CN111489196A

CN111489196A - Prediction method and device based on deep learning network, electronic equipment and medium

Info

Publication number: CN111489196A
Application number: CN202010270195.7A
Authority: CN
Inventors: 麻泽武
Original assignee: Alipay Hangzhou Information Technology Co Ltd
Current assignee: Alipay Hangzhou Information Technology Co Ltd
Priority date: 2020-04-08
Filing date: 2020-04-08
Publication date: 2020-08-04
Anticipated expiration: 2040-04-08
Also published as: CN111489196B

Abstract

The embodiment of the specification discloses a prediction method based on a deep learning network, which is used for acquiring N user related data of a target user, wherein N is an integer not less than 2; processing the N user related data and M groups of creative data corresponding to M creatives by using a factorization model to obtain a user feature combination of the target user and M creative feature combinations, wherein M is an integer not less than 2; performing dot product processing on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations; and combining and inputting the user feature vectors corresponding to the N user related data, the M creative feature vectors corresponding to the M groups of creative data and the dot product feature into a deep neural network model, and determining a target creative matched with the target user from the M creatives.

Description

Prediction method and device based on deep learning network, electronic equipment and medium

Technical Field

The embodiment of the specification relates to the technical field of deep learning, in particular to a prediction method and device based on a deep learning network, electronic equipment and a medium.

Background

With the rapid development of mobile electronic equipment, the applications on the mobile electronic equipment are more and more, the popularization and application on the electronic equipment are promoted to be more and more extensive, and when a user opens a certain application, products which the user is interested in, such as shoes, clothes, novels and the like, can be automatically recommended to the user, so that the user experience is improved.

In the prior art, when predicting a product in which a user is interested, a linear model logistic regression algorithm is generally used to create a prediction model, the prediction model is trained by constructing a large number of features, such as more than 100 ten thousand features, and then prediction is performed through the trained prediction model.

Disclosure of Invention

The embodiment of the specification provides a prediction method, a prediction device, an electronic device and a prediction medium based on a deep learning network, which can improve prediction efficiency under the condition of effectively ensuring prediction accuracy.

A first aspect of an embodiment of the present specification provides a prediction method based on a deep learning network, including:

acquiring N user related data of a target user, wherein N is an integer not less than 2;

processing the N user related data and M groups of creative data corresponding to M creatives by using a factorization model to obtain a user feature combination of the target user and M creative feature combinations, wherein M is an integer not less than 2;

performing dot product processing on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations;

and combining and inputting the user feature vectors corresponding to the N user related data, the M creative feature vectors corresponding to the M groups of creative data and the dot product feature into a deep neural network model, and determining a target creative matched with the target user from the M creatives.

A second aspect of the embodiments of the present specification provides a prediction apparatus based on a deep learning network, including:

the system comprises a user data acquisition unit, a data processing unit and a data processing unit, wherein the user data acquisition unit is used for acquiring N user related data of a target user, and N is an integer not less than 7;

the FM model processing unit is used for processing the N pieces of user related data and M groups of creative data corresponding to M creatives by using a factorization model to obtain a user feature combination of the target user and M creative feature combinations, wherein M is an integer not less than 7; performing dot product processing on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations;

and the DNN model processing unit is used for inputting the user feature vectors corresponding to the N sets of user related data, the M creative feature vectors corresponding to the M sets of creative data and the dot product feature combination into a deep neural network model, and determining a target creative matched with the target user from the M creative.

The third aspect of the embodiments of the present specification further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the steps of the prediction method based on the deep learning network when executing the program.

The fourth aspect of the embodiments of the present specification also provides a computer-readable storage medium, on which a computer program is stored, which when executed by a processor, performs the steps of the prediction method based on the deep learning network.

