CN113076484A

CN113076484A - Product recommendation method, device, equipment and storage medium based on deep learning

Info

Publication number: CN113076484A
Application number: CN202110459644.7A
Authority: CN
Inventors: 林思涵; 张雷妮; 张奕宁; 卓全娇; 曾璐; 张文新
Original assignee: China Construction Bank Co ltd Shenzhen Branch
Current assignee: China Construction Bank Co ltd Shenzhen Branch
Priority date: 2021-04-27
Filing date: 2021-04-27
Publication date: 2021-07-06

Abstract

The application provides a product recommendation method, a product recommendation device, product recommendation equipment and a storage medium based on deep learning, wherein the method comprises the steps of determining a candidate product library corresponding to a target user; screening a plurality of products to be recommended from a candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule; analyzing the product characteristics of each product to be recommended and the user characteristics of a target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model to obtain the predicted response probability of each product to be recommended; and recommending part of products to be recommended to the user according to the predicted response probability. According to the scheme, the user characteristics of the target user are analyzed by using the deep learning technology, the potential preference of the specific user to each product can be deeply mined, and therefore the product recommendation accuracy is improved.

Description

Product recommendation method, device, equipment and storage medium based on deep learning

Technical Field

The invention relates to the technical field of deep learning, in particular to a product recommendation method, device, equipment and storage medium based on deep learning.

Background

Banks often need to recommend various products to users, such as credit cards, financing, funds and the like, and the current product recommendation method generally adopts a traditional classification algorithm to classify different users according to the attributes of the users, such as dividing different intervals according to income per year, classifying according to different professions, and then recommending corresponding products for users of specific classes.

The method has the problems that the preferences of different users in the same category to various products are generally different, the traditional classification algorithm can only predict the products which are possibly preferred by the users in each category, the personalized preferences of each user cannot be predicted, the accuracy of the finally obtained recommendation result is low, and the recommended products are often not in accordance with the actual preferences of the users.

Disclosure of Invention

In view of the problems in the prior art, the present application provides a product recommendation method, apparatus, device and storage medium based on deep learning, so as to improve the accuracy of the product recommendation technology.

The application provides a product recommendation method based on deep learning in a first aspect, which comprises the following steps:

determining a candidate product library corresponding to a target user; wherein the candidate product library contains a plurality of products not purchased by the target user;

screening a plurality of products to be recommended from the candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule;

for each product to be recommended, analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by utilizing a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model to obtain the predicted response probability of the product to be recommended; the user characteristics are determined according to a first element representing basic information of a target user, a second element representing product holding conditions of the target user and a third element representing consumption preferences of the target user;

sequencing the products to be recommended according to the predicted response probability of the products to be recommended from large to small to obtain a list to be recommended;

and pushing product information of the first N products to be recommended to the target user according to the disturbance-free strategy set by the target user.

Optionally, the screening, by using the word vector model, the collaborative filtering technology, the implicit semantic model and the preset service recommendation rule, of the candidate product library to obtain a plurality of products to be recommended includes:

training to obtain a word vector model by taking the purchased product of the target user as a training corpus;

calculating by using the word vector model to obtain a product vector corresponding to each product in the candidate product library;

selecting a first recommended product from the candidate product library according to the similarity between the product vector of each product in the candidate product library and the product vector of the purchased product; wherein a similarity between the product vector of the first recommended product and the product vector of the purchased product is greater than a first threshold;

determining a plurality of similar users similar to the target user;

determining the products purchased by the similar users in the candidate product library as second recommended products;

obtaining interaction records of a plurality of users including the target user for a part of products in the candidate product library, and generating a product scoring matrix according to the interaction records; wherein the product scoring matrix comprises scores of the plurality of users for a portion of the products in the candidate product library;

performing matrix decomposition on the product scoring matrix to obtain a user matrix and a product matrix;

calculating to obtain the score of the target user for each product of the candidate product library according to the user matrix and the product matrix;

determining products with the scores of the target users larger than a second threshold value in the candidate product library as third recommended products;

determining a business line to which each product in the candidate product library belongs;

determining the product of which the business line is a priority recommendation line specified by the business recommendation rule in the candidate product library as a fourth recommendation product;

and determining the first recommended product, the second recommended product, the third recommended product and the fourth recommended product as products to be recommended.

