vector is calculated and redundant features are removed based on the similarity. For the selected feature vectors, the Pearson correlation coefficient between each feature vector and the customer classification result is calculated. The first quantitative feature is selected as the salient feature according to the ranking, and the second quantitative feature with the highest positive correlation is selected as the core feature from the salient features. Customers are divided based on the salient features to generate customer stratification results. The core features and customer stratification results are stored in a structured knowledge base. In response to business queries, the core features are retrieved from the structured knowledge base and input into a large language model to generate strategy suggestions for business queries. In this process, redundant features were eliminated through similarity calculations to reduce the computational complexity of subsequent steps. Significant and core features were selected using Pearson correlation coefficients to ensure that the features used for customer segmentation and strategy generation have strong predictive capabilities. The core features and segmentation results were stored in a structured knowledge base to achieve efficient reuse of feature data. By inputting core features into a large language model to generate strategy suggestions, the structured data was transformed into natural language. This effectively solved the problems of low processing efficiency, difficulty in extracting and interpreting complex feature information, and difficulty in transforming complex data into actionable strategy suggestions caused by the reliance on manual analysis in traditional customer profile reports. As a result, the accuracy and reliability of marketing strategy generation were improved.
[0061] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0062] Figure 1 A flowchart illustrating the strategy generation method provided in this application is shown below. Figure 2 As shown, the method includes:
[0063] S101. Obtain multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data.
[0064] In one possible embodiment, the raw feature data obtained comes from the bank's customer relationship management system, including customer age, average monthly transaction amount over the past three months, deposit balance, occupation type, and customer service call records. Taking the feature "average monthly transaction amount over the past three months" as an example, this feature is first binned into three bin intervals: "[0, 1000)", "[1000, 5000)", and "[5000, +∞)". The total number of customers, the number of positive samples (e.g., those who have purchased wealth management products), and the number of negative samples (e.g., those who have not purchased wealth management products) within each bin are counted. Based on the above statistical data, the evidence weight value and information value of each bin are calculated, where the evidence weight value of the j-th bin is... The calculation formula is:
[0065]
[0066] Information value of the j-th sub-box The calculation formula is:
[0067]
[0068] The information value of a feature is obtained by summing the information values of all bins within that feature, and the evidence weight value of that feature is obtained by summing the evidence weight values of all bins within that feature. This method is then used to iterate through all the original feature data, calculating the evidence weight value and information value for each feature.
[0069] S102. Vectorize the multiple original feature data separately to obtain the feature vector corresponding to each original feature data. Calculate the similarity between each feature vector and remove redundant features based on the similarity to obtain the filtered feature vector.
[0070] More specifically, the multiple raw feature data include structured behavioral data and unstructured text data. Any one of the raw feature data is vectorized to obtain a corresponding feature vector. Specifically, if the raw feature data is structured behavioral data, the structured behavioral data is binned, and a feature vector corresponding to the structured behavioral data is constructed based on the evidence weight value, information value, and positive / negative sample ratio of each bin. If the raw feature data is unstructured text data, text feature indicators are extracted through natural language processing, and the text feature indicators are binned. A feature vector corresponding to the unstructured text data is constructed based on the evidence weight value, information value, and positive / negative sample ratio of each bin.
[0071] In one possible implementation, taking the structured behavioral data "age" as an example, the feature is divided into four bins: "[18,25]", "[26, 35]", "[36, 45]", and "[46, +∞]". The number of people in each bin, the percentage of people in that bin relative to the total number of people in all bins for that feature (i.e., the percentage of people in that bin), the number of positive samples, and the number of negative samples are calculated. The evidence weight, information value, and positive sample percentage for each bin are also calculated. These statistics for each bin (i.e., the number of people in that bin, the percentage of people in that bin, the number of positive samples, the number of negative samples, the positive sample percentage, the evidence weight, and the information value) are combined into a multi-dimensional vector, which serves as the feature vector for that bin. The feature vectors of all bins are concatenated to form the overall feature vector for the "age" feature.
