WO2021189976A1

WO2021189976A1 - Product information pushing method and apparatus, device, and storage medium

Info

Publication number: WO2021189976A1
Application number: PCT/CN2020/136326
Authority: WO
Inventors: 王淦
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-03-25
Filing date: 2020-12-15
Publication date: 2021-09-30
Also published as: CN111553759A

Abstract

A product information pushing method and apparatus, a device, and a storage medium. The method comprises: obtaining user feature information from a target user terminal, the user feature information comprising numerical features and taxonomic features (S101); performing data preprocessing on the numerical features and the taxonomic features to construct a fusion vector (S102); processing the fusion vector on the basis of a deep neural network sub-model in a product recommendation model to obtain a deep feature vector, and processing the fusion vector on the basis of a cross neural network sub-model in the product recommendation model to obtain a cross feature vector (S103); constructing a fusion feature vector according to the deep feature vector and the cross feature vector, and predicting, according to the fusion feature vector, respective recommendation scores of several products to be recommended (S104); and sending, according to the recommendation scores of the products to be recommended, to the target terminal the products to be recommended (S105). More types of user attributes can be integrated to participate in decision-making, thus improving product recommendation accuracy.

Description

Method, device, equipment and storage medium for pushing product information

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office, the application number is CN202010220066.7, and the title is "a method, device, equipment and storage medium for pushing product information" on March 25, 2020, and its entire contents Incorporated in this application by reference.

Technical field

This application relates to the field of data analysis, and in particular to a method, device, equipment and storage medium for pushing product information.

Background technique

The traditional product recommendation system judges the product that the user will buy next by analyzing the user's product historical purchase and consulting situation by the salesperson. For manual product recommendation methods, a large amount of labor is required, which increases labor costs and at the same time brings a poor user experience. Due to the large number of users, manual recommendation methods can no longer meet the needs of current users, and intelligent product automatic recommendation solutions have also emerged.

Existing intelligent product automatic recommendation schemes generally determine the correlation between user attributes and product attributes through big data analysis, and then determine the products that users may be interested in based on user attributes, product attributes and their correlation. However, the types of user attributes that can participate in the decision-making are limited, and a lot of reference information will be missed in the recommendation decision-making process, so that the determined product may not be of real interest to the user, resulting in a low recommendation accuracy rate.

Therefore, how to improve the accuracy of product recommendation becomes an urgent problem to be solved.

Summary of the invention

This application provides a method for pushing product information, and the method includes:

Obtain user characteristic information from the target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics;

Perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

Process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and process the fusion vector based on the cross neural network sub-model in the product recommendation model to obtain a cross feature vector ；

Constructing a fusion feature vector according to the depth feature vector and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

According to the respective recommendation scores of the several products to be recommended, product information of at least one product to be recommended is sent to the target user terminal.

This application also provides a product information push device, the device includes:

A data acquisition module for acquiring user characteristic information from a target user terminal, the user characteristic information including numerical characteristics and sub-type characteristics;

A data processing module, configured to perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

The vector processing module is used to process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and perform processing on the fusion vector based on the cross neural network sub-model in the product recommendation model Processing to get the cross feature vector;

The score prediction module is configured to construct a fusion feature vector according to the depth feature vector and the cross feature vector, and predict the respective recommendation scores of several products to be recommended according to the fusion feature vector;

The product recommendation module is configured to send the product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the products to be recommended.

The application also provides a computer device, the computer device includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and realizes when the computer program is executed The following steps:

The present application also provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the following steps:

Description of the drawings

In order to explain the technical solutions of the embodiments of the present application more clearly, the following will briefly introduce the drawings used in the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present application. Ordinary technicians can obtain other drawings based on these drawings without creative work.

FIG. 1 is a schematic flowchart of a method for pushing product information according to an embodiment of the present application;

Figure 2 is a schematic diagram of a scenario of a product information push method applied to a server;

FIG. 3 is a schematic diagram of a sub-process of performing data preprocessing on the numerical feature and the sub-type feature in FIG. 1 to construct a fusion vector;

4 is a schematic block diagram of the structure of a product recommendation model provided by an embodiment of the present application;

FIG. 5 is a schematic flowchart of a product recommendation model training process provided by an embodiment of the present application;

FIG. 6 is a schematic block diagram of a product information pushing device provided by an embodiment of the present application;

FIG. 7 is a schematic block diagram of another product information pushing device provided by an embodiment of the present application;

FIG. 8 is a schematic block diagram of the structure of a computer device according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

The flowchart shown in the drawings is only an example, and does not necessarily include all contents and operations/steps, nor does it have to be executed in the described order. For example, some operations/steps can also be decomposed, combined or partially combined, so the actual execution order may be changed according to actual conditions.

It should be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in the specification of this application and the appended claims, unless the context clearly indicates other circumstances, the singular forms "a", "an" and "the" are intended to include plural forms.

It should also be understood that the term "and/or" used in the specification and appended claims of this application refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

The embodiments of the present application provide a method, device, computer equipment, and storage medium for pushing product information. The product information push method can be applied to a server, and is used for processes such as pushing suitable product information for the user according to the user characteristic information of the user, so as to improve the accuracy of product recommendation.

