Product recommendation method and device, equipment, medium and program product

Abstract

The invention provides a product recommendation method which can be applied to the technical field of artificial intelligence or the field of financial science and technology. The method comprises the steps of performing feature extraction on obtained user related information through a preset auto-encoder to obtain user features and interaction features, and performing splicing processing on the user features and the interaction features to obtain a user product attribute feature matrix; performing degree prediction processing based on the user product attribute feature matrix through a preset degree prediction model to obtain a degree prediction value; performing graph convolution calculation based on the user features, the interaction features and the degree prediction value through a preset graph neural network to obtain user feature representation; and performing weighted mapping according to the user feature representation to obtain a product adaptive score, and generating product recommendation information based on the product adaptive score. The invention further provides a product recommendation device, equipment, a medium and a program product.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, specifically to matching recommendation methods, and more specifically to a product recommendation method, apparatus, device, medium, and program product. Background Technology

[0002] With the rapid development of technology, providing a solution for personalized financial product recommendations is particularly crucial. Intelligent automated product recommendation methods can avoid interference from human factors, ensuring that financial product recommendations are more objective and accurate, thereby improving customer experience and enhancing the competitiveness of bank wealth management business.

[0003] Current personalized recommendation systems based on matrix factorization and collaborative filtering still have at least the following shortcomings: First, matrix factorization and collaborative filtering do encounter bottlenecks when handling large-scale data. For example, when the number of users and products increases to millions or even tens of millions, the dimensionality of the matrix becomes extremely large. Factoring such a matrix requires significant computational resources and time, resulting in memory consumption, excessively long training times, and difficulties in real-time updates. Second, similarity calculations in collaborative filtering (such as cosine similarity or Pearson coefficients) become less effective when data is sparse. Simultaneously calculating the similarity matrix for all users or items is highly complex, making it impractical for large-scale systems. Third, matrix factorization primarily focuses on the direct interactions between the latent feature vectors of users and items, such as dot products. However, real-world user behavior may involve more complex patterns. For example, user A might like product B because product B possesses certain features that are indirectly related to user A's historical behavior. Higher-order associations may include third-order user-item-user relationships or more complex graph structures. For instance, graph neural networks can capture multi-order relationships, but because they focus on uncovering higher-order associations, they tend to overlook local relationships.

[0004] Therefore, there is an urgent need for a product recommendation solution that can reduce memory consumption, shorten training time, reduce the computational complexity of similarity, and simultaneously focus on high-order global and local information, thereby improving the accuracy of product recommendations. Summary of the Invention

[0005] In view of the above problems, this application provides product recommendation methods, apparatus, devices, media and program products.

[0006] According to a first aspect of this application, a product recommendation method is provided, comprising: extracting features from acquired user-related information using a pre-defined autoencoder to obtain user features and interaction features, and concatenating the user features and interaction features to obtain a user-product attribute feature matrix; the user-related information includes user information and interaction information related to the user information; performing degree prediction processing based on the user-product attribute feature matrix using a pre-defined degree prediction model to obtain a degree prediction value; performing graph convolution calculation based on user features, interaction features, and degree prediction value using a pre-defined graph neural network to obtain a user feature representation; performing weighted mapping based on the user feature representation to obtain a product adaptive score, and generating product recommendation information based on the product adaptive score.

[0007] According to an embodiment of this application, a user-related information is extracted using a preset autoencoder to obtain user features and interaction features. The user features and interaction features are then concatenated to obtain a user product attribute feature matrix. This includes applying activation function operations to the user features and interaction features based on a preset encoding weight matrix and encoding bias coefficient in the autoencoder, and concatenating the user features and interaction features to form a user product attribute feature matrix.

[0008] According to an embodiment of this application, a degree prediction value is obtained by performing degree prediction processing based on the user product attribute feature matrix using a preset degree prediction model, including: performing matrix decomposition and degree calculation on the user product attribute feature matrix based on the prediction weight matrix and prediction bias coefficient of the degree prediction model to obtain the transmission node degree; and performing dimensional transformation on the transmission node degree using a preset compression function to obtain the degree prediction value.

[0009] According to an embodiment of this application, a user feature representation is obtained by performing graph convolution calculation based on user features, interaction features, and degree prediction values ​​using a preset graph neural network. This includes: establishing a user-product adjacency matrix based on user features, interaction features, and degree prediction values; normalizing and exponentially calculating the user-product adjacency matrix to obtain attribute feature vectors of users or products in specific layers of the graph neural network; performing an adaptive summation operation based on the attribute feature vectors in all layers of the graph neural network to obtain user representations and product representations; and using the user representations and product representations as the user feature representation.

[0010] According to embodiments of this application, an autoencoder, a degree prediction model, and a graph neural network are integrated into a pre-trained product recommendation model. The pre-training of the product recommendation model includes: capturing user data and product data in a preset data warehouse, and performing data preprocessing operations based on the user data and product data to obtain a product feature matrix and a customer-product interaction graph; repeatedly training a preset fusion architecture based on the product feature matrix and the customer-product interaction graph until the fusion loss value related to the fusion architecture reaches a preset fusion threshold, then using the optimal fusion architecture as the product recommendation model; wherein the fusion architecture includes an encoder structure, a feedforward neural network, and a graph convolutional network; the fusion loss value is obtained by weighted summation of the reconstruction loss value related to the encoder structure, the degree prediction loss value related to the feedforward neural network, and the cross-entropy loss value related to the graph convolutional network.

[0011] According to an embodiment of this application, the autoencoder is repeatedly trained by the encoder; wherein, until the reconstruction loss value reaches a preset reconstruction threshold, the optimal encoder is taken as the autoencoder; wherein, obtaining the reconstruction loss value includes: performing dimensionality reduction processing on the product feature matrix and the customer feature matrix extracted from the customer product interaction graph to obtain an initial matrix; constructing a customer and product mapping relationship based on the initial matrix to form a customer product splicing matrix; decoding the customer product splicing matrix to obtain a reconstruction matrix corresponding to the initial matrix; and using a preset reconstruction algorithm to calculate the loss value between the reconstruction matrix and the initial matrix to obtain the reconstruction loss value.

