A product recommendation method, apparatus, medium, and device

Abstract

This application relates to the financial and computer fields, and discloses a product recommendation method, apparatus, medium, and device. The method includes: constructing a target structure graph based on information about each car dealer and each product; using a multi-layer feedforward neural network layer in the target recommendation model to extract features from the attribute information corresponding to each node in the target structure graph, generating an initial feature vector for each node; using a feature update layer in the target recommendation model to update the initial feature vector of each node based on the initial feature vector of each node and the initial feature vectors of each node's neighboring nodes, obtaining the target feature vector of each node; and using a recommendation layer in the target recommendation model to determine the recommendation order of each candidate product and perform product recommendation based on the target feature vector of the target car dealer and the target feature vectors of each candidate product. This method can be applied to product recommendation scenarios in the fintech field, improving the rationality and efficiency of product recommendations.

Description

Technical Field

[0001] This application relates to the fields of finance and computer technology, and in particular to a product recommendation method, apparatus, medium and device. Background Technology

[0002] Cars are becoming increasingly popular as an important means of transportation. Due to their high price, car buyers often opt for installment payments. In other words, in the auto finance business, finance companies typically establish partnerships with local car dealers, providing them with financial services such as installment payments and car loans to customers. Consequently, finance companies usually recommend corresponding financial products to different car dealers.

[0003] In the current process of recommending financial products, the main reliance is on human experience or simple rule engines for judgment. However, this human-based product recommendation method suffers from problems such as unreasonable recommendation results and low recommendation efficiency. Summary of the Invention

[0004] In view of this, this application provides a product recommendation method, apparatus, medium and equipment, the main purpose of which is to solve the problems of unreasonable product recommendation and low recommendation efficiency in the current situation.

[0005] To address the above problems, this application provides a product recommendation method, comprising: Based on information about each car dealer and each product, a target structure graph containing several nodes and several edges is constructed. Each node is configured with the attribute information of the car dealer or the attribute information of the product. Using the multi-layer feedforward neural network in the pre-trained target recommendation model, feature extraction is performed on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. Using the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, the initial feature vector of each node is updated to obtain the target feature vector of each node. Using the recommendation layer in the target recommendation model, the recommendation order of each candidate product is determined and product recommendations are made based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product.

[0006] Optionally, the dealer information includes any one or more of the following: dealer attribute information, historical cooperation information with the product, and association with the dealer; The product information includes any one or more of the following: product attribute information and its association with similar products; The process of constructing a target structure graph containing several nodes and edges based on information from various car dealers and products specifically includes: Create dealer nodes for each car dealer and product nodes for each product; Configure attribute information for dealer nodes based on the attribute information of each dealer, and configure attribute information for product nodes based on the attribute information of each product. Based on the historical cooperation information of products and the relationship with car dealers in the car dealer information, configure edges for the corresponding car dealer nodes and product nodes, configure edges for the corresponding car dealer nodes and car dealer nodes, and configure association information for each edge; Based on the association information of similar products in the product information, edges are configured for the corresponding product nodes and product nodes, and association information is configured for each edge to construct the target structure graph.

[0007] Optionally, the attribute information of the car dealer includes any one or more of the following: capital size, city, showroom area, main brands, years of operation, and mainstream car price range; The product's attribute information includes any one or more of the following: yield, term, down payment ratio, service fee, and risk level.

[0008] Optionally, before generating the initial feature vectors for each node, the method further includes: Normalization or one-hot encoding is performed on each attribute information corresponding to the same node in sequence to obtain each original attribute vector; The concatenated vector is obtained by concatenating the original attribute vectors corresponding to the same node. The method of utilizing the multi-layer feedforward neural network in the pre-trained target recommendation model to extract features from the attribute information corresponding to each node in the target structure graph and generate the initial feature vector for each node specifically includes: By utilizing the multi-layer feedforward neural network in the pre-trained target recommendation model, the concatenated vectors of each node are subjected to non-linear transformation to obtain the initial feature vector.

[0009] Optionally, the target recommendation model includes at least two feature update layers; When performing vector updates, each feature update layer specifically includes: updating the target node's vector based on the target node's first initial feature vector or the target node's first historical feature vector output by the previous feature update layer, the second initial feature vectors of each neighboring node adjacent to the target node or the second historical feature vectors of each neighboring node output by the previous feature update layer, and the association information between the target node and each of the neighboring nodes, to obtain the target node's first current feature vector, thereby obtaining the current feature vectors of each node.

[0010] Optionally, the vector update of the target node is performed based on the first initial feature vector of the target node or the first historical feature vector of the target node output by the previous feature update layer, the second initial feature vectors of each neighboring node adjacent to the target node or the second historical feature vectors of each neighboring node output by the previous feature update layer, and the association information between the target node and each neighboring node, to obtain the first current feature vector of the target node, thereby obtaining the current feature vector of each node, specifically including: Based on the association information between the target node and each neighbor node in the target structure graph, the weight of each neighbor node is determined; Based on the first initial feature vector of the target node or the first historical feature vector of the target node output by the previous feature update layer, the second initial feature vector of each neighboring node adjacent to the target node or the second historical feature vector of each neighboring node output by the previous feature update layer, and the weight of each neighboring node, vector aggregation is performed to obtain the aggregated feature vector corresponding to the target node. The aggregated feature vector is subjected to a nonlinear transformation to obtain the first current feature vector of the target node, thereby obtaining the current feature vector of each node.