[0072] In one possible implementation, taking unstructured text data "customer service call records" as an example, the text of the customer service call records is first segmented to remove stop words. Then, the TF-IDF algorithm is used to extract keywords, and a sentiment analysis model is used to calculate the sentiment score (range 0 to 1) for each customer service call record. The sentiment score is used as a text feature indicator and divided into three bins: "[0, 0.3)", "[0.3, 0.6)", and "[0.6, 1.0]". The number of participants, the percentage of participants, the number of positive samples, and the number of negative samples in each bin are counted, and the evidence weight, information value, and positive sample percentage of each bin are calculated. These statistics for each bin are combined into a multi-dimensional vector as the feature vector for that bin. The feature vectors of all bins are concatenated to generate the overall feature vector of the "customer service call records".
[0073] This embodiment unifies the original feature data of different types into a binned vector format with evidence weight value, information value and positive and negative sample ratio as the core, so that heterogeneous data can be similarity calculated and feature filtered in the same framework, laying the foundation for subsequent feature deduplication.
[0074] In one possible embodiment, after obtaining the feature vector corresponding to each original feature data, the feature vector of "average monthly transaction amount in the past 3 months" is used. and the feature vector of "average monthly deposit balance in the past 3 months" For example, calculate the cosine similarity between two feature vectors. The formula for calculating cosine similarity is:
[0075]
[0076] The similarity threshold is set to 0.85. When the cosine similarity between two feature vectors is greater than 0.85, these two features are considered redundant. All features are iterated through to remove redundant features, resulting in a set of filtered feature vectors.
[0077] Optionally, after completing the removal of redundant features, features with information value lower than a preset information value threshold are also removed.
[0078] S103. For the filtered feature vectors, calculate the Pearson correlation coefficient between each feature vector and the customer classification result. Select the first quantitative feature as the significant feature according to the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significant features as the core feature.
[0079] In one possible implementation, taking the remaining 50 feature vectors after filtering as an example, the classification label for each customer in the training dataset is obtained. This classification label indicates whether the customer belongs to the target group (e.g., high-risk customer). For each feature vector, the proportion of positive samples in each bin and the customer's classification label are substituted into the formula for the Pearson correlation coefficient r:
[0080]
[0081] in, Let be the percentage of positive samples in the bin for the i-th customer on this feature. For the classification result label of the i-th customer, and These are the means of the two. The Pearson correlation coefficient between each feature vector and the classification result is calculated. The absolute value of the Pearson correlation coefficient represents the strength of the linear correlation between the feature and the classification result, and the positive or negative sign indicates a positive or negative correlation.
[0082] In one possible embodiment, features are sorted from largest to smallest based on the absolute value of the Pearson correlation coefficient, and the top 20 features (i.e., the number of first quantitative feature indicators is 20) are selected as significant features. Then, from these 20 significant features, the top 5 features with the highest positive Pearson correlation coefficients (i.e., the number of second quantitative feature indicators is 5) are selected as core features.
[0083] S104. Based on salient features, segment customers and generate customer segmentation results; store core features and customer segmentation results in a structured knowledge base; when responding to business queries, retrieve core features from the structured knowledge base, input the retrieved core features into a large language model, and generate strategy suggestions for business queries.
[0084] More specifically, customer segmentation is performed based on salient features to generate customer stratification results. This includes: taking the salient feature data of each customer as input and performing clustering using the binary K-means algorithm; treating the input salient feature data as an initial cluster, calculating the total variance of the initial cluster, and using the K-means++ initialization method to select the feature vector with the smallest total variance from all feature vectors within the initial cluster as the cluster center point for binary splitting, generating new cluster centers and reassigning customers, recursively repeating the center point selection and splitting steps until the preset number of clusters is reached; and assigning a stratification label describing the corresponding group characteristics to each cluster to generate customer stratification results.
[0085] In one possible implementation, 20 salient feature data points from 10,000 customers are acquired to construct a feature matrix M. Each row of M corresponds to a customer, and each column corresponds to a salient feature. Each customer is represented as a 20-dimensional vector. This feature matrix M is then used as input data for a clustering algorithm. First, the entire feature matrix M is considered as an initial cluster C0, and the total variance of cluster C0 is calculated. The calculation formula is:
[0086]
[0087] in, Let i be the 20-dimensional vector of the i-th customer. Let C0 be the center point of the initial cluster, which is the mean of all customer vectors. If the preset number of clusters k=5, then when the current number of clusters is less than k, a binary search is performed on each existing cluster. For each cluster C, the K-means++ initialization method is used to select two cluster centers: that is, a customer vector in C is randomly selected as the first cluster center. For each remaining customer vector in C ,calculate The square of the shortest distance to the selected cluster centers The next cluster center is selected according to the following probability formula. This ensures that the initial cluster centers are far apart from each other.
