Product recommendation method and apparatus, computer device, and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a product recommendation method and device, computer equipment and a storage medium, which comprise the following steps: obtaining original object information of a target object; performing feature conversion on the original object information to obtain object features; performing cluster quantity calculation based on the object features to obtain a target cluster quantity; performing clustering processing on the object features based on the target cluster quantity to obtain target clustering features; generating target portrait data of the target object based on the target clustering features and the original object information; performing product screening on candidate products in a product resource library based on the target portrait data to obtain target recommended products, and recommending the target recommended products to the target object. The application can be applied to a financial product recommendation scene of financial technology, and improves the accuracy of product recommendation.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence technology and data analysis technology, and is applicable to the financial technology field, particularly to a product recommendation method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Product recommendation methods can recommend relevant products to users based on their personal information. For example, in a financial context, sales personnel can recommend financial products such as wealth management products, insurance products, and financial leasing products based on customer information.

[0003] Currently, product recommendations in financial scenarios primarily employ customer segmentation. This involves first classifying customers based on their information, and then recommending matching financial products to different customer groups. However, customer information in the financial industry exhibits significant variations in numerical characteristics such as financing amount, registered capital, and annual revenue. Customer segmentation is easily affected by extreme values, and customer information also includes category characteristics, which traditional clustering algorithms struggle to accurately handle. This leads to reduced accuracy in customer segmentation, which in turn affects the accuracy of product recommendations. Summary of the Invention

[0004] This application provides a product recommendation method, apparatus, computer equipment, and storage medium to solve the technical problem that the accuracy of product recommendations is affected by the inability to accurately classify customer groups.

[0005] Firstly, a product recommendation method is provided, including: Obtain the original object information of the target object; The original object information is subjected to feature transformation to obtain object features; The number of clusters is calculated based on the object characteristics to obtain the target cluster number. The object features are clustered based on the number of target clusters to obtain target cluster features; Based on the target clustering features and the original object information, target profile data of the target object is generated; Based on the target profile data, candidate products in the product resource library are screened to obtain target recommended products, and the target recommended products are recommended to the target object.

[0006] Secondly, a product recommendation device is provided, comprising: The object information acquisition module is used to obtain the original object information of the target object; The object feature conversion module is used to perform feature conversion on the original object information to obtain object features; The cluster count calculation module is used to calculate the number of clusters based on the object characteristics to obtain the target cluster count. A clustering processing module is used to perform clustering processing on the object features based on the number of target clusters to obtain target cluster features; The portrait generation module is used to generate target portrait data of the target object based on the target clustering features and the original object information; The recommended product filtering module is used to filter candidate products in the product resource library based on the target profile data, obtain target recommended products, and recommend the target recommended products to the target object.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the product recommendation method described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the product recommendation method described above.

[0009] The aforementioned product recommendation method, apparatus, computer equipment, and storage medium, specifically designed for financial product recommendation scenarios within the financial business, firstly acquires the original object information of the target object and performs feature transformation on this information, converting unstructured data into quantifiable and analyzable object features. Next, based on the object features, the number of clusters is calculated to obtain the target cluster number, allowing for dynamic determination of classification dimensions and avoiding biases caused by fixed clustering parameters. Furthermore, based on the target cluster number, the object features are clustered to obtain target cluster features, which can uncover potential group characteristics of the object and improve the rationality of user segmentation. Finally, target profile data of the target object is generated based on the target cluster features and the original object information. Based on this target profile data, candidate products in the product resource library are screened to obtain target recommended products, which are then recommended to the target object. This solves the technical problem of inaccurate customer group classification affecting the accuracy of product recommendations, improves the alignment between recommended products and users' actual needs, and enhances the accuracy of product recommendations. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for a product recommendation method according to an embodiment of this application; Figure 2This is a flowchart illustrating a product recommendation method in one embodiment of this application; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 4 yes Figure 3 A flowchart illustrating a specific implementation of step S21; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 6 yes Figure 5 A flowchart illustrating a specific implementation of step S33; Figure 7 yes Figure 2 A schematic diagram of a specific implementation of step S40; Figure 8 yes Figure 2 A schematic diagram of a specific implementation method for step S50; Figure 9 This is a schematic diagram of the product recommendation device in one embodiment of this application; Figure 10 This is a schematic diagram of the structure of a computer device according to one embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0014] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). It is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, intent recognition, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0015] The K-Prototypes algorithm is a clustering method for handling mixed-type data. It can simultaneously analyze both numerical (continuous) and categorical (discrete) features in a dataset. K-Prototypes combines the core ideas of K-Means and K-Modes: for numerical features, K-Means uses Euclidean distance to calculate the distance between samples and cluster centers; for categorical features, K-Modes uses Hamming distance (based on mismatch counts) to measure differences. A weighted combined distance metric then integrates these two methods, effectively handling mixed data in a single clustering process. The goal of K-Prototypes is to iteratively optimize the distribution of samples into different clusters and update the "prototype" (the centroid containing both the numerical mean and the categorical mode) of each cluster to minimize the overall clustering loss. This makes K-Prototypes a practical tool for analyzing real-world data containing both continuous variables (such as income and age) and discrete variables (such as gender and occupation).

[0016] The Elbow Method is a method used to determine the optimal number of clusters in cluster analysis. It calculates the Within-Cluster Sum of Squares (WCSS) for different numbers of clusters (K values). WCSS is the sum of squared errors within each cluster from the sample to its cluster center. As K increases, WCSS decreases rapidly and then levels off. By plotting the WCSS as a function of different numbers of clusters, the inflection point (similar to the "elbow") is generally considered the optimal number of clusters. After this point, the rate of decrease in WCSS slows significantly, indicating diminishing returns from further increasing K.