[0011] Optionally, the step of using the recommendation layer in the target recommendation model to determine the recommendation order of each candidate product based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product specifically includes: Using the recommendation layer in the target recommendation model, the matching degree is calculated based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product, so as to obtain the matching degree of each candidate product and determine the recommendation order of each candidate product based on the matching degree.

[0012] To address the above problems, this application provides a product recommendation device, comprising: The construction module is used to build a target structure graph containing several nodes and several edges based on information about each car dealer and each product. Each node is configured with the attribute information of the car dealer or the attribute information of the product. The generation module is used to extract features from the attribute information corresponding to each node in the target structure graph by using the multi-layer feedforward neural network layer in the pre-trained target recommendation model, and generate the initial feature vector of each node. The update module is used to update the initial feature vector of each node by utilizing the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, so as to obtain the target feature vector of each node. The recommendation module is used to determine the recommendation order of each candidate product and make product recommendations based on the recommendation layer in the target recommendation model, the first target feature vector corresponding to the target car dealer, and the second target feature vector corresponding to each candidate product.

[0013] To address the aforementioned problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of any of the product recommendation methods described above.

[0014] To address the aforementioned problems, this application provides a computer device, comprising at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of any of the aforementioned product recommendation methods.

[0015] This application discloses a product recommendation method, apparatus, medium, and device. By constructing a target structure graph containing several nodes and configuring corresponding attribute information for each node in the graph, it can associate each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating each initial feature vector by combining the features of neighboring nodes, it can capture the potential associations between car dealers and products, between products, and between car dealers, making the updated target feature vectors more consistent with the actual association logic. By determining the recommendation order based on the target feature vectors of target car dealers and candidate products, it can achieve accurate matching between car dealers and products, effectively avoid blind recommendations, ensure that the recommended products are highly consistent with the attributes and needs of the target car dealers, improve the rationality of recommendations, and increase product recommendation efficiency.

[0016] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an application environment for a product recommendation method in one embodiment of this application; Figure 2 This is a flowchart illustrating a product recommendation method according to an embodiment of this application; Figure 3 This is a structural block diagram of a product recommendation device according to another embodiment of this application; Figure 4 This is a structural block diagram of a computer device according to an embodiment of this application; Figure 5 This is a structural block diagram of another computer device according to an embodiment of this application. Detailed Implementation

[0018] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0019] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0020] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0021] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0022] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0023] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0024] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0025] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0026] The product recommendation method provided in this application embodiment can be applied to, for example, Figure 1In this application environment, the client communicates with the server via a network. Users can input information about car dealers and products on the client, which then submits this information to the server. The server constructs a target structure graph based on the dealer and product information. Then, using a pre-trained multi-layer feedforward neural network in the target recommendation model, it encodes the attribute information corresponding to each node in the target structure graph, generating initial feature vectors for each node. Using the feature update layer in the target recommendation model, based on the initial feature vectors of each node and the initial feature vectors of its neighboring nodes in the target structure graph, it updates the initial feature vectors of each node to obtain the target feature vectors for each node. Finally, using the recommendation layer in the target recommendation model, based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product, it determines the recommendation order of the candidate products and makes product recommendations. In this application, for complex product recommendation scenarios in financial product recommendation businesses, the product recommendation method of this application can be adopted. This involves constructing a target structure graph containing nodes for each car dealer and each product. Then, a model is used to update the initial feature vectors of each node in this structure graph. During the feature vector update, the feature vectors of each neighboring node are incorporated. This not only preserves the original attribute features of the nodes themselves but also deeply obtains the features of their multi-hop neighbors through the network topology, thus generating a high-order semantic vector of the complex business relationship network. This accurately captures the implicit associations between each car dealer and financial products, thereby obtaining the target feature vector. This lays the foundation for subsequent accurate product recommendations based on the first target feature vector of the target car dealer and the second target feature vectors of each product. In this application, 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 of specific embodiments further illustrates this application.

[0027] This application provides a product recommendation method that can be applied to computer devices such as terminals and servers. Taking a server used in a lending company in the financial sector as an example, the lending company's server can be a standalone server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Figure 2 As shown, the method in this embodiment includes the following steps: Step S101: Based on the information of each car dealer and each product, construct a target structure graph containing several nodes and several edges. Each node is configured with the attribute information of the car dealer or the attribute information of the product. In this step, each node is either a dealer node or a product node, and therefore each node is configured with dealer attribute information or product attribute information. In this target structure graph, two nodes are connected by edges. Specifically, corresponding two nodes can be connected with edges based on the relationships between dealers, between dealers and products, and between products.

[0028] Step S102: Using the multi-layer feedforward neural network layer in the pre-trained target recommendation model, feature extraction processing is performed on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. In the specific implementation process of this step, since each node may correspond to multiple attribute information, that is, each car dealer has multiple attribute information and each product has multiple attribute information, when extracting features from the attribute information of the same node, we can first normalize or perform one-hot encoding on each attribute information to obtain each original attribute vector. Then, we can concatenate the original attribute vectors corresponding to the same node to obtain the concatenated vector. Finally, we can use a multi-layer feedforward neural network to perform feature extraction processing on the concatenated vectors corresponding to each node to obtain the initial feature vector.