[0088]
[0089] Optionally, based on the two selected cluster centers and All customer vectors in cluster C are assigned to the nearest cluster center, forming two sub-clusters C1 and C2. The assignment rule is as follows: If... Then If it is assigned to C1, then it is assigned to C2. . Indicates the first The sample set of each cluster It is the first The center of each cluster.
[0090] Optionally, the center points μ1 and μ2 of C1 and C2 are calculated respectively, which are the mean values of the vectors within each sub-cluster, and μ1 and μ2 are used as the new cluster centers. The formula for calculating these center points is:
[0091]
[0092] Optionally, the above steps are repeated, with the assignment and center point update steps performed repeatedly, until the clustering results converge, i.e., the center points no longer change. At this point, the sum of the total variances of C1 and C2 is calculated. .
[0093] Optionally, by traversing each existing cluster, the sum of the total variances after bisection is calculated for each cluster. The cluster that minimizes the sum of the total variances after bisection is selected as the actual cluster for this split. This cluster is divided into two sub-clusters, replacing the original cluster and increasing the current cluster size by 1. The above bisection attempt and actual splitting process is repeated until the current cluster size reaches the preset cluster size k=5. At this point, the algorithm terminates, resulting in 5 final clusters.
[0094] In one possible implementation, after obtaining five final clusters, a hierarchical label is assigned to each final cluster. That is, based on the distribution of cluster centroid values across various saliency features, the typical characteristics of each cluster are identified. For example, if the centroid of the first cluster has a high value for the "average monthly transaction amount over the past three months" feature and a low value for the "risk preference score" feature, and the average age of customers within the cluster is greater than 45 years old, then this cluster is labeled as "high-net-worth, stable customer group"; if the centroid of the second cluster has a low value for the "age" feature and a high value for the "monthly active days of mobile banking" feature, then it is labeled as "young, growing customer group"; if the centroid of the third cluster has a high value for the "age" feature and a low value for the "number of wealth management products held" feature, then it is labeled as "elderly, conservative customer group". Through this method, hierarchical labels describing the characteristics of the corresponding groups are generated for each of the five final clusters, forming the customer segmentation results.
[0095] This embodiment combines the binary K-means algorithm with the K-means++ initialization method, which solves the problem of unstable results caused by the random initialization of traditional K-means, improves the clustering quality, and enhances the accuracy and stability of customer stratification results.
[0096] Optionally, the core features and customer segmentation results are stored in a structured knowledge base. Specifically, this includes: storing the core features, salient features, and customer segmentation results in a standardized format in the structured knowledge base; when a new customer is acquired, performing feature calculation, feature extraction, and clustering segmentation steps on the new customer, and storing the updated results in the structured knowledge base; when querying existing customers, retrieving the corresponding data from the structured knowledge base.
[0097] In one possible implementation, after generating customer segmentation results, a structured knowledge base is constructed, implemented using a relational database. Twenty salient features, along with their corresponding feature names, binning intervals, evidence weights, information values, and Pearson correlation coefficients, are stored in a feature table. Five core features are individually labeled and stored in a core feature table. The customer segmentation results are stored in a segmentation result table, where each record contains a customer identifier and a corresponding segmentation label. When the banking system adds 1000 new customers, the complete feature calculation, feature extraction, and clustering segmentation steps described above are performed on these 1000 new customers, and the feature data and incremental segmentation results for the new customers are inserted into the corresponding table in the knowledge base. When business personnel query existing customers, the system directly retrieves the customer's feature data and segmentation labels from the corresponding table in the knowledge base, eliminating the need for recalculation.
[0098] This embodiment achieves persistent storage and efficient reuse of feature data and hierarchical results by constructing a structured knowledge base and distinguishing the processing logic between new and old customers. This avoids redundant calculations, reduces system resource consumption, and improves data timeliness.
