Product recommendation model training, information push method, device, equipment and medium

The problem of low product recommendation accuracy in the prior art is solved by processing user and product feature vectors through vector embedding layer and hidden layer, and more efficient model training and recommendation accuracy are achieved.

CN115033793BInactive Publication Date: 2024-11-26CHINA PING AN LIFE INSURANCE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210680718.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-16
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the prior art, the vector layer of the product recommendation model only uses one user or product feature vector representation, and cannot deeply explore the correlation between the user and the product, resulting in a low accuracy of product recommendation.

Method used

The vector embedding layer converts the sample user information and sample product information into feature vectors, and inactivates the user feature vectors and product feature vectors in the hidden layer to avoid overfitting, and iteratively updates the recommended model until the total loss value reaches convergence condition.

Benefits of technology

It improves the efficiency of model data processing, avoids overfitting, and improves the training accuracy and recommendation accuracy of product recommendation models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115033793B_ABST
    Figure CN115033793B_ABST
Patent Text Reader

Abstract

The present invention discloses a product recommendation model training, information push method, device, equipment and medium. The method inputs sample user information and sample product information into a preset recommendation model; determines a user feature vector corresponding to the sample user information and a product feature vector corresponding to the sample product information through a vector embedding layer; performs inactivation processing on the user feature vector and the product feature vector through a hidden layer to obtain a user inactivation vector and a product inactivation vector; determines a predicted product label based on the user inactivation vector and the product inactivation vector, and determines a total loss value according to the recommended product label and the predicted product label; when the total loss value does not reach a preset convergence condition, iteratively updates the initial parameters of the preset recommendation model until the total loss value reaches the convergence condition, and records the preset recommendation model after convergence as a product recommendation model. The present invention avoids overfitting of the preset recommendation model during the training process, and improves the accuracy of model training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of prediction model technology, and in particular to a product recommendation model training, information push method, device, equipment and medium. Background Art

[0002] With the development of science and technology, more and more offline sales businesses are gradually exploring online sales methods, such as the insurance industry, e-commerce industry or service industry.

[0003] In the prior art, a recommendation model obtained through pre-training generally makes recommendations to users based on the characteristics of users or products. However, the vector layer of the recommendation model generally only uses one feature vector representation of the user or product. This makes it impossible to deeply explore the relationship between users and products, resulting in a low accuracy rate of product recommendations. Summary of the invention

[0004] The embodiments of the present invention provide a product recommendation model training, information push method, device, equipment and medium to solve the problem of low accuracy of product recommendation in the prior art.

[0005] A product recommendation model training method, comprising:

[0006] Receive a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; one sample user information is associated with one recommended product tag;

[0007] Obtaining a preset recommendation model including initial parameters, and inputting the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer;

[0008] Determining a user feature vector corresponding to the sample user information through the vector embedding layer, and determining a product feature vector corresponding to the sample product information through the vector embedding layer;

[0009] Performing inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, and performing inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector;

[0010] Determine a predicted product label corresponding to the sample user information based on the user deactivation vector and the product deactivation vector, and determine a total loss value according to the recommended product label and the predicted product label;

[0011] When the total loss value does not reach the preset convergence condition, the initial parameters of the preset recommendation model are iteratively updated until the total loss value reaches the convergence condition, and the preset recommendation model after convergence is recorded as the trained product recommendation model.

[0012] An information push method, comprising:

[0013] Acquire target user information and a list to be recommended; the list to be recommended includes at least one piece of information to be recommended;

[0014] Inputting the target user information and the information to be recommended into a product recommendation model, and filtering the information to be recommended that matches the target user information from the list to be recommended by the product recommendation model; the product recommendation model is obtained according to the product recommendation model training method;

[0015] The information to be recommended is pushed to the client.

[0016] A product recommendation model training device, comprising:

[0017] An instruction receiving module, used to receive a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; one sample user information is associated with one recommended product tag;

[0018] A data input module, used to obtain a preset recommendation model including initial parameters, and input the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer;

[0019] a feature mapping module, configured to determine a user feature vector corresponding to the sample user information through the vector embedding layer, and to determine a product feature vector corresponding to the sample product information through the vector embedding layer;

[0020] An inactivation processing module, configured to perform inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, and perform inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector;

[0021] a label prediction module, configured to determine a predicted product label corresponding to the sample user information based on the user deactivation vector and the product deactivation vector, and determine a total loss value according to the recommended product label and the predicted product label corresponding to the same sample user information;

[0022] The model updating module is used to iteratively update the initial parameters of the preset recommendation model when the total loss value does not reach the preset convergence condition, until the total loss value reaches the convergence condition, and record the preset recommendation model after convergence as the trained product recommendation model.