[0012] According to an embodiment of this application, the degree prediction model is repeatedly trained by a feedforward neural network, which is communicatively connected to the encoder. The optimal feedforward neural network is selected as the degree prediction model until the degree prediction loss value reaches a preset degree prediction threshold. Obtaining the degree prediction loss value includes: extracting features from the customer product interaction graph to obtain a customer product adjacency matrix; calculating the degree of the customer product adjacency matrix to obtain a label degree value; calculating the degree of the customer product splicing features output by the encoder using the feedforward neural network to obtain an initial degree value; performing dimensionality transformation on the initial degree value using a preset compression function to obtain a training degree value; and calculating the loss based on the training degree value and the label degree value to obtain the degree prediction loss value.

[0013] According to embodiments of this application, the graph neural network is repeatedly trained from a graph convolutional network; the last layer of the graph convolutional network communicates with the feedforward neural network through the last layer of the encoder, and each layer of the graph convolutional network is connected to each layer of the encoder in a one-to-one correspondence; wherein, obtaining the cross-entropy loss value includes: establishing a reinforced adjacency matrix based on the training degree value, the product feature matrix, and the customer feature matrix; calculating different powers of the reinforced adjacency matrix, and normalizing the reinforced adjacency matrix based on the different powers of the reinforced adjacency matrix to obtain a normalized matrix; calculating different powers of the normalized matrix, and performing convolution processing based on the different powers of the normalized matrix to obtain node attribute feature vectors; performing an adaptive summation operation based on the node attribute feature vectors to obtain the corresponding representation of the customer product; performing probability prediction on the corresponding representation of the customer product through a preset probability prediction function to obtain the training prediction probability; and continuing to calculate the loss based on the training prediction probability and the pre-acquired sample probability to obtain the cross-entropy loss value.

[0014] According to embodiments of this application, weighted mapping based on user feature representation to obtain product adaptive scores, and generating product recommendation information based on product adaptive scores, includes: performing a dot product calculation on the product embedding vector of the target product and the user feature representation to obtain a dot product vector; performing numerical processing on the dot product vector through a preset mapping function to obtain an initial value; scaling up the initial value to obtain the score of the target product; and performing sorting and corresponding operations on the scores of each target product to form product recommendation information.

[0015] A second aspect of this application provides a product recommendation device, comprising: a feature extraction module, used to extract features from acquired user-related information using a preset autoencoder to obtain user features and interaction features, and to concatenate the user features and interaction features to obtain a user-product attribute feature matrix; the user-related information includes user information and interaction information related to the user information; a degree prediction module, used to perform degree prediction processing based on the user-product attribute feature matrix using a preset degree prediction model to obtain a degree prediction value; a feature fusion module, used to perform graph convolution calculation based on user features, interaction features, and degree prediction value using a preset graph neural network to obtain a user feature representation; and a data generation module, used to perform weighted mapping based on the user feature representation to obtain a product adaptive score, and to generate product recommendation information based on the product adaptive score.

[0016] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0017] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0018] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] The above one or more embodiments have the following beneficial effects: First, a preset autoencoder is used to extract features from the acquired user-related information to obtain user features and interaction features. Then, the user features and interaction features are concatenated to obtain a user product attribute feature matrix. Then, a degree prediction model plays a connecting role, performing degree prediction processing based on the user product attribute feature matrix to obtain degree prediction values ​​representing local information. Finally, a graph neural network is used to perform graph convolution calculation based on the user features, interaction features, and degree prediction values ​​to obtain a user feature representation that focuses on both global and local information. This solves the technical problem in the prior art where graph neural networks only focus on global features and ignore local features, resulting in poor recommendation accuracy. Thus, the acquired user feature representation can not only focus on global information but also on local information, improving information focus and thereby improving the accuracy of product recommendations. Attached Figure Description

[0020] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 The illustrations depict application scenarios of product recommendation methods, apparatus, devices, media, and program products according to embodiments of this application.

[0022] Figure 2 A flowchart illustrating a product recommendation method according to an embodiment of this application is shown schematically;

[0023] Figure 3 This illustration shows a flowchart of the data preprocessing and data pre-processing operations before the product recommendation model is pre-trained in the product recommendation method according to an embodiment of this application;

[0024] Figure 4 This illustration schematically shows the specific structure and data flow involved in the pre-trained product recommendation model in the product recommendation method according to an embodiment of this application;

[0025] Figure 5 A schematic diagram illustrating the structure of a product recommendation device according to an embodiment of this application is shown; and

[0026] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a product recommendation method according to an embodiment of this application. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] In the technical solution of this application, the application data related to users / customers is collected, stored, used, processed, transmitted, provided, disclosed and applied with the authorization of users / customers, which complies with the provisions of relevant laws and regulations, takes necessary confidentiality measures and does not violate public order and good morals.

[0032] Users' risk preferences can change rapidly, which traditional methods cannot capture in time, leading to the recommendation of unsuitable products and increasing the risk of user churn. Therefore, a combination of graph neural networks and autoencoders is used for intelligent product recommendation. However, due to the fundamental processing logic of graph neural networks, which focuses on capturing high-order features, it is easy to ignore local features, resulting in poor recommendation accuracy. Furthermore, when the number of users and products increases to millions or even tens of millions, the dimension of the matrix becomes very large. Decomposing these large matrices requires a lot of computing resources and time, resulting in problems such as memory consumption, excessively long training time, and difficulty in real-time updates.

[0033] This application provides a product recommendation method, apparatus, device, medium, and program product. The product recommendation method includes: extracting features from acquired user-related information using a preset autoencoder to obtain user features and interaction features; concatenating the user features and interaction features to obtain a user-product attribute feature matrix; the user-related information includes user information and interaction information related to the user information; performing degree prediction processing based on the user-product attribute feature matrix using a preset degree prediction model to obtain a degree prediction value; performing graph convolution calculation based on user features, interaction features, and the degree prediction value using a preset graph neural network to obtain a user feature representation; performing weighted mapping based on the user feature representation to obtain a product adaptive score; and generating product recommendation information based on the product adaptive score. This addresses the technical problem in the prior art where graph neural networks only focus on global features and ignore local features, leading to poor recommendation accuracy. Therefore, the acquired user feature representation can not only focus on global information but also on local information, improving information focus and thus improving product recommendation accuracy.