The beneficial effects of the embodiment of the specification are as follows:

based on the technical scheme, after the user feature combination and the user feature vector of the target user are obtained, the user feature combination and M creative feature combinations corresponding to M creatives are directly subjected to dot product processing; after the user feature vectors are obtained, the user feature vectors, the M creative feature vectors and the dot product feature combinations are input into a deep neural network model, and a target creative matched with a target user is determined from the M creatives; so that the characteristic part of the target user only needs to be calculated once; when the target user and the M creatives are predicted, the characteristic part of the target user is usually calculated for M times, so that the technical scheme can reduce the calculation times of the characteristic part of the target user, and the characteristic part of the target user is large in quantity and large in calculation amount, so that the prediction efficiency can be effectively improved; in addition, the deep neural network model and the factorization model fused DeepFM model adopted by the technical scheme are used for prediction, and the prediction accuracy of the DeepFM model is higher, so that the prediction accuracy can be effectively ensured.

Drawings

FIG. 1 is a flowchart illustrating a method for deep learning network-based prediction in an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram of an improved deep FM model in an embodiment of the present specification;

FIG. 3 is a schematic structural diagram of a prediction apparatus based on a deep learning network in an embodiment of the present disclosure;

fig. 4 is a schematic structural diagram of an electronic device in an embodiment of this specification.

Detailed Description

In order to better understand the technical solutions, the technical solutions of the embodiments of the present specification are described in detail below with reference to the drawings and specific embodiments, and it should be understood that the specific features of the embodiments and embodiments of the present specification are detailed descriptions of the technical solutions of the embodiments of the present specification, and are not limitations of the technical solutions of the present specification, and the technical features of the embodiments and embodiments of the present specification may be combined with each other without conflict.

In a first aspect, as shown in fig. 1, an embodiment of the present specification provides a prediction method based on a deep learning network, including:

s102, obtaining N user related data of a target user, wherein N is an integer not less than 2;

s104, processing N user related data and M groups of creative data corresponding to M creatives by using a factorization model to obtain a user feature combination of a target user and M creative feature combinations, wherein M is an integer not less than 2;

s106, performing dot product processing on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations;

and S108, combining and inputting the user feature vectors corresponding to the N pieces of user related data, the M creative feature vectors corresponding to the M groups of creative data and the dot product feature into the deep neural network model, and determining a target creative matched with the target user from the M creatives.

In this embodiment of the present specification, the N pieces of user-related data may be user-related feature data of a target user, and may include basic attributes of the user, interest preferences of the user, historical click statistical information of the user, and the like, and in order to differentiate the users as detailed as possible, the number N of the N pieces of user-related data may be larger, and is usually an integer greater than 10, for example, 10,20,30, 50, 100, and the like.

Further, the basic attributes of the user typically include information such as name, gender, date of birth, height, native place, address, and school calendar.

Step S102 is executed first, when a target user opens an application, information such as basic attributes, interest preferences, and user history click statistical information of the target user is obtained as N user related data of the target user, and the target start application may be an application that starts a browser to open a web page or opens an APP on a smart phone, such as a certain treasure, a certain letter, and a certain group.

Specifically, when a target user starts a browser to open a web page, such as an internet-accessible homepage, N user-related data of the target user may be obtained from a corresponding server, where the corresponding server may be a cloud server or a local server, and the target user-related data is stored in the corresponding server, so that all or part of the target user-related data is searched from the corresponding server as the N user-related data according to a user ID of the target user.

After acquiring the N user-related data, step S104 is executed, first, a preset created Factorization Machine (FM) model and an acquisition method of M groups of creative data corresponding to M creatives are acquired, and then the M groups of creative data and the N user-related data are respectively input into the FM model to obtain M creative feature combinations corresponding to the M groups of creative data and user feature combinations corresponding to the N user-related data.

In the embodiment of the specification, the creative idea may be an advertisement, a design drawing, a color layout drawing, and the like, and when the creative idea is a product advertisement, a certain product advertisement with a higher target user click rate may be predicted from M product advertisements according to N user-related data of the target user; when the creative idea is a building design drawing, a certain design drawing which is more matched with a target user can be predicted from M design drawings according to the related data of the N users and the building structure; when the creative idea is a color layout, a certain color layout more matched with the target user can be predicted from the M color layouts according to the N user-related data and the object needing color layout, and the specification is not particularly limited.