[0023] An information push device, comprising:

[0024] An information acquisition module, used to acquire target user information and a list to be recommended; the list to be recommended includes at least one piece of information to be recommended;

[0025] An information matching module, used for inputting the target user information and the information to be recommended into a product recommendation model, and filtering the information to be recommended that matches the target user information from the list to be recommended by the product recommendation model; the product recommendation model is obtained according to the product recommendation model training method;

[0026] The information push module is used to push the information to be recommended to the client.

[0027] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned product recommendation model training method when executing the computer program, or implements the above-mentioned information push method when executing the computer program.

[0028] A computer-readable storage medium stores a computer program, which implements the above-mentioned product recommendation model training method when executed by a processor, or implements the above-mentioned information push method when executed by a processor.

[0029] The above-mentioned product recommendation model training, information push method, device, computer equipment and storage medium, the product recommendation model training method converts sample user information and sample product information into the form of feature vectors through the vector embedding layer of the preset recommendation model. Thus, the sample user information and sample product information are converted into formalized feature vectors, which improves the efficiency of model data processing. Part of the feature vectors in the user feature vector and the product feature vector are directly inactivated through the hidden layer of the preset recommendation model to directly shield part of the feature vectors in the user feature vector or the product feature vector. Avoid overfitting of the preset recommendation model during the training process, and improve the training accuracy of the preset recommendation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0031] Figure 1 Schematic diagram of an application environment of a product recommendation model training method in one embodiment of the present invention;

[0032] Figure 2 is a flow chart of a product recommendation model training method in one embodiment of the present invention;

[0033] Figure 3 is a principle block diagram of a product recommendation model training device in one embodiment of the present invention;

[0034] Figure 4 is a schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The product recommendation model training method provided by the embodiment of the present invention can be applied as follows: Figure 1 Specifically, the product recommendation model training method is applied in a product recommendation model training system, and the product recommendation model training system includes: Figure 1 The client and server shown in the figure communicate with each other through the network to solve the problem of low accuracy of product recommendation in the prior art. Among them, the client, also known as the user end, refers to a program corresponding to the server that provides local services to customers. The client can be installed on but not limited to various personal computers, laptops, smart phones, tablets and portable wearable devices. The server can be implemented as an independent server or a server cluster composed of multiple servers. Among them, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms.

[0037] In one embodiment, if Figure 2 As shown, a product recommendation model training method is provided, and the method is applied in Figure 1 The server in the example is used as an example to illustrate the process, including the following steps:

[0038] S10: receiving a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; one sample user information is associated with one recommended product tag.

[0039] It is understandable that the product recommendation instruction can be sent directly to the server by the user through the client, or it can be automatically generated after the user sends sample user information and sample product information through the client. Among them, the sample user information is the information that describes the user's characteristics. Exemplarily, the sample user information may include but is not limited to the user's age, personality preferences or economic status. Sample product information is the information that describes the product characteristics. Exemplarily, the sample product information may include but is not limited to the type, price or applicable population of the product. The recommended product label is a label of the recommended product that is pre-set corresponding to each sample user information. The recommended product label can be implemented by manual labeling.

[0040] S20: Obtain a preset recommendation model including initial parameters, and input the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer.

[0041] It can be understood that the preset recommendation model can be pre-built based on a neural network. The preset recommendation model in this embodiment includes a vector embedding layer and a hidden layer (in addition, it also includes conventional network layers such as an input layer and an output layer). Among them, the vector embedding layer is used to perform vector conversion on sample user information or sample product information. The hidden layer is used to inactivate some vectors in the feature vector obtained by the vector embedding layer, so as to prevent the preset recommendation model from overfitting during the training process. The initial parameters are the model parameters when constructing the preset recommendation model. In the subsequent steps, during the training process of the preset recommendation model, the initial parameters are continuously iteratively updated to improve the recommendation accuracy of the preset recommendation model.

[0042] S30: Determine a user feature vector corresponding to the sample user information through the vector embedding layer, and determine a product feature vector corresponding to the sample product information through the vector embedding layer.

[0043] It can be understood that since both the sample user information and the sample product information are described in an informal manner, such as text description, digital description or letter description, the sample user information and the sample product information are converted into vectors through the vector embedding layer.