[0034] Figure 1 The illustration shows an application scenario diagram of the product recommendation method according to an embodiment of this application.

[0035] like Figure 1 As shown, the application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105.

[0036] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0037] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0038] Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. Network 104 serves as the medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0039] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0040] It should be noted that the product recommendation method provided in this application embodiment can generally be executed by server 105. Correspondingly, the product recommendation device provided in this application embodiment can generally be located in server 105. The product recommendation method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the product recommendation device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0041] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0042] The following will be based on Figure 1 The described scene, through Figures 2-4 The product recommendation method according to the embodiments of this application will be described in detail.

[0043] Figure 2 A flowchart illustrating a product recommendation method according to an embodiment of this application is shown.

[0044] like Figure 2As shown, the product recommendation method in this embodiment includes steps S210 to S230, as detailed below:

[0045] Operation S210 extracts features from the acquired user-related information using a pre-defined autoencoder to obtain user features and interaction features, and concatenates these features to obtain a user-product attribute feature matrix. User-related information includes user information and interaction information related to that information. Operation S220 uses a pre-defined degree prediction model to perform degree prediction processing based on the user-product attribute feature matrix to obtain a degree prediction value. Operation S230 uses a pre-defined graph neural network to perform graph convolution calculations based on user features, interaction features, and the degree prediction value to obtain a user feature representation. Operation S240 performs a weighted mapping based on the user feature representation to obtain a product adaptive score, and generates product recommendation information based on the product adaptive score.

[0046] It should be noted that the product can be any product that can be recommended, such as financial products including but not limited to insurance products and wealth management products, or other products or items that can be recommended by information technology.

[0047] As an example, the autoencoder, degree prediction model, and graph neural network are integrated into a pre-trained product recommendation model. Based on this, a complete product recommendation operation is performed, with the degree prediction model playing a crucial role. It performs degree prediction processing based on the user product attribute feature matrix to obtain degree prediction values ​​representing local information. Then, the graph neural network performs graph convolution calculations based on the user features, interaction features, and degree prediction values ​​to obtain a user feature representation that focuses on both global and local information. This solves the technical problem in existing technologies where graph neural networks only focus on global features and ignore local features, resulting in poor recommendation accuracy. As a result, the obtained user feature representation can not only focus on global information but also on local information, improving information focus and thus improving the accuracy of product recommendations.

[0048] In this embodiment, during operation S210, activation functions are applied to user features and interaction features based on the preset encoding weight matrix and encoding bias coefficients in the autoencoder. This concatenates the user features and interaction features to form a user product attribute feature matrix. Therefore, when applying the preset autoencoder, no parameter adjustments are needed to directly obtain the user product attribute feature matrix based on the preset encoding weight matrix and encoding bias coefficients. This matrix is ​​then used by the following degree prediction model, thereby laying the foundation for improving product recommendation accuracy and overall data processing efficiency.

[0049] For example, the user feature is the feature obtained by extracting features from user information in user-related information, the interaction feature is the feature obtained by extracting features from interaction information related to user information in user-related information, and the user product attribute feature matrix is ​​a feature matrix formed by performing related matrix operations on the user feature and the interaction feature, such as activation function operations, which can be ReLU functions, etc.

[0050] The encoding weight matrix and encoding bias coefficients are parameters that are pre-set when training the autoencoder. The specific training method will be explained in detail in the following examples.

[0051] In this embodiment, operation S220 extracts features from the acquired user-related information using a preset autoencoder to obtain user features and interaction features, and concatenates these features to obtain a user product attribute feature matrix. This includes: performing degree prediction processing on the user product attribute feature matrix using a preset degree prediction model to obtain degree prediction values, including: performing matrix decomposition and degree calculation on the user product attribute feature matrix based on the prediction weight matrix and prediction bias coefficient of the degree prediction model to obtain the transmission node degree; and performing dimensional transformation on the transmission node degree using a preset compression function to obtain the degree prediction values. In this embodiment, the preset compression function is used to perform dimensional transformation on the transmission node degree to obtain the degree prediction values. This overcomes the mismatch between the transmission node degree obtained based on the user product attribute feature matrix and the degree value required by the subsequent graph neural network, allowing the generated degree prediction values ​​to be directly used as input to the subsequent graph neural network. This improves the inclusiveness of the autoencoder and graph neural network for the degree prediction model, thereby laying the foundation for improving the accuracy of product recommendations.

[0052] For example, the prediction weight matrix and prediction bias coefficient are parameters that are obtained in advance when training and generating the degree prediction model. The specific calculation methods for matrix decomposition and degree calculation, as well as the specific calculation methods for dimensional transformation of the transmission node degree, are also calculation logics that are set in advance when training and generating the degree prediction model. The specific calculation logic will be illustrated in detail in the subsequent operation of training and generating the degree prediction model.

[0053] In this embodiment, operation S230 uses a preset graph neural network to perform graph convolution calculations based on user features, interaction features, and degree prediction values ​​to obtain user feature representations. This includes: establishing a user-product adjacency matrix based on user features, interaction features, and degree prediction values; normalizing and exponentially calculating the user-product adjacency matrix to obtain attribute feature vectors of users or products in specific layers of the graph neural network; performing adaptive summation operations based on attribute feature vectors in all layers of the graph neural network to obtain user representations and product representations; and using user representations and product representations as user feature representations.

[0054] For example, the input of the graph neural network includes not only the user features and interaction features output by each layer of the autoencoder corresponding to each layer of the graph neural network, but also the degree prediction value output by the degree prediction model. Based on the newly added input (degree prediction value), an adaptive summation operation is performed, so as to not only focus on the global features of users and related products, but also on local features, thereby improving information attention and improving the accuracy of product recommendations.

[0055] It should be noted that the specific execution algorithms / logic for establishing the user-product adjacency matrix based on user characteristics, interaction characteristics, and degree prediction values; the specific execution algorithms / logic for normalizing and exponentially calculating the user-product adjacency matrix to obtain the attribute feature vectors of users or products in specific layers of the graph neural network; and the specific algorithms / logic for adaptively summing the attribute feature vectors in all layers of the graph neural network to obtain user and product representations are all pre-set logic / algorithms when training the graph neural network. The specific formulas and algorithms used will be explained in detail in the subsequent examples of training the graph neural network.

