CN115482019A

CN115482019A - Activity attention prediction method and device, electronic equipment and storage medium

Info

Publication number: CN115482019A
Application number: CN202110597318.2A
Authority: CN
Inventors: 樊鹏
Original assignee: Tencent Cloud Computing Beijing Co Ltd
Current assignee: Tencent Cloud Computing Beijing Co Ltd
Priority date: 2021-05-31
Filing date: 2021-05-31
Publication date: 2022-12-16

Abstract

The application relates to the technical field of computers, in particular to the technical field of machine learning, and provides an activity attention prediction method, an activity attention prediction device, electronic equipment and a storage medium, which are used for improving the prediction accuracy of the attention of an object to an activity. The method comprises the following steps: extracting account characteristics corresponding to the target object based on historical behavior data of the target object aiming at target activities, and extracting image characteristics corresponding to the target object based on image data related to the target object; performing feature splicing on the account features and the portrait features to obtain high-dimensional sparse features corresponding to the target object; predicting the attention of the target object to the target activity based on the high-dimensional sparse feature to obtain a predicted probability value corresponding to the target object; and determining the attention of the target object to the target activity based on the estimated probability value. According to the method and the device, the activity attention is predicted based on the high-dimensional sparse feature obtained by feature splicing of the account feature and the portrait feature, and the accuracy of attention prediction can be effectively improved.

Description

Activity attention prediction method and device, electronic equipment and storage medium

Technical Field

The application relates to the technical field of computers, in particular to the technical field of machine learning, and provides an activity attention prediction method, an activity attention prediction device, electronic equipment and a storage medium.

Background

In the related art, the technical scheme for predicting activity attention mainly comprises the following steps: and determining a data rule based on manual experience, and predicting activity attention based on the determined data rule. Taking a car show activity as an example, the product operation is based on business experience and a mode of combining the attributes of the car show, the division rules of potential people with high willingness on the car show are set, for example, advertisements of high-end car shows, and medium-high-end mobile phone devices and male groups are selected to be released. However, the method of determining the data rule based on manual experience not only uses a very limited number of rules, but also has low prediction accuracy.

Disclosure of Invention

The embodiment of the application provides an activity attention prediction method and device, electronic equipment and a storage medium, which are used for improving the prediction accuracy of the attention of an object to an activity.

The activity attention prediction method provided by the embodiment of the application comprises the following steps:

extracting account features corresponding to a target object based on historical behavior data of the target object for target activities, and extracting portrait features corresponding to the target object based on portrait data related to the target object;

performing feature splicing on the account features and the portrait features to obtain high-dimensional sparse features corresponding to the target object;

predicting the attention of the target object to the target activity based on the high-dimensional sparse feature to obtain a predicted probability value corresponding to the target object;

determining a degree of attention of the target object to the target activity based on the pre-estimated probability value.

An activity attention prediction device provided by the embodiment of the application comprises:

the characteristic extraction unit is used for extracting account characteristics corresponding to a target object based on historical behavior data of the target object aiming at target activities, and extracting portrait characteristics corresponding to the target object based on portrait data related to the target object;

the feature splicing unit is used for performing feature splicing on the account features and the portrait features to obtain high-dimensional sparse features corresponding to the target object;

the prediction unit is used for predicting the attention degree of the target object to the target activity based on the high-dimensional sparse feature to obtain a prediction probability value corresponding to the target object;

a determining unit, configured to determine a degree of attention of the target object to the target activity based on the pre-estimated probability value.

Optionally, the estimating unit is specifically configured to:

inputting the high-dimensional sparse features into a trained attention estimation model, extracting features of the high-dimensional sparse features based on the attention estimation model, and obtaining and outputting estimation probability values;

the attention estimation model is obtained based on training of a training sample data set, training samples in the training sample data set comprise high-dimensional sparse features, the high-dimensional sparse features are obtained by splicing account features and portrait features of sample objects, sample labels are marked on the training samples and used for representing whether the sample objects pay attention to the target activities, the account features are extracted based on historical behavior data of the sample objects on the target activities, and the portrait features are extracted based on portrait data related to the sample objects.

Optionally, the pre-estimation model of the degree of interest includes a shallow sub-network and a deep sub-network; the estimation unit is specifically configured to:

inputting the high-dimensional sparse features into a shallow sub-network in the attention prediction model, and performing cross feature extraction on the high-dimensional sparse features based on the shallow sub-network to obtain corresponding target cross vectors, wherein the high-dimensional sparse features comprise a plurality of sparse feature fields, and each sparse feature field is used for representing the account features or the portrait features;

and inputting the target cross vector into the deep subnetwork, and mapping the target cross vector based on the deep subnetwork to obtain the estimated probability value output by the deep subnetwork.

Optionally, the shallow sub-network includes an embedded layer, a punishment compression layer, a bilinear cross layer, and a merging layer; the estimation unit is specifically configured to:

inputting the high-dimensional sparse features into the shallow sub-network, and performing embedding representation on the high-dimensional sparse features based on an embedding layer in the shallow sub-network to obtain corresponding low-dimensional dense embedding vectors;

inputting the low-dimensional dense embedded vector into a compression reward and punishment layer in the shallow sub-network, and extracting the importance degree between sparse feature fields in the high-dimensional sparse feature based on the compression reward and punishment layer to obtain a corresponding importance degree weighting vector;

inputting the low-dimensional dense embedded vector and the importance weighting vector into a bilinear crossing layer in the shallow sub-network, and performing cross feature extraction on the low-dimensional dense embedded vector and the importance weighting vector based on the bilinear crossing layer to obtain a first cross vector corresponding to the low-dimensional dense embedded vector and a second cross vector corresponding to the importance weighting vector;

and splicing the first cross vector and the second cross vector based on a merging layer in the shallow sub-network to obtain the target cross vector.

Optionally, the estimating unit is specifically configured to:

based on a compression reward and punishment layer in the shallow sub-network, performing compression representation on the low-dimensional dense embedded vector to obtain a corresponding compression statistical vector;

extracting the importance degree among sparse feature fields in the high-dimensional sparse features based on the compressed statistical vector to obtain corresponding feature group weight vectors, wherein elements in the feature group weight vectors are used for expressing the importance degree corresponding to the sparse feature fields in the high-dimensional sparse features;

and taking the product of the low-dimensional dense embedding vector and the feature set weight vector as the importance degree weight vector output by the embedding layer.