[0017] The silhouette score is an internal evaluation metric used to assess the quality of clustering results. It comprehensively measures the compactness of a sample within its own cluster and its separation from other clusters. For each sample, the silhouette score calculates its average distance to other samples within its cluster (reflecting intra-cluster cohesion) and its average distance to all samples in its nearest neighbor cluster (reflecting inter-cluster separation). The difference between these two distances is divided by the maximum value to obtain the silhouette score, which ranges from -1 to 1. The silhouette scores of all samples are averaged to obtain the overall clustering silhouette score. A score closer to 1 indicates compactness within clusters and good inter-cluster separation; a score closer to 0 indicates cluster overlap; and a negative value suggests the sample may have been assigned to an incorrect cluster. Therefore, the silhouette score is often used to select the optimal number of clusters K, typically choosing the value of K that maximizes the silhouette score.

[0018] The Davies-Bouldin index (DB index) is an internal evaluation metric used to assess the effectiveness of clustering algorithms. The basic idea of ​​the Davies-Bouldin index is as follows: for each pair of clusters in the clustering results, calculate the ratio of the sum of the average intra-cluster distances (i.e., the average distances from samples within a cluster to the cluster center) to the distance between the two cluster centers. Then, take the ratio corresponding to the cluster most similar to the first cluster (i.e., the one with the largest ratio). Finally, average these maximum values ​​across all clusters. The Davies-Bouldin index ranges from 0 to positive infinity. A smaller value indicates denser intra-cluster samples, higher inter-cluster separation, and a more ideal clustering effect. Therefore, the Davies-Bouldin index is widely used to compare the quality of results under different numbers of clusters K, helping to determine the optimal number of clusters.

[0019] Simple Matching Distance (SMD) is a distance metric used to measure the degree of difference between two categorical samples. For two samples containing multiple categorical features, the SMD is defined as the number of features where the samples differ in corresponding values. Specifically, if the two samples have the same value on the j-th categorical feature, the contribution is 0; if they differ, the contribution is 1. The SMD is obtained by summing the comparison results of all features. The value of this distance ranges from 0 to the total number of features; a smaller value indicates that the two samples are more similar in terms of categorical features. SMD is commonly used in clustering algorithms (such as K-means clustering). Modes or K Prototypes handles categorical data, avoiding the high-dimensional sparsity and distance distortion problems caused by numericalizing categorical features (such as one-hot encoding), and can intuitively and efficiently measure the degree of mismatch between categorical features.

[0020] Product recommendation methods can recommend relevant products to users based on their personal information. For example, in a financial context, sales personnel can recommend financial products such as wealth management products, insurance products, and financial leasing products based on customer information.

[0021] Currently, product recommendations in financial scenarios primarily employ customer segmentation. This involves first classifying customers based on their information, and then recommending matching financial products to different customer groups. However, customer information in the financial industry exhibits significant variations in numerical characteristics such as financing amount, registered capital, and annual revenue. Customer segmentation is easily affected by extreme values, and customer information also includes category characteristics, which traditional clustering algorithms struggle to accurately handle. This leads to reduced accuracy in customer segmentation, which in turn affects the accuracy of product recommendations.

[0022] In addition, commonly used customer clustering schemes usually rely on a single evaluation index (such as using only the elbow rule or only the silhouette coefficient) when determining the optimal number of clusters K. A single index is easily affected by the characteristics of data distribution and will cause bias, resulting in unstable and unreliable selection of K value, which in turn leads to inaccurate customer classification and will also affect the accuracy of product recommendations.

[0023] Based on this, embodiments of this application provide a product recommendation method, apparatus, computer equipment, and storage medium to solve the technical problem that the accuracy of product recommendations is affected by the inability to accurately classify customer groups.

[0024] The product recommendation method, apparatus, computer equipment, and storage medium provided in this application are specifically described through the following embodiments. First, the product recommendation method in this application is described.

[0025] The product recommendation method provided in this application embodiment can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain the original object information of the target object from the client; perform feature transformation on the original object information to obtain object features; calculate the number of clusters based on the object features to obtain the number of target clusters; perform clustering processing on the object features based on the number of target clusters to obtain target cluster features; generate target profile data of the target object based on the target cluster features and the original object information; filter candidate products in the product resource library based on the target profile data to obtain target recommended products. The target recommended products are then fed back to the client to recommend target recommended products to the target object.

[0026] In this application, targeting financial product recommendation scenarios within financial services, the process begins by first acquiring the original object information of the target user and performing feature transformation on this information. This transforms unstructured data into quantifiable and analyzable object features. Next, the number of clusters is calculated based on the object features to obtain the target cluster number, allowing for dynamic determination of classification dimensions and avoiding biases caused by fixed clustering parameters. Furthermore, the object features are clustered based on the target cluster number to obtain target cluster features, which can uncover potential group characteristics and improve the rationality of user segmentation. Finally, target profile data for the target user is generated based on the target cluster features and the original object information. Based on this target profile data, candidate products in the product resource library are screened to obtain target recommended products, which are then recommended to the target user. This solves the technical problem of inaccurate customer group classification affecting the accuracy of product recommendations, improves the alignment between recommended products and users' actual needs, and enhances the accuracy of product recommendations.

[0027] The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description uses specific embodiments to illustrate this application.

[0028] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the product recommendation method provided in this application embodiment includes the following steps: S10. Obtain the original object information of the target object; The target audience is the client(s) to whom you wish to recommend financial products (such as individual investors or corporate clients). The original target information consists of relevant information about the target audience. If the target audience is an individual investor, the original target information may include their age, gender, income, transaction history, and risk assessment results. If the target audience is a corporate client, the original target information may include the amount of financing, registered capital, annual revenue, company type, overdue payment stage, and client rating.

[0029] Financial products include, but are not limited to, wealth management products, insurance products, financial leasing products, and loan products.

[0030] It is understandable that the original object information includes numeric data and categorical data, and different types need to be processed accordingly in subsequent processing.