[0099] Optionally, the retrieved core features are input into a large language model to generate strategy suggestions for business queries. Specifically, this includes: determining the corresponding prompt word template based on the retrieved core features; inputting the core features and prompt word template into the large language model so that the large language model outputs strategy suggestions, wherein the prompt word template is used to guide the large language model to generate strategy suggestions based on structured data; verifying whether the strategy suggestions output by the large language model meet the requirements of text accuracy and logical consistency, and regenerating them if they do not.
[0100] In one possible implementation, after retrieving the core features, the business personnel input the business query "Please generate a marketing strategy for high-net-worth, stable clients." The system retrieves five core feature data corresponding to this client group from the structured knowledge base, including: "Average monthly transaction amount in the past 3 months" is in the high range (greater than 50,000 yuan), "Number of wealth management products held" is greater than 3, "Risk preference score" is low risk, "Customer service call sentiment score" is greater than 0.8, and "Age" is 45 years or older. The system determines the "strategy suggestion template" corresponding to the current query type as the prompt word template based on the preset prompt word mapping relationship. The prompt word template content is: "You are a financial strategy analysis expert. Generate a marketing strategy for [client group name] based on the following structured data: [Core feature data]". The system fills the core feature data into the above strategy suggestion template, thereby generating a complete prompt word instruction, and inputs it into the large language model. The large language model outputs strategy suggestions based on core features and the prompt word instruction, such as "For high-net-worth, conservative clients, we recommend low-risk fixed-income financial products. Combined with their high trading activity and proactive customer service feedback, one-on-one asset allocation services can be provided through dedicated account managers." The system then validates the strategy suggestion text, checking for factual errors (e.g., mismatch between feature data and the risk level of the suggested product) and logical contradictions (e.g., recommending high-risk products to clients with low-risk preferences). If the validation passes, the strategy suggestion is output; otherwise, it is regenerated.
[0101] This embodiment works in collaboration between a structured knowledge base and a large language model to transform precise structured feature data into natural language strategy suggestions, solving the problems of low efficiency in manual interpretation and difficulty in transforming data into business strategies.
[0102] Optionally, the structured knowledge base is updated regularly, including adding new feature data, removing outdated feature data, and merging duplicate feature data.
[0103] In one possible implementation, the structured knowledge base is updated monthly. First, newly generated feature data is detected, such as the newly added "monthly active days for mobile banking" feature. The system calculates the evidence weight, information value, and Pearson correlation coefficient for this feature and adds it and its statistical indicators to the feature table. Second, outdated feature data is detected, such as the discontinued "number of transactions at branch counters" feature. The system marks this feature as invalid and removes it from the feature table. Finally, a duplicate check is performed on the feature table. If two features with different names but a feature vector similarity exceeding 0.95 (i.e., a similarity threshold of 0.95) and similar information values (e.g., a difference less than a preset information value difference) are found, they are determined to be duplicate features. The two features are merged to remove duplicates, retaining one feature and updating the relevant reference relationships.
[0104] This embodiment improves the accuracy and timeliness of data in the structured knowledge base through a periodic update mechanism, avoiding the problem of strategy recommendations becoming invalid due to outdated data.
[0105] The strategy generation method provided in this application reduces the computational complexity by calculating the evidence weight and information value of the original feature data, vectorizing the original feature data, and then calculating similarity to remove redundant features. It also reduces the computational complexity by using Pearson correlation coefficients to screen significant and core features, ensuring that the features used for customer segmentation and strategy generation have strong predictive capabilities. Furthermore, it improves the accuracy and stability of customer segmentation by combining the binary K-means algorithm with the K-means++ initialization method for clustering. Finally, it generates strategy suggestions by inputting core features into a large language model, enabling business personnel to directly obtain actionable marketing strategies without relying on manual analysis, thus solving the problems of low interpretation efficiency and difficulty in transforming complex data into business strategies using traditional methods.