Optionally, the low-dimensional dense embedding vector and the importance weighting vector are in a matrix form; the estimation unit is specifically configured to:

based on a preset parameter matrix in the bilinear cross layer, performing cross feature extraction between different sparse features in the low-dimensional dense embedded vector to obtain the first cross vector; and

and performing cross feature extraction on different sparse features in the importance weighting vector based on a preset parameter matrix in the bilinear cross layer to obtain the second cross vector.

Optionally, the deep subnetwork includes a plurality of hidden layers, and the pre-estimating unit is specifically configured to:

inputting the target cross vector into the deep sub-network, and performing at least one mapping process on the target cross vector based on a plurality of hidden layers in the deep sub-network to obtain the estimated probability value output by the deep sub-network.

Optionally, the apparatus further comprises a training unit:

the training unit is used for training and obtaining the attention degree estimation model in the following way:

according to the training samples in the training sample data set, performing loop iterative training on the attention prediction model, and outputting the trained attention prediction model when the training is finished; wherein, the following operations are executed in the one-time loop iteration training process:

selecting at least one training sample from the training sample data set, and inputting the at least one training sample into the attention prediction model;

respectively extracting the features of the high-dimensional sparse features in the at least one training sample based on the attention prediction model to obtain each prediction probability value output by the attention prediction model;

and adjusting parameters of the attention prediction model based on the prediction probability values and the sample labels marked on the corresponding training samples.

Optionally, the apparatus further comprises:

the acquisition unit is used for dividing a first set time period into different time intervals according to different time dimensions before the feature extraction unit extracts the account features corresponding to the target object based on the historical behavior data of the target object for the target activity;

and acquiring historical behavior data of the target object on the target activity in each time interval.

Optionally, when the portrait characteristics include behavior portrait characteristics of the target object, the feature extraction unit is specifically configured to:

acquiring network connection track information related to login equipment used by the target object and application flow use behavior information;

determining a network behavior mode corresponding to the target object by analyzing the network connection track information;

and performing feature extraction through the network behavior mode and the application flow using behavior information to obtain behavior portrait features corresponding to the target object.

Optionally, when the image feature includes an apparatus image feature of a login apparatus used by the target object, the feature extraction unit is specifically configured to:

and extracting the characteristics of the basic attribute information related to the login equipment used by the target object to obtain the equipment portrait characteristics corresponding to the target object.

Optionally, when the portrait features include a feedback portrait feature of the target object on the target activity within a second set time period, the feature extraction unit is specifically configured to:

acquiring feedback behavior data of the target object generated aiming at the target activity in a second set time period;

and performing feature extraction on the feedback behavior data to acquire feedback portrait features corresponding to the target object.

An electronic device provided by an embodiment of the present application includes a processor and a memory, where the memory stores program codes, and when the program codes are executed by the processor, the processor is caused to execute any one of the steps of the activity attention prediction method.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the electronic device to perform the steps of any of the activity attention prediction methods described above.

An embodiment of the present application provides a computer-readable storage medium, which includes program code for causing an electronic device to perform any of the steps of the activity attention prediction method described above when the program product runs on the electronic device.

The beneficial effect of this application is as follows:

the embodiment of the application provides an activity attention prediction method and device, electronic equipment and a storage medium. According to the method, the historical behavior data generated by the target object on the target activity and the portrait characteristics related to the target object are used for carrying out feature extraction and feature splicing, high-dimensional sparse characteristics are constructed, prediction is carried out based on the characteristics, and the attention of the target object on the target activity is obtained through analysis.

Additional features and advantages of the application will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the application. The objectives and other advantages of the application may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:

fig. 1 is an alternative schematic diagram of an application scenario in an embodiment of the present application;

FIG. 2 is a flowchart illustrating an activity attention prediction method according to an embodiment of the present application;

FIG. 3 is a schematic structural diagram of an attention estimation model in an embodiment of the present application;

fig. 4 is a schematic flowchart of a method for obtaining a target cross vector in an embodiment of the present application;

fig. 5 is a flowchart illustrating a method for obtaining an importance weighting vector according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a method of computing cross-signatures in an embodiment of the present application;

FIG. 7 is a schematic diagram of a training method of an attention estimation model in an embodiment of the present application;

FIG. 8A is a diagram of a training and application architecture for a model in an embodiment of the present application;

FIG. 8B is a process diagram of an offline partial solution in an embodiment of the present application;

FIG. 9 is a detailed flowchart of an activity attention prediction method in an embodiment of the present application;

FIG. 10A is a graph illustrating a comparative analysis of model effects according to an embodiment of the present application;

fig. 10B is a diagram illustrating a business effect comparison analysis in the embodiment of the present application;

FIG. 11 is a schematic diagram illustrating a structure of an activity attention prediction apparatus according to an embodiment of the present application;

fig. 12 is a schematic diagram of a hardware component structure of an electronic device to which an embodiment of the present application is applied;

fig. 13 is a schematic diagram of a hardware component structure of another electronic device to which the embodiment of the present application is applied.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some embodiments, but not all embodiments, of the technical solutions of the present application. All other embodiments obtained by a person skilled in the art based on the embodiments described in the present application without any creative effort belong to the protection scope of the technical solution of the present application.

Some concepts related to the embodiments of the present application are described below.

Portrait features: the method refers to user portrait, and the user portrait is an effective tool for sketching target users and connecting user appeal and design direction. The image characteristics in the embodiment of the present application include basic image characteristics of the target object, i.e. the age, sex, occupation, etc. of the target object. In addition, the portrait characteristics may further include at least one of a behavior portrait characteristic of the target object, a device portrait characteristic of a login device used by the target object, a feedback portrait characteristic of the target object for the target activity within a second set time period, and the like.

Bilinear function (bilinear function): is a generalization of linear functions. Let V1, V2 be a linear space on the domain P, and the bilinear mapping φ of V1V 2 to P is called a bilinear function on V1V 2. In particular, when V1= V2= V, Φ is referred to as a bilinear function on V. A semi-bilinear function (sesquilinear function) is a generalization of bilinear functions.