[0056] The autoencoder, degree prediction model, and graph neural network used in the above operations S210~S230 are all pre-trained. When the user uses them, they can be directly combined and applied. The specific algorithms and logic used in the application are similar to those used in the training. The specific steps for training and generating the autoencoder, degree prediction model, and graph neural network will be described in detail below.

[0057] In this embodiment, the autoencoder, degree prediction model, and graph neural network are integrated into a pre-trained product recommendation model. The pre-training of the product recommendation model includes: capturing user data and product data in a preset data warehouse, and performing data preprocessing operations based on the user data and product data to obtain a product feature matrix and a customer-product interaction graph; repeatedly training a preset fusion architecture based on the product feature matrix and the customer-product interaction graph until the fusion loss value related to the fusion architecture reaches a preset fusion threshold, then using the optimal fusion architecture as the product recommendation model; wherein the fusion architecture includes an encoder structure, a feedforward neural network, and a graph convolutional network; the fusion loss value is obtained by weighted summation of the reconstruction loss value related to the encoder structure, the degree prediction loss value related to the feedforward neural network, and the cross-entropy loss value related to the graph convolutional network.

[0058] In this embodiment, a pre-acquired product feature matrix and customer product interaction graph are used to repeatedly train a preset fusion architecture until the fusion loss value associated with the fusion architecture reaches a preset fusion threshold. The optimal fusion architecture is then used as the product recommendation model. That is, data preprocessing is required before training to obtain sample data containing the product feature matrix and customer product interaction graph. Prior to data preprocessing, a data preprocessing process is also included, capturing user data and product data from a data warehouse.

[0059] Figure 3 This diagram illustrates the workflow of data preprocessing and pre-processing operations before pre-training a product recommendation model. For example, before obtaining the sample data containing the product feature matrix and customer-product interaction graph, the following preprocessing steps can be performed: For instance, continuous features of customers and products often have multiple dimensions, so dimensionless processing is required first. Since incomplete customer information may exist, in this example, missing value handling strategies such as mean imputation can be used during preprocessing. For example, first, the first data is obtained from the data warehouse, and then categorized according to the data type to obtain the second data. For instance, continuous features are binarized, and discontinuous features are binarized; specific encoding is used for categorical features, such as OneHot encoding (an encoding technique that converts categorical variables into binary vectors).

[0060] The following specific example illustrates the processing procedure in detail: First, pre-defined computer statements are used to retrieve customer and product characteristic information from the data warehouse. This information can be exported in a pre-defined format, such as CSV. Then, a pre-defined specific database is used to read the information with the pre-defined format, thus obtaining unprocessed second data. A partial example of this first data is shown below:

[0061]

[0062] It should be noted that the above data is only for illustrating the data format and has no other illustrative effect. For the continuous features in this first set of data, such as age, annual income, and credit score, the continuous feature of age is binarized. This involves converting the continuous value into a binary value (0 / 1) based on a threshold, where the threshold needs to be combined with business knowledge. Example: Is the age greater than 40? Processing result:

[0063]

[0064] Binning is performed on the discontinuous characteristic of annual income, such as equal-width binning. Example: Annual income binning (interval width = 200,000 yuan), bin boundaries: [0, 20), [20, 40), [40, 60), [60, 80). Tags: low income = 0, middle income = 1, upper-middle income = 2, high income = 3. The resulting processing is as follows:

[0065]

[0066] Specific encoding processing is applied to categorical features, such as converting each category value into an independent binary feature column. Example: Original occupation categories: Engineer, Teacher, Freelancer, Doctor. After conversion, four new columns are added, one for each category. The processing result is as follows:

[0067]

[0068] This yields a clearly structured second set of data. The second data concerning users is designated as user data, and the second data concerning products is designated as product data. Further data preprocessing is then performed on this user and product data to obtain a customer feature matrix, a product feature matrix (such as a financial product feature matrix), and a customer-product interaction diagram. This customer-product interaction diagram can be a financial product interaction diagram. In one specific embodiment, this diagram can first be generated by transforming the interaction relationship between the customer and the product to obtain interaction transformation information, and then using a software library (such as the Networkx library) to construct the diagram based on this interaction transformation information. After obtaining the product feature matrix and customer product interaction graph, the process of pre-training the product recommendation model is carried out. For example, it is necessary to first establish a fusion architecture including an encoder structure, a feedforward neural network, and a graph convolutional network, and obtain the loss function for each framework. For example, it is necessary to obtain the reconstruction loss value for the encoder structure, the degree prediction loss value for the feedforward neural network, and the cross-entropy loss value for the graph convolutional network. Then, the three loss values ​​are weighted to obtain the first weight, the second weight, and the third weight corresponding to the three loss values. Finally, the fusion loss value of the entire product recommendation model (fusion architecture) is obtained by weighted summation based on the first weight, the second weight, and the third weight.

[0069] Figure 4 The diagram illustrates the specific structure and data flow involved in the pre-trained product recommendation model in the product recommendation method according to an embodiment of this application. The following will describe in turn the specific operation process of how to obtain the reconstruction loss value (autoencoder reconstruction loss) of the encoder structure, the degree prediction loss value of the feedforward neural network, and the cross-entropy loss value of the graph convolutional network.

[0070] In this embodiment, the autoencoder is trained repeatedly by an encoder; the optimal encoder is selected as the autoencoder once the reconstruction loss value reaches a preset reconstruction threshold; obtaining the reconstruction loss value includes: performing dimensionality reduction processing on the product feature matrix and the customer feature matrix extracted from the customer product interaction graph to obtain an initial matrix; constructing a customer-product mapping relationship based on the initial matrix to form a customer-product splicing matrix; decoding the customer-product splicing matrix to obtain a reconstruction matrix corresponding to the initial matrix; and calculating the loss value between the reconstruction matrix and the initial matrix using a preset reconstruction algorithm to obtain the reconstruction loss value.