[0029] Step S103: Using the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, the initial feature vector of each node is updated to obtain the target feature vector of each node. In the specific implementation process, this step updates the initial feature vector of a node by combining the feature vectors of its neighboring nodes. This not only preserves the original attribute features of the node itself, but also deeply obtains the features of its multi-hop neighbors through the network topology, realizing in-depth mining of the implicit associations between various car dealers and financial products. This leads to the updated target feature vector with more comprehensive information, laying the foundation for subsequent accurate product recommendations based on the first target feature vector of the target car dealer and the second target feature vectors of each product.

[0030] Step S104: Using the recommendation layer in the target recommendation model, based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product, determine the recommendation order of each candidate product and make product recommendations.

[0031] In the specific implementation process, this step can calculate the matching degree / matching score based on the first target feature vector and the second target feature vector to obtain the matching score / matching degree of each candidate product, and finally make product recommendations based on the matching score / matching degree of each candidate product.

[0032] The product recommendation method in this embodiment constructs a target structure graph containing several nodes and configures corresponding attribute information for each node in the graph. This associates each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating these initial feature vectors with the features of neighboring nodes, the method captures potential associations between car dealers and products, between products, and between car dealers. This makes the updated target feature vectors more closely aligned with actual association logic. Determining the recommendation order based on the target feature vectors of the target car dealer and candidate products achieves precise matching between car dealers and products, effectively avoiding blind recommendations and ensuring that the recommended products highly match the attributes and needs of the target car dealers. This improves the rationality and efficiency of product recommendations.

[0033] Based on the above embodiments, another embodiment of this application provides a product recommendation method: In this embodiment, when constructing the target structure diagram, that is, when performing step S101, the following steps are specifically included: Step S101-1: Create dealer nodes for each dealer and product nodes for each product; In this step, specific nodes representing car dealers / products can be created based on the dealer's logo, dealer's name, product logo, and product name.

[0034] The products in this step are financial products, such as loan financial products. That is, different loan financial products have different down payment ratios, service fees, and terms.

[0035] Step S101-2: Configure attribute information for car dealer nodes based on the attribute information of car dealers in each car dealer information, and configure attribute information for product nodes based on the attribute information of products in each product information; In this step, the dealer information includes one or more of the following: dealer attribute information, historical cooperation information with the product, and association with the dealer. The dealer attribute information includes one or more of the following: capital size, city, showroom area, main brands, years of operation, and mainstream car price range.

[0036] Historical cooperation information with the product includes: number of collaborations and success rate. Relationships with car dealers can include any one or more of the following: geographical distance, business cooperation, equity affiliation, etc.

[0037] Product information includes one or more of the following: product attribute information and its association with similar products. The product attribute information includes one or more of the following: rate of return, term, down payment ratio, service fee, and risk level.

[0038] The association with similar products can include the cosine similarity with similar products. Specifically, similarity can be calculated based on the characteristics of each product to determine the association between the two products.

[0039] In this step, by obtaining the information of each car dealer and each product, we lay the foundation for quickly and accurately constructing the target structure diagram, which includes each car dealer and each product as nodes, car dealers and car dealers as edges, car dealers and products as edges, and products and products as edges.

[0040] Step S101-3: Based on the historical cooperation information of products and the association relationship with car dealers in the car dealer information, configure edges for the corresponding car dealer nodes and product nodes, configure edges for the corresponding car dealer nodes and car dealer nodes, and configure association information for each edge. In other words, the edges in this embodiment include car dealer-car dealer edges and car dealer-product edges. Among them, the car dealer-product edge represents historical cooperation records (number of cooperations, success rate).

[0041] Car dealer-car dealer side: indicates geographical distance, business cooperation, equity relationship, etc.

[0042] Step S101-4: Based on the association information of similar products in the product information, configure edges for the corresponding product nodes and product nodes, and configure association information for each edge to construct and obtain the target structure graph.

[0043] In this embodiment, the edges in the target structure graph also include product-product edges, which represent the similarity between products and can be based on the cosine similarity of feature vectors. As a similarity, where p i p represents the feature vector of product i. j This represents the feature vector of product j. Represents the eigenvector p i With eigenvector p j Cosine similarity between them.

[0044] In this embodiment, by constructing the target structure graph using the aforementioned dealer and product information, the construction of the target structure graph becomes more reasonable and accurate, enabling multi-dimensional data coverage. This lays the foundation for quickly and accurately recommending corresponding financial products to the target dealer based on the target structure graph. Furthermore, by configuring association information for each node in the target structure graph, the weights of feature aggregation for each neighboring node can be reasonably and accurately determined using the association information when updating the feature vectors of each node, ensuring reasonable and accurate updates of the feature vectors.

[0045] Based on the above embodiments, another embodiment of this application provides a product recommendation method. In this embodiment, by utilizing the multi-layer feedforward neural network layer in the pre-trained target recommendation model, feature extraction processing is performed on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. That is, before executing step S102, normalization processing or one-hot encoding processing can be performed on each attribute information corresponding to the same node in sequence to obtain each original attribute vector; the original attribute vectors corresponding to the same node are concatenated to obtain the concatenated vector; finally, the multi-layer feedforward neural network layer in the pre-trained target recommendation model is used to perform nonlinear transformation processing on the concatenated vector of each node to obtain the initial feature vector.