[0044] Specifically, after the sample user information and the sample product information are input into the preset recommendation model, the sample user information is converted into a feature vector through the vector embedding layer to obtain a user feature vector. The sample product information is converted into a feature vector through the vector embedding layer to obtain a product feature vector.

[0045] S40: performing inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, and performing inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector.

[0046] Understandably, since the sample user information or sample product information contains many features, if the preset recommendation model is trained directly based on the user feature vector and the product feature vector, overfitting may occur. As a result, the recommendation accuracy of the preset recommendation model in the training phase is much higher than the recommendation accuracy of the preset recommendation model in the testing phase. Therefore, in this embodiment, the user feature vector or the product feature vector is directly deactivated to directly shield part of the feature vectors in the user feature vector or the product feature vector.

[0047] It should be noted that since the user feature vector is a multi-dimensional feature vector, there are c-dimensional feature vectors as described above. That is, the user feature vector includes multiple feature vectors, so in this embodiment, some feature vectors in the user feature vector are deactivated. However, the user feature vector in the prior art is often a feature vector of one dimension, so in the prior art, the elements in the feature vector are deactivated. That is, it is equivalent to deactivating the elements in the multi-dimensional feature vector in the user feature vector in this embodiment. The deactivation processing objects of the two are different.

[0048] Specifically, after the user feature vector corresponding to the sample user information is determined by the vector embedding layer, since the user feature vector is a multidimensional feature vector, part of the feature vectors in the user feature vector are inactivated by the hidden layer, that is, part of the feature vectors in the user feature vector are randomly shielded to obtain the user inactivation vector. After the product feature vector corresponding to the sample product information is determined by the vector embedding layer, since the product feature vector is a multidimensional feature vector, part of the feature vectors in the product feature vector are inactivated by the hidden layer, that is, part of the feature vectors in the product feature vector are randomly shielded to obtain the product inactivation vector.

[0049] S50: Based on the user deactivation vector and the product deactivation vector, determine the predicted product label corresponding to the sample user information, and determine the total loss value according to the recommended product label and the predicted product label.

[0050] It can be understood that the predicted product label is the label corresponding to the recommended product that best matches the sample user information by fitting the preset recommendation model. Exemplarily, assuming that the product recommendation model training method is applied to a movie recommendation scenario, the predicted product label can be the label corresponding to the movie recommended by the preset recommendation model (such as using the movie name as the predicted product label). Assuming that the product recommendation model training method is applied to an insurance recommendation scenario, the predicted product label can be the label corresponding to the insurance product recommended by the preset recommendation model (such as using the name corresponding to the insurance type as the predicted product label). The total loss value represents the gap between the recommended product label and the predicted product label.

[0051] Specifically, after the user feature vector is inactivated through the hidden layer to obtain the user inactivation vector, and the product feature vector is inactivated through the hidden layer to obtain the product inactivation vector, the user inactivation vector and the product inactivation vector are interactively processed to obtain the predicted product label that best matches the sample user information. The total loss value is then determined through the recommended product label and the predicted product label.

[0052] S60: When the total loss value does not reach the preset convergence condition, iteratively update the initial parameters of the preset recommendation model until the total loss value reaches the convergence condition, and record the preset recommendation model after convergence as the trained product recommendation model.

[0053] It can be understood that the convergence condition can be the condition that the total loss value is less than the set threshold, that is, when the total loss value is less than the set threshold, the training is stopped; the convergence condition can also be the condition that the total loss value is very small after 10,000 calculations and will not decrease anymore, that is, when the total loss value is very small after 10,000 calculations and will not decrease, the training is stopped, and the preset recommendation model after convergence is recorded as the trained product recommendation model.

[0054] Furthermore, after determining the total loss value according to the recommended product label and the predicted product label, when the total loss value does not reach the preset convergence condition, the initial parameters of the initial prediction model are adjusted according to the total loss value, and the sample user information and the sample product information are re-input into the preset recommendation model after adjusting the initial parameters, so that when the total loss value corresponding to the sample user information reaches the preset convergence condition, another sample user information is selected, and the above steps S20 to S50 are executed to obtain the total loss value corresponding to the sample user information, and when the total loss value does not reach the preset convergence condition, the initial parameters of the preset recommendation model are adjusted again according to the total loss value, so that the total loss value corresponding to the sample user information reaches the preset convergence condition.