For example, the product information push method is used for the server; the server can be an independent server or a server cluster. The following embodiments will introduce in detail the product information push method applied to the server.

Hereinafter, some embodiments of the present application will be described in detail with reference to the accompanying drawings. In the case of no conflict, the following embodiments and features in the embodiments can be combined with each other.

Please refer to FIG. 1. FIG. 1 is a schematic flowchart of a method for pushing product information according to an embodiment of the present application.

Figure 2 is a schematic diagram of a scenario of a product information push method applied to a server.

As shown in Figure 1 and Figure 2, the product information push method specifically includes: Step S101 to Step S105.

Step S101: Obtain user characteristic information from a target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics.

Exemplarily, the user characteristic information may include the user's vehicle information. In some embodiments, the user can input vehicle information on the terminal he uses, and then the terminal sends the vehicle information to the server.

Exemplarily, if the user clicks the product recommendation button on the terminal, the vehicle information input interface can be displayed on the terminal; the user can input vehicle information on the input interface, and then the terminal sends the vehicle information to the server.

In some embodiments, the vehicle information may include purchase price, number of seats, exhaust steam volume, agency code, broker code, agent code, car dealer code, vehicle family, brand, nature, vehicle type, new energy vehicle logo , Gender of the owner of the driving license, etc.

Among them, the purchase price, the number of seats, and the amount of exhaust steam are quantitative values, so vehicle information such as the purchase price, number of seats, and amount of exhaust steam can be determined as numerical features. Institution code, broker code, agent code, car dealer code, car family, brand, nature of belonging, vehicle type, new energy vehicle logo, driving license owner’s gender and other unquantifiable features, such as numbers and/or letter codes Therefore, these features can be determined as sub-type features.

Step S102: Perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector.

Since the numerical feature and the sub-type feature are different data types, different preprocessing methods need to be used to process the numerical feature and the sub-type feature respectively.

In some embodiments, referring to FIG. 3, the data preprocessing of the numerical feature and the sub-category feature to construct a fusion vector specifically includes: step S1021 to step S1023.

Step S1021: Perform standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature.

In some embodiments, the numerical features are normalized by a max-min normalization method.

Max-min normalization is also called dispersion standardization, which is a linear transformation of the original data, so that the result value is mapped to [0, 1]. The conversion function is as follows: x'=(x-min)÷(max-min). Wherein, max represents the maximum value of the numeric feature, min represents the minimum value of the numeric feature, x represents the numeric value of the numeric feature in the vehicle information obtained from the target user terminal, and x'represents the numeric data corresponding to the numeric feature.

For example, if the maximum value of the purchase price of a preset vehicle is 1.1 million RMB and the minimum value is 100,000 RMB, if the purchase price in the vehicle information obtained from the target user terminal is 300,000 RMB, the purchase price will be standardized. The numerical data obtained is 0.2.

In other embodiments, the numerical characteristics are standardized so that the numerical data corresponding to the numerical characteristics satisfy a normal distribution.

Exemplarily, the conversion function of the standardization process can be expressed as: x'=(x-μ)÷σ, where μ represents the mean value of the numerical feature sample data, σ represents the standard deviation of the numerical feature sample data, and x represents from The numerical value of the numerical feature in the vehicle information acquired by the target user terminal, and x'represents the numerical data corresponding to the numerical feature.

By standardizing numerical features, it is possible to avoid the problem of poor accuracy of product recommendation due to large differences in the values of different numerical features, and the problem of poor accuracy of product recommendations when the numerical features are small; the numerical features with large differences will be treated by standardization. The value of is compressed to a smaller range, that is, the difference between the values of different numeric features can be maintained, and the difference between the values of different numeric features can also be reduced, so that the product information push method can be maintained in more scenarios Sufficient accuracy.

Exemplarily, the purchase price, the number of seats, and the exhaust steam volume in the vehicle information are standardized, and the corresponding numerical data obtained are D1, D2, and D3, respectively.

Step S1022: Perform vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature.

In some embodiments, the server pre-stores a mapping table of the sub-type characteristics and the sub-type vectors, and the sub-type vectors corresponding to the sub-type characteristics can be obtained by looking up the table.

In other embodiments, the sub-type feature is vectorized based on the vector embedding layer in the product recommendation model to obtain the sub-type vector corresponding to the sub-type feature. Through the product recommendation model trained by the model, the vector embedding layer learns the correspondence between the sub-type features and the sub-type vectors in the training samples, so that the corresponding sub-type vectors can be determined according to the sub-type features.

Exemplarily, the agency code, broker code, agent code, car dealer code, and car series in the vehicle information are respectively vectorized, and the sub-type vectors E1, E2, E3, E4, and E5 are obtained accordingly.

Step S1023: Construct a fusion vector according to the numerical data and the classification vector.

In some embodiments, the numerical data and the sub-type vector are spliced end to end as the input vector of the product recommendation model. Exemplarily, the fusion vector is represented as D1, D2, D3, E1, E2, E3, E4, E5.