[0031] S20. Perform feature transformation on the original object information to obtain object features; It's important to understand that feature transformation here refers to the process of characterizing the original object information, eliminating differences in data units, making different features comparable, and improving the accuracy of subsequent clustering algorithms. For example, Figure 3 As shown, in step S20, the object features include object category features and object numerical features. Feature transformation is performed on the original object information to obtain the object features, including the following steps: S21. Perform data cleaning on the original object information to obtain the initial object information; It's important to understand that "issue recall" here refers to operations such as checking, filtering, correcting, and deleting original object information to remove noise, errors, duplicates, and inconsistencies from the data, thereby improving data quality and reliability. For example, Figure 4 As shown, step S21, which involves data cleaning of the original object information to obtain initial object information, includes the following steps: S211. Identify missing values ​​in the original object information to obtain field missing information; S212. Filter the original object information based on the missing field information to obtain alternative secondary object information; S213. Perform anomaly identification on the candidate object information to obtain field anomaly information; S214. Based on the field exception information, perform row filtering on the candidate object information to obtain the initial object information.

[0032] For steps S211 to S213, by identifying missing values ​​in the original object information, data problems can be accurately located, and missing field information can be obtained. Based on the missing field information, column filtering of the original object information can be performed to remove columns with severe missing values, reduce data redundancy, and obtain candidate object information. Next, anomaly identification is performed on the candidate object information to find abnormal data in the candidate object information, obtain field anomaly information, and row filtering of the candidate object information is performed based on the field anomaly information to remove rows containing abnormal values, ensuring data integrity and obtaining initial object information. This effectively improves data quality and provides an accurate and reliable data foundation for subsequent applications such as feature transformation and financial product recommendation.

[0033] Specifically, by traversing every field and record of the original object information, it checks for missing values ​​such as null or undefined values. For example, checking the user's age field to see if any records have an empty age value. For any missing values ​​found, the field name, record number, and other information are recorded to form field missing information. In essence, field missing information records which fields in the original object information have missing values, along with details such as the number and location of these missing values. Identifying missing values ​​helps assess data quality and understand the completeness of the data. If an important field has too many missing values, it may be necessary to reconsider the use of that field or re-collect data.

[0034] Next, for each column in the original object information, the missing rate is calculated based on the field's missing information. Then, columns with missing rates exceeding a preset missing value threshold are pruned (column filtering) to obtain candidate object information. For example, the missing value threshold can be set to 40%, meaning that if the missing rate of elements in a column exceeds 40%, that column is filtered out. Understandably, column filtering reduces data redundancy, improves data processing efficiency, and prevents missing values ​​from having an excessive impact on subsequent analysis.

[0035] Furthermore, after obtaining the candidate object information, outlier detection is performed on it. Data in the candidate object information that does not conform to normal patterns or business logic, such as records containing negative, zero, or null values, is identified as field anomaly information. This field anomaly information records detailed information about the presence of outliers in each field of the candidate object information, such as which fields are abnormal, the type of outlier, and the number of outliers. By identifying and handling outliers, it is possible to prevent abnormal data from misleading subsequent analysis and recommendations.

[0036] Finally, rows containing outliers in the candidate object information are deleted based on field anomaly information (row filtering). For example, if a row contains an outlier in a field, that row will be deleted. This further improves data quality, ensuring that the data in the initial object information is more accurate and reliable, providing a solid data foundation for subsequent financial product recommendations.

[0037] S22. Classify the initial object information by field to obtain the initial numerical field information and the initial category field information; It's understandable that the initial object information is obtained by processing the original object information, which includes both numeric and categorical data. Similarly, the initial object information also includes both numeric and categorical data. Therefore, it's necessary to categorize the initial object information by field, examining the data characteristics of each field to determine whether it's numeric or categorical. For example, the "age" field contains specific numbers and is a numeric type; the "gender" field contains "male" or "female" and is a categorical type. Furthermore, numeric fields are grouped together to form initial numeric field information, and categorical fields are grouped together to form initial categorical field information, facilitating targeted processing for different types in subsequent steps.

[0038] S23. Perform logarithmic transformation on the initial numerical field information to obtain initial numerical features, and then standardize the initial numerical features to obtain object numerical features. It should be noted that for some values ​​in the initial numerical fields, such as financing amount, registered capital, annual revenue and other highly skewed numerical features, if used directly, the data distribution may be too large and cannot be effectively clustered. Therefore, it is necessary to perform logarithmic transformation on the initial numerical fields to compress the numerical span, transform the nonlinear relationship into an approximate linear one, improve the data distribution, and obtain the initial numerical features.

[0039] In some embodiments, the transformation can be performed using the logarithmic transformation formula log(1+x), where x is the value of the initial numerical field information.

[0040] Understandably, by performing numerical transformation and standardization on the initial numerical fields, financial numerical features spanning several orders of magnitude (such as financing amounts ranging from tens of thousands to hundreds of millions) can be compressed into a reasonable numerical range, preventing extreme values ​​from dominating the clustering results and making the contribution of each feature to customer group segmentation more balanced and reasonable.

[0041] Furthermore, after obtaining the initial numerical features, a feature standardization algorithm (such as the zero-mean unit variance standardization method) is used to standardize the initial numerical features, eliminate the scale differences between features of different dimensions, and obtain the object numerical features, which facilitates the fair measurement of the contribution of each feature when performing clustering in the future.

[0042] S24. Perform feature processing on the initial category field information to obtain object category features.

[0043] It's important to understand that the feature processing in step S24 refers to converting the initial category field information into a numerical form or other feature representation suitable for model processing. Common methods include one-hot encoding and label encoding.

[0044] For example, one-hot encoding converts each category value into a binary vector where only one position is 1 and the rest are 0. For instance, for the "Financial Product Type" field, with three categories: stocks, funds, and bonds, stocks can be represented as [1,0,0], funds as [0,1,0], and bonds as [0,0,1]. Tag encoding maps each category value to an integer. For example, for the "User Gender" field, "male" is mapped to 0, and "female" is mapped to 1.

[0045] By using feature processing, categorical data can be recognized and processed by the model, improving data usability and model performance while preserving categorical information.

[0046] It is understandable that object characteristics include object category characteristics and object numerical characteristics.