[0055] In this way, after the preset recommendation model is trained with all sample user information, the results output by the preset recommendation model can continue to approach the accurate results, making the prediction accuracy higher and higher, until the total loss values ​​corresponding to all sample user information reach the preset convergence conditions, and the preset recommendation model after convergence is recorded as the trained product recommendation model.

[0056] In this embodiment, the sample user information and sample product information are converted into the form of feature vectors through the vector embedding layer of the preset recommendation model. Thus, the sample user information and sample product information are converted into formalized feature vectors, which improves the efficiency of model data processing. Part of the feature vectors in the user feature vector and the product feature vector are directly deactivated through the hidden layer of the preset recommendation model to directly shield part of the feature vectors in the user feature vector or the product feature vector. This avoids overfitting of the preset recommendation model during the training process and improves the training accuracy of the preset recommendation model.

[0057] In one embodiment, in step S30, that is, determining the user feature vector corresponding to the sample user information through the vector embedding layer, includes:

[0058] (1) Perform feature mapping on the sample user information to obtain a user mapping vector.

[0059] It is understandable that the above description points out that the sample user information is described in an informal manner. Therefore, in this embodiment, the sample user information is feature mapped by One-hot encoding to obtain a user mapping vector. Furthermore, since the sample user information contains user features of multiple different types (such as age type, economic status type, or occupation type, etc.), the user mapping vector is a high-dimensional vector. The user mapping vector is represented as: u =R u×d . Among them I u Refers to the user mapping vector of the uth sample user information. d is the dimension of One-hot encoding. That is, the dimension of each user mapping vector is 1×d.

[0060] (2) Initializing three-dimensional matrix parameters, and performing tensor product processing on the user mapping vector based on the three-dimensional matrix parameters to obtain the user feature vector.

[0061] It can be understood that the three-dimensional matrix parameters are preset parameters, and the three-dimensional matrix parameters can be M∈R d×n×c. Wherein, c is a hyperparameter that can be adjusted by developers or debuggers. n is a parameter that is set much smaller than d. Specifically, after feature mapping of sample user information is performed to obtain a user mapping vector, the three-dimensional matrix parameters are initialized, and the three-dimensional matrix parameters and the user mapping vector are tensor-producted to obtain a user feature vector.

[0062] The tensor product is processed as follows: E u ∈R u×n×c .

[0063] The determining, through the vector embedding layer, a product feature vector corresponding to the sample product information includes:

[0064] (3) Perform feature mapping on the sample product information to obtain a product mapping vector.

[0065] It is understandable that the sample product information is also described in an informal manner as mentioned in the above description. Therefore, in this embodiment, the sample product information is feature mapped by One-hot encoding to obtain a product mapping vector. Furthermore, since the sample product information contains multiple different types of product features (such as product type, product price, or applicable population, etc.), the product mapping vector is also a high-dimensional vector. The user mapping vector is represented as: P i =R i×d . Among them, P i Refers to the product mapping vector of the i-th sample product information. d is the dimension of One-hot encoding. That is, the dimension of each product mapping vector is 1×d.

[0066] (4) Performing tensor product processing on the product mapping vector based on the three-dimensional matrix parameters to obtain the product feature vector.

[0067] Specifically, after feature mapping is performed on the sample product information to obtain the product mapping vector, a tensor product process is performed on the three-dimensional matrix parameters and the product mapping vector to obtain the product feature vector. The tensor product process is as follows: E i ∈R i×n×c .

[0068] In one embodiment, in step S40, that is, performing inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, includes:

[0069] (1) Obtaining a preset deactivation ratio and determining a first feature quantity of a first feature vector included in the user feature vector.

[0070] It is understandable that the preset deactivation ratio can be set according to the number of user feature vectors or product feature vectors. Exemplarily, the preset deactivation ratio can be set to 1 / 10 or 1 / 5, etc. The first feature quantity refers to the total amount of first feature vectors contained in the user feature vector. The first feature vector is the feature vector of different dimensions in the user feature vector.

[0071] (2) determining a first deactivation quantity according to the preset deactivation ratio and the first feature quantity, and selecting a first feature vector of the first deactivation quantity from the user feature vectors.

[0072] Specifically, after obtaining the preset deactivation ratio and determining the first feature quantity of the first feature vector contained in the user feature vector, the product of the preset deactivation ratio and the first feature quantity is used as the first deactivation quantity, and then the first feature vector of the first deactivation quantity is randomly selected from the user feature vector.