Optionally, the analyzing, by using a pre-constructed logistic regression model, a gradient boosting decision tree and a deep learning model, the product characteristics of the product to be recommended and the user characteristics of the target user to obtain the predicted response probability of the product to be recommended includes:

analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model to obtain a first response probability of the product to be recommended;

analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed gradient lifting decision tree to obtain a second response probability of the product to be recommended;

analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed deep learning model to obtain a third response probability of the product to be recommended;

and carrying out weighted average on the first response probability, the second response probability and the third response probability of the product to be recommended to obtain the predicted response probability of the product to be recommended.

Optionally, the pushing, to the target user, product information of the first N products to be recommended in the list to be recommended according to the disturbance-free policy set by the target user includes:

judging whether the current time is within a pushable time period specified by a disturbance-free strategy set by the target user;

if the current time is within the push time period, calculating the time interval from the previous push time to the current time;

if the time interval is greater than or equal to an interval threshold value specified by the disturbance-free strategy, acquiring product information of any one non-recommended product in the list to be recommended; the non-recommended products refer to products of which the product information is not pushed to the target user in the first N products of the list to be recommended;

and pushing the obtained product information to the target user through a preset pushing channel so as to finish one-time pushing.

Optionally, after pushing the product information of each product to be recommended in the list to be recommended to the target user one by one according to the disturbance-free policy set by the target user, the method further includes:

acquiring feedback information of the target user; the feedback information is used for representing the feedback condition of the target user to the pushed product information;

and updating the logistic regression model, the gradient lifting decision tree and the deep learning model according to the feedback information of the target user.

The second aspect of the present application provides a product recommendation device based on deep learning, including:

the determining unit is used for determining a candidate product library corresponding to the target user; wherein the candidate product library contains a plurality of products not purchased by the target user;

the screening unit is used for screening a plurality of products to be recommended from the candidate product library by utilizing a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule;

the analysis unit is used for analyzing the product characteristics of the products to be recommended and the user characteristics of the target users by utilizing a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model aiming at each product to be recommended to obtain the predicted response probability of the products to be recommended; the user characteristics are determined according to a first element representing basic information of a target user, a second element representing product holding conditions of the target user and a third element representing consumption preferences of the target user;

the sorting unit is used for sorting the products to be recommended according to the predicted response probability of the products to be recommended in a descending order to obtain a list to be recommended;

and the pushing unit is used for pushing the product information of the front N products to be recommended in the list to be recommended to the target user according to the disturbance-free strategy set by the target user.

Optionally, the analyzing unit analyzes the product features of the product to be recommended and the user features of the target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model, and when obtaining the predicted response probability of the product to be recommended, the analyzing unit is specifically configured to:

carrying out weighted average on the first response probability, the second response probability and the third response probability of the product to be recommended to obtain the predicted response probability of the product to be recommended

Optionally, the apparatus further includes an updating unit, configured to:

A third aspect of the present application provides a computer storage medium for storing a computer program, which when executed is specifically configured to implement the deep learning based product recommendation method provided in any one of the first aspects of the present application.

A fourth aspect of the present application provides an electronic device comprising a memory and a processor;

wherein the memory is for storing a computer program;

the processor is configured to execute the computer program, and in particular, to implement the deep learning based product recommendation method provided in any one of the first aspects of the present application.

The application provides a product recommendation method, a product recommendation device, product recommendation equipment and a storage medium based on deep learning, wherein the method comprises the steps of determining a candidate product library corresponding to a target user; wherein the candidate product library contains a plurality of products not purchased by the target user; screening a plurality of products to be recommended from a candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule; for each product to be recommended, analyzing the product characteristics of the product to be recommended and the user characteristics of a target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model to obtain the predicted response probability of the product to be recommended; the user characteristics are determined according to a first element representing basic information of a target user, a second element representing product holding conditions of the target user and a third element representing consumption preferences of the target user; according to the predicted response probability of each product to be recommended, sequencing each product to be recommended in a descending order to obtain a list to be recommended; and pushing product information of the first N products to be recommended to the target user according to a disturbance-free strategy set by the target user. According to the scheme, the user characteristics of the target user are analyzed by using the deep learning technology, the potential preference of the specific user to each product can be deeply mined, and therefore the product recommendation accuracy is improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.