[0106] Figure 2 A schematic diagram of the structure of the strategy generation system provided in this application is shown below. Figure 2 As shown, the strategy generation system includes a data source module, a processing engine module, a knowledge base module, a large language model, and an application module. The data source module collects raw customer feature data, including but not limited to basic customer information, transaction behavior data, and customer service dialogue records, and sends this raw feature data to the processing engine module. The processing engine module receives the raw feature data from the data source module, performs feature engineering and clustering / stratification processing on the raw feature data, generates salient features, core features, and customer segmentation results, and stores the processing results in the knowledge base module. The knowledge base module receives and stores the salient features, core features, and customer segmentation results sent by the processing engine module. When business personnel initiate a business query through the application module, the knowledge base module responds to the business query, retrieves the corresponding core features and customer segmentation results, and sends the retrieved data to the large language model. The large language model receives the core features and customer segmentation results from the knowledge base module, determines the corresponding prompt word template based on the retrieved core features, and generates a natural language strategy suggestion based on the prompt word template, returning the generated strategy suggestion to the application module. This application module is used to send the business query input by the business personnel to the knowledge base module when it receives the business query, and to receive the strategy suggestions returned by the large language model and display the strategy suggestions to the business personnel.
[0107] Figure 3 A schematic diagram of the strategy generation device provided in this application is shown below. Figure 3 As shown, the strategy generation device 30 provided in this embodiment includes:
[0108] The acquisition module 301 is used to acquire multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data.
[0109] Processing module 302 is used to vectorize multiple original feature data separately to obtain the feature vector corresponding to each original feature data, calculate the similarity between each feature vector, remove redundant features based on the similarity, and obtain the filtered feature vector.
[0110] The processing module 302 is also used to calculate the Pearson correlation coefficient between each feature vector and the customer classification result for the filtered feature vectors, select the first quantitative feature as the significance feature according to the sorting of the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significance features as the core feature.
[0111] The processing module 302 is also used to segment customers based on salient features and generate customer segmentation results; store the core features and customer segmentation results in a structured knowledge base; retrieve the core features from the structured knowledge base in response to business queries; input the retrieved core features into a large language model to generate strategy suggestions for business queries.
[0112] Optionally, the processing module 302 is also used to take the salient feature data of each customer as input and perform clustering using the binary K-means algorithm;
[0113] The input salient feature data is regarded as an initial cluster. The total variance of the initial cluster is calculated. Using the K-means++ initialization method, the feature vector with the smallest total variance of all feature vectors in the initial cluster is selected as the cluster center point for binary splitting. New cluster centers are generated and customers are reassigned. The center point selection and splitting steps are recursively repeated until the preset number of clusters is reached.
[0114] Assign a hierarchical label describing the characteristics of the corresponding group to each cluster generated in order to produce customer stratification results.
[0115] Optionally, the processing module 302 is also used to store the core features, salient features and customer segmentation results into a structured knowledge base in a standardized format;
[0116] When a new customer is acquired, feature calculation, feature extraction, and clustering stratification steps are performed on the new customer, and the updated results are stored in a structured knowledge base; when an existing customer is queried, the corresponding data is retrieved from the structured knowledge base.
[0117] Optionally, the processing module 302 is also used to determine the corresponding prompt word template based on the retrieved core features;
[0118] The core features and prompt word templates are input into the large language model so that the large language model can output policy suggestions. The prompt word templates are used to guide the large language model to generate policy suggestions based on structured data.
[0119] Verify whether the strategy suggestions output by the large language model meet the requirements of text accuracy and logical consistency; if not, regenerate the model.
[0120] Optionally, the multiple original feature data include structured behavioral data and unstructured text data; the processing module 302 is also used to bin the structured behavioral data when the original feature data is structured behavioral data, and to construct the feature vector corresponding to the structured behavioral data based on the evidence weight value, information value and positive and negative sample ratio of each bin.
[0121] If the original feature data is unstructured text data, then text feature indicators are extracted through natural language processing, the text feature indicators are binned, and feature vectors corresponding to the unstructured text data are constructed based on the evidence weight value, information value and positive and negative sample ratio of each bin.
[0122] Optionally, the processing module 302 is also used to periodically update the structured knowledge base, including adding new feature data, removing outdated feature data, and merging duplicate feature data.
[0123] The strategy generation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0124] Figure 4 A schematic diagram of the structure of the electronic device provided in this application. Figure 4 As shown, the electronic device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 40 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.
[0125] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.