[0073] (3) Performing inactivation processing on the selected first feature vector to obtain the user inactivation vector.

[0074] Specifically, after selecting the first deactivation number of first feature vectors from the user feature vector, the selected first feature vectors are deactivated. That is, all the selected first feature vectors are changed to 0, thereby achieving the effect of shielding the selected first feature vectors. Then, the user feature vector after the selected first feature vector is deactivated is recorded as the user deactivation vector.

[0075] The step of performing inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector includes:

[0076] (4) Determine the second feature quantity of the second feature vector included in the product feature vector.

[0077] It can be understood that the second feature quantity refers to the total amount of second feature vectors contained in the product feature vector. The second feature vector is the feature vector of different dimensions in the product feature vector.

[0078] (5) determining a second deactivation quantity according to the preset deactivation ratio and the second feature quantity, and selecting a second feature vector of the second deactivation quantity from the product feature vector.

[0079] Specifically, after determining the second feature quantity of the second feature vector included in the product feature vector, the product of the preset inactivation ratio and the second feature quantity is used as the second inactivation quantity, and then the second feature vector of the second inactivation quantity is randomly selected from the product feature vector.

[0080] (6) Performing inactivation processing on the selected second eigenvector to obtain the product inactivation vector.

[0081] Specifically, after selecting the second inactivation number of second feature vectors from the product feature vector, the selected second feature vectors are inactivated. That is, all the selected second feature vectors are changed to 0, thereby achieving the effect of shielding the selected second feature vectors. Then, the product feature vector after the inactivation process of the selected second feature vector is recorded as the user inactivation vector.

[0082] In this embodiment, part of the feature vectors in the user feature vector and the product feature vector are directly deactivated through the hidden layer in the preset recommendation model, thereby shielding part of the feature vectors in the user feature vector and the product feature vector, and preventing the preset recommendation model from overfitting during the training process.

[0083] In one embodiment, one sample product information is associated with one sample product label; in step S50, that is, determining a predicted product label corresponding to the sample user information based on the user inactivation vector and the product inactivation vector, includes:

[0084] Perform vector concatenation on the user inactivation vector and the product inactivation vector to obtain a feature concatenation vector.

[0085] It can be understood that vector concatenation is the process of merging and concatenating the user inactivation vector and the product inactivation vector to form a matrix. The feature concatenation vector can be regarded as the matrix obtained by merging and concatenating the user inactivation vector and the product inactivation vector.

[0086] A preset interaction function is obtained, and the feature splicing vector is interactively processed through the preset interaction function to obtain a matching value between the user deactivation vector and each of the product deactivation vectors; one product deactivation vector corresponds to one matching value.

[0087] It can be understood that the preset interaction function is used to mine the matching relationship between the user inactivation vector and the product inactivation vector. Exemplarily, the preset interaction function can be an inner product interaction function.

[0088] Specifically, after vector splicing is performed on the user inactivation vector and the product inactivation vector to obtain the feature splicing vector, the matching value between the user inactivation vector and the product inactivation vector is determined by a preset interaction function. It can be understood that in the above steps, a sample user information can be input into the preset recommendation model, and one or all sample product information can be input into the preset recommendation model. Then, after the feature splicing vector is interactively processed by the preset interaction function, the matching value between the user inactivation vector corresponding to the sample user information and the product inactivation vector corresponding to each sample product information is obtained.

[0089] The sample product information corresponding to the product deactivation vector corresponding to the maximum matching value is recorded as the predicted product information, and the sample product label associated with the predicted product information is recorded as the predicted product label.

[0090] Specifically, after interactively processing the feature concatenation vector through a preset interactive function to obtain the matching value between the user deactivation vector and each product deactivation vector, the sample product information corresponding to the product deactivation vector corresponding to the largest matching value is selected and recorded as the predicted product information, and the sample product label associated with the predicted product information is recorded as the predicted product label. It can be understood that the sample product label is obtained by pre-labeling each sample product information.

[0091] In one embodiment, in step S50, that is, determining the total loss value according to the recommended product label and the predicted product label, includes:

[0092] According to the recommended product label and the predicted product label corresponding to the same sample user information, the predicted loss value corresponding to the sample user information is determined; one sample user information corresponds to one predicted loss value.

[0093] It can be understood that the predicted product label output by the preset recommendation model corresponds to the sample user information. Therefore, based on the recommended product label and the predicted product label corresponding to the same sample user information, the predicted loss value corresponding to the sample user information can be determined. The predicted loss value represents the gap between the recommended product label and the predicted product label corresponding to the same sample user information.

