CN115795144A

CN115795144A - Product recommendation method and device and electronic equipment

Info

Publication number: CN115795144A
Application number: CN202211378276.4A
Authority: CN
Inventors: 徐世界; 刘昊骋; 王天祺; 徐靖宇; 田建
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2022-11-04
Filing date: 2022-11-04
Publication date: 2023-03-14

Abstract

The disclosure provides a product recommendation method and device and electronic equipment, relates to the technical field of computers, and particularly relates to the technical fields of machine learning, intelligent recommendation and the like. The specific implementation scheme is as follows: determining candidate recommended users who own the first product and do not own the second product; inputting the user characteristics of the candidate recommending users into a product recommending model to obtain the interest degree of each candidate recommending user output by the product recommending model to a second product, wherein the product recommending model is obtained by training the user characteristics of sample users in advance, and the sample users comprise negative sample users who possess the first product and do not possess the second product in a historical time window and positive sample users who possess the first product and the second product in the historical time window; determining a target recommended user from the candidate recommended users according to the interest degree; and recommending the second product to the target recommending user. Therefore, the target recommendation user can obtain reasonable product recommendation, and user experience is improved.

Description

Product recommendation method and device and electronic equipment

Technical Field

The present disclosure relates to the field of computer technology, and in particular, to the field of machine learning, intelligent recommendation, and the like.

Background

In the related art, a product company generally needs to recommend a suitable product for a user who wants to purchase the product.

Disclosure of Invention

The disclosure provides a product recommendation method and device and electronic equipment.

According to a first aspect of the present disclosure, there is provided a product recommendation method including:

determining candidate recommended users who own the first product and do not own the second product;

inputting the user characteristics of the candidate recommending users into a product recommending model to obtain the interest degree of each candidate recommending user output by the product recommending model to a second product, wherein the product recommending model is obtained by training the user characteristics of sample users in advance, and the sample users comprise negative sample users who own a first product and do not own the second product in a historical time window and positive sample users who own the first product and the second product in the historical time window;

determining a target recommended user from the candidate recommended users according to the interest degree;

and recommending the second product to the target recommending user.

According to a second aspect of the present disclosure, there is provided a training method of a product recommendation model, including:

obtaining user characteristics of sample users, the sample users including negative sample users who own a first product and do not own a second product within a historical time window, and positive sample users who own the first product and the second product within the historical time window;

inputting the user characteristics into an original training model to obtain the interest degree of each sample user in the second product output by the original training model;

constructing a training loss according to the interest degree, wherein the training loss is negatively correlated with the interest degree of the positive sample user and positively correlated with the interest degree of the negative sample user;

and adjusting the original training model according to the training loss to obtain a product recommendation model.

According to a third aspect of the present disclosure, there is provided a product recommendation device including:

the candidate determining module is used for determining candidate recommending users who own the first product and do not own the second product;

an interest obtaining module, configured to input user characteristics of the candidate recommending users to a product recommending model, and obtain a degree of interest of each candidate recommending user in a second product output by the product recommending model, where the product recommending model is obtained by training user characteristics of sample users in advance, and the sample users include negative sample users who have a first product and do not have the second product in a historical time window, and positive sample users who have the first product and the second product in the historical time window;

the target determining module is used for determining a target recommending user from the candidate recommending users according to the interest degree;

and the product recommending module is used for recommending the second product to the target recommending user.

According to a fourth method of the present disclosure, there is provided a training apparatus for a product recommendation model, including:

a characteristic obtaining module for obtaining user characteristics of sample users, wherein the sample users comprise negative sample users who possess a first product and do not possess a second product in a historical time window, and positive sample users who possess the first product and the second product in the historical time window;

the initial training module is used for inputting the user characteristics into an original training model to obtain the interest degree of each sample user output by the original training model in the second product;

a model adjustment module for constructing a training loss according to the interest level, wherein the training loss is negatively correlated to the interest level of the positive sample user and positively correlated to the interest level of the negative sample user;

and the module obtaining module is used for adjusting the original training model according to the training loss to obtain a product recommendation model.

According to a fifth aspect of the present disclosure, there is also provided an electronic device, including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of the preceding first and second aspects.

According to a sixth aspect of the present disclosure, there is also provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method according to any of the preceding first and second aspects.