The prediction method in the embodiment of the specification can be applied to an online click rate prediction scene, and automatically sends an advertisement request to an advertisement platform under the condition that a user opens a corresponding application, such as a webpage or an APP, in the online click rate prediction scene; after receiving the advertisement request, the advertisement platform predicts the target advertisement matched with the target user on line by the prediction method and pushes the target advertisement, so that the target advertisement can be clicked by the target user with higher probability, and the click rate can be effectively improved by the prediction method.

Specifically, when M groups of creative data are obtained, M creatives can be selected from a preset creative library according to N user-related data; then acquiring a group of creative data corresponding to each creative in the M creatives to obtain M groups of creative data; of course, M creatives may be randomly selected from a preset creative library.

Specifically, when M creatives are selected from a preset creative library according to N user-related data, the M creatives may be selected from the preset creative library according to some or all of the N user-related data; for example, M creatives may be selected from a preset creative library according to historical click records of the user, or M creatives may be selected from a preset creative library according to a scholarly calendar and interest preferences in the basic attributes of the user.

For example, the user historical click records of the target user show that shoes, perfume and backpacks are sequentially arranged at the highest historical click times of the target user, and then a plurality of creatives related to the shoes, perfume and backpacks are selected from a preset creative library as M creatives according to the user historical click records.

Specifically, after N pieces of user related data are obtained, embedding layer processing is carried out on the N pieces of user related data to obtain a user feature vector of a target user, and an FM (frequency modulation) model is used for processing the user feature vector to obtain a user feature combination; and after M groups of creative data are obtained, performing embedded layer processing on the group of creative data aiming at each group of creative data in the M groups of creative data to obtain creative feature vectors of the group of creative data, and processing the creative feature vectors of the group of creative data by using an FM (frequency modulation) model to obtain creative feature combinations corresponding to the group of creative data.

In the process of carrying out embedding layer processing on the N user-related data, the N user-related data can be subjected to embedding layer processing to obtain a user feature vector; and after the user characteristic vector is obtained, processing the user characteristic vector by using an FM model to obtain a user characteristic combination.

Correspondingly, in the process of embedding layer processing on the set of creation data aiming at each set of creative data in the M sets of creation data, the embedded layer processing can be carried out on the set of special creative data to obtain creative feature vectors of the set of creative data; and then, processing the creative feature vectors of the group of creative data by using an FM model to obtain a creative feature combination corresponding to the group of creative data. Therefore, the operation is carried out on each group of creative data, creative feature combinations corresponding to each group of creative data are obtained, and then M groups of creative feature combinations are obtained.

The embedding layer processing (embedding) in the embodiment of the present specification is to map a plurality of features into one feature which is a K-dimensional vector, where K is an integer not less than 2. For example, if N user-related data have 1000 features, and the feature dimension corresponding to the embedding layer processing is 100 dimensions, 1000 features are embedded into a 100-dimensional vector by using an embedding layer processing method, and the vector obtained by embedding is a user feature vector with a 100-dimensional vector; correspondingly, if a set of creative data has 500 features, the set of creative data is embedded into a 100-dimensional vector by the embedding layer processing method, and the creative feature vector of the set of creative data, which is embedded into the vector, also has a 100-dimensional vector. In this way, the dimension of the vector contained by the user feature vector and the dimension of the vector contained by each creative feature vector are the same.

For example, if M groups of creative data are a1, a2 and a3 in sequence, and N user-related data are b1, embedding layer processing is performed on a1, a2 and a3 to obtain 3 creative feature vectors which are a11, a21 and a31 in sequence; correspondingly, b1 is subjected to embedding layer processing, and the obtained user feature vector is b11, wherein the vector dimensions contained in b11, a11, a21 and a31 are all the same, for example, the vector dimensions include a 50-dimensional vector, a 100-dimensional vector, a 200-dimensional vector, and the like.

After the user feature combinations and the M creative feature combinations are obtained, step S106 is executed to perform dot product processing on the M creative feature combinations and the user feature combinations to obtain dot product feature combinations.

Specifically, the sum of creative feature combinations of the M creative feature combinations may be obtained; multiplying every two creative feature combinations in the M creative feature combinations to obtain a product vector of every two creative feature combinations, and obtaining the sum of all the product vectors as the sum of the product vectors; and finally, obtaining a dot product feature combination according to the sum of the creative feature combinations and the sum of the product vectors.