[0094] A preset regular expression is obtained, and the total loss value is determined based on the preset regular expression and all the predicted loss values.

[0095] It can be understood that the preset regular expression is used to prevent the preset recommendation model from overfitting during the training process.

[0096] Specifically, the total loss value can be determined by the following expression:

[0097]

[0098] Among them, L(φ) is the total loss value; is the predicted product label corresponding to the i-th sample user information; y i is the recommended product label corresponding to the i-th sample user information; is the predicted loss value corresponding to the i-th sample user information; Ω(f k) is a preset regular expression; k represents the number of iterations of the initial prediction model. It can be understood that one training of the preset recommendation model for one sample user information is recorded as one initial parameter iteration of the preset recommendation model. Furthermore, k represents the number of times the preset regular expression is used to prevent overfitting of the sample user information when the preset recommendation model is trained for one sample user information.

[0099] In one embodiment, a method for pushing information is provided, including:

[0100] Target user information and a list to be recommended are obtained; the list to be recommended includes at least one piece of information to be recommended.

[0101] It can be understood that the target user information is the personal information of the target user. Exemplarily, the target user information can be the age, personality preference or economic status of the target user. The list to be recommended can be set according to different application scenarios. Exemplarily, it is assumed that the information push method in the present embodiment is applied to a movie recommendation scenario. The list to be recommended is a list of movies to be recommended, and the information to be recommended is the information of the movies to be recommended (such as the name, theme type or starring role of the movie to be recommended). It is assumed that the information push method in the present embodiment is applied to an insurance recommendation scenario. The list to be recommended is a list of insurance products to be recommended, and the information to be recommended is the information of the insurance products to be recommended (such as the type of insurance, insurance amount or insurance conditions of the insurance products to be recommended).

[0102] The target user information and the information to be recommended are input into a product recommendation model, and the information to be recommended that matches the target user information is screened out from the list to be recommended by the product recommendation model. The product recommendation model is obtained according to the product recommendation model training method.

[0103] The information to be recommended is pushed to the client.

[0104] Specifically, after obtaining the target user information and the list of products to be recommended, the target user information and all the products to be recommended can be input into the product recommendation model. The target user information and all the products to be recommended are matched by the product recommendation model. Then, the products to be recommended that best match the target user information are selected from all the products to be recommended. The products to be recommended are pushed to the client. The client can be the client of the target user corresponding to the target user information, or it can be the client of other users associated with the target user (such as relatives of the target user).

[0105] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0106] In one embodiment, a product recommendation model training device is provided, and the product recommendation model training device corresponds to the product recommendation model training method in the above embodiment. Figure 3 As shown, the product recommendation model training device includes an instruction receiving module 10, a data input module 20, a feature mapping module 30, an inactivation processing module 40, a label prediction module 50 and a model updating module 60. The functional modules are described in detail as follows:

[0107] The instruction receiving module 10 is used to receive a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; one sample user information is associated with one recommended product tag;

[0108] A data input module 20 is used to obtain a preset recommendation model including initial parameters, and input the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer;

[0109] A feature mapping module 30, configured to determine a user feature vector corresponding to the sample user information through the vector embedding layer, and to determine a product feature vector corresponding to the sample product information through the vector embedding layer;

[0110] The deactivation processing module 40 is used to perform deactivation processing on the user feature vector through the hidden layer to obtain a user deactivation vector, and to perform deactivation processing on the product feature vector through the hidden layer to obtain a product deactivation vector;

[0111] A label prediction module 50, configured to determine a predicted product label corresponding to the sample user information based on the user deactivation vector and the product deactivation vector, and determine a total loss value based on the recommended product label and the predicted product label corresponding to the same sample user information;

[0112] The model updating module 60 is used to iteratively update the initial parameters of the preset recommendation model when the total loss value does not reach the preset convergence condition, until the total loss value reaches the convergence condition, and record the preset recommendation model after convergence as the trained product recommendation model.

[0113] Preferably, the feature mapping module 30 comprises:

[0114] A user information mapping unit, used to perform feature mapping on the sample user information to obtain a user mapping vector;

[0115] A user vector processing unit, used for initializing three-dimensional matrix parameters, and performing tensor product processing on the user mapping vector based on the three-dimensional matrix parameters to obtain the user feature vector;

[0116] A product information mapping unit, used to perform feature mapping on the sample product information to obtain a product mapping vector;

[0117] A product vector processing unit is used to perform tensor product processing on the product mapping vector based on the three-dimensional matrix parameters to obtain the product feature vector.