[0071] As an example, since the feature vectors of customer and financial products have high dimensionality, an autoencoder is used for feature dimensionality reduction in this embodiment. The following example operation can be adopted:

[0072] ;

[0073] in, It is an activation function, which can be used in this example. The function is used as an activation function. The ReLU (Rectified Linear Unit) function is a commonly used activation function in artificial neural networks. and These are the encoders in the first... The trainable weight matrix and bias of the layer. and These are the encoders Layers and The nodes of the layer are represented, and Set as the initial matrix . The specific definitions are as follows:

[0074]

[0075] in, and This represents the attribute feature matrix of customers and wealth management products, where M is the number of customers, N is the number of wealth management products, and r is the dimension of the feature vectors of customers and wealth management products. This represents the customer product concatenation matrix obtained by concatenating the customer feature matrix and the financial feature matrix along the column direction.

[0076] In this example, a decoder is used to obtain the reconstruction loss value. Specifically, the decoder algorithm is defined as follows:

[0077] ;

[0078] The input to the decoder is... , and These are the trainable weight matrix and bias of the l-th layer in the decoder, respectively. and These are the node embeddings at layers l-1 and l in the decoder, respectively. The output of the decoder is the original feature matrix. The reconstruction matrix is ​​obtained. The loss value between the reconstruction matrix and the initial matrix is ​​then calculated using the following reconstruction algorithm to obtain the reconstruction loss value:

[0079] .

[0080] It should be noted that the application process of the autoencoder is similar to the training process described above, and the formulas used are also the same. The decoder's operation only occurs during the training phase. When applying the trained autoencoder, the decoder process is unnecessary; only the step of obtaining the user product splicing matrix (corresponding to the customer product splicing matrix in the training phase) needs to be executed. During the training phase of the autoencoder, all sample data (product feature matrix and customer product interaction graph) undergo dimensionality reduction processing before subsequent feature extraction and mapping. This addresses the current technical problems of extremely large matrix dimensions when the number of customers and products increases to millions or even tens of millions, requiring significant computational resources and time to decompose such matrices, resulting in excessive memory consumption, long training times, and difficulties in real-time updates. Therefore, it achieves the beneficial effects of "reducing matrix dimensions, lowering memory consumption, and shortening training time."

[0081] Similarly, in this embodiment, the degree prediction model (also called the degree prediction module) is repeatedly trained by a feedforward neural network, which is connected to the encoder. The optimal feedforward neural network is selected as the degree prediction model once the degree prediction loss value reaches a preset degree prediction threshold. Obtaining the degree prediction loss value includes: extracting features from the customer product interaction graph to obtain a customer product adjacency matrix; calculating the degree of the customer product adjacency matrix to obtain a label degree value; calculating the degree of the customer product splicing features output by the encoder using the feedforward neural network to obtain an initial degree value; performing dimensionality transformation on the initial degree value using a preset compression function to obtain a training degree value; and calculating the loss based on the training degree value and the label degree value to obtain the degree prediction loss value.

[0082] As an example, the node degree prediction model aims to encourage the autoencoder to learn better local structural information. During the training process of generating the node degree prediction model, the parameters of the autoencoder are continuously optimized by lowering the joint loss function, causing both the autoencoder and the graph neural network to tend to learn local information. This allows the autoencoder to learn better local information, which is then passed to the graph neural network, enabling the graph convolutional network to also learn local information more effectively.

[0083] As an example, firstly, feature extraction is performed on the customer product interaction graph to obtain the customer product adjacency matrix; then, the degree of the customer product adjacency matrix is ​​calculated to obtain the node degree. This pre-calculated node degree can be used as the training label for the degree prediction model, i.e., the label degree value. The process of obtaining the label degree value (node ​​degree) is defined as follows:

[0084]

[0085]

[0086] Among them, the customer product interaction diagram corresponds to the customer product interaction matrix. ,in, Let represent the interaction relationship between the i-th customer and the j-th product, including the number of interactions between the customer and the product. Let A represent the customer-product adjacency matrix. If an element is non-zero, it indicates that the i-th customer has purchased the j-th financial product. Represents a node The degree is then determined. Then, the specific representation of the k-th layer autoencoder module is transformed using a feedforward neural network to obtain the initial degree value. The initial degree value is as follows:

[0087] ;

[0088] in, and This represents the trainable weight matrix and biases of the feedforward neural network. Since the initial degree values ​​differ from the pre-calculated label dimensions, a compression function is used to change the dimension of the transformed node degrees, resulting in the trained degree values:

[0089] ;

[0090] Here, squeeze represents a parameterless compression function, i.e. After compression, the matrix of training degree values ​​can be transformed into a corresponding vector, allowing for the calculation of the corresponding node degree prediction loss.

[0091] ;

[0092] like Figure 4 As shown, by optimizing the above loss function, local information representing graph structure information can be incorporated into the specific representation of the autoencoder. Then, the specific representation of the autoencoder needs to be passed to the graph convolutional network, so that the graph convolutional network also incorporates local structure information. Overall, it can be seen that the degree prediction loss not only incorporates local structure information into the autoencoder, but also allows the graph convolution to better incorporate local structure information, thus enabling it to pay attention to both global and local information at the same time, thereby improving the accuracy of product recommendation.

[0093] In this embodiment, the graph neural network is repeatedly trained from a graph convolutional network. The last layer of the graph convolutional network communicates with the feedforward neural network through the last layer of the encoder, and each layer of the graph convolutional network is connected to each layer of the encoder in a one-to-one correspondence. The process of obtaining the cross-entropy loss value includes: establishing a reinforced adjacency matrix based on the training degree value, the product feature matrix, and the customer feature matrix; calculating different powers of the reinforced adjacency matrix and normalizing it based on these powers to obtain a normalized matrix; calculating different powers of the normalized matrix and performing convolution processing based on these powers to obtain node attribute feature vectors; performing an adaptive summation operation based on the node attribute feature vectors to obtain the corresponding representation of the customer product; performing probability prediction on the corresponding representation of the customer product using a preset probability prediction function to obtain the training prediction probability; and continuing loss calculation based on the training prediction probability and the pre-acquired sample probability to obtain the cross-entropy loss value.