[0046] For example, when the attribute information of a node includes "fund size" and "city," the "fund size" attribute, being a continuous value (e.g., 50 million), can be standardized / normalized (scaled to a floating-point number between 0 and 1). Conversely, for discrete features like "city," one-hot encoding can be used. Therefore, after performing normalization or one-hot encoding on each attribute of the same node to obtain the original attribute vectors, these vectors can be concatenated to obtain a concatenated vector. Finally, the multi-layer feedforward neural network / multi-layer perceptron (MLP) in the pre-trained target recommendation model is used to concatenate the vectors of each node. Feature extraction is performed to obtain an initial feature vector, which is achieved by concatenating the vectors corresponding to the original attributes using a multilayer perceptron (MLP). Encode as a low-dimensional vector The specific formula is as follows:

[0047] in, It represents a specific node in the diagram (it could be "a specific car dealer" or "a specific financial product"); The concatenated vector for node v is... The attribute comprehensive feature set can be obtained by preprocessing all the dimensional attributes (numerical, categorical, etc.) of a node (such as normalization, one-hot encoding), and then concatenating them end to end to form a long one-dimensional array. This array is the concatenated vector of the original attribute vectors. ; Let v be the initial feature vector of node v.

[0048] In this embodiment, by first normalizing and performing one-hot encoding on the attribute information of the same node using the above-mentioned method, and then concatenating the information, it can be ensured that the concatenated vector obtained by concatenation is in the same feature dimension, thereby enabling subsequent processing of the concatenated vector. The process involves "purification" and "dimensionality reduction" to transform it into a dense initial embedding vector containing high-order nonlinear features. This serves as the initial feature vector for each node in subsequent iterations. This lays the foundation for accurate and reasonable vector updates.

[0049] Based on the above embodiments, another embodiment of this application provides a product recommendation method. In this embodiment, the target recommendation model includes at least two feature update layers. When each feature update layer performs vector update, it specifically includes: updating the target node's vector based on the target node's first initial feature vector or the target node's first historical feature vector output by the previous feature update layer, the target node's second initial feature vector or the target node's second historical feature vector output by the previous feature update layer, and the association information between the target node and each of the neighboring nodes, to obtain the target node's first current feature vector, thereby obtaining the current feature vector of each node. Specifically, each feature update layer in the target recommendation model can determine the weight of each neighbor node based on the association information between the target node and each neighbor node in the target structure graph; based on the first initial feature vector of the target node or the first historical feature vector of the target node output by the previous feature update layer, the second initial feature vector of each neighbor node adjacent to the target node or the second historical feature vector of each neighbor node output by the previous feature update layer, and the weight of each neighbor node, vector aggregation is performed to obtain the aggregated feature vector corresponding to the target node; the aggregated feature vector is subjected to nonlinear transformation processing to obtain the first current feature vector of the target node, thereby obtaining the current feature vector of each node.

[0050] In this implementation, the formula used for feature aggregation in each feature update layer is as follows:

[0051] Where k represents the k-th feature update layer; This represents the aggregated feature vector output by the k-th feature update layer for node v; This represents the feature vector output by the (k-1)th feature update layer after updating the features of node v; This represents the feature vector output by the (k-1)th feature update layer after updating the features of the neighboring nodes of node v; and All represent the edge matrix, i.e., the model parameters in the k-th feature update layer; This represents the attention weight, which is the weight corresponding to the target node v and its neighbor node u.

[0052] Each feature update layer undergoes a nonlinear transformation, using the following formula:

[0053] in, This is the ReLU activation function.

[0054] For example, taking a target recommendation model containing two feature update layers (a first feature update layer and a second feature update layer), and considering any node v in the target structure graph as the target node, the feature update process for the target node v is as follows: Step 1: Using the first feature update layer, determine the weights of each neighbor node based on the association information between the target node and its neighbor nodes. ; In this step, an attention mechanism can be used to adaptively calculate a weight value based on the association information between the target node and its neighboring nodes. This weight represents the neighboring nodes. Information for the target node The importance of.

[0055] Step 2: Based on the first initial feature vector of the target node v The second initial feature vector of each neighboring node u adjacent to the target node. Based on the weights of each neighboring node, vector aggregation is performed on the target node v to obtain the aggregated feature vector corresponding to the target node. ; Specifically, the aggregation formula is:

[0056] in, for The set of neighboring nodes, and The model training parameters in the first feature update layer (edge ​​matrix) are the learnable parameters.

[0057] Step 3: Aggregate feature vectors Perform nonlinear transformation to obtain the first current feature vector of the target node v. Similarly, steps one through three are repeated with the other nodes as the target nodes in turn to perform vector update processing and obtain the current feature vectors of the other nodes. That is, to obtain the current feature vector of each node; Specifically, the formula for the nonlinear transformation is:

[0058] in, This is the ReLU activation function.