Specifically, the sum of the creative feature combinations may be added to the sum of the product vectors to obtain a dot product feature combination.

After the dot product feature combination is obtained, step S108 is performed.

In step S108, the user feature vectors and M creative feature vectors are obtained, and at this time, when the user feature vectors and M creative feature vectors are obtained, since the user feature vectors and M creative feature vectors are obtained in step S104, the user feature vectors and M creative feature vectors can be directly read from the cache.

After the user feature vectors and the M creative feature vectors are obtained, obtaining user feature weights and M creative weights corresponding to the M creatives; then, obtaining the product of the user characteristic vector and the user characteristic weight as the product of the user characteristic vector; obtaining the product of each creative feature vector in the M creative feature vectors and the corresponding creative weight; and inputting the product of the user feature vectors, each creative feature vector and the product and dot product feature combination of the corresponding creative weight into a Deep Neural Network (DNN) model, and determining a target creative matched with the target user from the M creatives.

Specifically, after the user feature vector product, the product of each creative feature vector and the corresponding creative weight and the dot product feature combination are input into the DNN model, the DNN model predicts the prediction probability of the target user and each creative in the M creatives, and the creative with the maximum prediction probability is selected as the target creative. For example, M creatives are 3 creatives, the target user and the 3 creatives are predicted by the DNN model to have prediction probabilities of 0.85,0.1 and 0.05 in sequence, and then the creatives with the prediction probabilities of 0.85 are used as the target creatives.

For example, taking the user feature vector as b11 and the M creative feature vectors as a11, a21 and a31 as examples, if the user feature weight is w₀The 3 creative weights corresponding to the M creative feature vectors are w in sequence₁，w₂And w₃At this time, the product of the feature vectors of the users is obtained as w₀× b11, and the product of each creative feature vector and the corresponding creative weight is in turn w₁×a11，w₂× a21 and w₃× a31, then w₀×b11，w₁×a11，w₂×a21，w₃× a31 and the dot product feature combination are used as the input of a Deep Neural Network (DNN) model, the DNN model carries out data processing, the prediction probability of the target user and each creative idea in M creative ideas is predicted, and the creative idea with the maximum prediction probability is used as the target creative idea.

Thus, in the actual application process, when the target user opens an application, it is equivalent to initiate a request online, at this time, the target user is clear, and the user characteristic data of the target user is prompted to be fixed. At this moment, when the server recommends an intention for the target user, M creatives can be selected from a preset creatives library, and the selected creatives are different, so that the creatives feature parts are also different, and a plurality of creatives have M creatives feature parts.

When M creatives are used for predicting a target user in the prior art, traditional DeepFM can perform batch prediction calculation, M groups of creative data can walk M Directed Acyclic Graph (DAG) links, parallel acceleration can be performed inside the links, and parts irrelevant to data can be calculated simultaneously under multiple threads, but in fact, the results are only faster than that of cyclic prediction of each piece of data by less than 5%. According to the technical scheme adopted by the embodiment of the specification, when M creatives are used for predicting the target user, only one user characteristic vector and M creative characteristic vectors are required to be input into the DNN model, M groups of creative data can be simultaneously calculated in parallel without mutual dependence, and dot product processing is finally carried out; aiming at the user part of the target user, after the user characteristic vector is determined, the input vector of the DNN model and the processing result of the FM model are calculated according to the user characteristic vector, and then the input vector, the processing result and the dot product characteristics of M creatives are combined for subsequent calculation.

Because the characteristic part of the target user in the technical scheme adopted by the embodiment of the specification only needs to be calculated once, and the M groups of creative data can be simultaneously and parallelly calculated, compared with the prior art, if the three-layer network width of the DNN is L1, L2 and L3, the calculated amount of M ×L 2 × (L1 + L3) is reduced, the calculated amount is effectively reduced, and the accuracy of prediction can be effectively ensured by using a DeepFM model, so that the technical scheme adopted by the embodiment of the specification can improve the prediction efficiency on the basis of ensuring the prediction accuracy.