[0094] Based on this, the trained graph neural network can simultaneously focus on global and local features, thereby improving the overall attention of the large model to user and product information, and thus improving the accuracy of product recommendations.

[0095] As an example, to obtain better local structural information, an autoencoder-specific representation containing local structural information needs to be taken as input. This autoencoder-specific representation with local structural information is then passed to the corresponding layers of the graph convolutional network.

[0096] First, a reinforced adjacency matrix is ​​established based on the training degree values, product feature matrix, and customer feature matrix. Then, different powers of the reinforced adjacency matrix are calculated, and the adjacency matrix is ​​normalized.

[0097] ;

[0098] in, This indicates the adjacency matrix of customer product interactions. The normalized matrix, D, represents the degree matrix corresponding to the training degree values. The degree matrix has only non-zero diagonal elements, for example... Let I represent the total number of times the i-th customer has purchased financial products. Let I represent an identity matrix with all 1s on its diagonal.

[0099] After normalization, the result is exponentially calculated, such as: , , Then, a graph convolutional network is used for representation learning to obtain the node attribute feature vectors: ;in, This represents the attribute feature vector corresponding to the i-th customer or financial product at the l-th propagation layer. Indicates the first The feature matrix output by the layer autoencoder.

[0100] To fully utilize the node representations learned from different graph convolutional layers, an adaptive summation operation is used to process the features to obtain the representation corresponding to the customer product. The formula for obtaining this representation is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] in, This represents the final customer and product representation. The weights represent the features of different graph convolutional layers. Represents cosine similarity. With the initial propagation feature matrix The greater the similarity, the less smoothed the adjacency matrix is, and the more obvious the differences in representation between nodes, making it more suitable for distinguishing nodes. Therefore, in calculation... At that time, a probabilistic prediction function is used to obtain the node representation. Then, the node prediction probability is obtained through the probability prediction function, also known as the training prediction probability. For example, the Softmax function (normalization function) is shown in the following formula:

[0105] ;

[0106] Based on this, the loss is further calculated using the training prediction probability and the pre-acquired sample probability to obtain the cross-entropy loss value: .

[0107] The fusion loss is the combination of the three losses mentioned above:

[0108]

[0109] in, and These are the weighting coefficients that balance out these losses.

[0110] This allows the input of the trained graph neural network to include not only the user features and interaction features output by each layer of the trained autoencoder, but also the degree prediction values ​​output by the trained degree prediction model. Furthermore, an adaptive summation operation is performed based on the degree prediction values, thus focusing not only on the global features of users and related products, but also on local features, thereby improving information focus and enhancing the accuracy of product recommendations.

[0111] The above describes the complete training and usage process of a trained product recommendation model, including a trained autoencoder, a trained degree prediction model, and a trained graph neural network. When applying this trained graph neural network product recommendation model, after obtaining user feature representations that contain both global and local information through the trained graph neural network product recommendation model, operation S240 is executed. Operation S240 performs a weighted mapping based on the user feature representation to obtain an adaptive product score, and generates product recommendation information based on the adaptive product score. This includes: calculating the dot product vector by performing a dot product calculation on the product embedding vector of the target product and the user feature representation to obtain a dot product vector; numerically processing the dot product vector using a preset mapping function to obtain an initial value; proportionally amplifying the initial value to obtain the score of the target product; and sorting and corresponding operations on the scores of each target product to form product recommendation information. Thus, the step of proportionally amplifying the initial value is introduced, which makes it easier for the backend program to distinguish score intervals, thereby improving the processing efficiency and accuracy of generating product recommendation information.

[0112] As an example, similar to the training process, after the model is optimized using the loss function described above to obtain a trained product recommendation model, user feature representations can be obtained. If it's necessary to calculate whether to recommend product j to customer i, the embedding vector of product j and the embedding vector of customer i can be multiplied by a dot product and then processed through a mapping function, such as the sigmoid function, to obtain an initial value. In this example, the initial value is scaled up to obtain the score of the target product, with the score ranging from [0,1]. For example, it can then be determined whether the score is greater than a preset threshold of 0.5. If it is greater than 0.5, a recommendation is made; if it is less than 0.5, no recommendation is made. The higher the recommended score, the more suitable the financial product is for the customer, according to the model.

[0113]

[0114] in, This represents the sigmoid function (a type of S-shaped logic function). , and The vector above represents customer i and financial product j. For a given customer, all candidate products are scored. If the goal is to recommend the top ten products, then the top ten products with the highest scores are recommended to that customer.

[0115] As described above, the product recommendation method provided in this embodiment first extracts features from the acquired user-related information using a preset autoencoder to obtain user features and interaction features. These user features and interaction features are then concatenated to obtain a user-product attribute feature matrix. A degree prediction model then plays a crucial role, performing degree prediction processing based on the user-product attribute feature matrix to obtain degree prediction values ​​representing local information. Finally, a graph neural network performs graph convolution calculations based on the user features, interaction features, and degree prediction values ​​to obtain a user feature representation that focuses on both global and local information. This addresses the technical problem in existing technologies where graph neural networks only focus on global features and ignore local features, leading to poor recommendation accuracy. The method ensures that the obtained user feature representation not only focuses on global information but also on local information, improving information focus and thus enhancing product recommendation accuracy.

[0116] Based on the above product recommendation method, this application also provides a product recommendation device. The following will be combined with... Figure 5 The device is described in detail.

[0117] Figure 5 A schematic block diagram of a product recommendation device according to an embodiment of this application is shown.

[0118] like Figure 5 As shown, the product recommendation device 500 of this embodiment includes a feature extraction module 510, a degree prediction module 520, a feature fusion module 530, and a data generation module 540.

[0119] The feature extraction module 510 can perform operation S210, which is used to extract features from the acquired user-related information through a preset autoencoder to obtain user features and interaction features, and to concatenate the user features and interaction features to obtain a user product attribute feature matrix; the user-related information includes user information and interaction information related to user information.

[0120] The degree prediction module 520 can perform operation S220 to perform degree prediction processing based on the user product attribute feature matrix through a preset degree prediction model to obtain the degree prediction value.

[0121] The feature fusion module 530 can perform operation S230, which is used to perform graph convolution calculation based on user features, interaction features, and degree prediction values ​​through a preset graph neural network to obtain user feature representation.