According to a seventh aspect of the present disclosure, there is also provided a computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of the first and second aspects.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 is a schematic flow diagram of a product recommendation method according to the present disclosure;

FIG. 2 is a schematic flow chart diagram of a method of training a product recommendation model according to the present disclosure;

FIG. 3 is another flow diagram of a method of training a product recommendation model according to the present disclosure;

FIG. 4 is another flow diagram of a method of training a product recommendation model according to the present disclosure;

FIG. 5 is another flow diagram of a method of training a product recommendation model according to the present disclosure;

FIG. 6 is a schematic diagram of one configuration of a product recommendation device according to the present disclosure;

FIG. 7 is a schematic diagram of an arrangement of a training apparatus for a product recommendation model according to the present disclosure;

FIG. 8 is a block diagram of an electronic device for implementing a product recommendation method and a training method for a product recommendation model according to embodiments of the present disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the disclosure are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

Based on the fact that product manufacturing or marketing companies often need to recommend suitable products to users, so that the users can conveniently obtain products meeting the needs of the users, and often the related art cannot accurately determine which products the users are interested in and cannot recommend the suitable products to the users, the present disclosure provides a product recommendation method, as shown in fig. 1, the method includes:

s101, determining candidate recommended users who own the first product and do not own the second product.

S102, inputting the user characteristics of the candidate recommending users into the product recommending model to obtain the interest degree of each candidate recommending user in the second product, wherein the interest degree of each candidate recommending user is output by the product recommending model.

S103, determining a target recommending user from the candidate recommending users according to the interest degree.

And S104, recommending the second product to the target recommending user.

According to the embodiment, the user who purchases the first product is used as the candidate recommending user, the user characteristics of the candidate recommending user are input into the product recommending model, the product recommending model outputs the interest degree of the candidate recommending user for the second product, the product recommending model is obtained through user characteristic training of the sample user, the sample user is the user who owns the first product but does not own the user characteristics of the second product, and the user who owns both the first product and the second product, therefore, the product recommending model obtained based on the sample user training can learn the interest degree of the candidate recommending user for the second product, the interest degree of the candidate recommending user for the second product can be predicted according to the user characteristics of the candidate recommending user, the target recommending user in the candidate recommending user is determined according to the interest degree, the second product is recommended for the target recommending user, the target user obtains reasonable product recommendation, and user experience is improved.

In S101, it can be understood that the candidate recommended user in the present disclosure is a user already having a certain product, and for convenience of description, the insurance product is taken as an example below, in this example, the candidate recommended user is a user already purchasing a certain insurance product. Specifically, a database containing marketing information of insurance products, such as hive (a data warehouse tool), may be periodically queried to obtain a user who owns a certain insurance product (a first product) but does not own another insurance product (a second product) as a candidate recommending user.

Illustratively, since the premium of the long-term insurance product is high and the product expiration time is long, and the short-term insurance product is, for example, a gift, the premium is low and the product expiration time is short, selling the long-term insurance product has a high profit for the insurance company, while selling the short-term insurance product has a low profit for the insurance company, and further for the insurance company, the short-term insurance product may be used as a first product, the long-term insurance product may be used as a second product, a user who purchased the short-term insurance product but did not purchase the long-term insurance product may be used as a candidate recommending user, and a part of the candidate recommending users recommend the long-term insurance product, so as to improve the profit.

In S102, the product recommendation model is obtained by training user characteristics of sample users in advance, where the sample users include negative sample users who own the first product and do not own the second product in the historical time window, and positive sample users who own the first product and the second product in the historical time window.

The user who owns the first product and does not own the second product in the historical time window indicates that the user does not own the second product within a preset time after owning the first product, taking an insurance product as an example, that is, the user does not buy the second product insurance within the preset time after buying the insurance of the first product, and regarding such a user as a negative sample user. Correspondingly, the possession of the first product and the second product in the historical time window indicates that the user owns the second product within a preset time after owning the first product, taking an insurance product as an example, that is, the user purchases the insurance of the second product within the preset time after purchasing the insurance of the first product, and regarding such a user as a positive sample user. Specifically, if the preset time is three months, the user a buys a first product in month 1, and then buys a second product in month 2, the user a is taken as a positive sample user, the user B buys the first product in month 1, and the subsequent months 2, 3, and 4 do not buy the second product, and the user B is taken as a negative sample user.

The sample users comprise positive sample users and negative sample users, the positive sample users have second products and can be regarded as users interested in the second products, the negative sample users do not have the second samples and can be regarded as users interested in the second products, the mapping relation between the user characteristics of the sample users and whether the sample users are interested in the second products or not can be learned through a product recommendation model obtained through training according to the user characteristics of the sample users, and therefore whether the candidate recommendation users are interested in the second products or not is predicted according to the user characteristics of the candidate recommendation users, namely the interest degree of the candidate recommendation users in the second products is predicted. The user characteristics of the sample user may include the age, income, health condition, and the like of the user, and accordingly, the feature dimension of the obtained user characteristics of the candidate recommending user should be the same as the feature dimension of the user characteristics of the sample user. In one possible embodiment, the user characteristics of the sample users are generated using expert experience and/or automated feature generation tools based on information about each sample user, which may include product purchase records of the sample users, age, gender, etc.