Area under the Curve (Area under the dark, AUC): the area under the receiver operating characteristic curve (ROC) curve is the standard for judging the quality of the two-class prediction model. The characteristic curve of ROC receiver operation belongs to the signal detection theory. The abscissa of the ROC curve is the false positive rate (also called false positive rate), the ordinate is the true positive rate (true rate), and correspondingly, the true negative rate (true negative rate) and the false negative rate (false negative rate).

Network connection trajectory information: the time sequence information indicates a time sequence when the user connects to the network, and may indicate, for example, which Wireless-Fidelity (Wifi) the user connects to at which time point (time period).

Application traffic usage behavior information: the information refers to an application used by a user within a certain time, usage behavior, and the like.

Vehicle exhibition: the term "Auto show" refers to an automobile exhibition, such as an automobile product exhibition or trade meeting, an exhibition, etc., which is conducted in a professional exhibition hall or a center of a meeting place by an organization such as a government agency, a professional association, or a mainstream media.

The attention degree estimation model comprises the following steps: the model for estimating the attention of the user to a certain activity is provided in the embodiment of the application, for example, the model is used for estimating the attention of the user to a vehicle exhibition, and predicting whether the user is a user with high intention of the vehicle exhibition, namely a user group which has strong interest in off-line vehicle exhibition and has high click intention and high reservation intention on-line vehicle exhibition advertisements; for another example, the method is used for predicting the attention degree to the strategy game and predicting whether the user is a user with high downloading willingness to the strategy game, that is, a user group with high downloading willingness to the strategy game.

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the implementation method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject, and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

Machine Learning (ML) is a multi-domain cross subject, and relates to multiple subjects such as probability theory, statistics, approximation theory, convex analysis and algorithm complexity theory. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach to make computers have intelligence, and is applied in various fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning.

According to the method and the device, when the attention of the user to the target activity is predicted, a machine learning attention prediction model is adopted. The method for training the attention estimation model provided in the embodiment of the application can be divided into two parts, including a training part and an application part; the training part is used for training the attention degree estimation model through the machine learning technology, so that the attention degree estimation model is trained through a training sample data set given in the embodiment of the application, after the training sample passes through the attention degree estimation model, an output result of the attention degree estimation model is obtained, and model parameters are continuously adjusted through an optimization algorithm in combination with the output result; the application part is used for predicting the attention of the target object relative to the target activity by using the attention prediction model obtained by training in the training part.

The following briefly introduces the design concept of the embodiments of the present application:

the technical scheme for predicting the activity attention of the user in the related technology mainly comprises the following steps: the data rules are determined based on manual experience. However, the method for determining data rules based on artificial experience not only uses a very limited number of rules, but also cannot capture high-dimensional sparse feature information of interaction between the rules, and cannot determine optimal parameters of each rule, so that prediction accuracy is low.

In addition, the click probability of the current user on the activity-related advertisement can be predicted based on a non-deep learning data mining method, so that the activity attention of the user can be predicted. However, in some scenarios, such as a vehicle exhibition high willingness user prediction scenario, behavior characteristics are complex, characteristic information is difficult to express explicitly in data representation, and a traditional characteristic representation method and a non-deep learning model are not suitable for learning user characteristic information of the vehicle exhibition high willingness user prediction scenario.

In view of this, the present application provides an activity attention prediction method, an activity attention prediction apparatus, an electronic device, and a storage medium. According to the embodiment of the application, feature extraction and feature splicing are carried out on the basis of historical behavior data generated by the target object on the target activity and the image features related to the target object, high-dimensional sparse features are constructed, prediction is carried out on the basis of the features, and the attention of the target object on the target activity is obtained through analysis. The method realizes the capture of high-dimensional sparse characteristic information, and can effectively improve the prediction accuracy of the target object on the attention degree of the target activity.

The preferred embodiments of the present application will be described in conjunction with the drawings of the specification, it should be understood that the preferred embodiments described herein are only for illustrating and explaining the present application, and are not intended to limit the present application, and the embodiments and features of the embodiments in the present application may be combined with each other without conflict.

As shown in fig. 1, it is a schematic view of an application scenario of the embodiment of the present application. The application scenario diagram includes two terminal devices 110 and a server 120. The terminal device 110 and the server 120 may communicate with each other via a communication network.

In an alternative embodiment, the communication network is a wired network or a wireless network. The terminal 210 and the server 120 may be directly or indirectly connected through wired or wireless communication, and the application is not limited thereto.

In this embodiment, the terminal device 110 is an electronic device used by a user, and the electronic device may be an electronic device having a certain computing capability and running instant messaging software and a website or social contact software and a website, such as a personal computer, a mobile phone, a tablet computer, a notebook, an e-book reader, an intelligent car union, and an intelligent television. Each terminal device 110 is connected to the server 120 through a wireless Network, and the server 120 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a Network service, cloud communication, a middleware service, a domain name service, a security service, a Content Delivery Network (CDN), and a big data and artificial intelligence platform. The terminal may be, but is not limited to, a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and the like. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the application is not limited herein.

The attention estimation model may be deployed on the server 120 for training, and a large number of training samples may be stored in the server 120 for training the attention estimation model. Optionally, after the attention estimation model is obtained based on the training method in the embodiment of the present application through training, the trained attention estimation model may be directly deployed on the server 120 or the terminal device 110. Generally, the attention prediction model is directly deployed on the server 120, and in the embodiment of the present application, the attention prediction model is mainly used for predicting the attention of the object to the target activity.

In a possible application scenario, the training samples in the present application may be stored by using a cloud storage technology. A distributed cloud storage system (hereinafter, referred to as a storage system) refers to a storage system that integrates a large number of storage devices (storage devices are also referred to as storage nodes) of different types in a network through application software or an application interface to cooperatively work by using functions such as cluster application, grid technology, and a distributed storage file system, and provides a data storage function and a service access function to the outside.

In the following, the activity attention prediction method provided by the exemplary embodiment of the present application is described with reference to the accompanying drawings in combination with the application scenarios described above, it should be noted that the application scenarios described above are only shown for the convenience of understanding the spirit and principle of the present application, and the embodiments of the present application are not limited in this respect.