The embodiment of the present specification further provides an improved deep fm model, which is specifically shown in fig. 2. The input part of the model of the refined DeepFM comprises user shared data 10, a creative 20, a creative 21 and a creative 22, wherein the user shared data 10 is N user-related data of a target user; mapping the user shared data 10 into embedding to obtain a user characteristic vector 11; then, FM processing is performed on the user feature vector 11 to obtain a user feature combination 111. Correspondingly, 3 groups of creative data of the creative 20, the creative 21 and the creative 22 are mapped into the embedding to obtain a creative feature vector 201 of the creative 20, a creative feature vector 211 of the creative 21 and a creative feature vector 221 of the creative 22, and then the creative feature vector 201, the creative feature vector 211 and the creative feature vector 221 are subjected to FM processing to obtain a creative feature combination 202, a creative feature combination 212 and a creative feature combination 222 in sequence.

Wherein, the user feature vector 11, the creative feature vector 201, the creative feature vector 211 and the creative feature vector 221 are sequentially x₀，x₁，x₂And x₃Representation, and the user feature combination 111, creative feature combination 202, creative feature combination 212, and creative feature combination 222 are sequentially x₀₁，x₁₁，x₂₁And x₃₁Indicate that for x₀₁，x₁₁，x₂₁And x₃₁Performing dot product processing to obtain x for the dot product feature combination 23₄Is shown, and x₄＝x₀₁+x₁₁+x₂₁+x₃₁+x₁₁×x₂₁+x₁₁×x₃₁+x₂₁×x₃₁。

And if the user characteristic weight corresponding to the user characteristic vector 11 is reused by w₀Representing and sequentially using w creative weights corresponding to the creative feature vector 201, the creative feature vector 211 and the creative feature vector 221₁，w₂And w₃Indicate, then x₀×w₀，x₁×w₁，x₂×w₂，x₃×w₃And x₄As output of the first layer of the DNN model, and x₀×w₀，x₁×w₁，x₂×w₂，x₃×w₃And x₄As an input to the second layer of the DNN model, 3 output results are obtained, respectively, as predicted probability 30, predicted probability 31, and predicted probability 32, by the DNN model processing.

Wherein x is₀×w₀，x₁×w₁，x₂×w₂，x₃×w₃Indicated in sequence in figure 2 by 24, 25,26 and 27, prediction probability 30 characterizes the prediction probability of the predicted target user corresponding to creative 20, prediction probability 31 characterizes the prediction probability of the predicted target user corresponding to creative 21, and prediction probability 32 characterizes the prediction probability of the predicted target user corresponding to creative 22. Then, the creative idea with the maximum probability is selected from the prediction probability 30, the prediction probability 31 and the prediction probability 32 to serve as the target creative idea, and if the prediction probability 31 is maximum, the target creative idea is determinedThe creative 21 is intended, whereby it can be determined that the target user has the greatest probability of clicking on the creative 21, and thus the creative 21 is pushed to the target user, increasing the user's click rate.

In this way, after the user feature combination and the user feature vector of the target user are obtained, the user feature combination and M creative feature combinations corresponding to M creatives are directly subjected to dot product processing; after the user feature vectors are obtained, the user feature vectors, the M creative feature vectors and the dot product feature combinations are input into a deep neural network model, and a target creative matched with a target user is determined from the M creatives; the feature part of the target user only needs to be calculated once, and when the target user and M creatives are predicted in the prior art, the feature part of the target user needs to be calculated M times, so that the technical scheme can reduce the calculation times of the feature part of the target user, and the feature part of the target user is large in quantity and large in calculated amount, so that the prediction efficiency can be effectively improved; and the prediction is carried out by the deep FM model adopted by the technical scheme, and the prediction accuracy can be effectively determined according to the advantages of the deep FM model.