[0122] The data generation module 540 can perform operation S240 to perform weighted mapping based on user feature representation to obtain product adaptive score, and generate product recommendation information based on product adaptive score.

[0123] Specifically, in this embodiment, the feature extraction module 510 applies activation function operations to user features and interaction features based on the preset encoding weight matrix and encoding bias coefficient in the autoencoder, so as to concatenate the user features and interaction features to form a user product attribute feature matrix.

[0124] The degree prediction module 520 performs matrix decomposition and degree calculation on the user product attribute feature matrix based on the prediction weight matrix and prediction bias coefficient of the degree prediction model to obtain the degree of the transmission node; and uses a preset compression function to perform dimensional transformation on the degree of the transmission node to obtain the degree prediction value.

[0125] The feature fusion module 530 includes: a matrix building unit, used to build a user-product adjacency matrix based on user features, interaction features, and degree prediction values; a vector calculation unit, used to normalize and exponentiate the user-product adjacency matrix to obtain attribute feature vectors of users or products in specific layers of the graph neural network; and an adaptive operation unit, used to perform adaptive summation operations based on attribute feature vectors in all layers of the graph neural network to obtain user representations and product representations; and uses user representations and product representations as user feature representations.

[0126] The aforementioned autoencoder, degree prediction model, and graph neural network are integrated into a pre-trained product recommendation model. The pre-training of the product recommendation model includes: capturing user and product data in a pre-defined data warehouse, and performing data preprocessing operations based on the user and product data to obtain a product feature matrix and a customer-product interaction graph; repeatedly training a pre-defined fusion architecture based on the product feature matrix and the customer-product interaction graph until the fusion loss value related to the fusion architecture reaches a pre-defined fusion threshold, then using the optimal fusion architecture as the product recommendation model; wherein the fusion architecture includes an encoder structure, a feedforward neural network, and a graph convolutional network; the fusion loss value is obtained by weighted summation of the reconstruction loss value related to the encoder structure, the degree prediction loss value related to the feedforward neural network, and the cross-entropy loss value related to the graph convolutional network.

[0127] The autoencoder is trained repeatedly by an encoder; the optimal encoder is selected as the autoencoder when the reconstruction loss value reaches a preset reconstruction threshold. Obtaining the reconstruction loss value includes: performing dimensionality reduction on the product feature matrix and the customer feature matrix extracted from the customer-product interaction graph to obtain an initial matrix; constructing a customer-product mapping relationship based on the initial matrix to form a customer-product splicing matrix; decoding the customer-product splicing matrix to obtain a reconstruction matrix corresponding to the initial matrix; and calculating the loss value between the reconstruction matrix and the initial matrix using a preset reconstruction algorithm to obtain the reconstruction loss value.

[0128] The degree prediction model is trained repeatedly by a feedforward neural network, which communicates with the encoder. The optimal feedforward neural network is selected as the degree prediction model once the degree prediction loss value reaches a preset degree prediction threshold. Obtaining the degree prediction loss value includes: extracting features from the customer product interaction graph to obtain the customer product adjacency matrix; calculating the degree of the customer product adjacency matrix to obtain the label degree value; calculating the degree of the customer product splicing features output by the encoder using the feedforward neural network to obtain the initial degree value; performing dimensionality transformation on the initial degree value using a preset compression function to obtain the training degree value; and calculating the loss based on the training degree value and the label degree value to obtain the degree prediction loss value.

[0129] This graph neural network is repeatedly trained from a graph convolutional network. The last layer of the graph convolutional network communicates with the feedforward neural network through the last layer of the encoder, and each layer of the graph convolutional network is connected to each layer of the encoder in a one-to-one correspondence. The process of obtaining the cross-entropy loss value includes: establishing a reinforced adjacency matrix based on training degree values, product feature matrix, and customer feature matrix; calculating different powers of the reinforced adjacency matrix and normalizing it based on these powers to obtain a normalized matrix; calculating different powers of the normalized matrix and performing convolution processing based on these powers to obtain node attribute feature vectors; performing adaptive summation based on the node attribute feature vectors to obtain the corresponding representation of the customer product; performing probability prediction on the corresponding representation of the customer product using a preset probability prediction function to obtain the training prediction probability; and continuing loss calculation based on the training prediction probability and the pre-acquired sample probability to obtain the cross-entropy loss value.

[0130] The data generation module 540 includes: an embedding calculation unit, used to perform dot product calculation on the product embedding vector of the target product and the user feature representation to obtain a dot product vector; a numerical processing unit, used to perform numerical processing on the dot product vector through a preset mapping function to obtain an initial value; and to scale up the initial value to obtain the score of the target product; and an information generation unit, used to perform sorting and corresponding operations on the scores of each target product to form product recommendation information.

[0131] Furthermore, according to embodiments of this application, any multiple modules among the feature extraction module 510, degree prediction module 520, feature fusion module 530, and data generation module 540 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the feature extraction module 510, degree prediction module 520, feature fusion module 530, and data generation module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the feature extraction module 510, degree prediction module 520, feature fusion module 530, and data generation module 540 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0132] It should be noted that the implementation methods, technical problems solved, functions achieved, and technical effects of each module in the device embodiment are the same as or similar to the implementation methods, technical problems solved, functions achieved, and technical effects of each corresponding step in the method embodiment, and will not be repeated here.

[0133] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a product recommendation method according to an embodiment of this application.

[0134] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0135] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0136] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0137] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0138] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0139] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the product recommendation method provided in the embodiments of this application.

[0140] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0141] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0142] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0143] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0145] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A product recommendation method, characterized in that, The method includes: The acquired user-related information is processed by a preset autoencoder to extract user features and interaction features. The user features and interaction features are then concatenated to obtain a user product attribute feature matrix. The user-related information includes user information and interaction information related to the user information. The degree prediction value is obtained by performing degree prediction processing based on the user product attribute feature matrix using a preset degree prediction model. The user feature representation is obtained by performing graph convolution calculation based on the user features, the interaction features, and the degree prediction value using a preset graph neural network. A weighted mapping is performed based on the user feature representation to obtain a product adaptive score, and product recommendation information is generated based on the product adaptive score.