It should be noted that the activity attention prediction method in the embodiment of the present application may be executed by the server or the terminal device alone, or may be executed by both the server and the terminal device.

Referring to fig. 2, a flowchart of an implementation of the activity attention prediction method according to the embodiment of the present application is illustrated by taking a server as an execution subject, and the specific implementation flow of the method is as follows:

s21: the server extracts account features corresponding to the target object based on historical behavior data of the target object aiming at target activities, and extracts image features corresponding to the target object based on image data related to the target object;

wherein, the acquisition of the related data is also required before the step S21 is executed.

In an alternative embodiment, historical behavior data may be collected based on:

dividing a first set time period into different time intervals according to different time dimensions;

and acquiring historical behavior data of the target object on target activities in each time interval.

For example, the online behavior data generated by the user for the target activity in different time windows (last half year, last three months, last month, last 1 week, last 3 days), different time periods (rest time, activity time, weekend time, working day time, etc.) is counted offline at regular time as the historical behavior data of the target object for the target activity. Further, based on the historical behavior data, account features corresponding to the target object are extracted.

In addition, image data related to the target object needs to be acquired, and further, image features corresponding to the target object are extracted based on the image data, which will be described in detail below.

S22: the server carries out feature splicing on the account features and the portrait features to obtain high-dimensional sparse features corresponding to the target object;

s23: the server predicts the attention of the target object to the target activity based on the high-dimensional sparse feature to obtain an estimated probability value corresponding to the target object;

in an alternative embodiment, the above steps may be implemented in an artificial intelligence manner.

Specifically, a machine learning model is constructed: and (5) estimating a model of the attention. Inputting the high-dimensional sparse features into a trained attention degree estimation model, extracting the features of the high-dimensional sparse features based on the attention degree estimation model, obtaining estimation probability values and outputting the estimation probability values.

The method comprises the steps that an attention degree estimation model is obtained based on training of a training sample data set, the training sample in the training sample data set comprises a high-dimensional sparse feature, the high-dimensional sparse feature is obtained by splicing account features and portrait features of sample objects, sample labels are marked on the training sample and used for representing whether the sample objects concern target activities, the account features are extracted based on historical behavior data of the sample objects on the target activities, and the portrait features are extracted based on portrait data related to the sample objects.

S24: and the server determines the attention of the target object to the target activity based on the pre-estimated probability value.

In the embodiment of the present application, the attention of the target object to the target activity may be set as: 0 (not paying attention), 1 (paying attention), in this way, if the estimated probability value is greater than the preset threshold value, the corresponding attention degree is 1, otherwise, the corresponding attention degree is 0.

Or, different attention levels can be set, correspondingly, different probability intervals are divided, and one interval corresponds to one level, so that the attention level corresponding to the probability interval to which the estimated probability value belongs can be determined according to the probability interval.

It should be noted that the above-listed methods for determining the attention based on the estimated probability value are only examples, and any determination method is applicable to the embodiments of the present application, and is not specifically limited herein.

In the above embodiment, feature extraction and feature concatenation are performed based on historical behavior data of the target object on the target activity and the image features related to the target object, high-dimensional sparse features are constructed, prediction is performed based on the features, and the attention of the target object on the target activity is obtained through analysis. The method realizes the capture of high-dimensional sparse characteristic information, and can effectively improve the prediction accuracy of the target object on the attention degree of the target activity.

The following describes the image feature extraction process in detail:

in embodiments of the present application, the portrait features may include, in addition to base portrait features of the target object, at least one of:

the behavior portrait characteristics of the target object, the device portrait characteristics of the login device used by the target object, and the feedback portrait characteristics of the target object to the target activity in a second set time period.

The following respectively describes the extraction processes of different image features:

in an optional implementation manner, when the portrait features include behavior portrait features of the target object, the portrait features corresponding to the target object are extracted by:

acquiring network connection track information related to login equipment used by a target object and application flow use behavior information; determining a network behavior mode corresponding to a target object by analyzing the network connection track information; and performing feature extraction by using the behavior information through a network behavior mode and application flow to acquire behavior portrait features corresponding to the target object.

The network connection track information may refer to track information generated by the user connecting to the WIFI, that is, a time sequence corresponding to the user connecting to the WIFI: WIFI1 is connected at the time of T1, WIFI2, \ 8230is connected at the time of T2, and WIFIN is connected at the time of TN.

Based on the network connection track information, a corresponding WIFI behavior mode, namely a network behavior mode, can be obtained through analysis. In the embodiment of the present application, the network behavior pattern corresponding to the target object specifically refers to: when to go out, entertain, when to go home, etc.

The APPlication traffic behavior information may refer to a traffic usage behavior sequence of a user using different types of applications (applications, APPs), for example, a traffic usage behavior sequence of a user using a social APP.

In the embodiment, based on information such as user network connection tracks and application flow behaviors, behavior habits and interests of users can be effectively known, and more accurate user portrayal can be constructed.

In an optional implementation manner, when the image features include device image features of a login device used by the target object, the image features corresponding to the target object are extracted as follows:

The login device used by the target object refers to a terminal device, for example: mobile phones, tablets, computers, and the like. Taking the login device as a mobile phone as an example, the basic attribute information related to the login device specifically refers to: a brand of a mobile phone, a Read-Only Memory (ROM) size of the mobile phone, a system version of the mobile phone, and the like.

In the embodiment, based on the characteristic information of the terminal equipment used by the user, the hardware requirement and the living standard of the user can be indirectly reflected, so that a more accurate user portrait can be constructed.

In an optional implementation manner, when the image feature includes a feedback image feature of the target object moving to the target within a second set time period, the image feature corresponding to the target object is extracted by:

acquiring feedback behavior data of the target object generated aiming at the target activity in a second set time period; and performing feature extraction on the feedback behavior data to acquire feedback portrait features corresponding to the target object.

Taking a target activity as an example of a vehicle exhibition activity, the feedback behavior data mainly refers to: the feedback behavior of the user on the exposed vehicle exhibition advertisement comprises the following steps: whether to click on an advertisement, whether to pre-order by real name, etc.