The degree of interest output by the product recommendation model may be represented by a score, and a higher score indicates that the candidate recommended user is more interested in the second product, in other examples, the product recommendation model may also be represented by a rating, specifically, a rating 1 and a rating 0 are output, where the rating 1 indicates that the candidate recommended user is interested in the second product, and the rating 0 indicates that the candidate recommended user is not interested in the second product, and so on, and the method for representing the degree of interest in the present disclosure is not limited.

In one possible embodiment, before the user features are input into the product recommendation model, the user features of the category type may be subjected to label encoder (tag encoder) encoding, where the category type features are features that are not originally represented by numerical values, such as gender: for men and women, to process these features for the product recommendation model, the features need to be numerically coded before being input into the model, for example, if male is coded as 2, female is coded as 3.

In S103, according to the interest degree, a candidate recommended user with a large interest degree of the second product is taken as a target recommended user, where the interest degree is expressed by a score, for example, if the score of the interest degree of one candidate recommended user output by the product recommendation model is greater than a preset threshold, the candidate recommended user is taken as the target recommended user, or the score of the interest degree of one candidate recommended user output by the product recommendation model is in the score ranking of the interest degrees of all candidate recommended users, and the ranking order from large to small is a preset ranking order, the candidate recommended user is taken as the target recommended user, so that each target recommended user is determined from all candidate recommended users.

In S104, it can be understood that the target recommending user is a user who is interested in the second product, so that the recommending success rate for recommending the second product to the target recommending user is high, and the target recommending user can conveniently obtain the relevant information of the product that the target recommending user is interested in, thereby improving the user experience.

In the foregoing method, a target recommended user is determined based on the interest level output by the product recommendation model, and a second product is recommended for the target recommended user, it can be understood that the accuracy of the interest level output by the candidate recommended user by the product recommendation model directly affects whether the second product can be recommended for a suitable user, based on which, the present disclosure provides a training method of the product recommendation model to improve the accuracy of the product recommendation model, as shown in fig. 2, the method includes:

s201, obtaining user characteristics of sample users.

S202, inputting the user characteristics into the original training model to obtain the interest degree of each sample user in the second product output by the original training model.

And S203, constructing training loss according to the interest degree.

And S204, adjusting the original training model according to the training loss to obtain a product recommendation model.

Wherein, in S201, the sample users include a negative sample user who owns the first product and does not own the second product within the historical time window, and a positive sample user who owns the first product and the second product within the historical time window.

The sample user has already been explained in the foregoing S102, and reference may be made to the related description of the foregoing S102, which is not described herein again.

In S202, the original training model is a model that has not been trained yet, and may be an XGBoost (eXtreme Gradient Boosting) model or a LightGBM (Light Gradient Boosting Machine, a distributed Gradient Boosting framework based on a decision tree algorithm), which is not limited by the present disclosure. The user characteristics of the sample users are input into the original training model, so that the original training model outputs the interest level of each sample user in the second product, and the interest level can be represented by a score or other forms as described above.

Wherein, in S203, the training loss is negatively correlated with the interest level of the positive sample user, and is positively correlated with the interest level of the negative sample user.

It is understood that the positive sample user is a user who owns both the first product and the second product in the historical time window, and who necessarily has a greater interest level in the second product to purchase the second product on the basis of owning the first product, while the negative sample user is a user who owns only the first product in the historical time window, and who will not subsequently purchase the second product due to the lesser interest level in the second product. It can be seen that, if the original training model is accurate, the output interest level of the positive sample user for the second product is relatively large, and the output interest level of the negative sample user for the second product is relatively small, in other words, if the interest level corresponding to the positive sample user is relatively small, the original training model is not accurate enough, the training loss is relatively large, correspondingly, if the interest level corresponding to the negative sample user is relatively large, the original training model is not accurate enough, the training loss is relatively large, and based on this, the training loss is constructed. In particular, the training loss may be represented by a loss function, such as a huber loss function, a softmax loss function, and so on, which is not limited by this disclosure.

In S204, the relevant parameters of the original training model are adjusted according to the training loss until the training loss constructed based on the interest level output by the adjusted original training model is smaller than a certain threshold, for example, the output positive sample user has a higher interest level in the second product, and the output negative sample user has a lower interest level in the second product, which may be regarded as that the adjusted original training model may accurately predict the interest level of the candidate recommended user in the second product, and then the original training model is used as the product recommendation model for subsequent use.