[0047] S30. Calculate the number of clusters based on object characteristics to obtain the target cluster number; It's important to understand that the cluster count calculation here refers to evaluating the clustering effect under different K values ​​using specific assessment metrics (such as the Elbow Method, Silhouette Score, Davies-Bouldin index, etc.). Selecting the optimal value (i.e., the target number of clusters) avoids over-segmentation or merging caused by a fixed K value, ensuring reasonable user segmentation and improving the targeting of subsequent recommendations. For example, Figure 5 As shown, step S30, which involves calculating the number of clusters based on object features to obtain the target cluster number, includes the following steps: S31. Obtain the preset number of multiple candidate clusters; It should be noted that the preset number of candidate clusters refers to a series of possible cluster counts set in advance. The number of candidate clusters is the K value, for example, 2 to 15, which includes 14 candidate clusters.

[0048] Understandably, setting multiple candidate cluster numbers provides a variety of options, avoiding the limitations that a fixed number of clusters might introduce. For example, in financial product recommendations within fintech scenarios, different user groups may be suited to different numbers of customer segmentation. Setting only a fixed number of clusters might not accurately reflect the diverse needs of users. Presetting multiple candidate cluster numbers, such as 5, 10, or 15, increases the likelihood of finding the most suitable number of clusters.

[0049] S32. For each candidate cluster number, perform clustering processing on the object features based on the candidate cluster number to obtain multiple candidate cluster features; Specifically, for each number of candidate clusters, a preset clustering algorithm (K-Prototypes algorithm) is applied to cluster the object features based on the selected number of candidate clusters, resulting in candidate cluster features corresponding to each number of candidate clusters. This allows for a direct visualization of the classification of object features under different numbers of candidate clusters. It is understandable that the number of candidate clusters corresponds to the number of candidate cluster features.

[0050] In the clustering process, firstly, candidate cluster centers are initialized based on the number of candidate clusters. Next, Euclidean distance is calculated between the candidate cluster centers and the object's numerical features to obtain candidate numerical distances. Then, simple matching distance is calculated between the candidate cluster centers and the object's class features to obtain candidate class distances. These candidate numerical and class distances are then weighted and fused to obtain candidate distance data. This candidate distance data represents the distance between the object's features and each candidate cluster center. Finally, the K-Prototypes algorithm is used to cluster the object features based on the candidate distance data to obtain candidate cluster features.

[0051] S33. Based on multiple candidate clustering features, perform clustering evaluation to obtain target cluster data.

[0052] It's important to understand that cluster evaluation here refers to using specific evaluation metrics (such as the Elbow Method, Silhouette Score, and Davies-Bouldin index) to assess the characteristics of candidate clusters with different numbers of candidate clusters (K-value). This allows for an objective and quantitative comparison of clustering performance with varying numbers of candidate clusters, thereby selecting the optimal number of target clusters. For example, Figure 6 As shown, step S33, which involves evaluating clustering based on multiple candidate clustering features to obtain the target cluster data, includes the following steps: S331. Perform elbow rule inflection point detection on multiple candidate clustering features to obtain the number of the first clusters; S332. Maximize the silhouette coefficient of multiple candidate clustering features to obtain the number of second clusters; S333. Minimize the inter-cluster separation of multiple candidate clustering features to obtain the number of third clusters; S334. Based on the number of clusters in the first cluster, the number of clusters in the second cluster, and the number of clusters in the third cluster, a voting process is performed to obtain the target cluster data.

[0053] For steps S331 to S334, firstly, the elbow rule is used to detect the inflection points of multiple candidate clustering features to obtain the number of the first cluster. Then, the silhouette coefficient is used to maximize the silhouette coefficient of multiple candidate clustering features to obtain the number of the second cluster, measuring the clustering effect by the degree of feature compactness within clusters and separation between clusters. Next, the Davies-Bouldin index is used to minimize the separation between clusters of multiple candidate clustering features to measure the separation between clusters, obtaining the number of the third cluster. Then, a voting process is performed based on the number of the first, second, and third clusters to obtain the target cluster data. Using three complementary clustering evaluation metrics, and integrating multiple evaluation results, the optimal number of target clusters is automatically determined through a majority voting mechanism. This overcomes the selection bias of a single metric, reduces manual intervention and subjective judgment, and improves the robustness and reliability of the K-value decision.

[0054] In step S331 of some embodiments, the elbow method is used to calculate the candidate clustering features corresponding to each number of candidate clusters, calculate the sum of squared errors within the cluster from the object feature to its candidate cluster center, form WCSS curves corresponding to different numbers of candidate clusters, and then find the inflection point (i.e., the elbow) in the WCSS curve. The number of candidate clusters corresponding to the inflection point is taken as the first number of clusters.

[0055] In step S332 of some embodiments, the candidate clustering features corresponding to each number of candidate clusters are calculated. The average distance between a certain object feature and other object features within the same cluster (reflecting intra-cluster cohesion) and the average distance between a certain object feature and all object features in its nearest neighbor clusters (reflecting inter-cluster separation) are calculated. Then, the silhouette coefficient value of the object feature is obtained by dividing the difference between the two average distances by the maximum value of the two average distances. Further, the silhouette coefficient values ​​of all object features are averaged to obtain the silhouette coefficient of the candidate clustering feature corresponding to that number of candidate clusters. Finally, the candidate clustering feature with the largest value is selected from the silhouette coefficients corresponding to all candidate clustering features, and the number of candidate clusters corresponding to the maximum value is taken as the second number of clusters.

[0056] In step S333 of some embodiments, the candidate clustering features corresponding to each candidate cluster number are calculated. The ratio of the sum of the average intra-cluster distances (i.e., the average distance from samples within a cluster to the cluster center) of two clusters to the distance between the two cluster centers is calculated. Then, the ratio corresponding to the cluster most similar to this cluster (i.e., with the largest ratio) is taken, and the average of these maximum values ​​across all clusters is calculated to obtain the Davies-Bouldin index of the candidate clustering features corresponding to the candidate cluster number. Finally, the smallest Davies-Bouldin index is selected from all the Davies-Bouldin indices of the candidate clustering features corresponding to the candidate cluster number, and the candidate cluster number corresponding to the smallest value is taken as the third cluster number.