Fig. 1 is a flowchart of a deep learning-based product recommendation method according to an embodiment of the present application;

fig. 2 is a schematic structural diagram of a deep learning-based product recommendation device according to an embodiment of the present application;

fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

At present, the operation mode of customers still stays in a mode taking products as centers, and respective marketing channels mastered by business departments cause repeated marketing to customers, and simultaneously, the customers are difficult to master comprehensive insight on user requirements. There is a lack of a customer-centric intelligent recommendation engine. Therefore, a unified intelligent bank product recommendation engine with customers as centers needs to be constructed, customer experience is optimized, and marketing cost is reduced.

In the aspect of technical implementation, product recommendation at the present stage mainly utilizes a traditional classification algorithm to classify users according to partial attributes, predicts the products preferred by the users in each category, and recommends the preferred products to all the users in the category, so that generalization capability is weak and flexibility is poor. Therefore, it is necessary to explore the application of deep learning models in recommendation systems to improve the generalization and automatic learning capabilities of the models.

In addition, in the current product recommendation technology, the marketing process records of the customers in all channels are not uniformly integrated, the process monitoring and effect evaluation of marketing are not performed, the user feedback of marketing at each time is not mastered, and the basis for optimizing marketing models and strategies is not provided. Therefore, it is necessary to construct a unified marketing process recording standard, incorporate the recording standard into a data warehouse, form a unified full-channel historical marketing client library, and update the client feedback situation in real time, so as to fully support the iterative optimization of marketing strategies.

To solve the problems in the prior art, an embodiment of the present application provides a product recommendation method based on deep learning, and the method provided by the present application has the following effects:

on one hand, aiming at products such as credit cards, stages, individual credits, deposits, financing and funds, a multi-product intelligent recommendation engine taking a client as a center is provided, an optimal marketing scheme and a pushing strategy are automatically generated for the client, diversified demands and preferences of the client are met, the client is served for the public, and homogenization competition is avoided while the client experience is upgraded.

In the second aspect, a Wide & Deep learning model (a Deep learning model fusing a shallow model and a Deep model) derived from Google is innovatively applied to a recommendation algorithm, the Deep level behavior characteristics of customers are automatically mined, potential patterns in data are deeply explored, and the generalization and learning capability of the model is greatly improved.

And in the third aspect, the data of the whole channel is opened up in a comprehensive manner, a marketing effect monitoring closed-loop system is set up, the marketing activity effect is comprehensively developed in an all-round manner, and an important feedback effect is played for optimizing the model and the strategy. And an automatic marketing closed loop from modeling, marketing and feedback to continuous optimization is realized.

Referring to fig. 1, the product recommendation method based on deep learning provided by the present application may specifically include the following steps:

and S101, determining a candidate product library corresponding to the target user.

Wherein the candidate product library contains a plurality of products not purchased by the target user.

The target user can be any user needing product recommendation.

The candidate product library may be determined as follows:

for a target user needing product recommendation, all products purchased by the user are acquired first, then all the products purchased by the target user are deleted from all the products currently operated, and the remaining products form a candidate product library of the target user, that is, a set of products not purchased by the target user is determined as the candidate product library of the target user.

S102, screening a plurality of products to be recommended from a candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule.

In step S102, the candidate product library may be screened in four ways, i.e., a word vector model, a collaborative filtering technology, a hidden semantic model, and a preset service recommendation rule, respectively, and all products obtained after the four ways are screened are determined as products to be recommended.

That is, through step S102, the method may perform coarse screening on the candidate product library by using a word vector model, a collaborative filtering, an implicit semantic model (LFM) model, and a service recommendation rule preset by each service line of the bank, for the candidate product library corresponding to the client, so as to quickly and accurately screen out irrelevant articles, thereby saving the calculation resources consumed by the subsequent sorting layer.

The word vector model may be understood as an embedding (embedding) technology, which refers to taking historical purchase sequence data of a target user (including all products purchased by the target user) as a corpus, training a word2vec (word-vector) model, where the trained word2vec model includes an embedding (embedding) vector of each product purchased by the target user, and using the trained word2vec model to further calculate embedding vectors corresponding to all products in a candidate product library, where the embedding vector corresponding to each product may also be referred to as a product vector of the product.

Finally, by calculating the similarity of the product vectors of every two products, a plurality of products which are possibly interested by the target user (namely, the client) can be screened from the candidate product library, namely, Top-N products which are most interested by the client are generated and are taken as part of the recommended products.

S103, analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model to obtain the predicted response probability of the product to be recommended.