The following describes the activity attention prediction process in detail with reference to the attention prediction model listed in fig. 3:

fig. 3 is a schematic structural diagram of an attention estimation model according to an embodiment of the present application. Wherein the model is divided into Deep Part and Shallow Part based on the dotted line.

In an optional implementation manner, when feature extraction is performed on the high-dimensional sparse features based on the attention prediction model to obtain and output prediction probability values, the specific process is as follows:

firstly, inputting high-dimensional sparse features into a shallow sub-network in an attention estimation model, and performing cross feature extraction on the high-dimensional sparse features based on the shallow sub-network to obtain corresponding target cross vectors; the target cross vector is input into the deep sub-network, and the target cross vector is subjected to mapping processing based on the deep sub-network, so that an estimated probability value output by the deep sub-network is obtained.

The shallow sub-network is mainly used for learning the importance degree between features and learning hidden information after the features are crossed.

The high-dimensional sparse features in the embodiment of the present application include a plurality of sparse feature fields, where each sparse feature field is used to represent an account feature or a portrait feature, that is, different fields represent different features, and the features may be referred to as account features or portrait features. For example, the high-dimensional sparse feature is composed of f fields, field1 representing the age of the user in the representation feature, field2 representing the gender of the user in the representation feature, \ 8230, and Field df representing the account feature.

Wherein, the shallow sub-network can be further divided into: the device comprises an embedded layer, a compression reward and punishment layer, a bilinear cross layer and a merging layer.

An optional implementation manner is that, according to the flowchart shown in fig. 4, cross feature extraction on high-dimensional sparse features may be implemented, and a corresponding target cross vector is obtained, which specifically includes the following steps:

s401: the server inputs the high-dimensional sparse features into the shallow sub-network, and embedding representation is carried out on the high-dimensional sparse features on the basis of an embedding layer in the shallow sub-network to obtain corresponding low-dimensional dense embedding vectors;

specifically, the server inputs the high-dimensional Sparse feature into the shallow sub-network based on the Sparse input Layer (Sparse input Layer), and embeds (Embedding) the high-dimensional Sparse feature based on the Embedding Layer (Embedding Layer) to obtain a low-dimensional dense Embedding vector (Embedding).

S402: the server inputs the low-dimensional dense embedded vector into a compression reward and punishment layer in the shallow sub-network, and extracts the importance degree between sparse feature fields in the high-dimensional sparse feature based on the compression reward and punishment layer to obtain a corresponding importance degree weighting vector;

as shown in fig. 3, that is, the low-dimensional dense embedded vector is input into the compression award or punishment Layer (sentet Layer), and a corresponding importance weighting vector (sentet-Like Embeddings) is obtained.

S403: the server inputs the low-dimensional dense embedded vectors and the importance degree weighted vectors into a bilinear crossing layer in a shallow sub-network, and performs cross feature extraction on the low-dimensional dense embedded vectors and the importance degree weighted vectors based on the bilinear crossing layer to obtain first cross vectors corresponding to the low-dimensional dense embedded vectors and second cross vectors corresponding to the importance degree weighted vectors;

as shown in the combination of FIG. 3, the low-dimensional dense embedding vector Embeddings and the importance weighting vector SENTET-Like Embeddings are input into the Bilinear intersection Layer. Based on the bilinear crossing layer, respectively carrying out crossing feature extraction on Embeddings and SENET-Like Embeddings to obtain respective corresponding crossing vectors, namely a first crossing vector and a second crossing vector.

S404: and the server splices the first cross vector and the second cross vector based on the merging layer in the shallow sub-network to obtain a target cross vector.

With reference to fig. 3, vector splicing is performed on the first cross vector and the second cross vector based on a Combination Layer (Combination Layer), so as to obtain a target cross vector.

And finally, inputting the target cross vector into Deep Part, and performing feature mapping based on the Deep sub-network to obtain a final estimated probability value.

In the embodiment of the application, the main function of the SENET Layer is to learn an importance degree of different features, weight important features and weaken the features containing small amount of information. The specific method comprises the following steps: using the Embedding vector (i.e. low-dimensional dense embedded vector) of the feature set as an input, a feature set weight vector a = [ a1,. A, ai,. A,. Af ] is generated, and the original feature set Embedding vector E is multiplied by a to obtain a new Embedding vector V = [ V1,. Vi,. Vf ], which is the importance weighting vector in the embodiment of the present application.

In an alternative embodiment, S402 can be implemented according to the flowchart shown in fig. 5, and includes the following steps:

s501: the server performs compression representation on the low-dimensional dense embedded vector based on a compression reward and punishment layer in the shallow sub-network to obtain a corresponding compression statistical vector;

this step mainly involves the Squeeze (compression) operation, which is mainly an operation of summarizing statistics on embedding vectors in each feature group. In the embodiment of the present application, a pooling operation is used to perform a compressed representation on an original feature set embedding vector (i.e., a low-dimensional dense embedding vector) E = [ E1,..,. Ei,. Ef ], so as to obtain a compressed statistical vector Z = [ Z1,..,. Zi,. Zf ], where zi represents global information of an ith feature.

S502: the server extracts the importance degree among sparse feature fields in the high-dimensional sparse features based on the compressed statistical vector to obtain corresponding feature group weight vectors, wherein elements in the feature group weight vectors are used for expressing the importance degree corresponding to the sparse feature fields in the high-dimensional sparse features;

this step mainly involves an Excitation operation. This step learns the importance weights of the feature set based on the compressed statistical vectors of the feature set, and the embodiment of the present application learns using a neural network with two layers, the first layer being a dimension reduction layer and the second layer being a dimension promotion layer, and the obtained feature set weight vector a = [ a1,.. A,. Ai,. A.

S503: and the server takes the product of the low-dimensional dense embedding vector and the feature set weight vector as an importance degree weighting vector output by the embedding layer.

This step is the Re-Weight operation. This step re-weights the original feature set Embedding vector by using the feature importance weight obtained from the Excitation operation (i.e., multiplies the feature set weight vector a by the low-dimensional dense Embedding vector E), and obtains an importance weight vector V = [ V1.,. Vi.,. Vf ].