The fusion vector constructed according to the numerical data and the sub-type vector, because the user's numerical characteristics and relevant information of the sub-type characteristics are retained at the same time, such as purchase price, number of seats, exhaust steam volume, agency code, broker code , Agent code, car dealer code, car family, brand, nature, vehicle type, new energy vehicle logo, driving license owner’s gender, etc., can better integrate more types of user attributes to participate in decision-making, so as to obtain more Accurate portraits of user interests. Recommending products for users based on user interest portraits can improve the accuracy of product recommendations. In some embodiments, before constructing a fusion vector based on the numerical data and the sub-type vector, the method further includes the following steps:

The product category information is obtained from the target user terminal, and the product category information is vectorized to obtain the product category vector corresponding to the product category information.

Exemplarily, the user also enters the category of the product that needs to be recommended in the target user terminal, so that the server can obtain product category information according to the user's input, such as auto insurance, which can be life insurance, auto insurance, etc.

In some embodiments, the performing vectorization processing on the product category information to obtain the product category vector corresponding to the product category information specifically includes the following steps: obtaining the difference between the preset product category information and the product category vector Mapping relationship data; performing vectorization processing on the product category information according to the mapping relationship data to obtain a product category vector corresponding to the product category information. The server pre-stores a mapping table of product category information and product category vectors, and the product category vector corresponding to the product category information can be obtained by looking up the table.

In other embodiments, the performing vectorization processing on the product category information to obtain the product category vector corresponding to the product category information specifically includes the following steps: performing vectorization processing on the product based on the vector embedding layer in the product recommendation model The category information is vectorized to obtain the corresponding product category vector. Through the product recommendation model trained by the model, the vector embedding layer learns the corresponding relationship between the product category information and the product category vector in the training sample, so that the corresponding product category vector can be determined according to the product category information.

In some embodiments, the constructing a fusion vector according to the numerical data and the classification vector specifically includes the following steps:

According to the numerical data, the classification vector, and the product category vector, a fusion vector is constructed.

Exemplarily, the product category information obtained from the target user terminal is auto insurance, and the product category vector corresponding to the auto insurance is F1, and the numerical data, the sub-type vector, and the product category vector are spliced head to tail to construct a fusion vector, that is, as The input vector of the product recommendation model, for example, the fusion vector is represented as D1, D2, D3, E1, E2, E3, E4, E5, F1.

According to the numerical data, the sub-type vector, and the product category vector, the fusion vector constructed, because in addition to retaining the user’s numerical characteristics and sub-type characteristics related information, it also retains product category information, which can be used in the user On the basis of the interest profile, according to the corresponding category characteristics of the product, further analyze the products that users may be interested in, so as to improve the accuracy of product recommendation.

In some embodiments, before constructing a fusion vector based on the numerical data and the sub-type vector, the method further includes the following steps:

Obtain the historical product purchase record corresponding to the target user terminal, perform self-attention mechanism processing on the historical product purchase record, and obtain the purchase record vector corresponding to the historical product purchase record.

Exemplarily, according to the user information of the target user terminal, the product that the user has purchased is obtained from the customer database, and the historical product purchase record corresponding to the user is generated.

If the target user is a renewal user, and the target user has purchased more auto insurance types in history, it is also more likely to continue to purchase the auto insurance types next time. In the recommendation of auto insurance types for renewal users, the importance of each type of insurance can be evaluated based on historical product purchase records. Therefore, for renewal users, according to historical auto insurance purchase records, the attention (self-attention) mechanism is used to rank the importance of insurance types. If the user has purchased more types of auto insurance in the history, the type of insurance will get a higher score. At the same time, the result processed by the attention mechanism can be used as the input of the product recommendation model, so that the prediction of the model can be optimized by using the attention mechanism.

According to the numerical data, the classification vector, the product category vector and/or the purchase record vector, a fusion vector is constructed.

Exemplarily, the product category information obtained from the target user terminal is auto insurance, the product category vector corresponding to auto insurance is F1, the purchase record vector corresponding to the historical product purchase record is G1, and the numerical data, the category vector, and The product category vector and/or the purchase record vector are spliced from the beginning to the end as the input vector of the product recommendation model. For example, the fusion vector can be expressed as D1, D2, D3, E1, E2, E3, E4, E5, F1, G1, Or D1, D2, D3, E1, E2, E3, E4, E5, G1.

Step S103: Process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and process the fusion vector based on the cross neural network sub-model in the product recommendation model to obtain Cross feature vector.

In some embodiments, as shown in FIG. 4, the product recommendation model includes a deep neural network sub-model deep network and a cross neural network sub-model cross network. Among them, the core idea of crossover network is to apply explicit feature crossover in an effective way. The crossover network is composed of crossover layers, and each crossover layer has the following formula:

Among them, X _l and X _l+1 are column vectors, respectively representing the output of the cross layer from the lth layer and the (l+1)th layer, X ₀ is the preset vector, and W _l and b _l are the lth layer. The connection parameter between the two cross layers of the layer and the (l+1)th layer, f() represents feature crossing, which is used to fit the residual of the output of this layer and the output of the previous layer, namely X _l The residual of ₊₁ -X _l.