By adopting the embodiment, a loss function can be established for the original training model based on the characteristics of the positive sample user and the negative sample user in the sample users, the original training model is adjusted according to the loss function, and the adjusted original training model is guided to accurately output the interest degrees of the positive sample user and the negative sample user in the sample users on the second product, so that the product recommendation model capable of accurately predicting the interest degrees of the candidate recommendation users on the second product is obtained.

In the training process of the product recommendation model, the determination of the user characteristics is also important, if the user characteristics of the selected sample user change greatly with time, it can be understood that the user characteristics are not stable enough, and the relevant conditions of the user cannot be embodied accurately, and further, the product recommendation model obtained by training according to the user characteristics may not accurately predict the interest degree of the candidate recommended user in the second product, based on which, the disclosure also provides a training method of the product recommendation model, as shown in fig. 3, the method comprises:

s301, acquiring sample feature distribution of a sample user on each original feature dimension and cross-time feature distribution of a cross-time test user on each original feature dimension.

S302, the characteristic dimension of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets the preset condition is screened out from the original characteristic dimension to obtain a target characteristic dimension.

S303, obtaining the user characteristics of the sample user on the target characteristic dimension.

S304, inputting the user characteristics into the original training model to obtain the interest degree of each sample user in the second product output by the original training model.

S305, constructing training loss according to the interest degree.

And S306, adjusting the original training model according to the training loss to obtain a product recommendation model.

In S301, the cross-time test user is a user who owns the first product within the cross-time window, and the cross-time window does not overlap with the historical time window.

Because the cross-time window is not overlapped with the historical time window, the time when the cross-time test user owns the first product is not overlapped with the time when the sample user owns the first product, for example, if the historical time window corresponding to the sample user is 1-6 months, the cross-time window corresponding to the cross-time test user may be a month which is not overlapped with 1-6 months, such as 7 months or 8 months.

The original feature dimension is a feature dimension where all original features of the sample user obtained at the beginning are located, and may include dimensions of age, gender, health condition, income, and the like, the original features of each sample user in each original feature dimension have a corresponding distribution, and the distribution is used as the sample feature distribution of each sample user in each original feature dimension. And the cross-time feature distribution is the distribution of the original features of the user on the original feature dimension in the cross-time test.

Here, in S302, it can be understood that, since the sample user and the cross-time testing user are not in one time window, the initial possession time of the first product by the sample user and the cross-time testing user is greatly different. And if the characteristic dimension with larger distribution difference exists in the sample characteristic distribution and the cross-time characteristic distribution, the characteristic of the characteristic dimension can be greatly changed along with the change of time, and further the characteristic of the characteristic dimension is not stable enough, so that the characteristic dimension can be screened out. Therefore, the larger the difference between the sample feature distribution and the feature distribution in a certain feature dimension across time is, the more likely the feature dimension meets the preset condition, and for example, the preset condition may be that the distribution difference is greater than a specific threshold. And (4) screening out the feature dimensions meeting the preset conditions in the original feature dimensions, namely screening out unstable original features to obtain stable user features on the target feature dimensions capable of reflecting relevant characteristics of the sample user.

In one possible embodiment, the feature dimension satisfying the preset condition may be determined and screened by the following method:

s3021, evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the original characteristic dimension through the stability index to obtain a first evaluation result.

S3022, screening out the characteristic dimension of which the difference between the sample characteristic distribution and the cross-time characteristic distribution in the original characteristic dimension meets a first preset condition according to the first evaluation result, and obtaining a first characteristic dimension.

And S3023, evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the first characteristic dimension in a countercheck verification mode to obtain a second evaluation result.

And S3024, screening out the characteristic dimension in which the difference between the sample characteristic distribution and the cross-time characteristic distribution in the first characteristic dimension meets a second preset condition according to the second evaluation result, and obtaining a target characteristic dimension.

In S3021, the Stability indicator may reflect the Stability of the distribution of the sample feature distribution in each fraction segment (each feature dimension) and the distribution of the feature distribution across time, and the Stability indicator may be a PSI (position Stability Index) indicator, and a distribution difference between the sample feature distribution and the distribution of the feature distribution across time in the original feature dimension may be estimated by using the Stability indicator, so as to obtain a first estimation result of the distribution difference of the feature distribution in each feature dimension.

In S3022, according to the first evaluation result, feature dimensions that satisfy a first preset condition in the original feature dimensions are determined to be screened, so as to obtain a first feature dimension. The first preset condition may be that the difference of the feature distribution on the original feature dimension is greater than a preset threshold, and it can be understood that the larger the difference of the distribution of the sample feature distribution and the distribution of the cross-time feature distribution on a certain original feature dimension is, the more likely the original feature dimension meets the first preset condition.