[0118] Preferably, the deactivation processing module 40 includes:

[0119] A first quantity determining unit, configured to obtain a preset deactivation ratio and determine a first feature quantity of a first feature vector included in the user feature vector;

[0120] A first vector selection unit, configured to determine a first deactivation quantity according to the preset deactivation ratio and the first feature quantity, and select a first feature vector of the first deactivation quantity from the user feature vector;

[0121] A first deactivation processing unit, configured to perform a deactivation process on the selected first feature vector to obtain the user deactivation vector;

[0122] A second quantity determining unit, used to determine a second feature quantity of a second feature vector included in the product feature vector;

[0123] A second vector selection unit, configured to determine a second deactivation quantity according to the preset deactivation ratio and the second feature quantity, and select a second feature vector of the second deactivation quantity from the product feature vector;

[0124] The second inactivation processing unit is used to perform inactivation processing on the selected second feature vector to obtain the product inactivation vector.

[0125] Preferably, the tag prediction module 50 includes:

[0126] A vector concatenation unit, used for concatenating the user deactivation vector and the product deactivation vector to obtain a feature concatenation vector;

[0127] An interactive processing unit, used to obtain a preset interactive function, and interactively process the feature splicing vector through the preset interactive function to obtain a matching value between the user deactivation vector and each of the product deactivation vectors; one product deactivation vector corresponds to one matching value;

[0128] The label prediction unit is used to record the sample product information corresponding to the product deactivation vector corresponding to the maximum matching value as the predicted product information, and record the sample product label associated with the predicted product information as the predicted product label.

[0129] Preferably, the tag prediction module 50 further includes:

[0130] A predicted loss value determination unit, configured to determine a predicted loss value corresponding to the sample user information according to the recommended product label and the predicted product label corresponding to the same sample user information; one sample user information corresponds to one predicted loss value;

[0131] The total loss value determining unit is used to obtain a preset regular expression and determine the total loss value based on the preset regular expression and all the predicted loss values.

[0132] For the specific limitations of the product recommendation model training device, please refer to the limitations of the product recommendation model training method above, which will not be repeated here. Each module in the above-mentioned product recommendation model training device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0133] In one embodiment, an information push device is provided, including:

[0134] An information acquisition module, used to acquire target user information and a list to be recommended; the list to be recommended includes at least one piece of information to be recommended;

[0135] An information matching module, used for inputting the target user information and the information to be recommended into a product recommendation model, and filtering the information to be recommended that matches the target user information from the list to be recommended by the product recommendation model; the product recommendation model is obtained according to the product recommendation model training method;

[0136] The information push module is used to push the information to be recommended to the client.

[0137] For the specific definition of the information push device, please refer to the definition of the information push method above, which will not be repeated here. Each module in the above-mentioned information push device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0138] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the data used by the product recommendation model training method or the information push method in the above-mentioned embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a product recommendation model training method is implemented, or when the computer program is executed by the processor, an information push method is implemented.

[0139] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the product recommendation model training method in the above-mentioned embodiment is implemented, or when the processor executes the computer program, the information push method in the above-mentioned embodiment is implemented.