Due to the small number of parameters of the crossover network, the capabilities of the model are limited. In order to capture high-order nonlinear intersections, a deep network is introduced in parallel in the product recommendation model. The deep network is a fully connected feedforward neural network, and each deep layer has the following formula:

h _l+1 = f(W _l H _l + b _l )

Among them, h _l and h _l+1 are the output of the hidden layer of the lth layer and the (l+1)th layer, respectively, and W _l and b _l are the two layers of the lth layer and the (l+1)th layer. The connection parameter between the deep layers, f() represents the activation function—ReLU function.

The product recommendation model is obtained by jointly training the cross network cross network and the deep neural network deep network, which not only retains the ability of the deep neural network deep network to capture complex feature combinations, but also in the cross network cross network, each layer has feature cross feature crossing, can learn cross features, and does not require manual feature engineering. Therefore, the recommended model of this product has a small memory, is very efficient when learning features of a specific order combination, and the additional complexity introduced is also minimal.

Step S104: Construct a fusion feature vector according to the depth feature vector and the cross feature vector, and predict the respective recommendation scores of several products to be recommended based on the fusion feature vector.

Exemplarily, the cross feature vector output by the cross neural network sub-model in the product recommendation model is X _l+1 , the depth feature vector output by the deep neural network sub-model is h _l+1 , and the depth in the product recommendation model The outputs of the neural network sub-model deep network and the cross neural network sub-model cross network are spliced to obtain a fusion feature vector, which can be expressed as X _l+1 , h _l+1 , for example.

In some embodiments, the product recommendation model further includes a logistic regression sub-model, and the logistic regression sub-model includes a linear network layer and a sigmoid activation layer.

The fusion feature vector is input into the logistic regression sub-model, and the linear network layer processes the fusion feature vector to obtain an output vector containing several elements. Exemplarily, the several elements have a one-to-one correspondence with several products to be recommended, wherein the several elements are in a proportional relationship with the recommended scores of the products to be recommended, that is, the larger the element corresponding to a product to be recommended, the larger the product to be recommended The higher the recommended score for the product.

Exemplarily, the output vector of the linear network layer is passed through the sigmoid activation function to calculate the recommended score corresponding to each product to be recommended, so that the recommended score can be processed as a value between 0 and 1.

Step S105: Send product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the plurality of products to be recommended.

Exemplarily, the product to be recommended with the largest recommended score may be determined as the target product, and then product information of the target product, such as the auto insurance agreement, price, etc., is obtained, and then the product information is sent to the target user terminal.

In some embodiments, the sending product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the products to be recommended includes the following steps:

If the recommendation score of the product to be recommended is greater than the preset threshold, the product to be recommended is sent to the target user terminal.

The lowest recommended score can be preset, such as 0.75, if the recommended score of the product to be recommended is greater than 0.75, the product information of the product to be recommended is sent to the target user terminal; if there is no product to be recommended If the recommended score is greater than 0.75, the product information corresponding to the product to be recommended with the largest recommended score is sent to the target user terminal.

In some embodiments, please refer to FIG. 5, the product information pushing method further includes a training process of a product recommendation model, for example, it may include step 201 to step S206.

It is understandable that the training process of the product recommendation model can be implemented by a terminal or a server; the product recommendation model obtained after the training can be deployed on the terminal or server used to implement the product information pushing method of the foregoing embodiment.

Step 201: Obtain training sample data. The training sample data includes product information of several users and products of interest, and the product information includes numerical features and sub-type features.

The product information includes the user's vehicle information, and the vehicle information includes purchase price, number of seats, exhaust steam volume, agency code, broker code, agent code, car dealer code, car series, brand, nature of belonging, and vehicle type , New energy vehicle logo, gender of owner of driving license, etc. Among them, the purchase price, the number of seats, and the amount of exhaust steam are quantitative values, so vehicle information such as the purchase price, number of seats, and exhaust amount can be determined as numerical features. Institution code, broker code, agent code, car dealer code, car family, brand, nature of belonging, vehicle type, new energy vehicle logo, driving license owner’s gender and other unquantifiable features, such as numbers and/or letter codes Therefore, these features can be determined as sub-type features.

Step 202: Perform standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature, and perform vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature.

In some embodiments, the numerical features are normalized by a max-min normalization method. The reason is that if the data difference between multiple features is large, the convergence speed will be very slow. In the model involving distance calculation, if the numerical difference between multiple features is large, then the effect of the feature with a small value on the distance will be Very small, it will affect the accuracy of the model.

The sub-type features are randomly initialized to obtain the corresponding sub-type vector, and the sub-type vector is updated during the model training process. Exemplarily, L2 regularization is used to process sub-type features, so that model complexity and instability are reduced in the process of model training and learning, so as to avoid the risk of overfitting.

Exemplarily, for each auto insurance type, a corresponding auto insurance type feature vector is initialized, and the sub-type variable is used to express the auto insurance type information, and the initial dimension of the feature of the sub-type variable is 10 dimensions. At the same time, in the process of model training, the eigenvector values of the sub-types are updated, and the initial eigenvalues are between 0 and 0.001. When the eigenvalue is between 0 and 0.001, the model training has the fastest convergence speed.