[0059] Step 4: Convert the current feature vector of each node ( , ) as historical feature vector; Step 5: Using the second feature update layer, determine the weight of each neighbor node based on the association information between the target node v and each neighbor node m. Step 6: Based on the first historical feature vector of target node v (Feature vectors output by the first feature update layer), and the second initial feature vectors of each neighbor node u. (u=1,2,...), and the weights of each neighbor node u. Vector aggregation is performed on the target node v to obtain the aggregated feature vector corresponding to the target node. ; Specifically, the aggregation formula is:

[0060] in, for The set of neighboring nodes, and The model training parameters in the second feature update layer (edge ​​matrix) are the learnable parameters; This represents the attention weight, which is the weight corresponding to the target node v and its neighbor node u.

[0061] Step 7: Aggregate feature vectors Perform nonlinear transformation to obtain the first current feature vector of the target node v. That is, to obtain the target feature vector of the target node v. Similarly, steps five to seven are repeated with other nodes as target nodes in turn to perform vector update processing, so as to obtain the current feature vector of each node, that is, to obtain the target feature vector corresponding to each node. Specifically, the formula for the nonlinear transformation is:

[0062] in, This is the ReLU activation function.

[0063] In this embodiment, by updating the initial feature vectors of each node using the above method, the updated feature vectors become more comprehensive and accurate. That is, determining the aggregation weights based on the association between the target node and its neighboring nodes ensures a more reasonable and accurate aggregation result. By employing multiple feature update layers, in addition to aggregating feature information from directly adjacent neighboring nodes, it also aggregates feature information from other multi-hop neighboring nodes, i.e., aggregating feature information from indirect neighboring nodes within a predetermined number of hops. This makes the final updated target feature vector more comprehensive, laying the foundation for subsequent reasonable and accurate product recommendations to the target car dealer based on the target feature vector.

[0064] Another embodiment of this application provides a product recommendation method. In this embodiment, when using the recommendation layer in the target recommendation model to determine the recommendation order of each candidate product and perform product recommendation based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product, that is, when executing step S104, it specifically includes: using the recommendation layer in the target recommendation model to calculate the matching degree based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product, to obtain the matching degree of each candidate product, so as to determine the recommendation order of each candidate product based on the matching degree. That is, after updating the initial feature vector of each node in the target structure graph in the above manner and obtaining the target feature vector of each node, the first node corresponding to the target car dealer c can be determined, and then the target feature vector corresponding to the first node can be used as the first target feature vector. For example, the first target feature vector is determined as... Similarly, determine the second node corresponding to a candidate product p, and then use the target feature vector corresponding to this second node as the second target feature vector. For example, determine the second target feature vector as... Then, the matching degree / matching score between the first target feature vector and the second target feature vector can be calculated to obtain the matching degree / matching score between the target car dealer and each candidate product. Finally, based on the matching degree / matching score of each candidate product, the candidate products are sorted in descending order, thus determining the recommendation order of each candidate product. In this embodiment, the formula for calculating the matching degree / matching score s is: ,in, The value of T represents the matching degree between the target car dealer c and the candidate product p; T represents the vector transpose. This represents the first target feature vector corresponding to the target car dealer c; This represents the second target feature vector corresponding to candidate product p.

[0065] In this embodiment, by constructing a target structure graph containing several nodes and configuring corresponding attribute information for each node in the graph, it is possible to associate each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating each initial feature vector by combining the features of neighboring nodes, it is possible to capture the potential associations between car dealers and products, between products, and between car dealers, making the updated target feature vectors more consistent with the actual association logic. By determining the recommendation order based on the target feature vectors of the target car dealers and candidate products, accurate matching between car dealers and products can be achieved, effectively avoiding blind recommendations, ensuring that the recommended products are highly compatible with the attributes and needs of the target car dealers, improving the rationality of recommendations, and increasing product recommendation efficiency.

[0066] Another embodiment of this application provides a product recommendation method. In this embodiment, a target recommendation model can be pre-trained before product recommendation. The specific model training process is as follows: Step 1: Collect information on each sample car dealer and each sample product to obtain information on each sample car dealer, each sample product, and the corresponding recommended product tag information for each sample car dealer; Step 2: Based on the information of each sample car dealer and each sample product, construct a sample structure graph containing several sample nodes and several sample edges. Each node is configured with the sample attribute information of the car dealer or the sample attribute information of the product. Step 3: Using the multi-layer feedforward neural network in the initial recommendation model, perform feature extraction processing on the sample attribute information corresponding to each sample node in the sample structure graph to generate the initial feature vector of each node. Step 4: Using the feature update layer in the initial recommendation model, based on the initial feature vector of each sample node and the initial feature vector of the sample neighbor nodes corresponding to each sample node in the sample structure graph, the initial feature vector of each sample node is updated to obtain the target feature vector of each sample node. In this step, the initial recommendation model contains at least two initial feature update layers. Specifically, based on the initial first feature update layer, the sample weights of each neighboring node are determined according to the sample association information between the target node and each neighboring node. Based on the first initial feature vector of the target node, the second initial feature vectors of each neighboring node, and the sample weights of each neighboring node, vector aggregation is performed on the target node to obtain the aggregated feature vector corresponding to the target node. A non-linear transformation is then applied to the aggregated feature vector to obtain the first current feature vector of the target node. Similarly, the above process is repeated for other nodes as target nodes to perform vector update processing and obtain the current feature vectors of other nodes. This means obtaining the current feature vector of each sample node.