The user characteristics are determined according to a first element representing basic information of the target user, a second element representing product holding conditions of the target user and a third element representing consumption preferences of the target user.

It should be noted that, in step S103, the execution of step S102 is performed on each item of product to be recommended, that is, each item of product to be recommended determines a predicted response probability through step S103.

The first element may specifically include personal information such as occupation, address, age, name, gender, and physical health condition of the target user in the recent period.

The second element may include product information of each product currently held by the target user, for example, which products the target user specifically holds, where the categories of the products are distributed, for example, there are several credit card products, there are several fund products, etc., and may further include consumption or usage of each product by the user, for example, for a credit card product, how much amount the target user has consumed with a credit card in the last several months may be included in the second element, and for a fund product, the amount of each fund purchased by the target user may be included in the second element.

The third element may include information capable of reflecting the consumption preference of the target user, for example, information about products collected by the target user in a recent period of time, top K products viewed by the target user most frequently in a recent month, topics of articles viewed by the target user in the recent month, and the like.

In step S103, product features of the product to be recommended and user features of the target user may be analyzed by using three algorithms, namely, a logistic regression model, a gradient boosting decision tree and a deep learning model, respectively, and then analysis results of the three algorithms are integrated to obtain a predicted response probability of the product to be recommended.

The predicted response probability of the product to be recommended may be considered as a probability that the product information of the product to be recommended is interested by the target user, or a probability that the target user browses the product information after pushing the product information of the product to be recommended to the target user.

Through the step S103, the scheme can construct a logistic regression model, a GBDT model and a Deep learning Wide & Deep model based on factors such as basic information of customers, product holding conditions and consumption preference, and train to obtain the response probability of the target user to each product, so that the accurate sequencing of multiple products is realized.

S104, sorting the products to be recommended according to the predicted response probability of the products to be recommended in a descending order to obtain a list to be recommended.

After the sorting is finished, the first product in the list to be recommended is the product with the highest predicted response probability in all the products to be recommended, the predicted response probability of the second product is only the first product, and the subsequent products are sorted in the same way.

And S105, pushing the product information of the first N products to be recommended in the list to be recommended to the target user according to the disturbance-free strategy set by the target user.

N is a preset positive integer, and may be specifically set according to the number of products in the candidate product library, for example, when the candidate product library includes 20 products, N may be set to 5.

In the above embodiment, step S102 is to obtain a plurality of products to be recommended by screening from the candidate product library by using the word vector model, the collaborative filtering technology, the implicit semantic model and the preset service recommendation rule, and may be specifically executed according to the following process.

Firstly, the process of screening products by using a word vector model is as follows:

in the process, firstly, the purchased product of the target user is used as a training corpus to train to obtain a word vector model, and a specific training method can refer to the related prior art, which is not limited herein.

And after the training is finished, calculating by using a word vector model to obtain a product vector corresponding to each product in the candidate product library.

And selecting a first recommended product from the candidate product library according to the similarity between the product vector of each product in the candidate product library and the product vector of the purchased product.

The product vector for the purchased product may be obtained from the trained word vector model.

Wherein a similarity between the product vector of the first recommended product and the product vector of the purchased product is greater than a first threshold.

Specifically, the following determination process may be performed for each product of the candidate product library:

calculating the similarity between the product vector of the product and the product vector of each product purchased by the target user to obtain a plurality of similarities, if the target user purchases P products, calculating the similarity between the product and the product vectors of the P purchased products, namely obtaining P similarities, then judging whether the similarities are greater than a first threshold value, if so, determining the product in the candidate product library as a first recommended product, and if not, determining the product is not the first recommended product.

The method for screening the product by utilizing the collaborative filtering technology comprises the following steps:

first, a plurality of similar users similar to the target user are determined.

For example, a user who works in the same enterprise as the target user and has a similar annual income may be determined as a similar user of the target user.

And determining the products purchased by the similar users in the candidate product library as second recommended products.

Specifically, the products purchased by each similar user in the last several months may be obtained, then the products purchased by the similar users may be searched one by one in the candidate product library, and then the searched product may be determined as the second recommended product.

For example, the candidate product library contains three products a, B and C, and the product B can be determined as the second recommended product by finding that the product B is a product that is recently purchased by a similar user through the above steps.