Exemplarily, the purchase price, the number of seats, and the exhaust steam volume in the vehicle information of a certain user in the training sample data are standardized, and the numerical data are correspondingly obtained as D1, D2, and D3, and the user’s vehicle information is The agency code, broker code, agent code, car dealer code, and car series are respectively vectorized, and the sub-type vectors E1, E2, E3, E4, and E5 are obtained accordingly.

Step 203: Construct a fusion vector according to the user's numerical data and the classification vector.

Step 204: Input the fusion vector of the user into the deep neural network sub-model and the cross neural network sub-model in the product recommendation model, respectively, to obtain a deep feature vector and a cross feature vector.

Exemplarily, after the user's fusion vector is input to the cross neural network sub-model in the product recommendation model, the output cross feature vector is X _l+1 ; the user's fusion vector is input to the product After recommending the deep neural network sub-model in the model, the output depth feature vector is h _l+1 .

Step 205: Construct a fusion feature vector according to the depth feature vector of the user and the cross feature vector, and predict the respective recommendation scores of several products to be recommended based on the fusion feature vector.

In some embodiments, the product recommendation model further includes a logistic regression sub-model, wherein the logistic regression sub-model includes a linear network layer and a sigmoid activation layer. The fusion feature vector is input into the logistic regression sub-model, and the linear network layer processes the fusion feature vector to obtain an output vector containing several elements. The output vector of the linear network layer is passed through the sigmoid activation function to calculate the recommended score corresponding to each product to be recommended, so that the recommended score can be processed as a value between 0 and 1.

Step 206: Calculate a loss value according to the respective recommendation scores of the plurality of products to be recommended corresponding to the user and the product of interest of the user, and adjust the model parameters of the product recommendation model according to the loss value.

In some embodiments, the acquired data set is divided into a training data set, a verification data set, and a test data set. The training data set is used for model training, the verification data set is used for model selection, and the test data set is used for model testing.

In some embodiments, the training problem can be treated as a dual problem or a dual problem. The non-dual problem is that in the constructed positive and negative sample pairs, if the user purchases some types of insurance, the sample of the insurance is a positive sample, then We label this type of insurance as 1, and if some insurance types are not purchased, and the sample of this type of insurance is a negative sample, then we will label these types of insurance as 0. In the dual problem, each sample pair corresponds to a positive sample and a negative sample. As a sorting problem, we make the score of the positive sample higher than the negative sample in the training process of the model to achieve the sorting effect of the model.

In some embodiments, during the model training process, the deep neural network is first calculated. The number of deep neural network layers can be adjusted as needed. For example, a two-layer linear neural network is selected, and each neural network layer corresponds to a weight and a bias. Bias, during model training, the weight and Bias values are updated. The dimension of the hidden layer can be adjusted as needed, such as setting the dimension of the hidden layer to 50. The final input dimension of the model is 100. In order to achieve the desired effect of the model, the output layer dimension of the middle layer is set to 50, and the input layer dimension is set to 10, and the output layer dimension can be adjusted as needed. Through the deep neural network, the model can learn feature information in the deep sense.

In some embodiments, in the deep neural network, the Batch Normalization mechanism can be selectively used. During the training of the neural network, as the depth deepens, the input value distribution will shift and approach the upper and lower ends of the value range. In order to solve this problem, Batch Normalization uses certain normalization methods to force the distribution of the input values of each layer of neural network back to the standard normal distribution with a mean of 0 and a variance of 1, making the distribution back to a nonlinear function more sensitive to the input The loss function can be changed greatly, such as the gradient becomes larger, so as to avoid the problem of the disappearance of the gradient. At the same time, the larger the gradient can speed up the convergence speed of the model and increase the training speed.

In some embodiments, in the model training process, in order to learn the cross information between the features, the cross neural network mechanism is used. The number of cross neural network layers can be adjusted as needed. The implementation mechanism of the cross neural network is based on the previous one each time. The result of the layer cross neural network is multiplied by a matrix of the same dimension, and the output result of the previous layer cross neural network and a corresponding offset are added at the same time.

Exemplarily, an optimizer is selected for model training, where the optimizer can be selected, such as the sgd optimizer, and the learning rate of the optimizer can be adjusted according to actual needs to improve the model training effect. The sgd optimizer has the characteristics of fast training speed and not easy to fall into the local optimal solution.

Exemplarily, a dynamic learning rate scheme can be used in the product recommendation model. For example, for every 100 training iterations of the model, the learning rate is attenuated by 10 times. The purpose of setting the dynamic learning rate is to make the model more accessible to the optimal solution.

The product information push method provided by the foregoing embodiment obtains user characteristic information from a target terminal, and the user characteristic information includes numerical characteristics and sub-type characteristics. The numerical characteristics and sub-type characteristics are processed separately to obtain a fusion vector. The deep neural network sub-model and the cross neural network sub-model in the pre-trained product recommendation model process the fusion vector to obtain the deep feature vector and the cross feature vector, and construct the fusion feature vector according to the deep feature vector and the cross feature vector. The fusion feature vector predicts the respective recommendation scores of several products to be recommended, and finally, according to the respective recommendation scores of the recommended products, sends the product information of at least one product to be recommended to the target user terminal, which can integrate more types User attributes participate in decision-making, thereby improving the accuracy of product recommendations.