[0057] Finally, a vote is taken on the number of clusters in the first, second, and third clusters. The number of times the same number of clusters appears in the first, second, and third clusters is counted. The cluster with the most occurrences is selected as the final result, which is the target cluster data. This target cluster data is used to indicate the final number of clusters that are considered appropriate after the voting process.

[0058] For example: Suppose the first cluster has 5 clusters, the second cluster has 6 clusters, and the third cluster has 5 clusters. After voting, the highest K value is 5, so 5 is selected as the final target cluster data.

[0059] In some embodiments, if the number of the first cluster, the second cluster, and the third cluster are all different, a rule can be pre-set to select any one of the three clusters as the target cluster data. Alternatively, the median of the three clusters can be selected as the target cluster data.

[0060] S40. Cluster the object features based on the number of target clusters to obtain the target cluster features; It's important to understand that the clustering process here refers to iteratively optimizing cluster centers using the K-Prototypes algorithm to uncover potential group characteristics among target objects. For example, Figure 7 As shown, step S40, which involves clustering the object features based on the number of target clusters to obtain the target cluster features, includes the following steps: S41. Initialize the cluster centers based on the target number of clusters to obtain the initial cluster centers; S42. Calculate the distance between the initial cluster centers and the numerical features of the objects to obtain numerical distance data. Calculate the distance between the initial cluster centers and the category features of the objects to obtain category distance data. Perform a weighted fusion of the numerical distance data and the category distance data to obtain the target distance data. The target distance data is used to characterize the distance between the object features and each initial cluster center. S43. Cluster the object features based on the target distance data to obtain the target cluster features.

[0061] For steps S41 to S43, initial cluster centers are obtained based on the target number of clusters, providing a reasonable starting point for subsequent clustering and avoiding local optima caused by random initialization, thus improving convergence efficiency. Next, distances are calculated between the initial cluster centers and the object's numerical features to obtain numerical distance data, and between the initial cluster centers and the object's categorical features to obtain categorical distance data. The numerical and categorical distance data are then weighted and fused to obtain the target distance data. This target distance data characterizes the distance between the object's features and each initial cluster center, comprehensively considering the impact of different types of features on clustering, making the distance measurement more comprehensive and accurate. Finally, the object features are clustered based on the target distance data to obtain target cluster features, which more accurately classifies objects into appropriate clusters, thereby improving the accuracy and reliability of the clustering results.

[0062] In step S41 of some embodiments, the initial cluster center positions are set randomly or according to certain rules (such as uniform distribution in the feature space) based on the number of target clusters. For example, if the number of target clusters is 8, 8 points are randomly set as initial cluster centers to provide reasonable starting points for subsequent clustering processes. Specifically, the initial cluster centers contain multiple central features. Assuming that the object features include 4 object numerical features and 3 object category features, then the initial cluster centers contain 4 numerical central features and 3 category central features.

[0063] For example: Suppose there are three initial cluster centers, cluster 0 center: {funding amount = 5 million, annual revenue = 20 million, enterprise type = state-owned enterprise, industry = manufacturing}, cluster 1 center: {funding amount = 500,000, annual revenue = 2 million, enterprise type = private enterprise, industry = trade}, cluster 2 center: {funding amount = 20 million, annual revenue = 100 million, enterprise type = foreign-invested enterprise, industry = technology}.

[0064] In step S42 of some embodiments, for the numerical features of the object, the numerical center features in the initial cluster center are used to calculate the Euclidean distance with the numerical features of the object to obtain a result, namely, numerical distance data.

[0065] For object category features, a simple matching distance calculation is performed between the category center features in the initial cluster centers and the object category features to obtain a result, namely the category distance data.

[0066] Finally, by dynamically balancing the contribution weights of numerical features and categorical features in the clustering process using Gamma weights (e.g., setting gamma=0.75), the dimensionality explosion and distance distortion problems caused by one-hot encoding are avoided, achieving native unified processing of mixed numerical and categorical features. Specifically, the numerical distance data and categorical distance data are weighted and fused based on the gamma weights to obtain the target distance data, for example, calculated using the following formula: Target distance data = Numerical distance data + Gamma weight * Categorical distance data.

[0067] After identifying the target distance data, the K-Prototypes algorithm is used to perform an iterative clustering process based on the target distance data, ultimately obtaining the target clustering features.

[0068] Understandably, clustering can group objects with similar characteristics into one category, allowing for personalized financial product recommendations based on the characteristics of each cluster. For example, for users in the aggressive cluster, high-risk, high-return equity funds and other financial products can be recommended, improving the accuracy of recommendations and user satisfaction.

[0069] S50. Generate target profile data of the target object based on target clustering features and original object information; It's important to understand that profile generation here refers to fusing individual user characteristics with the central characteristics of their cluster to generate a profile that includes both "group tags" and "individual tags." For example... Figure 8 As shown, step S50, which involves generating target profile data for the target object based on the target clustering features and the original object information, includes the following steps: S51. Determine cluster labels based on target clustering features, and group the original object information based on cluster labels to obtain grouped object information; S52. Classify the grouped object information by field to obtain the grouped numerical field information and the grouped category field information; S53. Perform numerical statistics on the grouped numerical fields to obtain the grouped numerical statistical characteristics; S54. Perform category statistics on the grouping category field information to obtain the grouping category statistical characteristics; S55. Based on the statistical characteristics of group numerical data and the statistical characteristics of group category, classify the grouped object information to obtain customer group classification information; S56. Based on customer group classification information, construct a profile to obtain customer group profile data, and confirm the customer group profile data as the target profile data of the target object.