In S3023, the countermeasure verification method (adaptive verification) is a method for determining whether the distributions of the training set (sample feature distribution) and the test set (cross-time feature distribution) are consistent, and when the countermeasure verification is used as a feature screening method, the feature with obvious timing fluctuation can be found, which helps us quickly find the feature of the feature dimension that is unstable in the original feature dimension. Therefore, the first feature dimension which is already subjected to the stability index evaluation and screening can be evaluated again based on the countermeasure verification mode, and a second evaluation result of the distribution difference of the feature distribution of each feature dimension in the first feature dimension can be obtained.

In S3024, according to the second evaluation result, feature dimensions that satisfy a second preset condition in each first feature dimension are determined to be screened, so as to obtain a target feature dimension. The second preset condition may be that the difference of the feature distribution in the first feature dimension is greater than a preset threshold, and it can be understood that the larger the difference between the distribution of the sample feature distribution and the distribution of the cross-time feature distribution in a certain feature dimension is, the more likely the feature dimension satisfies the second preset condition.

By selecting the embodiment, the stability index and the countermeasure verification mode can accurately evaluate the stability of the feature distribution, the distribution difference between the sample feature distribution and the cross-time feature distribution on the original feature dimension is evaluated through the stability index and the countermeasure verification mode, the target feature dimension is screened out according to the evaluation result, and the stability of the user features acquired on the target feature dimension subsequently is further improved.

In S303, it can be understood that the user characteristics of the sample user in the target characteristic dimension are stable, the user characteristics of the relevant characteristics of the sample user can be accurately reflected, and the original recommendation model is trained using the user characteristics, so that the interest level output by the product recommendation model obtained through subsequent training is more accurate.

In S304, the step S202 is the same, and reference may be specifically made to the related description of S202, which is not described herein again.

In S305, the step S203 is the same, which may specifically refer to the related description of S203, and is not described herein again.

In S306, the step S204 is the same, which may specifically refer to the related description of S204, and is not described herein again.

By selecting the embodiment, the user characteristics stably distributed in the target characteristic dimension are obtained purposefully by screening out the characteristic dimension with unstable characteristics in the original characteristic dimension, so that the product recommendation model obtained by subsequent training according to the user characteristics of the stable sample users is more accurate.

In order to determine the accuracy of the interest degree of the candidate recommended user predicted by the trained product recommendation model in the second product, the product recommendation model may be verified, and based on this, the present disclosure further provides a verification method of the product recommendation model, as shown in fig. 4, the method includes:

s401, obtaining the user characteristics of the verification sample.

S402, inputting the user characteristics of the verification samples into the product recommendation model to obtain the interest degree of each verification sample output by the product recommendation model.

S403, determining a matching degree between the interest degree and the negative sample user in the verification sample and the positive sample user in the verification sample.

And S404, verifying the product recommendation model according to the matching degree to obtain a verification result.

Wherein in S401, the verification sample comprises a negative sample user having the first product and not having the second product within the historical time window, and a positive sample user having the first product and the second product within the historical time window.

The verification sample is similar to the sample user and includes a negative sample user who owns the first product and does not own the second product in the historical time window and a positive sample user who owns the first product and the second product in the historical time window, but specifically, the sample user is a sample for training the product recommendation model, and the verification sample is a sample for verifying the product recommendation model, and the sample user and the verification sample include different specific samples, and if the sample user includes the user a, the user a is not included in the verification sample.

In S402, the step is similar to S102, except that the user characteristic of the candidate recommended user is input in S102, the interest level of the candidate recommended user in the second product is obtained, and the user characteristic of the verification sample is input in S402, and the interest level of the verification sample in the second product is obtained, which may specifically refer to the related description of S102, and is not described herein again.

In S403, as described above, the negative sample user does not have the second product in the historical time window, and is a user with a low interest level in the second product, so if the product recommendation model is accurate, the interest level of the negative sample user in the second product is low, the positive sample user has the second product in the historical time window, and is a user with a high interest level in the second product, and therefore, if the product recommendation model is accurate, the interest level of the negative sample user in the second product is high. Therefore, whether each interest degree is matched with the characteristics of each positive sample user and each negative sample user in the second product can be judged based on the interest degree of each positive sample user and each negative sample user in the verification sample output by the product recommendation model, that is, if the interest degree of the positive sample user output by the product recommendation model to the second product is high, the interest degree is matched with the positive sample user, and if the interest degree of the negative sample user output by the product recommendation model to the second product is low, the interest degree is matched with the positive sample user. Based on this, the degree of matching of each validation sample to the degree of interest is determined.

In one possible embodiment, the present disclosure provides a method of determining a degree of interest to match a verification sample, the method comprising:

and S4031, performing equal-frequency binning on the verification samples according to the interest degree to obtain each box body.