The process of screening using the implied semantic model is as follows:

firstly, interaction records of a plurality of users including a target user for a part of products in a candidate product library are obtained, and a product scoring matrix is generated according to the interaction records.

The product scoring matrix comprises scores of a plurality of users for a part of products in the candidate product library. The user's score for a product may be determined according to whether the user purchases the product and the number of purchases, whether the user browses product information and the number of browsing times of the product, and whether the user collects the product, and the score interval may be set to 1 to 10, and if the user purchases a product many times, the user's score for the product is set to 10, if the user does not purchase the product, but browses product information of the product many times and collects the product, the score is set to 8, and if the user does not purchase the product at all, the score is set to 1.

Assuming that the obtained users are user 1 to user 5 in turn, and the products included in the candidate product library are product a to product E in turn, the scoring matrix can be represented by the following table 1:

TABLE 1

	Product A	Product B	Product C	Product D	Product E
						User 1	2		3
User 2
						User 3		6
User 4				8
						User 5	7		9

And the numerical value in each cell represents the grade of the user in the corresponding row for the product in the corresponding column. If the blank cell indicates that the corresponding user is not pushed the product information of the corresponding product.

And performing matrix decomposition on the product scoring matrix to obtain a user matrix and a product matrix.

For the scoring matrix shown in table 1, blank cells in the scoring matrix may be filled with 0, so as to obtain a complete matrix, and then the scoring matrix may be decomposed by using the existing matrix decomposition algorithm, so as to obtain a user matrix and a product matrix.

Taking table 1 as an example, the matrix shown in table 1 can be decomposed into a user matrix with 5 rows and L columns and a product matrix with L rows and 5 columns, where L is a positive number determined by the matrix decomposition algorithm.

Each row in the user matrix corresponds to a user and each column in the product matrix corresponds to a product.

And calculating to obtain the score of the target user for each product in the candidate product library according to the user matrix and the product matrix.

Taking table 1 as an example, assuming that user 2 is a target user, when calculating the score, the row corresponding to user 2 in the user matrix may be multiplied by the column corresponding to product a in the product matrix, specifically, the numerical values of the corresponding positions in the row and the column are multiplied and then the multiple products are added, the obtained result is taken as the score of user 2 on product a, similarly, the row corresponding to user 2 and the column corresponding to product B may be multiplied to obtain the score of user 2 on product B, and so on, the scores of user 2 on products a to E may be finally obtained.

Finally, products in the candidate product library with the score larger than the second threshold value of the target user can be determined as third recommended products.

For example, the second threshold may be set to 5, and then a product having a score greater than 5 for the target user may be determined as the third recommended product.

The screening process according to the business recommendation rule is as follows:

and determining the business line of each product in the candidate product library.

The business line may specifically be a credit card, a fund, a personal loan, etc.

And determining the product of which the business line belongs to the candidate product library is the priority recommendation line specified by the business recommendation rule as a fourth recommended product.

That is, some service lines may be specified as preferred recommendation lines in the service recommendation rule, for example, a credit card line may be specified as a preferred recommendation line, and then, in this process, all products belonging to the credit card class in the candidate product library may be determined as fourth recommended products.

After the first recommended product, the second recommended product, the third recommended product and the fourth recommended product are obtained through screening in the four ways, the first recommended product, the second recommended product, the third recommended product and the fourth recommended product can be determined as products to be recommended.

It is understood that there may be multiple items of the first to fourth recommended products, and there may be overlap between the items, for example, one item may be both the first recommended product and the second recommended product, so that in the implementation, the first to fourth recommended products may be regarded as four product sets, and the products included in the union of the four product sets are determined as the products to be recommended.

Optionally, in step S103, analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model, a gradient boosting decision tree and a deep learning model to obtain the predicted response probability of the product to be recommended, which may be specifically executed in the following manner.

The product characteristics may include the category to which the product belongs, the product vector generated by using the word vector model in step S102, the accumulated number of purchasers of the product, and other data.

Firstly, analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model to obtain a first response probability of the product to be recommended.

The logistic regression model may be constructed from a plurality of recommended samples. The recommended sample may be generated from push actions that have been performed in the past. For example, each time product information of a product is pushed to a user, the user characteristics of the user and the product characteristics of the product can be combined into a recommendation sample, meanwhile, interaction behaviors (such as browsing information, collecting information, purchasing the product, ignoring information, deleting information and the like) between the user and the pushed product information are recorded as recommendation results of the recommendation sample, and then, the logistic regression model can be constructed by using a plurality of recommendation samples and corresponding recommendation results.

And analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed gradient lifting decision tree to obtain a second response probability of the product to be recommended.

The gradient boosting decision tree can also be constructed by the plurality of recommendation samples and the corresponding recommendation results.

And analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed deep learning model to obtain a third response probability of the product to be recommended.

Specifically, the Deep learning model may be a Wide & Deep model, the model is formed by fusing a shallow model (i.e., Wide model) and a Deep model (Deep model), when analyzing product features of a product to be recommended and user features of a target user by using the Wide & Deep model, discrete features and manual cross features in the product to be recommended, such as customer product holding conditions, large preference and the like, may be specifically input into the shallow model, continuous variables in the features and dense features after being subjected to Embedding mapping, and product vectors of the product to be recommended generated by using the word vector model in the foregoing steps may be input into the Deep model, and the shallow model and the Deep model are jointly trained, so that the model has both memory capability and generalization capability.

Specifically, corresponding weight coefficients may be allocated to the three analysis methods in advance, for example, a logistic regression model is allocated with a weight coefficient of 0.2, a gradient boosting decision tree is allocated with a weight coefficient of 0.2, a deep learning model is allocated with a weight coefficient of 0.6, and finally, a weighted average of the first response probability, the second response probability and the third response probability is calculated according to the weight coefficients of the corresponding analysis methods to obtain a predicted response probability of the product to be recommended.

The specific construction methods of the logistic regression model, the gradient boosting decision tree and the deep learning model can refer to the related prior art, and are not limited herein.

It should be noted that, in step S103, each product to be recommended obtains a predicted response probability, and if products a, B, and C to be recommended are assumed, in step S103, the product features of the product a and the user features of the target user are analyzed to obtain the predicted response probability of the product a, the product features of the product B and the user features of the target user are analyzed to obtain the predicted response probability of the product B, and the product features of the product C and the user features of the target user are analyzed to obtain the predicted response probability of the product C.

Optionally, the disturbance-free policy may include a preset pushable time period and an interval threshold, step S105, that is, a step of pushing product information of the first N products to be recommended in the list to be recommended to the target user according to the disturbance-free policy set by the target user, where a specific execution process may be:

judging whether the current time is within a pushable time period specified by a disturbance-free strategy set by a target user;

if the current time is within the pushing time period, calculating the time interval from the previous pushing time to the current time;

if the time interval is greater than or equal to the interval threshold value specified by the disturbance-free strategy, acquiring product information of any one non-recommended product in the list to be recommended; the non-recommended products refer to the products of which the product information is not pushed to the target user in the first N products of the list to be recommended;

and pushing the obtained product information to a target user through a preset pushing channel so as to finish one-time pushing.

In general, in the pushable time period, every time an interval threshold value is passed, a push is executed, and each push pushes product information of a product to be recommended to a target user.

The push channel may include a short message of a mobile phone, a short message of social media, a software popup, a video advertisement, and the like.

The pushing specifically means that product information of a product to be recommended is displayed on terminal equipment held by a target user through various pushing channels, such as a smart phone, a computer, and the like.

The pushable time period and the interval threshold in the disturbance-free strategy can be specified by a user, and a customer disturbance-free strategy can also be set by combining with historical marketing touch records.

For example, in the past, it is found that the product information pushed in the period from 10:00 pm to 11:00 pm is generally clicked by the user, and then the period from 10:00 pm to 11:00 pm can be set as the pushable period of the user.

Through the step S105, based on the predicted response probability of each product to be recommended after being sorted in the step S104, a customer disturbance-free strategy is set in combination with the historical marketing reach record, and a customer optimal product recommendation list is automatically generated. The method has the advantages that the fatigue of the customers is controlled while the customers are continuously operated, information cannot be continuously pushed to the target users, so that interference is generated on the target users, and the customers are continuously sensed in the marketing process to enrich the customer figures and update the iterative marketing model and strategy.

Optionally, after step S105, the method may further perform the following steps:

and S106, acquiring feedback information of the target user.

The feedback information is used for representing the feedback condition of the target user to the pushed product information.

The feedback condition may specifically include that the target user browses, deletes, ignores or collects the product information, and purchases the corresponding product, that is, the behavior of the target user on the pushed product information.