[0070] For steps S51 to S56, firstly, cluster labels are determined based on the target clustering features. Then, the original object information is grouped based on these labels to obtain grouped object information, thus orderly dividing all target objects. Next, the grouped object information is categorized by field to obtain grouped numerical field information and grouped category field information. Numerical statistics are performed on the grouped numerical fields to obtain grouped numerical statistical features, which can uncover key numerical characteristics. Category statistics are performed on the grouped category fields to obtain grouped category statistical features, which can obtain features such as category distribution. Finally, based on the grouped numerical and category statistical features, customer group classification is further performed on the grouped object information to obtain customer group classification information. Based on this customer group classification information, customer profiles are constructed to obtain customer profile data, which is then confirmed as the target profile data for the corresponding target objects in the grouped object information.

[0071] In step S51 of some embodiments, a cluster label corresponding to each cluster is determined based on the target clustering features. For example, the cluster labels include cluster 0, cluster 1, ..., cluster 8. Then, based on the cluster labels, objects belonging to the same cluster in the original object information are grouped together to obtain grouped object information. For example, all original object information labeled "cluster 0" is grouped together to form a single group.

[0072] It's understandable that the data in the grouped object information is raw data; the numeric data it contains is raw, not logarithmically transformed. Similarly, the categorical data it contains is also raw, not encoded. Therefore, it's necessary to categorize the fields in the grouped object information, examining the data characteristics of each field to determine whether it's numeric or categorical. For example, the "age" field contains specific numbers, belonging to the numeric type; the "gender" field contains "male" or "female," belonging to the categorical type. Furthermore, numeric fields are grouped together to form grouped numeric field information, and categorical fields are grouped together to form grouped categorical field information, facilitating targeted processing for different types in subsequent steps.

[0073] In step S53 of some embodiments, numerical statistics are performed on the grouping numerical fields in each grouping object information, such as calculating the mean, standard deviation, median, and quartiles. The resulting statistical results are the grouping numerical statistical characteristics. These characteristics can intuitively reflect the numerical features of the group of customers. In financial product recommendations, for example, by calculating the average and standard deviation of customer income, one can understand the overall income level and income disparities of the group of customers. If the average income is high and the standard deviation is small, it indicates that the group of customers has relatively stable and high income, and high-end financial products such as private banking services and high-end wealth management products can be recommended.

[0074] In step S54 of some embodiments, category statistics are performed on the grouping category field information in each grouping object information. For example, the frequency of occurrence of each category in the grouping category field information is counted, and its proportion to the total is calculated to obtain the grouping category statistical characteristics. The grouping category statistical characteristics can reveal the distribution of the group of customers in terms of categories. In financial product recommendation, for example, by statistically analyzing the category statistical characteristics of customer investment preferences, if it is found that most customers in the group prefer stock investment, then stock-related financial products, such as equity funds and individual stock recommendations, can be recommended to this group of customers to improve the accuracy of the recommendations.

[0075] In step S55 of some embodiments, for each group of object information, the target objects in the group of object information are further classified into customer groups based on the group numerical statistical characteristics and group category statistical characteristics. For example, the target objects in the group of object information are classified into customer groups based on quartiles and categories, and finally customer group classification information is obtained. By comprehensively considering numerical and category statistical characteristics for customer group classification, customer characteristics can be more comprehensively and accurately portrayed, and customers can be divided into groups with different needs and characteristics.

[0076] Finally, after clarifying the customer group classification information, we can use the trained large language model to generate business-interpretable customer group labels (such as high-value customer groups, micro and small enterprise customer groups, and standard customer groups) based on the customer group classification information, and output multi-dimensional feature profiles (mean, standard deviation, median, and category distribution) for each customer group. Then, the customer group profile data of each customer group is confirmed as the target profile data of the corresponding target object.

[0077] Understandably, customer profile data can intuitively and comprehensively showcase the characteristics and needs of each customer group, providing a clear basis for financial product recommendations. By identifying customer profile data as target profile data, precise marketing can be achieved, recommending suitable financial products based on the characteristics of different customer groups, thereby improving customer conversion rates and business efficiency.

[0078] For example, the generated customer group tags can be high-value core customer groups, high-financing-standard profit customer groups, high-financing-inefficient customer groups, medium-sized high-profit customer groups, standard customer groups, customer groups that need optimization, high-quality customer groups of micro and small enterprises, standard customer groups of micro and small enterprises, and low-value long-tail customer groups.

[0079] S60. Based on the target profile data, filter the candidate products in the product resource library to obtain the target recommended products, and recommend the target recommended products to the target audience.

[0080] Specifically, the process begins by acquiring a product resource library provided by a bank or financial institution. This library includes multiple candidate products that can be recommended, including but not limited to wealth management products, insurance products, financial leasing products, and loan products. Then, after clarifying the target profile data for each target individual, a pre-defined recommendation algorithm is used to filter the candidate products from the resource library to obtain the target recommended product, which is then recommended to the target individual.

[0081] In step S60 of some embodiments, that is, filtering candidate products based on target profile data to obtain target recommended products, the steps include the following: Obtain product information for candidate products; Product features are obtained by extracting features from product information; By matching product features with target profile data, user-product matching data is obtained. The target recommended product is obtained by filtering from candidate products based on user product matching data.

[0082] Specifically, by acquiring product information from candidate products and extracting features (such as risk level, return type, and maturity), unstructured product data can be transformed into quantifiable indicators. Furthermore, each product feature is matched with each target profile to obtain user-product matching data. Based on this matching data, target recommended products are selected from the candidate products. This ensures that the recommendation results meet users' personalized needs while improving recommendation efficiency, achieving accurate financial product recommendations.

[0083] It should be noted that product information is structured data used to describe the product attributes of candidate products, including basic attributes (such as product name, type, and term) and risk-return characteristics (such as expected return, volatility, and maximum drawdown).