S4032, the positive sample rate of the positive sample user and/or the negative sample rate of the negative sample user in each box body verification sample are counted.

S4033, the matching degree of the positive sample rate and/or the negative sample rate and the box body is determined, and the matching degree is used as the matching degree of the negative sample user in the verification sample and the positive sample user in the verification sample.

Among them, in S4031, equal frequency binning is a method of discretizing data, and equal frequency binning makes each bin include as many values as possible. In the present disclosure, the equally frequency binning of the verification samples according to the interest degree may be performed by binning the verification samples according to the corresponding interest degree thereof according to the interest degree, for example, if the interest degree is expressed by a score, the verification sample with the interest degree of 1 may be placed in one box, the verification sample with the interest degree of 2 may be placed in another box, and so on, and the equally frequency binning may be performed on each verification sample.

In S4032, it can be understood that each box may include positive sample users and/or negative sample users, a ratio of the number of the positive sample users to the number of all the verification samples of the box is a positive sample rate, and a ratio of the number of the negative sample users to the number of all the verification samples of the box is a negative sample rate. For example, if there are 5 verification samples in a box, of which 1 is a positive sample user and 4 negative sample users, the positive sample rate of the box is 20% and the negative sample rate is 80%.

In S4033, the matching degree between the positive sample rate and the box and/or the negative sample rate and the box is the matching degree between the positive sample rate and the interest degree corresponding to each verification sample in the box, in other words, the higher the positive sample rate of one box and the higher the interest degree corresponding to the positive sample rate are, the higher the matching degree between the positive sample rate and the box is, and correspondingly, the lower the negative sample rate of one box and the lower the interest degree corresponding to the negative sample rate are, the higher the matching degree between the negative sample rate and the box is.

The boxes are divided according to the interest degrees corresponding to the verification samples, so that for one box, the box corresponds to one interest degree. Taking the interesting degree as an example, if the score of the interesting degree corresponding to the box a is 1, that is, the interesting degree is low, it indicates that the product recommendation model predicts that the interesting degree of each verification sample in the box to the second product is low, and the score of the interesting degree corresponding to the box B is 6, that is, the interesting degree is high, it indicates that the product recommendation model predicts that the interesting degree of each verification sample in the box to the second product is high, it can be understood that, theoretically, the number of people in the box a who owns both the first product and the second product should be less than that in the box B, that is, the positive sample rate of the box a should be less than that of the box B, and if the positive sample rate in the box a is counted as 60%, and the positive sample rate in the box B is counted as 20%, it indicates that the positive sample rates of the box a and the box B are not matched with those in the box a and the box B. Therefore, the matching degree of the box body and the positive sample rate and/or the negative sample rate is used as the matching degree of the negative sample user in the verification sample and the positive sample user in the verification sample, and the obtained matching degree is more accurately determined.

By adopting the embodiment, the matching degree of the verification sample and the interest degree of the verification sample output by the product recommendation model to the second product can be accurately determined through a discrete statistical method, so that whether the product recommendation model is accurate or not is determined according to the matching degree subsequently to obtain the verification result, the accuracy of obtaining the verification result of the product recommendation model through the method is improved, and the recommendation accuracy of the product recommendation method is further improved.

In S404, it can be understood that, if the interest degree is not matched with the verification sample for a large amount, it indicates that the interest degree of the second product by the verification sample output by the product recommendation model is not accurate, and if the interest degree is matched with the verification sample for a large amount or all the verification samples are matched, the interest degree of the second product by the specification product recommendation model is accurate, and based on this, the verification result of the product recommendation model is obtained.

By adopting the embodiment, the accuracy of the product recommendation model is verified through the verification sample to obtain the verification result, so that the product recommendation model can be adjusted and then used according to the verification result, and the recommendation accuracy of the product recommendation method is improved.

In order to better explain the specific implementation process of the training method of the product recommendation model, the present disclosure further provides a flow diagram of the training method of the product recommendation model, as shown in fig. 5, which will be described below with reference to fig. 5:

s501, defining a sample and determining a sample user.

S502, data extraction, wherein the sample user and all the characteristics possibly used by the sample user are led into a modeling environment.

And S503, backtracking and generating the original features, and generating the original features corresponding to each sample user by using expert experience and an automatic feature generation tool.

And S504, feature screening, namely screening the original features to obtain the user features under the target feature dimension.

And S505, training a model, and training the original recommendation model through user characteristics to obtain a product recommendation model.

And S506, evaluating the model, namely evaluating the product recommendation model by a method of counting the positive sample rate through equal frequency binning, and determining the accuracy of the interest degree of the candidate recommendation user output by the product recommendation model.