2. The product recommendation method according to claim 1, characterized in that, The acquired user-related information is processed by a pre-defined autoencoder to extract user features and interaction features. These user features and interaction features are then concatenated to obtain a user product attribute feature matrix, including: Based on the preset encoding weight matrix and encoding bias coefficient in the autoencoder, activation function operations are applied to the user features and the interaction features to concatenate the user features and the interaction features to form a user product attribute feature matrix.

3. The product recommendation method according to claim 1, characterized in that, The degree prediction value is obtained by performing degree prediction processing based on the user product attribute feature matrix using a preset degree prediction model, including: Based on the prediction weight matrix and prediction bias coefficient of the degree prediction model, matrix decomposition and degree calculation are performed on the user product attribute feature matrix to obtain the transmission node degree. The degree of the transmission node is transformed using a preset compression function to obtain the degree prediction value.

4. The product recommendation method according to claim 1, characterized in that, The user feature representation is obtained by performing graph convolution calculations based on the user features, the interaction features, and the degree prediction value using a preset graph neural network, including: A user-product adjacency matrix is ​​established based on the user characteristics, the interaction characteristics, and the degree prediction value. The user-product adjacency matrix is ​​normalized and exponentially calculated to obtain the attribute feature vector of the user or product in a specific layer of the graph neural network. An adaptive summation operation is performed on the attribute feature vectors in all layers of the graph neural network to obtain user representation and product representation. The user representation and product representation are used as user feature representations.

5. The product recommendation method according to claim 1, characterized in that, The autoencoder, the degree prediction model, and the graph neural network are integrated into a pre-trained product recommendation model; Pre-training the product recommendation model includes: Capture user data and product data in a pre-defined data warehouse, and perform data preprocessing operations based on the user data and product data to obtain a product feature matrix and a customer product interaction diagram; The preset fusion architecture is repeatedly trained based on the product feature matrix and the customer product interaction graph until the fusion loss value associated with the fusion architecture reaches a preset fusion threshold. The optimal fusion architecture is then used as the product recommendation model. The fusion architecture includes an encoder structure, a feedforward neural network, and a graph convolutional network. The fusion loss value is obtained by weighted summation of the reconstruction loss value associated with the encoder structure, the degree prediction loss value associated with the feedforward neural network, and the cross-entropy loss value associated with the graph convolutional network.

6. The product recommendation method according to claim 5, characterized in that, The autoencoder is trained repeatedly by the encoder; wherein, until the reconstruction loss value reaches a preset reconstruction threshold, the optimal encoder is selected as the autoencoder; wherein, obtaining the reconstruction loss value includes: The product feature matrix and the customer feature matrix extracted from the customer product interaction graph are subjected to dimensionality reduction processing to obtain an initial matrix; Based on the initial matrix, a customer-product mapping relationship is constructed to form a customer-product splicing matrix; Decode the customer product splicing matrix to obtain a reconstructed matrix corresponding to the initial matrix; A preset reconstruction algorithm is used to calculate the loss value between the reconstruction matrix and the initial matrix to obtain the reconstruction loss value.

7. The product recommendation method according to claim 6, characterized in that, The degree prediction model is repeatedly trained by the feedforward neural network, which is communicatively connected to the encoder. The optimal feedforward neural network is selected as the degree prediction model until the degree prediction loss value reaches a preset degree prediction threshold. Obtaining the degree prediction loss value includes: Feature extraction is performed on the customer product interaction graph to obtain the customer product adjacency matrix; the degree of the customer product adjacency matrix is ​​calculated to obtain the label degree value; The feedforward neural network is used to calculate the degree of the customer product splicing features output by the encoder to obtain an initial degree value. The initial degree value is dimensionally transformed using a preset compression function to obtain the training degree value, and loss is calculated based on the training degree value and the label degree value to obtain the degree prediction loss value.

8. The product recommendation method according to claim 7, characterized in that, The graph neural network is trained repeatedly by the graph convolutional network; the last layer of the graph convolutional network communicates with the feedforward neural network through the last layer of the encoder, and each layer of the graph convolutional network is connected to each layer of the encoder in a one-to-one correspondence; wherein, obtaining the cross-entropy loss value includes: An enhanced adjacency matrix is ​​established based on the training degree value, the product feature matrix, and the customer feature matrix; Calculate different powers of the enhanced adjacency matrix, and normalize the enhanced adjacency matrix based on the different powers of the enhanced adjacency matrix to obtain a normalized matrix; Calculate different powers of the normalized matrix and perform convolution processing based on the different powers of the normalized matrix to obtain node attribute feature vectors; An adaptive summation operation is performed based on the node attribute feature vectors to obtain the corresponding representation of the customer's product. The training prediction probability is obtained by performing probability prediction on the corresponding representation of the customer product through a preset probability prediction function. Based on the training prediction probability and the pre-acquired sample probability, further loss calculation is performed to obtain the cross-entropy loss value.

9. The product recommendation method according to claim 1, characterized in that, Weighted mapping is performed based on the user feature representation to obtain a product adaptive score, and product recommendation information is generated based on the product adaptive score, including: The product embedding vector of the target product is multiplied by the user feature representation to obtain the dot product vector. The dot product vector is numerically processed using a preset mapping function to obtain an initial value; The initial value is scaled up proportionally to obtain the score of the target product; The scores of each target product are sorted and corresponding operations are performed to generate product recommendation information.

10. A product recommendation device, characterized in that, The device includes: The feature extraction module is used to extract features from the acquired user-related information through a preset autoencoder to obtain user features and interaction features, and to concatenate the user features and interaction features to obtain a user product attribute feature matrix; the user-related information includes user information and interaction information related to the user information; The degree prediction module is used to perform degree prediction processing based on the user product attribute feature matrix using a preset degree prediction model to obtain the degree prediction value. The feature fusion module is used to perform graph convolution calculation based on the user features, the interaction features, and the degree prediction value through a preset graph neural network to obtain the user feature representation; The data generation module is used to perform weighted mapping based on the user feature representation to obtain a product adaptive score, and to generate product recommendation information based on the product adaptive score.

11. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

Images