[0067] Then, using the k-th feature update layer, the sample weights of each neighbor node are determined based on the association information between the target node and each neighbor node. Based on the feature vector of the target node output by the previous feature update layer, the feature vectors of each neighbor node output by the previous feature update layer, and the weights of each neighbor node, vector aggregation and nonlinear transformation are performed again to obtain the current feature vector of the target node at the k-th feature update layer. Finally, the current feature vector output by the last feature update layer can be used as the target feature vector of the target node.

[0068] In this embodiment, when updating the model parameters in each feature update layer, an attention mechanism can be used to adaptively calculate the initial weights of each neighboring node based on the sample association information between the target node and each neighboring node. These initial weights are then used to aggregate the feature vectors of each neighboring node to obtain the current feature vector of the target node. The weights represent the neighboring nodes. Information for the target node The degree of importance. Then In other words, the attention mechanism can automatically learn and assign different weights to different edge types (different relationships configured on the edges). For example, car dealers can be connected by either "geographical distance" or "equity association." Business common sense suggests that parent and subsidiary companies with equity associations tend to have a high degree of consistency in their choice of financial products, while car dealers located only in the same city (geographically close) may have vastly different preferences. The model automatically discovers this pattern during training, thus assigning higher weights to "equity association edges." This assigns a lower weight to the "geographical distance edge." This not only improves accuracy but also gives the model strong business interpretability. In this embodiment, attention weights... Although it does not directly participate in the final score multiplication calculation, it determines the quality of the final feature vector at that time, indirectly and decisively affecting the final matching score.

[0069] Step 5: Using the recommendation layer in the initial recommendation model, determine the matching degree of each candidate product based on the first sample target feature vector corresponding to the target sample car dealer and the second sample target feature vector corresponding to each candidate product. Step 6: Based on the matching degree of each candidate product and the recommended product label corresponding to the same target sample car dealer, calculate the loss corresponding to each target sample car dealer; In this step, the loss The calculation formula is as follows:

[0070] in, For car dealers Recommended product tags (historical cooperative products, representing positive product samples). The recommended order of candidate products for each sample output by the model (negative product samples). This refers to the Sigmoid function.

[0071] Step 7: Based on the loss corresponding to each target sample car dealer, determine whether the predetermined model training stopping condition is met. If the model training condition is not met, update the model parameters in the feature update layer of the initial recommendation model to obtain the current recommendation model. Use the current recommendation model as the initial recommendation model and repeat Step 3 until the loss of each target sample car dealer meets the predetermined model training stopping condition. Then, use the current recommendation model as the target recommendation model.

[0072] In this step, when updating the model, the main focus is on the parameters in each feature update layer of the initial recommendation model. and parameters The model parameters are used for training to obtain the trained recommendation model.

[0073] In this embodiment, by first training the target recommendation model, the target recommendation model can be directly used to recommend products to target car dealers in the target structure graph. This not only ensures the accuracy of the recommendation results but also improves the efficiency of product recommendation.

[0074] Another embodiment of this application provides a product recommendation device, such as... Figure 3 As shown, it includes: Module 11 is used to construct a target structure graph containing several nodes and several edges based on information about each car dealer and each product. Each node is configured with attribute information of the car dealer or the product. The generation module 12 is used to extract features from the attribute information corresponding to each node in the target structure graph by using the multi-layer feedforward neural network layer in the pre-trained target recommendation model, and generate the initial feature vector of each node. The update module 13 is used to update the initial feature vector of each node by utilizing the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, so as to obtain the target feature vector of each node. The recommendation module 14 is used to utilize the recommendation layer in the target recommendation model to determine the recommendation order of each candidate product and make product recommendations based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product.

[0075] In this embodiment, the car dealer information includes any one or more of the following: the car dealer's attribute information, historical cooperation information with the product, and the association relationship with the car dealer; The product information includes any one or more of the following: product attribute information and its association with similar products; The process of constructing a target structure graph containing several nodes and edges based on information from various car dealers and products specifically includes: Create dealer nodes for each car dealer and product nodes for each product; Configure attribute information for dealer nodes based on the attribute information of each dealer, and configure attribute information for product nodes based on the attribute information of each product. Based on the historical cooperation information of products and the relationship with car dealers in the car dealer information, configure edges for the corresponding car dealer nodes and product nodes, configure edges for the corresponding car dealer nodes and car dealer nodes, and configure association information for each edge; Based on the association information of similar products in the product information, edges are configured for the corresponding product nodes and product nodes, and association information is configured for each edge to construct the target structure graph.

[0076] In this embodiment, the attribute information of the car dealer includes any one or more of the following: capital size, city, showroom area, main brands, years of operation, and mainstream car price range; the attribute information of the product includes any one or more of the following: rate of return, term, down payment ratio, service fee, and risk level.

[0077] In this embodiment, the product recommendation device further includes a processing module, which is used to: perform normalization processing or one-hot encoding processing on each attribute information corresponding to the same node in sequence to obtain each original attribute vector; and concatenate the original attribute vectors corresponding to the same node to obtain a concatenated vector. The generation module is specifically used to: utilize the multi-layer feedforward neural network layer in the pre-trained target recommendation model to perform non-linear transformation processing on the concatenated vector of each node to obtain the initial feature vector.