In the embodiment of the present specification, the FM model and the DNN model (i.e., the deep FM model) are usually obtained by training in an offline state, and may of course be obtained by training in an online state. When the DeepFM model is trained off-line, the training data is different from the on-line input data mode, the on-line calculation is carried out according to one request, however, each creative of one request has only one creative marked result at most, namely clicking or not clicking; offline data is the recovery and sampling of marking data and does not simulate online conditions. At this time, the modified DAG in the embodiment of the present specification may completely adapt to such a situation, and the creative input of the dynamic graph of the deep fm model is set to 1, and the DAG link and the original link at this time may be kept consistent, so that it is ensured that the training of the deep fm model does not cause loss, and the advantage of higher prediction accuracy of the deep fm model can be effectively ensured.

This means that each time a data block needs to be predicted there will be a portion of the repetition, i.e. the user-related feature. The part is characterized by large quantity and large calculation amount, and enough optimization space is provided for solving the acceleration problem. Only one user feature and N creative features are required to input the model for input, as shown on the far right of fig. 2. The N creative data calculation links are independent, parallel calculation can be simultaneously carried out in the DAG, mutual dependence is not needed, and finally superposition is carried out.

In a second aspect, based on the same technical concept, embodiments of the present specification provide a prediction apparatus based on a deep learning network, with reference to fig. 3, including:

a user data obtaining unit 301, configured to obtain N user-related data of a target user, where N is an integer not less than 7;

an FM model processing unit 302, configured to process, by using a factorization model, the N user-related data and M groups of creative data corresponding to the M creatives to obtain user feature combinations of the target user and M creative feature combinations, where M is an integer not less than 7; performing dot product processing on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations;

the DNN model processing unit 303 is configured to input, into a deep neural network model, the user feature vectors corresponding to the N pieces of user-related data, the M creative feature vectors corresponding to the M sets of creative data, and the dot product feature combination, and determine, from the M creatives, a target creative that matches the target user.

In an alternative embodiment, the FM model processing unit 302 is configured to obtain a sum of creative feature combinations of the M creative feature combinations; multiplying each two creative feature combinations in the M creative feature combinations to obtain a product vector of each two creative feature combinations, and obtaining the sum of all the product vectors as the sum of the product vectors; and obtaining the dot product feature combination according to the sum of the creative feature combinations and the product vector.

In an optional embodiment, the FM model processing unit 302 is configured to, for each set of creative data in the M sets of creative data, perform embedded layer processing on the set of creative data to obtain creative feature vectors of the set of creative data, and process the creative feature vectors of the set of creative data using the factorization model to obtain a creative feature combination corresponding to the set of creative data; and performing embedding layer processing on the N pieces of user related data to obtain a user characteristic vector of the target user, and processing the user characteristic vector by using the factorization machine model to obtain the user characteristic combination.

In an optional implementation manner, the DNN model processing unit 303 is configured to obtain a product of the user characteristic vector and a user characteristic weight as a user characteristic vector product; obtaining the product of each creative feature vector in the M creative feature vectors and the corresponding creative weight; and inputting the product of the user feature vectors, the product of each creative feature vector and the corresponding creative weight and the dot product feature combination into a deep neural network model to obtain the prediction result.

In an alternative embodiment, the presetting device further comprises:

the creative data acquisition unit is used for selecting the M creatives from a preset creative library according to the N user related data; and acquiring a group of creative data corresponding to each creative in the M creatives to obtain the M groups of creative data.

In a third aspect, based on the same inventive concept as the prediction method based on the deep learning network in the foregoing embodiment, an embodiment of the present specification further provides an electronic device, as shown in fig. 4, including a memory 404, a processor 402, and a computer program stored in the memory 404 and executable on the processor 402, where the processor 402 implements the steps of any one of the prediction methods based on the deep learning network when executing the program.

Where in fig. 4 a bus architecture (represented by bus 400) is shown, bus 400 may include any number of interconnected buses and bridges, and bus 400 links together various circuits including one or more processors, represented by processor 402, and memory, represented by memory 404. The bus 400 may also link together various other circuits such as peripherals, voltage regulators, power management circuits, and the like, which are well known in the art, and therefore, will not be described any further herein. A bus interface 405 provides an interface between the bus 400 and the receiver 401 and transmitter 403. The receiver 401 and the transmitter 403 may be the same element, i.e., a transceiver, providing a means for communicating with various other apparatus over a transmission medium. The processor 402 is responsible for managing the bus 400 and general processing, while the memory 404 may be used for storing data used by the processor 402 in performing operations.