Please refer to FIG. 6. FIG. 6 is a schematic block diagram of a product information pushing device provided by an embodiment of the present application, and the product information pushing device is used to execute the aforementioned product information pushing method. Wherein, the product information pushing device can be configured in a server or a terminal.

Among them, the server can be an independent server or a server cluster. The terminal can be an electronic device such as a mobile phone, a tablet computer, a notebook computer, a desktop computer, a personal digital assistant, and a wearable device.

As shown in FIG. 6, the product information pushing device 300 includes: a data acquisition module 301, a data processing module 302, a vector processing module 303, a score prediction module 304, and a product recommendation module 305.

The data acquisition module 301 is configured to acquire user characteristic information from a target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics.

The data processing module 302 is configured to perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector.

Among them, as shown in FIG. 7, the data processing module 302 includes a standardized processing sub-module 3021, a vectorization processing sub-module 3022 and a data fusion sub-module 3023.

Specifically, the standardization processing sub-module 3021 is used to perform standardization processing on the numerical features to obtain the numerical data corresponding to the numerical features; the vectorization processing sub-module 3022 is used to vectorize the sub-type features Through processing, the sub-type vector corresponding to the sub-type feature is obtained; the data fusion sub-module 3023 is configured to construct a fusion vector according to the numerical data and the sub-type vector.

The vector processing module 303 is configured to process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and perform processing on the fusion vector based on the cross neural network sub-model in the product recommendation model Process to get the cross feature vector.

The score prediction module 304 is configured to construct a fusion feature vector according to the depth feature vector and the cross feature vector, and predict the respective recommendation scores of several products to be recommended according to the fusion feature vector.

The product recommendation module 305 is configured to send product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the products to be recommended.

Among them, as shown in FIG. 7, the product recommendation module 305 includes a product judgment sub-module 3051 and a product push sub-module 3052.

Specifically, the product judgment sub-module 3051 is used to judge whether the recommended score of the product to be recommended is greater than a preset threshold; the product push sub-module 3052 is used to send the product to be recommended with the recommended score greater than the preset threshold to the target User terminal.

The product information push device 300 includes a model optimization module for obtaining training sample data, the training sample data includes product information of a number of users and products of interest, and the product information includes numerical features and scores. Type feature; standardize the numeric feature to obtain the numeric data corresponding to the numeric feature, and vectorize the sub-category feature to obtain the sub-category vector corresponding to the sub-category feature; The user’s numerical data and classification vectors are used to construct a fusion vector; the user’s fusion vector is input into the deep neural network sub-model and the cross neural network sub-model in the product recommendation model, respectively, to obtain the deep feature vector and cross Feature vector; construct a fusion feature vector according to the depth feature vector of the user and the cross feature vector, and predict the respective recommendation scores of several products to be recommended according to the fusion feature vector; according to the user’s corresponding The respective recommendation scores of a number of products to be recommended and the loss value of the user's product of interest are calculated, and the model parameters of the product recommendation model are adjusted according to the loss value.

It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the product information pushing device and each module described above can refer to the corresponding product information push method embodiment in the foregoing The process will not be repeated here.

The above-mentioned product information pushing device may be implemented in the form of a computer program, and the computer program may be run on a computer device as shown in FIG. 8.

Please refer to FIG. 8. FIG. 8 is a schematic block diagram of a structure of a computer device provided by an embodiment of the present application. The computer equipment can be a server or a terminal.

Referring to FIG. 8, the computer device includes a processor, a memory, and a network interface connected through a system bus, where the memory may be volatile or non-volatile.

The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any method for pushing product information.

The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any product information push method.

The network interface is used for network communication, such as sending assigned tasks. Those skilled in the art can understand that the structure shown in FIG. 8 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may Including more or fewer parts than shown in the figure, or combining some parts, or having a different arrangement of parts.

It should be understood that the processor may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), and application specific integrated circuits (Application Specific Integrated Circuits). Circuit, ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor.

Wherein, in an embodiment, the processor is used to run a computer program stored in a memory to implement the following steps:

In one embodiment, when the processor implements the data preprocessing of the numerical feature and the sub-category feature to construct a fusion vector, the processor is used to implement:

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature;

Performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

A fusion vector is constructed according to the numerical data and the classification vector.

In an embodiment, the processor is further configured to implement:

Obtaining product category information from the target user terminal, and performing vectorization processing on the product category information to obtain a product category vector corresponding to the product category information;

The constructing a fusion vector according to the numerical data and the classification vector includes:

In an embodiment, when the processor implements the vectorization processing on the product category information to obtain the product category vector corresponding to the product category information, the processor is configured to implement:

Obtain the mapping relationship data between the preset product category information and the product category vector;

According to the mapping relationship data, vectorization processing is performed on the product category information to obtain a product category vector corresponding to the product category information.