According to an embodiment of the present disclosure, there is also provided a product recommendation device, as shown in fig. 6, including:

a candidate determining module 601, configured to determine a candidate recommended user who owns the first product and does not own the second product;

an interest obtaining module 602, configured to input user characteristics of the candidate recommended users into a product recommendation model, and obtain a degree of interest of each candidate recommended user output by the product recommendation model in a second product, where the product recommendation model is obtained in advance through user characteristic training of sample users, and the sample users include negative sample users who own a first product and do not own the second product within a historical time window, and positive sample users who own the first product and the second product within the historical time window;

a target determining module 603, configured to determine a target recommended user from the candidate recommended users according to the interest level;

a product recommending module 604, configured to recommend the second product to the target recommending user.

According to an embodiment of the present disclosure, there is also provided a training apparatus for a product recommendation model, as shown in fig. 7, including:

a feature obtaining module 701, configured to obtain user features of sample users, where the sample users include a negative sample user who owns a first product and does not own a second product within a historical time window, and a positive sample user who owns the first product and the second product within the historical time window;

an initial training module 702, configured to input the user characteristics into an original training model, to obtain a degree of interest of each sample user in the second product, where the degree of interest is output by the original training model;

a model adjusting module 703 for constructing a training loss according to the interest level, wherein the training loss is negatively correlated to the interest level of the positive sample user and positively correlated to the interest level of the negative sample user;

a module obtaining module 704, configured to adjust the original training model according to the training loss to obtain a product recommendation model.

In a possible embodiment, further comprising:

the distribution acquisition module is used for acquiring sample feature distribution of sample users on each original feature dimension and cross-time feature distribution of cross-time test users on each original feature dimension, wherein the cross-time test users are users who own the first product in a cross-time window, and the cross-time window is not overlapped with the historical time window;

the characteristic screening module is used for screening out the characteristic dimension of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a preset condition from the original characteristic dimension to obtain a target characteristic dimension;

the feature obtaining module 701 is specifically configured to obtain a user feature of the sample user in the target feature dimension.

In one possible embodiment, the feature filtering module includes:

the first evaluation acquisition submodule is used for evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the original characteristic dimension through a stability index to obtain a first evaluation result;

the first screening submodule is used for screening out the characteristic dimension, in the original characteristic dimension, of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a first preset condition according to the first evaluation result to obtain a first characteristic dimension;

the second evaluation acquisition sub-module is used for evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the first characteristic dimension in a confrontation verification mode to obtain a second evaluation result;

and the second screening submodule is used for screening out the characteristic dimension of which the difference between the sample characteristic distribution and the cross-time characteristic distribution in the first characteristic dimension meets a second preset condition according to the second evaluation result to obtain a target characteristic dimension.

In a possible embodiment, further comprising:

a verification acquisition module for acquiring user characteristics of a verification sample, the verification sample including a negative sample user who owns the first product and does not own the second product within the historical time window, and a positive sample user who owns the first product and the second product within the historical time window;

the verification input module is used for inputting the user characteristics of the verification samples into the product recommendation model to obtain the interest degree of each verification sample output by the product recommendation model;

a verification matching module for determining a degree of matching of the degree of interest with a negative sample user in the verification sample and a positive sample user in the verification sample;

and the verification result module is used for verifying the product recommendation model according to the matching degree to obtain a verification result.

In one possible embodiment, the verification matching module comprises:

the box body submodule is used for performing equal-frequency box separation on the verification sample according to the interest degree to obtain each box body;

the sample submodule is used for counting the positive sample rate of the positive sample user and/or the negative sample rate of the negative sample user in the verification sample in each box body;

and the matching submodule is used for determining the matching degree of the positive sample rate and/or the negative sample rate and the box body, and the matching degree is used as the matching degree of the negative sample user in the verification sample and the positive sample user in the verification sample.

In the technical scheme of the disclosure, the collection, storage, use, processing, transmission, provision, disclosure and other processing of the personal information of the related user are all in accordance with the regulations of related laws and regulations and do not violate the good customs of the public order.

The present disclosure also provides an electronic device, a readable storage medium, and a computer program product according to embodiments of the present disclosure.

FIG. 8 illustrates a schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 8, the apparatus 800 includes a computing unit 801 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 802 or a computer program loaded from a storage unit 808 into a Random Access Memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The calculation unit 801, the ROM 802, and the RAM 803 are connected to each other by a bus 804. An input/output (I/O) interface 805 is also connected to bus 804.

A number of components in the device 800 are connected to the I/O interface 805, including: an input unit 806 such as a keyboard, a mouse, or the like; an output unit 807 such as various types of displays, speakers, and the like; a storage unit 808, such as a magnetic disk, optical disk, or the like; and a communication unit 809 such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information/data with other devices via a computer network such as the internet and/or various telecommunication networks.