[0140] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the product recommendation model training method in the above embodiment is implemented, or when the computer program is executed by a processor, the information push method in the above embodiment is implemented.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0142] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0143] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A product recommendation model training method, characterized in that: include: Receiving a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; One of the sample user information is associated with a recommended product tag; Obtaining a preset recommendation model including initial parameters, and inputting the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer; Determining a user feature vector corresponding to the sample user information through the vector embedding layer, and determining a product feature vector corresponding to the sample product information through the vector embedding layer; Performing inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, and performing inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector; Determine a predicted product label corresponding to the sample user information based on the user deactivation vector and the product deactivation vector, and determine a total loss value according to the recommended product label and the predicted product label; When the total loss value does not reach the preset convergence condition, iteratively update the initial parameters of the preset recommendation model until the total loss value reaches the convergence condition, and record the preset recommendation model after convergence as the trained product recommendation model; The determining, through the vector embedding layer, a user feature vector corresponding to the sample user information includes: Performing feature mapping on the sample user information to obtain a user mapping vector; Initializing three-dimensional matrix parameters, and performing tensor product processing on the user mapping vector based on the three-dimensional matrix parameters to obtain the user feature vector; The determining, through the vector embedding layer, a product feature vector corresponding to the sample product information includes: Performing feature mapping on the sample product information to obtain a product mapping vector; Performing tensor product processing on the product mapping vector based on the three-dimensional matrix parameters to obtain the product feature vector; The step of performing inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector includes: Obtaining a preset deactivation ratio, and determining a first feature quantity of a first feature vector included in the user feature vector; Determine a first deactivation quantity according to the preset deactivation ratio and the first feature quantity, and select a first feature vector of the first deactivation quantity from the user feature vector; Performing inactivation processing on the selected first feature vector to obtain the user inactivation vector; The step of performing inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector includes: Determining the second feature quantity of the second feature vector included in the product feature vector; Determine a second deactivation quantity according to the preset deactivation ratio and the second feature quantity, and select a second feature vector of the second deactivation quantity from the product feature vector; Performing inactivation processing on the selected second eigenvector to obtain the product inactivation vector; Associating one sample product information with a sample product label; and determining a predicted product label corresponding to the sample user information based on the user inactivation vector and the product inactivation vector, comprising: Performing vector concatenation on the user inactivation vector and the product inactivation vector to obtain a feature concatenation vector; Obtaining a preset interactive function, and interactively processing the feature splicing vector through the preset interactive function to obtain a matching value between the user deactivation vector and each of the product deactivation vectors; one product deactivation vector corresponds to one matching value; The sample product information corresponding to the product deactivation vector corresponding to the maximum matching value is recorded as the predicted product information, and the sample product label associated with the predicted product information is recorded as the predicted product label.

2. The product recommendation model training method according to claim 1, characterized in that: The determining of the total loss value according to the recommended product label and the predicted product label includes: According to the recommended product label and the predicted product label corresponding to the same sample user information, determining the predicted loss value corresponding to the sample user information; one sample user information corresponds to one predicted loss value; A preset regular expression is obtained, and the total loss value is determined based on the preset regular expression and all the predicted loss values.

3. An information push method, characterized in that: include: Obtain target user information and a list of users to be recommended; The to-be-recommended list includes at least one to-be-recommended information; Inputting the target user information and the information to be recommended into a product recommendation model, and filtering the information to be recommended that matches the target user information from the list to be recommended by the product recommendation model; The product recommendation model is obtained according to the product recommendation model training method according to claim 1 or 2; The information to be recommended is pushed to the client.

4. A product recommendation model training device, characterized in that: The product recommendation model training device is used to execute the product recommendation model training method according to claim 1 or 2, comprising: An instruction receiving module, used to receive a product recommendation instruction; the product recommendation instruction includes at least one sample user information and sample product information; one sample user information is associated with one recommended product tag; A data input module, used to obtain a preset recommendation model including initial parameters, and input the sample user information and the sample product information into the preset recommendation model; the preset recommendation model includes a vector embedding layer and a hidden layer; a feature mapping module, configured to determine a user feature vector corresponding to the sample user information through the vector embedding layer, and to determine a product feature vector corresponding to the sample product information through the vector embedding layer; An inactivation processing module, configured to perform inactivation processing on the user feature vector through the hidden layer to obtain a user inactivation vector, and perform inactivation processing on the product feature vector through the hidden layer to obtain a product inactivation vector; a label prediction module, configured to determine a predicted product label corresponding to the sample user information based on the user deactivation vector and the product deactivation vector, and determine a total loss value according to the recommended product label and the predicted product label corresponding to the same sample user information; The model updating module is used to iteratively update the initial parameters of the preset recommendation model when the total loss value does not reach the preset convergence condition, until the total loss value reaches the convergence condition, and record the preset recommendation model after convergence as the trained product recommendation model.

5. An information push device, characterized in that: An information acquisition module, used to acquire target user information and a list to be recommended; the list to be recommended includes at least one piece of information to be recommended; An information matching module, used for inputting the target user information and the information to be recommended into a product recommendation model, and filtering the information to be recommended that matches the target user information from the list to be recommended through the product recommendation model; The product recommendation model is obtained according to the product recommendation model training method according to claim 1 or 2; The information push module is used to push the information to be recommended to the client.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the product recommendation model training method as described in claim 1 or 2 is implemented, or when the processor executes the computer program, the information push method as described in claim 3 is implemented.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, the product recommendation model training method as described in claim 1 or 2 is implemented, or when the computer program is executed by the processor, the information push method as described in claim 3 is implemented.