And S107, updating a logistic regression model, and gradient-lifting a decision tree and a deep learning model according to the feedback information of the target user.

Specifically, in step S107, a differentiated evaluation index may be designed for the products to be recommended belonging to different business lines, and then each model is updated according to the feedback information of the products to be recommended of different business lines according to the differentiated evaluation index of each business line.

Through the steps of S106 and S107, the scheme can comprehensively integrate and get through the data of all channels, a marketing effect monitoring system is established, marketing effect records, data postback and the like are subjected to standard management through standardized marketing data management requirements, monitoring and effect evaluation of all-channel marketing processes are achieved, the feedback data is used as the basis, the optimized recommendation model, the operation strategy and the business process are continuously updated, and therefore closed-loop management of automatic marketing is achieved.

In the aspect of technical implementation, a word vector model is adopted to generate a product vector of each product to be recommended, then a Deep learning model, such as a Wide & Deep model, is used to analyze the product characteristics of the product to be recommended and the user characteristics of a target user, a Wide part of a single input layer is used to process a large number of sparse discrete characteristics, and the memory capacity of the model is improved; and (3) processing full-scale features by using Embedding and Deep part of multiple hidden layers, and performing Deep feature intersection to endow the model with generalization capability.

In the aspect of effect evaluation, namely in step S106 and step S107, a unified whole-channel business opportunity data effect evaluation mechanism is built, a one-key automatic business opportunity and effect registration mode is formed according to business line design differentiation evaluation indexes, the hit rate and the conversion rate of whole-channel business opportunities are monitored in real time, and a business opportunity data application closed loop is created.

The application provides a product recommendation method based on deep learning, which comprises the steps of determining a candidate product library corresponding to a target user; screening a plurality of products to be recommended from a candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule; analyzing the product characteristics of each product to be recommended and the user characteristics of a target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model to obtain the predicted response probability of each product to be recommended; and recommending part of products to be recommended to the user according to the predicted response probability. According to the scheme, the user characteristics of the target user are analyzed by using the deep learning technology, the potential preference of the specific user to each product can be deeply mined, and therefore the product recommendation accuracy is improved.

The product recommendation method based on deep learning has the following advantages:

aiming at bank products such as credit cards, stages, individual credits, deposits, financing, funds and the like, the potential will of a customer is deeply mined by utilizing a deep learning frontier technology, a multi-product intelligent recommendation engine taking the customer as a center is constructed, online and offline channels such as short messages, outbound calls, network points wifi and the like are linked, and the most appropriate product combination is automatically recommended to the customer through the most appropriate channel at the most appropriate time.

The method has the advantages that the data of all channels are integrated and communicated, a marketing effect monitoring closed-loop system is set up, the marketing activity effect is comprehensively integrated, and the important feedback effect is played on the optimization of models and strategies. And an automatic marketing closed loop from modeling, marketing and feedback to continuous optimization is realized.

With reference to fig. 2, the apparatus may specifically include the following units:

the determining unit 201 is configured to determine a candidate product library corresponding to the target user.

The screening unit 202 is configured to screen a plurality of products to be recommended from the candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule.

The analysis unit 203 is configured to analyze, for each product to be recommended, the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-established logistic regression model, a gradient lifting decision tree, and a deep learning model, so as to obtain a predicted response probability of the product to be recommended.

The sorting unit 204 is configured to sort, according to the predicted response probability of each product to be recommended, each product to be recommended in a descending order, so as to obtain a list to be recommended.

The pushing unit 205 is configured to push product information of the top N products to be recommended in the list to be recommended to the target user according to the disturbance-free policy set by the target user.

Optionally, the analysis unit 203 analyzes the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model, a gradient boost decision tree and a deep learning model, and when obtaining the predicted response probability of the product to be recommended, is specifically configured to:

Optionally, the apparatus further includes an updating unit 206, configured to:

acquiring feedback information of a target user; the feedback information is used for representing the feedback condition of the target user on the pushed product information;

and updating the logistic regression model, and gradient-lifting the decision tree and the deep learning model according to the feedback information of the target user.