After the importance weighting vector corresponding to the low-dimensional dense vector is calculated based on the above embodiment, it is necessary to extract cross features based on a Bilinear-Interaction Layer (Bilinear-Interaction Layer) shown in fig. 3. The traditional characteristic crossing mode widely adopts inner products and Hadamard products. And the two modes are difficult to effectively model the feature intersection on sparse data. The Biliner-Interaction Layer in the application proposes to combine the inner product and the Hadamard product and introduce an additional parameter matrix W (a preset parameter matrix) to learn the feature crossing.

In an alternative embodiment, the low-dimensional dense embedding vector and the importance weighting vector are in the form of a matrix; when extracting cross features based on bilinear cross layers, the specific process is as follows:

based on a preset parameter matrix in a bilinear cross layer, carrying out cross feature extraction on different sparse features in the low-dimensional dense embedded vector to obtain a first cross vector; and performing cross feature extraction on different sparse features in the importance weighting vector based on a preset parameter matrix in the bilinear cross layer to obtain a second cross vector.

Fig. 6 is a schematic diagram illustrating a method for calculating cross features according to an embodiment of the present disclosure. The process of calculating the cross feature by inner product, hadamard product, and the manner in the embodiment of the present application are respectively described.

Wherein, a mode a in fig. 6 introduces a process of calculating cross features based on Inner Product (Inner Product), a mode b introduces a process of calculating cross features based on Hadamard Product (Hadamard Product), a mode c introduces a process of calculating cross features based on Bilinear (Bilinear) cross layer in the present application, and a cross vector p _ij Is calculated by the formula：

p _ij ＝v _i W⊙v _j Wherein v is _i And v _j Representing different sparse features (also referred to as feature sets).

Finally, the original Embeddings-E and SENET-Like Embeddings-V are respectively processed by the double-linear cross layer to respectively obtain a cross vector p = [ p ] ₁ ，…，p _i ，…，p _n ]And q = [ q ] ₁ ，…，q _i ，…，q _n ]. The vector p is a first cross vector corresponding to the low-dimensional dense embedding vector, and the vector q is a second cross vector corresponding to the importance weighting vector.

Then, p = [ p ] by Combination Layer ₁ ，…，p _i ，…，p _n ]And q = [ q ] ₁ ，…，q _i ，…，q _n ]And (5) performing localization to obtain an output result as a target cross vector c.

In the above embodiment, both the Inner Product and the Hadamard Product are combined, and a weight matrix is inserted between two features to be crossed, so as to dynamically learn the combination relationship between the features, and effectively consider the importance degree of the features themselves.

In an alternative embodiment, the deep sub-network includes a plurality of hidden layers, and when predicting the predicted probability value based on the deep sub-network, the specific process is as follows:

and inputting the target cross vector into a deep sub-network, and performing at least one mapping treatment on the target cross vector based on a plurality of hidden layers in the deep sub-network to obtain an estimated probability value output by the deep sub-network.

It should be noted that the depth sub-network may be a multi-layer Perceptron (MLP), and the estimation of the estimated probability value may be obtained by inputting the target cross vector c into the MLP.

In an alternative embodiment, the attention prediction model is obtained by training in the following way:

according to training samples in the training sample data set, performing loop iterative training on the attention prediction model, and outputting the trained attention prediction model when the training is finished; referring to fig. 7, a schematic diagram of a training method for an attention estimation model in an embodiment of the present application is shown, where the following operations are performed in a one-cycle iterative training process, and the method specifically includes the following steps:

s701: the server selects at least one training sample from the training sample data set, and inputs the at least one training sample into the attention degree estimation model;

s702: the server respectively extracts the characteristics of high-dimensional sparse characteristics in at least one training sample based on the attention degree estimation model to obtain each estimation probability value output by the attention degree estimation model;

s703: and the server adjusts parameters of the attention prediction model based on the prediction probability values and the sample labels marked on the corresponding training samples.

In step S702, feature extraction is performed on the high-dimensional sparse features in at least one training sample based on the attention prediction model, and a specific process when the prediction probability value corresponding to each training sample is obtained is the same as the listed process for predicting the prediction probability value of the target object for the target activity, and repeated parts are not repeated.

In the following, taking a target activity as a vehicle exhibition activity as an example, a process of predicting a "vehicle exhibition will-wish user" based on the activity attention estimation method in the embodiment of the present application is described in detail.

Fig. 8A is a diagram illustrating a training and application architecture of a model according to an embodiment of the present application, which is specifically divided into an online portion and an offline portion.

In the online part, each service request of a user executes online memory (Decache) feature query, namely, real-time behavior feature data of the user is pulled from the online Decache and spliced into a real-time user feature vector. And then, performing feature splicing based on the offline trained attention estimation model and the multi-dimensional image features of the users, and judging whether the current users are the group of 'users with high automobile exhibition will'. And finally, recording the feedback of the online behavior of the highly-willing user output by the attention estimation model, namely the real-time feedback of the user on the online vehicle exhibition advertisement.

The off-line part comprises three parts of data accumulation, feature processing and model training. First, the data accumulation includes the following:

1) And outputting on-line behavior feedback of the user with high willingness by the model. And storing historical behavior data of the online vehicle exhibition advertisement output by the model and having high willingness to the online vehicle exhibition advertisement on a Distributed File System (HDFS). And (3) periodically and offline counting the online behavior characteristics of the user in different time windows (last half year, last three months, last month, last 1 week and last 3 days) and different time periods (rest time, activity time, weekend time, working day time and the like).

2) User portrait data. Constructing a rich user portrait based on portrait data related to a user, specifically including device portrait feature extraction, user behavior portrait feature extraction, advertisement feedback feature extraction of the user in the last N months and the like, wherein user portrait features obtained based on the feature extraction include but are not limited to the following parts or all of: user base attributes, device base attributes, geographic location attributes, software usage preferences, network connection attributes.

It should be noted that, the feature processing method of the offline experiment in the embodiment of the present application may include the following 8 aspects:

1) One-Hot Encoding (thermal Encoding). This processing mode is selected for characteristics such as the gender of the user.

2) Count Encoding. For example, for a Wifi Point of Interest (POI) feature of a user, count Encoding is used to identify the user and the Interest level of this POI. For example, the user has gone to the POI 'food-a national dish-city A' 3 times in the week.