In a fourth aspect, based on the inventive concept of the prediction method based on the deep learning network in the foregoing embodiments, the present specification embodiment further provides a computer readable storage medium, on which a computer program is stored, which when executed by a processor implements the steps of any one of the prediction methods based on the deep learning network.

The description has been presented with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the description. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

While preferred embodiments of the present specification have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all changes and modifications that fall within the scope of the specification.

It will be apparent to those skilled in the art that various changes and modifications may be made in the present specification without departing from the spirit and scope of the specification. Thus, if such modifications and variations of the present specification fall within the scope of the claims of the present specification and their equivalents, the specification is intended to include such modifications and variations.

Claims

1. A prediction method based on a deep learning network comprises the following steps:

2. The method of claim 1, wherein the performing a dot product on the user feature combinations and the M creative feature combinations to obtain dot product feature combinations comprises:

acquiring the sum of creative feature combinations of the M creative feature combinations;

multiplying each two creative feature combinations in the M creative feature combinations to obtain a product vector of each two creative feature combinations, and obtaining the sum of all the product vectors as the sum of the product vectors;

and obtaining the dot product feature combination according to the sum of the creative feature combinations and the product vector.

3. The method of claim 2, wherein the processing of the N user-related data and M sets of creative data corresponding to M creatives using a factorization model to obtain user feature combinations for the target user and M creative feature combinations comprises:

for each group of creative data in the M groups of creative data, performing embedded layer processing on the group of creative data to obtain creative feature vectors of the group of creative data, and processing the creative feature vectors of the group of creative data by using the factorization model to obtain creative feature combinations corresponding to the group of creative data;

and performing embedding layer processing on the N pieces of user related data to obtain a user characteristic vector of the target user, and processing the user characteristic vector by using the factorization machine model to obtain the user characteristic combination.

4. The method of claim 3, the inputting the user feature combinations, the M creative feature combinations, and the dot product feature combinations into a deep neural network model, the determining a creative from the M creatives as a predicted result of the target user, comprising:

obtaining the product of the user characteristic vector and the user characteristic weight as the product of the user characteristic vector;

obtaining the product of each creative feature vector in the M creative feature vectors and the corresponding creative weight;

and inputting the product of the user feature vectors, the product of each creative feature vector and the corresponding creative weight and the dot product feature combination into a deep neural network model to obtain the prediction result.

5. The method of claim 4, wherein the M creatives include M groups of creative data, and the M groups of creative data include:

selecting the M creatives from a preset creatives library according to the N user-related data;

and acquiring a group of creative data corresponding to each creative in the M creatives to obtain the M groups of creative data.

6. A prediction apparatus based on a deep learning network, comprising:

7. The apparatus of claim 6, the FM model processing unit to obtain a sum of creative feature combinations of the M creative feature combinations; multiplying each two creative feature combinations in the M creative feature combinations to obtain a product vector of each two creative feature combinations, and obtaining the sum of all the product vectors as the sum of the product vectors; and obtaining the dot product feature combination according to the sum of the creative feature combinations and the product vector.

8. The apparatus of claim 7, wherein the FM model processing unit is configured to, for each of the M sets of creative data, perform embedding layer processing on the set of creative data to obtain creative feature vectors of the set of creative data, and process the creative feature vectors of the set of creative data using the factorization model to obtain a creative feature combination corresponding to the set of creative data; and performing embedding layer processing on the N pieces of user related data to obtain a user characteristic vector of the target user, and processing the user characteristic vector by using the factorization machine model to obtain the user characteristic combination.

9. The device of claim 8, the DNN model processing unit to obtain a product of the user characteristic vector and a user characteristic weight as a user characteristic vector product; obtaining the product of each creative feature vector in the M creative feature vectors and the corresponding creative weight; and inputting the product of the user feature vectors, the product of each creative feature vector and the corresponding creative weight and the dot product feature combination into a deep neural network model to obtain the prediction result.

10. The apparatus of claim 9, further comprising:

11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the method of any one of claims 1-5 when executing the program.

12. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 5.