Computing unit 801 may be a variety of general and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and the like. The calculation unit 801 executes the respective methods and processes described above, such as the product recommendation method and the training method of the product recommendation model. For example, in some embodiments, the product recommendation method and the training method of the product recommendation model may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and/or installed onto device 800 via ROM 802 and/or communications unit 809. When loaded into RAM 803 and executed by the computing unit 801, the computer programs may perform one or more steps of the product recommendation method and training method of the product recommendation model described above. Alternatively, in other embodiments, the computing unit 801 may be configured to perform the product recommendation method and the training method of the product recommendation model in any other suitable manner (e.g., by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), system on a chip (SOCs), complex Programmable Logic Devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions/operations specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server with a combined blockchain.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present disclosure may be executed in parallel or sequentially or in different orders, and are not limited herein as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved.

The above detailed description should not be construed as limiting the scope of the disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A product recommendation method, comprising:

and recommending the second product to the target recommending user.

2. A training method of a product recommendation model comprises the following steps:

constructing a training loss according to the interest level, wherein the training loss is negatively correlated to the interest level of the positive sample user and positively correlated to the interest level of the negative sample user;

3. The method of claim 2, further comprising:

obtaining sample feature distribution of sample users on each original feature dimension and cross-time feature distribution of cross-time test users on each original feature dimension, wherein the cross-time test users are users who own the first product in a cross-time window, and the cross-time window is not overlapped with the historical time window;

screening out the characteristic dimension of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a preset condition from the original characteristic dimension to obtain a target characteristic dimension;

the acquiring of the user characteristics of the sample user comprises:

and acquiring the user characteristics of the sample user on the target characteristic dimension.

4. The method of claim 3, wherein the step of screening out the feature dimension, which is different from the sample feature distribution and the cross-time feature distribution by a preset condition, from the original feature dimension to obtain a target feature dimension comprises the steps of:

evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the original characteristic dimension through a stability index to obtain a first evaluation result;

screening out a characteristic dimension, in the original characteristic dimension, of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a first preset condition according to the first evaluation result to obtain a first characteristic dimension;

evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the first characteristic dimension in a countercheck verification mode to obtain a second evaluation result;

and screening out the characteristic dimension, in the first characteristic dimension, of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a second preset condition according to the second evaluation result to obtain a target characteristic dimension.

5. The method of claim 2, further comprising:

obtaining user characteristics of a verification sample comprising a negative sample user possessing the first product and not possessing the second product within the historical time window, and a positive sample user possessing the first product and the second product within the historical time window;

inputting the user characteristics of the verification samples into the product recommendation model to obtain the interest degree of each verification sample output by the product recommendation model;

determining a degree of matching of the degree of interest to a negative sample user in the verification sample and a positive sample user in the verification sample;

and verifying the product recommendation model according to the matching degree to obtain a verification result.

6. The method of claim 5, wherein said determining a degree of matching of said degree of interest to a negative sample user in said verification sample and a positive sample user in said verification sample comprises:

performing equal-frequency binning on the verification samples according to the interest degree to obtain box bodies;

counting the positive sample rate of the positive sample user and/or the negative sample rate of the negative sample user in the verification sample in each box;

and determining the matching degree of the positive sample rate and/or the negative sample rate and the box body as the matching degree of the negative sample user in the verification sample and the positive sample user in the verification sample.

7. A product recommendation device comprising:

the interest obtaining module is used for inputting the user characteristics of the candidate recommending users into a product recommending model to obtain the interest degree of each candidate recommending user output by the product recommending model to a second product, wherein the product recommending model is obtained by training the user characteristics of sample users in advance, and the sample users comprise negative sample users who possess a first product and do not possess the second product in a historical time window and positive sample users who possess the first product and the second product in the historical time window;

8. A training apparatus for a product recommendation model, comprising:

9. The apparatus of claim 8, further comprising:

the feature obtaining module is specifically configured to obtain user features of the sample user in the target feature dimension.

10. The apparatus of claim 9, wherein the feature screening module comprises:

the second evaluation obtaining sub-module is used for evaluating the distribution difference of the sample characteristic distribution and the cross-time characteristic distribution on the first characteristic dimension in a countermeasure verification mode to obtain a second evaluation result;

and the second screening submodule is used for screening out the characteristic dimension, in the first characteristic dimension, of which the difference between the sample characteristic distribution and the cross-time characteristic distribution meets a second preset condition according to the second evaluation result so as to obtain a target characteristic dimension.

11. The apparatus of claim 8, further comprising:

12. The apparatus of claim 11, wherein the verification matching module comprises:

13. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.

14. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-6.

15. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-6.