In an embodiment, before the processor implements the construction of a fusion vector based on the numerical data and the classification vector, the processor is configured to implement:

Acquiring a historical product purchase record corresponding to the target user terminal, and performing self-attention mechanism processing on the historical product purchase record to obtain a purchase record vector corresponding to the historical product purchase record;

In one embodiment, the processor is used to realize that when the processor realizes that the product information of at least one product to be recommended is sent to the target user terminal according to the respective recommendation scores of the several products to be recommended :

In an embodiment, when the processor implements the product information pushing method, it is further configured to implement:

Acquiring training sample data, where the training sample data includes product information of several users and products of interest, and the product information includes numerical features and sub-type features;

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature, and performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

Constructing a fusion vector according to the user's numerical data and the classification vector;

Inputting the fusion vector of the user into a deep neural network sub-model and a cross neural network sub-model in the product recommendation model, respectively, to obtain a deep feature vector and a cross feature vector;

Constructing a fusion feature vector according to the depth feature vector of the user and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

A loss value is calculated according to the respective recommendation scores of the plurality of products to be recommended corresponding to the user and the product of interest of the user, and the model parameters of the product recommendation model are adjusted according to the loss value.

The embodiments of the present application also provide a computer-readable storage medium, and the computer-readable storage medium may be volatile or non-volatile. The computer-readable storage medium stores a computer program, the computer program includes program instructions, and the processor executes the program instructions to implement any product information push method provided in the embodiments of the present application.

The computer-readable storage medium may be the internal storage unit of the computer device described in the foregoing embodiment, for example, the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a smart memory card (Smart Media Card, SMC), and a Secure Digital (SD) ) Card, Flash Card, etc.

The product information pushing device, storage medium, and computer equipment provided in the foregoing embodiments obtain user characteristic information from the target terminal, and the user characteristic information includes numerical characteristics and sub-type characteristics, and the numerical characteristics and sub-type characteristics are processed separately, To obtain the fusion vector, use the deep neural network sub-model and the cross neural network sub-model in the pre-trained product recommendation model to process the fusion vector to obtain the deep feature vector and the cross feature vector, which are constructed according to the deep feature vector and the cross feature vector Fusion feature vectors, thereby predicting the respective recommendation scores of several products to be recommended according to the fusion feature vector, and finally sending the product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the recommended products, More types of user attributes can be integrated to participate in decision-making, thereby improving the accuracy of product recommendations.

The above are only specific implementations of this application, but the protection scope of this application is not limited to this. Anyone familiar with the technical field can easily think of various equivalents within the technical scope disclosed in this application. Modifications or replacements, these modifications or replacements shall be covered within the protection scope of this application. Therefore, the protection scope of this application shall be subject to the protection scope of the claims.

Claims

A product information push method, which includes:

Obtain user characteristic information from the target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics;

Perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

Process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and process the fusion vector based on the cross neural network sub-model in the product recommendation model to obtain a cross feature vector ；

Constructing a fusion feature vector according to the depth feature vector and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

According to the respective recommendation scores of the several products to be recommended, product information of at least one product to be recommended is sent to the target user terminal.
5. The product information push method according to claim 1, wherein said performing data preprocessing on said numerical feature and said sub-category feature to construct a fusion vector comprises:

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature;

Performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

A fusion vector is constructed according to the numerical data and the classification vector.
5. The product information push method according to claim 1, wherein before said constructing a fusion vector according to said numerical data and said classification vector, the method further comprises:

Obtaining product category information from the target user terminal, and performing vectorization processing on the product category information to obtain a product category vector corresponding to the product category information;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, and the product category vector, a fusion vector is constructed.
The method for pushing product information according to claim 3, wherein said performing vectorization processing on said product category information to obtain a product category vector corresponding to said product category information comprises:

Obtain the mapping relationship data between the preset product category information and the product category vector;

According to the mapping relationship data, vectorization processing is performed on the product category information to obtain a product category vector corresponding to the product category information.
The product information push method according to any one of claims 1 to 4, wherein before said constructing a fusion vector according to said numerical data and said classification vector, it further comprises:

Acquiring a historical product purchase record corresponding to the target user terminal, and performing self-attention mechanism processing on the historical product purchase record to obtain a purchase record vector corresponding to the historical product purchase record;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, the product category vector and/or the purchase record vector, a fusion vector is constructed.
5. The product information push method according to claim 1, wherein said sending product information of at least one product to be recommended to said target user terminal according to respective recommendation scores of said plurality of products to be recommended comprises:

If the recommendation score of the product to be recommended is greater than the preset threshold, the product to be recommended is sent to the target user terminal.
The method for pushing product information according to claim 1, further comprising:

Acquiring training sample data, where the training sample data includes product information of several users and products of interest, and the product information includes numerical features and sub-type features;

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature, and performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

Constructing a fusion vector according to the user's numerical data and the classification vector;

Inputting the fusion vector of the user into a deep neural network sub-model and a cross neural network sub-model in the product recommendation model, respectively, to obtain a deep feature vector and a cross feature vector;

Constructing a fusion feature vector according to the depth feature vector of the user and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

A loss value is calculated according to the respective recommendation scores of the plurality of products to be recommended corresponding to the user and the product of interest of the user, and the model parameters of the product recommendation model are adjusted according to the loss value.
A product information push device, which includes:

A data acquisition module for acquiring user characteristic information from a target user terminal, the user characteristic information including numerical characteristics and sub-type characteristics;

A data processing module, configured to perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