[0084] It should be understood that feature extraction of product information here refers to transforming product information into quantifiable and comparable numerical feature vectors, unifying the dimensions and units of product features, and making different types of products (such as funds and insurance) comparable, thus providing a basis for subsequent product matching. The feature extraction process may include, but is not limited to, using methods such as data cleaning (removing invalid values), classification coding (e.g., converting "mixed" funds to 0 / 1 labels), and normalization (scaling the rate of return to the 0-1 range).

[0085] After obtaining the product features of each candidate product, a similarity calculation method can be used to calculate the correlation between the target profile data of each target object and the product features of each candidate product, thereby accurately quantifying the degree of association between the candidate product and the target object, and ensuring that the recommendation results are accurately aligned with user needs.

[0086] Finally, the top few candidate products with the largest matching data for the user product, or those with relatively large values, can be selected from the candidate products as the target recommended products.

[0087] Understandably, the number of products recommended can be one or more.

[0088] In some embodiments, after step S60, after obtaining the target recommended product, the target recommended product is pushed to the target object's client so that the target object can receive the recommendation and increase the conversion rate of the product recommendation.

[0089] As can be seen, in the above solution, for the financial product recommendation scenario under financial business, firstly, the original object information of the target object is obtained, and feature transformation is performed on the original object information to transform unstructured data into quantifiable and analyzable object features. Next, the number of clusters is calculated based on the object features to obtain the target cluster number, which can dynamically determine the classification dimension and avoid bias caused by fixed clustering parameters. Furthermore, based on the target cluster number, the object features are clustered to obtain target cluster features, which can uncover potential group characteristics of the object and improve the rationality of user segmentation. Finally, target profile data of the target object is generated based on the target cluster features and the original object information, and candidate products in the product resource library are screened based on the target profile data to obtain target recommended products. These target recommended products are then recommended to the target object. This solves the technical problem of the inability to accurately classify customer groups, which affects the accuracy of product recommendations, improves the fit between recommended products and users' actual needs, and enhances the accuracy of product recommendations.

[0090] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0091] In one embodiment, a product recommendation device is provided, which corresponds one-to-one with the product recommendation method in the above embodiments. For example... Figure 9 As shown, the product recommendation device includes an object information acquisition module 101, an object feature conversion module 102, a cluster quantity calculation module 103, a clustering processing module 104, a profile generation module 105, and a recommended product filtering module 106. Detailed descriptions of each functional module are as follows: Object information acquisition module 101 is used to acquire the original object information of the target object; The object feature conversion module 102 is used to perform feature conversion on the original object information to obtain object features; Cluster count calculation module 103 is used to calculate the number of clusters based on object characteristics to obtain the target cluster count; Clustering processing module 104 is used to perform clustering processing on object features based on the number of target clusters to obtain target cluster features; The portrait generation module 105 is used to generate target portrait data of the target object based on the target clustering features and the original object information. The recommended product filtering module 106 is used to filter candidate products in the product resource library based on target profile data, obtain target recommended products, and recommend target recommended products to target objects.

[0092] In one embodiment, the object features include object category features and object numerical features. The object feature conversion module 102 is specifically used for: Data cleaning is performed on the original object information to obtain the initial object information; The initial object information is categorized by field to obtain initial numeric field information and initial category field information; The initial numerical field information is logarithmically transformed to obtain initial numerical features, and the initial numerical features are then standardized to obtain object numerical features. The initial category field information is processed to obtain object category features.

[0093] In one embodiment, the object feature conversion module 102 is specifically used for: The missing value is identified in the original object information to obtain the field missing information; Based on the missing field information, the original object information is filtered by column to obtain the alternative secondary object information; Anomaly detection is performed on the candidate object information to obtain field anomaly information; Based on the field anomaly information, the candidate object information is filtered by row to obtain the initial object information.

[0094] In one embodiment, the cluster count calculation module 103 is further configured to: Obtain the number of multiple candidate clusters as preset; For each candidate cluster number, the object features are clustered based on the number of candidate clusters to obtain multiple candidate cluster features; Clustering evaluation is performed based on multiple candidate clustering features to obtain target cluster data.

[0095] In one embodiment, the cluster count calculation module 103 is specifically used for: Elbow rule inflection point detection is performed on multiple candidate clustering features to obtain the number of the first clusters; The number of second clusters is obtained by maximizing the silhouette coefficient of multiple candidate cluster features. The number of third clusters is obtained by minimizing the inter-cluster separation of multiple candidate clustering features. The target cluster data is obtained by voting based on the number of clusters in the first, second, and third clusters.

[0096] In one embodiment, the object features include object category features and object numerical features. The clustering processing module 104 is specifically used for: Initial cluster centers are obtained by initializing the number of target clusters; The numerical distance data is obtained by calculating the distance between the initial cluster centers and the numerical features of the objects. The categorical distance data is obtained by calculating the distance between the initial cluster centers and the categorical features of the objects. The numerical distance data and the categorical distance data are then weighted and fused to obtain the target distance data. The target distance data is used to characterize the distance between the object features and each initial cluster center. The object features are clustered based on the target distance data to obtain the target cluster features.

[0097] In one embodiment, the image generation module 105 is specifically used for: Cluster labels are determined based on the target clustering features, and the original object information is grouped based on the cluster labels to obtain grouped object information; The grouped object information is categorized by field to obtain grouped numerical field information and grouped category field information; Perform numerical statistics on the grouped numerical fields to obtain the statistical characteristics of the grouped numerical values; Perform category statistics on the grouping category field information to obtain the grouping category statistical characteristics; Based on the statistical characteristics of group numerical data and the statistical characteristics of group category, customer group classification is performed on the group object information to obtain customer group classification information; Based on customer group classification information, profiles are constructed to obtain customer group profile data, and this customer group profile data is then identified as the target profile data for the target object.