[0078] In this embodiment, the target recommendation model includes at least two feature update layers during implementation. When each feature update layer performs vector updates, the update module is specifically used to: update the target node's vector based on the target node's first initial feature vector or the target node's first historical feature vector output by the previous feature update layer, the target node's second initial feature vector or the target node's second historical feature vector output by the previous feature update layer, and the association information between the target node and each of the neighboring nodes, to obtain the target node's first current feature vector, thereby obtaining the current feature vector of each node.

[0079] In this embodiment, the update module is specifically used to: determine the weight of each neighboring node based on the association information between the target node and each neighboring node in the target structure graph; perform vector aggregation based on the first initial feature vector of the target node or the first historical feature vector of the target node output by the previous feature update layer, the second initial feature vector of each neighboring node adjacent to the target node or the second historical feature vector of each neighboring node output by the previous feature update layer, and the weight of each neighboring node, to obtain the aggregated feature vector corresponding to the target node; and perform nonlinear transformation processing on the aggregated feature vector to obtain the first current feature vector of the target node, thereby obtaining the current feature vector of each node.

[0080] In this embodiment, the recommendation module is specifically used to: utilize the recommendation layer in the target recommendation model to calculate the matching degree based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product, and obtain the matching degree of each candidate product, so as to determine the recommendation order of each candidate product based on the matching degree.

[0081] The device in this embodiment, by constructing a target structure graph containing several nodes and configuring corresponding attribute information for each node in the graph, can associate each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating each initial feature vector by combining the features of neighboring nodes, it can capture the potential associations between car dealers and products, between products, and between car dealers, making the updated target feature vectors more consistent with the actual association logic. By determining the recommendation order based on the target feature vectors of the target car dealer and candidate products, it can achieve accurate matching between car dealers and products, effectively avoiding blind recommendations, ensuring that the recommended products are highly compatible with the attributes and needs of the target car dealers, improving the rationality of recommendations, and increasing product recommendation efficiency.

[0082] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps: Step 1: Based on the information of each car dealer and each product, construct a target structure graph containing several nodes and several edges. Each node is configured with the attribute information of the car dealer or the attribute information of the product. Step 2: Using the multi-layer feedforward neural network in the pre-trained target recommendation model, perform feature extraction processing on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. Step 3: Using the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, update the initial feature vector of each node to obtain the target feature vector of each node. Step 4: Using the recommendation layer in the target recommendation model, based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product, determine the recommendation order of each candidate product and make product recommendations.

[0083] The specific implementation process of the above method steps can be found in the embodiments of the above-mentioned product recommendation methods, and will not be repeated here.

[0084] The storage medium in this application, by constructing a target structure graph containing several nodes and configuring corresponding attribute information for each node in the graph, can associate each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating each initial feature vector by combining the features of neighboring nodes, it can capture the potential relationships between car dealers and products, between products, and between car dealers, making the updated target feature vectors more consistent with the actual association logic. By determining the recommendation order based on the target feature vectors of the target car dealers and candidate products, it can achieve accurate matching between car dealers and products, effectively avoid blind recommendations, ensure that the recommended products are highly consistent with the attributes and needs of the target car dealers, improve the rationality of recommendations, and increase product recommendation efficiency.

[0085] Another embodiment of this application provides a computer device, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a product marketing strategy determination method on the server side.

[0086] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a product marketing strategy determination method.

[0087] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Step 1: Based on the information of each car dealer and each product, construct a target structure graph containing several nodes and several edges. Each node is configured with the attribute information of the car dealer or the attribute information of the product. Step 2: Using the multi-layer feedforward neural network in the pre-trained target recommendation model, perform feature extraction processing on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. Step 3: Using the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, update the initial feature vector of each node to obtain the target feature vector of each node. Step 4: Using the recommendation layer in the target recommendation model, based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product, determine the recommendation order of each candidate product and make product recommendations.

[0088] The specific implementation process of the above method steps can be found in the embodiments of the above-mentioned product recommendation methods, and will not be repeated here.

[0089] The computer device in this application, by constructing a target structure graph containing several nodes and configuring corresponding attribute information for each node in the graph, can associate each car dealer with each product, laying the foundation for subsequent product recommendation using a target recommendation model. By generating initial feature vectors based on the attribute information of each node using a pre-trained target recommendation model, and updating each initial feature vector by combining the features of neighboring nodes, it can capture the potential relationships between car dealers and products, between products, and between car dealers, making the updated target feature vectors more consistent with the actual association logic. By determining the recommendation order based on the target feature vectors of the target car dealer and candidate products, it can achieve accurate matching between car dealers and products, effectively avoid blind recommendations, ensure that the recommended products are highly consistent with the attributes and needs of the target car dealers, improve the rationality of recommendations, and increase product recommendation efficiency.

[0090] 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.