3) Category Embedding. According to data analysis, strong sparsity exists in many category characteristics, such as Wifi types. In order to avoid overfitting the model and improve the stability of the model, a neural network is introduced to convert a high-dimensional sparse classification variable into a low-dimensional dense Embedding variable.

4) NaN Embedding (category Embedding). For the processing of the missing value of the feature, methods such as 'removing', 'average filling' and 'missing mark' can be adopted, the experimental result is displayed in a PUSH (recommended) content scene, the missing value is converted into an Embedding expression mode, and the maximum positive benefit is achieved on the effect of the model.

5) Consolidation Encoding (merging Encoding). Multiple values under certain category variables can be summarized into the same information. For example, the multiple values of the system version features of the android mobile phone include three values of "4.2", "4.4" and "5.0", and the three values can be summarized into a "low-version android system" based on experience. Experiments prove that the processing mode of characteristic induction can bring greater forward benefits than that the characteristic one-hot of the android system version is directly used.

6) Feature Scaling. According to the distribution condition of the numerical characteristic, a proper normalization method is selected to eliminate dimensional difference between the characteristics, so that the model is more stable. For example, for features that fit or approximately fit a normal distribution, a gaussian normalization is chosen.

7) Wifi track Embedding. Based on a deep learning network, embelling is carried out on Wifi connection track data of a user, and information of Wi-Fi behavior patterns (Pattern) of the user, such as when the user goes out, when the user gets entertained, when the user goes home and the like, is captured.

8) APP Traffic Embedding (with Traffic Embedding). Based on a List-Embedding mode, carrying out Embedding extraction on flow use behavior sequences of different types of APP used by users, such as Traffic Embedding of social type APP, and obtaining low-dimensional and dense user behavior characteristics.

The modeling process of the offline experiment is specifically shown in fig. 7. The FiBiNet model is selected for model construction, and the model essentially learns the importance of the features dynamically by using a Squeeze-Excitation Network (compression reward and punishment Network) structure (for more important features, a larger weight is learned, and the weight of less important features is reduced) and better models the cross features by using a bilinear function.

In the attention estimation model in the embodiment of the application, a SENET layer is added in a traditional Embedding Stage (Embedding Stage) to embed the embedded features after the Embedding is finished, so that information related to Feature Importance (Feature Importance) is obtained. After combining the important information of the features with the original embedded features (Embedding features) to generate a new Feature vector, a new Bilinear intersection (Bilinear intersection) method combining the two features is selected to obtain the relation between the features without using the traditional inner product or Hadamard product method.

It should be noted that the characteristics of the description class in the embodiment of the present application may be processed into an integer Sequence/matrix (Int Sequence/matrix) as an Embedding Feature input to the model. In addition, it is preferable to set the reduction Ratio (Reduce Ratio) to 6 to 8.

Fig. 8B is a process diagram of an offline partial technical solution in the embodiment of the present application. The following describes in detail the main steps of the offline partial technical solution in the embodiment of the present application with reference to fig. 8B, where the main steps are divided into three parts: raw data accumulation, data feature engineering, model training and evaluation.

1. The raw data accumulation section mainly includes: reporting and storing log data, extracting original features and storing the original features.

This section is mainly used for: and accessing the log data requested by the user on line in real time to the HDFS for storage. In consideration of storage cost and subsequent computing efficiency, log key information is extracted based on Hive Structured Query Language (SQL), and redundant data is discarded.

2. The data characteristic engineering part mainly comprises: conventional feature processing and HDFS data storage are carried out based on a computing engine Spark, embedding feature processing and HDFS data storage are carried out based on TensorFlow (tensor).

The part mainly refers to: according to the data characteristics of the original features stored in the HDFS, appropriate feature processing is performed. In the embodiment of the application, spark and TensorFlow are selected for carrying out characteristic engineering, and the specific division of labor is as follows:

a) The conventional characteristic engineering method without Embedding calculates based on a Spark calculation engine, and stores the result in an HDFS;

b) The deep learning characteristic engineering method of Embedding calculates based on a TensorFlow calculation engine, and stores the result in an HDFS;

3. the model training and evaluating part mainly comprises: feature extraction and feature splicing, model training, model linear evaluation and model online evaluation.

Features are first read out locally from HDFS based on Hive SQL. Modeling was then performed based on the tensrflow. And finally, carrying out model evaluation based on a built-in mathematical evaluation method, wherein the evaluation result is divided into two conditions:

a) And if the evaluation index is positive, pushing the attention estimation model to the line, and performing AB Test (A/B Test). And if the A/B Test also displays the forward direction of each service index, all the flow is accessed for on-line formal use.

b) And (5) training the attention degree estimation model again when the evaluation index is negative until the effect of the model is in line with the expectation.

Fig. 9 is a detailed flowchart of an activity attention prediction method according to an embodiment of the present application. The specific implementation flow of the method is as follows:

step S901: and extracting the user account characteristics.

For example, spark or Hive SQL is used to extract user historical behavior data and historical online feedback data from HDFS. The main uses are: and constructing a positive and negative sample set and constructing a user basic portrait characteristic.

Step S902: and extracting user portrait characteristics.

From the self-built label system, a user image is extracted, and the specific dimensionality is as follows: user base attributes, device base attributes, geographic location attributes, software usage preferences, network connection attributes, and the like.

Wherein, the user basic attribute refers to: age, sex, etc.; the device grounding attribute refers to: mobile phone brand, mobile phone ROM memory size, etc.; the geographic location attribute refers to: the user's usual residence province/city, etc.; the software usage preference means: the number of times that the user used the social APP in the last month; the network connection attribute refers to a connection track of a user connecting with Wifi and the like.

Step S903: performing feature splicing on the account features and the portrait features;

the step mainly includes that a user account feature vector and a user portrait feature vector are spliced into a high-dimensional vector to be used as input of the attention prediction model. This is done by merging (Concat) all features into a high-dimensional vector by column.

Step S904: training a model;

and (3) based on the positive and negative samples marked with the labels, using TensorFlow to realize FiBiNet algorithm to carry out model training.