Optionally, the screening unit 202 is specifically configured to, when a plurality of products to be recommended are screened from the candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule, perform:

training to obtain a word vector model by taking a purchased product of a target user as a training corpus;

calculating by using a word vector model to obtain a product vector corresponding to each product in the candidate product library;

determining a plurality of similar users similar to the target user;

determining products purchased by similar users in the candidate product library as second recommended products;

acquiring interaction records of a plurality of users including a target user for a part of products in the candidate product library, and generating a product scoring matrix according to the interaction records; the product scoring matrix comprises scores of a plurality of users for part of products in the candidate product library;

calculating to obtain the score of the target user for each product in the candidate product library according to the user matrix and the product matrix;

determining a product which belongs to the business line in the candidate product library and is a priority recommendation line specified by the business recommendation rule as a fourth recommendation product;

Optionally, when the pushing unit 205 pushes the product information of the first N products to be recommended in the list to be recommended to the target user according to the disturbance-free policy set by the target user, the pushing unit is specifically configured to:

For the product recommendation device based on deep learning provided in the embodiments of the present application, specific working principles thereof may refer to relevant steps of the product recommendation method based on deep learning provided in any embodiment of the present application, and details are not repeated here.

The application provides a product recommendation device based on deep learning, wherein a determining unit 201 determines a candidate product library corresponding to a target user; wherein the candidate product library contains a plurality of products not purchased by the target user; the screening unit 202 screens a plurality of products to be recommended from a candidate product library by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset service recommendation rule; the analysis unit 203 analyzes the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model, a gradient lifting decision tree and a deep learning model for each product to be recommended to obtain the predicted response probability of the product to be recommended; the user characteristics are determined according to a first element representing basic information of a target user, a second element representing product holding conditions of the target user and a third element representing consumption preferences of the target user; the sorting unit 204 sorts the products to be recommended in a descending order according to the predicted response probability of the products to be recommended to obtain a list to be recommended; the pushing unit 205 pushes the product information of the top N products to be recommended in the list to be recommended to the target user according to the disturbance-free strategy set by the target user. According to the scheme, the user characteristics of the target user are analyzed by using the deep learning technology, the potential preference of the specific user to each product can be deeply mined, and therefore the product recommendation accuracy is improved.

The embodiment of the present application further provides a computer storage medium, which is used for storing a computer program, and when the computer program is executed, the computer program is specifically used for implementing the deep learning-based product recommendation method provided in any embodiment of the present application.

An electronic device is further provided in the embodiments of the present application, please refer to fig. 3, and the electronic device includes a memory 301 and a processor 302.

Wherein the memory 301 is used for storing computer programs;

the processor 302 is configured to execute a computer program, and is specifically configured to implement the deep learning based product recommendation method provided in any embodiment of the present application.

Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

It should be noted that the terms "first", "second", and the like in the present invention are only used for distinguishing different devices, modules or units, and are not used for limiting the order or interdependence relationship of the functions performed by the devices, modules or units.

Those skilled in the art can make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based product recommendation method is characterized by comprising the following steps:

2. The method of claim 1, wherein the screening of the candidate product library to obtain a plurality of products to be recommended by using a word vector model, a collaborative filtering technology, a hidden semantic model and a preset business recommendation rule comprises:

determining a plurality of similar users similar to the target user;

3. The method according to claim 1, wherein the analyzing the product characteristics of the product to be recommended and the user characteristics of the target user by using a pre-constructed logistic regression model, a gradient boosting decision tree and a deep learning model to obtain the predicted response probability of the product to be recommended comprises:

4. The method according to claim 1, wherein the pushing product information of the top N products to be recommended in the list to be recommended to the target user according to the do-not-disturb policy set by the target user comprises:

5. The method according to any one of claims 1 to 4, wherein after the pushing of the product information of each of the products to be recommended in the list to be recommended to the target user one by one according to the do-not-disturb policy set by the target user, the method further comprises:

6. A deep learning based product recommendation device, comprising:

7. The apparatus according to claim 6, wherein the analysis unit analyzes the product features of the product to be recommended and the user features of the target user by using a pre-constructed logistic regression model, a gradient boosting decision tree and a deep learning model, and when obtaining the predicted response probability of the product to be recommended, is specifically configured to:

8. The apparatus according to claim 6 or 7, wherein the apparatus further comprises an updating unit configured to:

9. A computer storage medium storing a computer program, which when executed is particularly adapted to implement the deep learning based product recommendation method according to any one of claims 1 to 5.

10. An electronic device comprising a memory and a processor;

wherein the memory is for storing a computer program;

the processor is configured to execute the computer program, in particular to implement the deep learning based product recommendation method according to any one of claims 1 to 5.