[0091] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application. Software tools or components not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

Claims

1. A product recommendation method, characterized in that, include: Based on information about each car dealer and each product, a target structure graph containing several nodes and several edges is constructed. Each node is configured with the attribute information of the car dealer or the attribute information of the product. Using the multi-layer feedforward neural network in the pre-trained target recommendation model, feature extraction is performed on the attribute information corresponding to each node in the target structure graph to generate the initial feature vector of each node. Using the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, the initial feature vector of each node is updated to obtain the target feature vector of each node. Using the recommendation layer in the target recommendation model, the recommendation order of each candidate product is determined and product recommendations are made based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product.

2. The method as described in claim 1, characterized in that, The car dealer information includes any one or more of the following: the car dealer's attribute information, historical cooperation information with the product, and the association relationship with the car dealer; The product information includes any one or more of the following: product attribute information and its association with similar products; The process of constructing a target structure graph containing several nodes and edges based on information from various car dealers and products specifically includes: Create dealer nodes for each car dealer and product nodes for each product; Configure attribute information for dealer nodes based on the attribute information of each dealer, and configure attribute information for product nodes based on the attribute information of each product. Based on the historical cooperation information of products and the relationship with car dealers in the car dealer information, configure edges for the corresponding car dealer nodes and product nodes, configure edges for the corresponding car dealer nodes and car dealer nodes, and configure association information for each edge; Based on the association information of similar products in the product information, edges are configured for the corresponding product nodes and product nodes, and association information is configured for each edge to construct the target structure graph.

3. The method as described in claim 1, characterized in that, The attribute information of the car dealers includes any one or more of the following: capital size, city, showroom area, main brands, years of operation, and mainstream car price range; The product's attribute information includes any one or more of the following: yield, term, down payment ratio, service fee, and risk level.

4. The method as described in claim 1, characterized in that, Before generating the initial feature vectors for each node, the method further includes: Normalization or one-hot encoding is performed on each attribute information corresponding to the same node in sequence to obtain each original attribute vector; The concatenated vector is obtained by concatenating the original attribute vectors corresponding to the same node. The method of utilizing the multi-layer feedforward neural network in the pre-trained target recommendation model to extract features from the attribute information corresponding to each node in the target structure graph and generate the initial feature vector for each node specifically includes: By utilizing the multi-layer feedforward neural network in the pre-trained target recommendation model, the concatenated vectors of each node are subjected to non-linear transformation to obtain the initial feature vector.

5. The method as described in claim 1, characterized in that, The target recommendation model includes at least two feature update layers; When performing vector updates, each feature update layer specifically includes: updating the target node's vector based on the target node's first initial feature vector or the target node's first historical feature vector output by the previous feature update layer, the second initial feature vectors of each neighboring node adjacent to the target node or the second historical feature vectors of each neighboring node output by the previous feature update layer, and the association information between the target node and each of the neighboring nodes, to obtain the target node's first current feature vector, thereby obtaining the current feature vectors of each node.

6. The method as described in claim 5, characterized in that, The process involves updating the target node's vector based on its first initial feature vector or the first historical feature vector output by the previous feature update layer, the second initial feature vectors of each neighboring node adjacent to the target node or the second historical feature vectors of each neighboring node output by the previous feature update layer, and the association information between the target node and each neighboring node. This process yields the first current feature vector of the target node, and ultimately, the current feature vectors of each node are obtained. Specifically, this includes: Based on the association information between the target node and each neighbor node in the target structure graph, the weight of each neighbor node is determined; Based on the first initial feature vector of the target node or the first historical feature vector of the target node output by the previous feature update layer, the second initial feature vector of each neighboring node adjacent to the target node or the second historical feature vector of each neighboring node output by the previous feature update layer, and the weight of each neighboring node, vector aggregation is performed to obtain the aggregated feature vector corresponding to the target node. The aggregated feature vector is subjected to a nonlinear transformation to obtain the first current feature vector of the target node, thereby obtaining the current feature vector of each node.

7. The method as described in claim 1, characterized in that, The step of utilizing the recommendation layer in the target recommendation model to determine the recommendation order of each candidate product based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product specifically includes: Using the recommendation layer in the target recommendation model, the matching degree is calculated based on the first target feature vector of the first node corresponding to the target car dealer and the second target feature vector of each second node corresponding to each candidate product, so as to obtain the matching degree of each candidate product and determine the recommendation order of each candidate product based on the matching degree.

8. A product recommendation device, characterized in that, include: The construction module is used to build a target structure graph containing several nodes and several edges based on information about each car dealer and each product. Each node is configured with the attribute information of the car dealer or the attribute information of the product. The generation module is used to extract features from the attribute information corresponding to each node in the target structure graph by using the multi-layer feedforward neural network layer in the pre-trained target recommendation model, and generate the initial feature vector of each node. The update module is used to update the initial feature vector of each node by utilizing the feature update layer in the target recommendation model, based on the initial feature vector of each node and the initial feature vector of the neighboring nodes corresponding to each node in the target structure graph, so as to obtain the target feature vector of each node. The recommendation module is used to utilize the recommendation layer in the target recommendation model to determine the recommendation order of each candidate product and make product recommendations based on the first target feature vector corresponding to the target car dealer and the second target feature vector corresponding to each candidate product.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the product recommendation method according to any one of claims 1-7.

10. A computer device, characterized in that, It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the product recommendation method according to any one of claims 1-7.

Images