[0126] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0127] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0128] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0129] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0130] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0131] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.
[0132] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.
[0133] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.
[0134] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.
[0135] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0136] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0137] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0138] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0139] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A strategy generation method, characterized in that, include: Obtain multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data; The original feature data are vectorized to obtain the feature vector corresponding to each original feature data. The similarity between each feature vector is calculated. Redundant features are removed based on the similarity to obtain the filtered feature vector. For the filtered feature vectors, calculate the Pearson correlation coefficient between each feature vector and the customer classification result. Select the first quantitative feature as the significance feature according to the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significance features as the core feature. Customers are segmented based on the salient features to generate customer segmentation results; the core features and the customer segmentation results are stored in a structured knowledge base; in response to business queries, the core features are retrieved from the structured knowledge base; the retrieved core features are input into a large language model to generate strategy suggestions for the business query.
2. The method according to claim 1, characterized in that, Based on the aforementioned salient features, customers are segmented to generate customer segmentation results, specifically including: The significant feature data of each customer is used as input, and the binary K-means algorithm is used for clustering. The input salient feature data is regarded as an initial cluster. The total variance of the initial cluster is calculated. Using the K-means++ initialization method, the feature vector with the smallest total variance of all feature vectors in the initial cluster is selected as the cluster center point for binary splitting. New cluster centers are generated and customers are reassigned. The center point selection and splitting steps are recursively repeated until the preset number of clusters is reached. Each cluster generated is assigned a hierarchical label describing the characteristics of the corresponding group to generate the customer segmentation results.
3. The method according to claim 1, characterized in that, The core features and customer segmentation results are stored in a structured knowledge base, specifically including: The core features, the salient features, and the customer segmentation results are stored in the structured knowledge base according to a standardized format. When a new customer is acquired, feature calculation, feature extraction, and clustering and hierarchical steps are performed on the new customer, and the updated results are stored in the structured knowledge base; when an existing customer is queried, the corresponding data is retrieved from the structured knowledge base.
4. The method according to claim 1, characterized in that, The retrieved core features are input into a large language model to generate strategy suggestions for the business query, specifically including: Determine the corresponding prompt word template based on the retrieved core features; The core features and the prompt word template are input into the large language model so that the large language model outputs policy suggestions, wherein the prompt word template is used to guide the large language model to generate policy suggestions based on structured data; Verify whether the strategy suggestions output by the large language model meet the requirements of text accuracy and logical consistency; if not, regenerate them.
5. The method according to claim 1, characterized in that, The multiple raw feature data include structured behavioral data and unstructured text data; Vectorizing any original feature data to obtain the feature vector corresponding to the original feature data specifically includes: If the original feature data is structured behavioral data, then the structured behavioral data is binned, and a feature vector corresponding to the structured behavioral data is constructed based on the evidence weight value, information value and positive and negative sample ratio of each bin. If the original feature data is unstructured text data, then text feature indicators are extracted through natural language processing, the text feature indicators are binned, and feature vectors corresponding to the unstructured text data are constructed based on the evidence weight value, information value, and positive and negative sample ratio of each bin.
6. The method according to claim 3, characterized in that, Also includes: The structured knowledge base is updated regularly, including adding new feature data, removing outdated feature data, and merging duplicate feature data.
7. A strategy generation apparatus, characterized in that, include: The acquisition module is used to acquire multiple raw feature data of the customer and calculate the evidence weight value and information value corresponding to each raw feature data. The processing module is also used to vectorize the multiple original feature data respectively to obtain the feature vector corresponding to each original feature data, calculate the similarity between each feature vector, and remove redundant features according to the similarity to obtain the filtered feature vector. The processing module is also used to calculate the Pearson correlation coefficient between each feature vector and the customer classification result for the filtered feature vectors, select the first quantitative feature as the significance feature according to the sorting of the Pearson correlation coefficient, and select the second quantitative feature with the highest positive correlation from the significance feature as the core feature. The processing module is further configured to segment customers based on the salient features and generate customer segmentation results; store the core features and the customer segmentation results in a structured knowledge base; retrieve the core features from the structured knowledge base in response to a business query; input the retrieved core features into a large language model to generate strategy suggestions for the business query.
8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.
10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.