The vector processing module is used to process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and perform processing on the fusion vector based on the cross neural network sub-model in the product recommendation model Processing to get the cross feature vector;

The score prediction module is configured to construct a fusion feature vector according to the depth feature vector and the cross feature vector, and predict the respective recommendation scores of several products to be recommended according to the fusion feature vector;

The product recommendation module is configured to send the product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the products to be recommended.
A computer device, wherein the computer device includes a memory and a processor;

The memory is used to store a computer program;

The processor is configured to execute the computer program and implement the following steps when the computer program is executed:

Obtain user characteristic information from the target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics;

Perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

Process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and process the fusion vector based on the cross neural network sub-model in the product recommendation model to obtain a cross feature vector ；

Constructing a fusion feature vector according to the depth feature vector and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

According to the respective recommendation scores of the several products to be recommended, product information of at least one product to be recommended is sent to the target user terminal.
9. The computer device according to claim 9, wherein said performing data preprocessing on said numerical feature and said sub-category feature to construct a fusion vector comprises:

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature;

Performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

A fusion vector is constructed according to the numerical data and the classification vector.
9. The computer device according to claim 9, wherein the following steps are further implemented when the computer program is executed:

Obtaining product category information from the target user terminal, and performing vectorization processing on the product category information to obtain a product category vector corresponding to the product category information;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, and the product category vector, a fusion vector is constructed.
11. The computer device according to claim 11, wherein said performing vectorization processing on said product category information to obtain a product category vector corresponding to said product category information comprises:

Obtain the mapping relationship data between the preset product category information and the product category vector;

According to the mapping relationship data, vectorization processing is performed on the product category information to obtain a product category vector corresponding to the product category information.
The computer device according to any one of claims 9 to 12, wherein the following steps are further implemented when the computer program is executed:

Acquiring historical product purchase records corresponding to the target user terminal, and performing self-attention mechanism processing on the historical product purchase records to obtain a purchase record vector corresponding to the historical product purchase record;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, the product category vector and/or the purchase record vector, a fusion vector is constructed.
9. The computer device according to claim 9, wherein the sending product information of at least one product to be recommended to the target user terminal according to the respective recommendation scores of the products to be recommended comprises:

If the recommendation score of the product to be recommended is greater than the preset threshold, the product to be recommended is sent to the target user terminal.
9. The computer device according to claim 9, wherein the following steps are further implemented when the computer program is executed:

Acquiring training sample data, where the training sample data includes product information of several users and products of interest, and the product information includes numerical features and sub-type features;

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature, and performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

Constructing a fusion vector according to the user's numerical data and the classification vector;

Inputting the fusion vector of the user into a deep neural network sub-model and a cross neural network sub-model in the product recommendation model, respectively, to obtain a deep feature vector and a cross feature vector;

Constructing a fusion feature vector according to the depth feature vector of the user and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

A loss value is calculated according to the respective recommendation scores of the plurality of products to be recommended corresponding to the user and the product of interest of the user, and the model parameters of the product recommendation model are adjusted according to the loss value.
A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor implements the following steps:

Obtain user characteristic information from the target user terminal, where the user characteristic information includes numerical characteristics and sub-type characteristics;

Perform data preprocessing on the numerical feature and the sub-category feature to construct a fusion vector;

Process the fusion vector based on the deep neural network sub-model in the product recommendation model to obtain a deep feature vector, and process the fusion vector based on the cross neural network sub-model in the product recommendation model to obtain a cross feature vector ；

Constructing a fusion feature vector according to the depth feature vector and the cross feature vector, and predicting respective recommendation scores of several products to be recommended according to the fusion feature vector;

According to the respective recommendation scores of the several products to be recommended, product information of at least one product to be recommended is sent to the target user terminal.
15. The computer-readable storage medium according to claim 16, wherein said performing data preprocessing on said numerical feature and said sub-category feature to construct a fusion vector comprises:

Performing standardization processing on the numerical feature to obtain numerical data corresponding to the numerical feature;

Performing vectorization processing on the sub-type feature to obtain a sub-type vector corresponding to the sub-type feature;

A fusion vector is constructed according to the numerical data and the classification vector.
15. The computer-readable storage medium of claim 16, wherein the following steps are further implemented when the computer program is executed:

Obtaining product category information from the target user terminal, and performing vectorization processing on the product category information to obtain a product category vector corresponding to the product category information;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, and the product category vector, a fusion vector is constructed.
18. The computer-readable storage medium according to claim 18, wherein said performing vectorization processing on said product category information to obtain a product category vector corresponding to said product category information comprises:

Obtain the mapping relationship data between the preset product category information and the product category vector;

According to the mapping relationship data, vectorize the product category information to obtain the product category vector corresponding to the product category information.
The computer-readable storage medium according to any one of claims 16 to 19, wherein the following steps are further implemented when the computer program is executed:

Acquiring a historical product purchase record corresponding to the target user terminal, and performing self-attention mechanism processing on the historical product purchase record to obtain a purchase record vector corresponding to the historical product purchase record;

The constructing a fusion vector according to the numerical data and the classification vector includes:

According to the numerical data, the classification vector, the product category vector and/or the purchase record vector, a fusion vector is constructed.