[0098] This application provides a product recommendation device for financial product recommendation scenarios in financial business. First, it acquires the original object information of the target object and performs feature transformation on this information, converting unstructured data into quantifiable and analyzable object features. Next, it calculates the number of clusters based on the object features to obtain the target cluster number, dynamically determining the classification dimension and avoiding bias caused by fixed clustering parameters. Further, it performs clustering processing on the object features based on the target cluster number to obtain target cluster features, which can uncover potential group characteristics of the object and improve the rationality of user segmentation. Finally, it generates target profile data of the target object based on the target cluster features and the original object information, and uses this target profile data to filter candidate products in the product resource library to obtain target recommended products. These target recommended products are then recommended to the target object. This solves the technical problem of inaccurate customer group classification affecting the accuracy of product recommendations, improves the alignment between recommended products and users' actual needs, and enhances the accuracy of product recommendations.

[0099] Specific limitations regarding the product recommendation device can be found in the limitations of the product recommendation method described above, and will not be repeated here. Each module in the aforementioned product recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0100] Please see Figure 10 , Figure 10 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 1002 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1002 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001 to perform the product recommendation method of the embodiments of this application, including: Obtain the original object information of the target object; The original object information is transformed to obtain object features; The number of clusters in the target cluster is calculated based on the characteristics of the objects. The object features are clustered based on the number of target clusters to obtain the target cluster features; Target profile data of the target object is generated based on target clustering features and original object information; Based on the target profile data, candidate products in the product resource library are screened to obtain target recommended products, which are then recommended to the target audience.

[0101] Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0102] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain the original object information of the target object; The original object information is transformed to obtain object features; The number of clusters in the target cluster is calculated based on the characteristics of the objects. The object features are clustered based on the number of target clusters to obtain the target cluster features; Target profile data of the target object is generated based on target clustering features and original object information; Based on the target profile data, candidate products in the product resource library are screened to obtain target recommended products, which are then recommended to the target audience.

[0103] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0104] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0105] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0106] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0107] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A product recommendation method, characterized in that, The method includes: Obtain the original object information of the target object; The original object information is subjected to feature transformation to obtain object features; The number of clusters is calculated based on the object characteristics to obtain the target cluster number. The object features are clustered based on the number of target clusters to obtain target cluster features; Based on the target clustering features and the original object information, target profile data of the target object is generated; Based on the target profile data, candidate products in the product resource library are screened to obtain target recommended products, and the target recommended products are recommended to the target object.

2. The product recommendation method according to claim 1, characterized in that, The calculation of the number of clusters based on the object features to obtain the number of target clusters includes: Obtain the number of multiple candidate clusters as preset; For each number of candidate clusters, the object features are clustered based on the number of candidate clusters to obtain multiple candidate cluster features; Clustering evaluation is performed based on multiple candidate clustering features to obtain the target cluster data.

3. The product recommendation method according to claim 2, characterized in that, The step of evaluating clustering based on multiple candidate clustering features to obtain the target cluster data includes: Elbow rule inflection point detection is performed on multiple candidate clustering features to obtain the number of the first clusters; The number of second clusters is obtained by maximizing the silhouette coefficient of multiple candidate clustering features. The number of third clusters is obtained by minimizing the inter-cluster separation of multiple candidate cluster features. The target cluster data is obtained by voting based on the number of the first cluster, the number of the second cluster, and the number of the third cluster.

4. The product recommendation method according to claim 1, characterized in that, The object features include object category features and object numerical features. The clustering process based on the number of target clusters to obtain target cluster features includes: Initial cluster centers are obtained by initializing the number of target clusters. Distance is calculated between the initial cluster centers and the numerical features of the objects to obtain numerical distance data. Distance is also calculated between the initial cluster centers and the category features of the objects to obtain category distance data. The numerical distance data and the category distance data are then weighted and fused to obtain target distance data. The target distance data is used to characterize the distance between the object features and each of the initial cluster centers. The object features are clustered based on the target distance data to obtain the target cluster features.

5. The product recommendation method according to any one of claims 1 to 4, characterized in that, The object features include object category features and object numerical features. The feature transformation of the original object information to obtain the object features includes: The original object information is cleaned to obtain initial object information; The initial object information is categorized by field to obtain initial numerical field information and initial category field information; The initial numerical field information is subjected to logarithmic transformation to obtain initial numerical features, and the initial numerical features are then standardized to obtain the object numerical features. The initial category field information is subjected to feature processing to obtain the object category features.

6. The product recommendation method according to claim 5, characterized in that, The step of cleaning the original object information to obtain initial object information includes: The original object information is subjected to missing value identification to obtain field missing information; Based on the missing information in the field, the original object information is filtered by column to obtain candidate secondary object information; Anomaly identification is performed on the candidate object information to obtain field anomaly information; Based on the abnormal information in the field, the candidate object information is filtered by row to obtain the initial object information.

7. The product recommendation method according to any one of claims 1 to 4, characterized in that, The process of generating target profile data for the target object based on the target clustering features and original object information includes: Cluster labels are determined based on the target clustering features, and the original object information is grouped based on the cluster labels to obtain grouped object information; The grouped object information is classified by field to obtain grouped numerical field information and grouped category field information; Perform numerical statistics on the grouped numerical fields to obtain grouped numerical statistical characteristics; Perform category statistics on the grouping category field information to obtain grouping category statistical features; Based on the group numerical statistical features and group category statistical features, the group object information is classified into customer groups to obtain customer group classification information. Based on the customer group classification information, a profile is constructed to obtain customer group profile data, and the customer group profile data is confirmed as the target profile data of the target object.

8. A product recommendation device, characterized in that, The device includes: The object information acquisition module is used to obtain the original object information of the target object; The object feature conversion module is used to perform feature conversion on the original object information to obtain object features; The cluster count calculation module is used to calculate the number of clusters based on the object characteristics to obtain the target cluster count. A clustering processing module is used to perform clustering processing on the object features based on the number of target clusters to obtain target cluster features; The portrait generation module is used to generate target portrait data of the target object based on the target clustering features and the original object information; The recommended product filtering module is used to filter candidate products in the product resource library based on the target profile data, obtain target recommended products, and recommend the target recommended products to the target object.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the product recommendation method as described in any one of claims 1 to 7.

Images