Step S905: evaluating under a model line;

in the embodiment of the application, the AUC index is selected for model evaluation under the model line. The larger the AUC value, the more likely the current classification algorithm is to rank positive samples ahead of negative samples, resulting in better classification results.

Because the AUC index is independent of the absolute value of the model prediction score, only the sequencing effect is concerned, and the method is more close to the requirement of actual business. In addition, the calculation method of the AUC simultaneously considers the classification capability of the learner on positive examples and negative examples, and can still reasonably evaluate the classifier under the condition of unbalanced samples.

Step S906: judging whether the model result reaches expectation, if so, executing step S907, otherwise, returning to step S904;

step S907: model online evaluation; the online model evaluation considers the help of the model to the specific service, and the evaluation method is based on A/B Test; the effect of the model was evaluated based on the on-line flow of the A/B Test. The indicators evaluated were: advertisement click rate, user real-name reservation rate, and the like.

Step S908: judging whether the model result reaches the expectation, if so, executing the step S909, otherwise, returning to the step S904;

in the embodiment of the present application, after returning to step S904 from step S906 or step S908, the model needs to be adjusted. That is, when either of the model-off-line evaluation and the model-on-line evaluation fails, the model is retrained until both evaluated metrics are passed. Wherein, the direction of model adjustment is at least one of the following: selecting positive and negative samples with different time windows and different drilling dimensions; and carrying out grid search on key parameters of the model, and selecting a parameter combination with the optimal effect.

Step S909: solidifying the model;

in the embodiment of the present application, after both the off-line evaluation and the on-line evaluation of the model are qualified, the trained model can be solidified based on the software () method of TensorFlow, and 4 files are generated in total:

a) The checkpoint text file records a path information list of the model file;

b) model, ckpt, data, recording network weight information;

c) model, ckpt, index, data and index are binary files and store variable weight information in the model;

step S910: an online service invocation model;

in the embodiment of the application, when the client calls the service interface, the server calls the cured model, and then returns the prediction result after pulling the user characteristics and the user real-time characteristics in the online Decache.

Step S911: and collecting on-line user behavior feedback of the vehicle exhibition will users predicted by the model.

In the embodiment of the application, the feedback of the model prediction result at the client by the user can be stored as the client log, so that the method and the device have direct help for improving the fine operation of the APP client flow.

It should be noted that, the above is only exemplified by taking the target activity as the vehicle exhibition, and in this embodiment, the attention estimation model is mainly used for predicting the "user with high intention of vehicle exhibition", but the target activity is not limited to the vehicle exhibition. The technical scheme in the embodiment of the application has strong reusability: firstly, the user type to which the positive sample belongs is changed, for example, a policy-type game downloads a user with high willingness, then the server accumulates corresponding log data, and finally, the results are output by using the same methods of feature splicing, feature processing and model training, which are not specifically limited herein.

Referring to fig. 10A, a model effect comparison analysis diagram in the embodiment of the present application is illustrated, which respectively introduces model effect comparison analysis obtained by identifying a vehicle exhibition will user based on an attention estimation model provided by artificially formulating strong rules, non-deep learning, and the present application:

from the effect of the AUC under the line, compared with other technical schemes, the scheme provided by the application is improved by 24.59% on average; from the effect of the on-line AUC, the scheme provided by the application is improved by 22.96% compared with other technical schemes on average.

Fig. 10B is a business effect comparison analysis diagram in the embodiment of the present application, which respectively introduces a business effect comparison analysis obtained by identifying a vehicle exhibition will user based on an attention estimation model provided by artificially formulating strong rules, non-deep learning, and the present application:

compared with other technical schemes, the scheme provided by the application is averagely improved by 199.28% from the aspect of the advertisement click rate; compared with other technical schemes, the scheme provided by the application is improved by 251.58% on average from the perspective of real-name reservation rate of users.

In conclusion, the attention prediction model based on the embodiment of the application is used for predicting the attention, so that the prediction accuracy can be effectively improved, and a set of recommendation system for identifying the interest personalized demands of the user can be rapidly established.

Based on the same inventive concept, the embodiment of the application also provides an activity attention prediction device. As shown in fig. 11, which is a schematic structural diagram of an activity attention prediction apparatus 1100 in the embodiment of the present application, the apparatus may include:

a feature extraction unit 1101, configured to extract an account feature corresponding to a target object based on historical behavior data of the target object for a target activity, and extract an image feature corresponding to the target object based on image data related to the target object;

the feature splicing unit 1102 is used for performing feature splicing on the account features and the portrait features to obtain high-dimensional sparse features corresponding to the target object;

the prediction unit 1103 is configured to predict a degree of attention of the target object to the target activity based on the high-dimensional sparse feature, and obtain a prediction probability value corresponding to the target object;

a determining unit 1104, configured to determine a degree of attention of the target object to the target activity based on the pre-estimated probability value.

Optionally, the estimation unit 1103 is specifically configured to:

inputting the high-dimensional sparse features into a trained attention estimation model, extracting the features of the high-dimensional sparse features based on the attention estimation model, obtaining estimation probability values and outputting the estimation probability values;

the attention estimation model is obtained based on training of a training sample data set, training samples in the training sample data set comprise high-dimensional sparse features, the high-dimensional sparse features are obtained by splicing account features and portrait features of sample objects, sample labels are marked on the training samples and used for representing whether the sample objects pay attention to target activities, the account features are extracted based on historical behavior data of the sample objects on the target activities, and the portrait features are extracted based on portrait data related to the sample objects.

Optionally, the attention prediction model includes a shallow subnetwork and a deep subnetwork; the estimation unit 1103 is specifically configured to:

inputting the high-dimensional sparse features into a shallow sub-network in the attention prediction model, performing cross feature extraction on the high-dimensional sparse features based on the shallow sub-network to obtain corresponding target cross vectors, wherein the high-dimensional sparse features comprise a plurality of sparse feature fields, and each sparse feature field is used for representing account features or portrait features;

and inputting the target cross vector into a deep sub-network, and mapping the target cross vector based on the deep sub-network to obtain an estimated probability value output by the deep sub-network.

Optionally, the shallow sub-network includes an embedded layer, a punishment compression layer, a bilinear cross layer, and a merging layer; the estimation unit 1103 is specifically configured to:

inputting the high-dimensional sparse features into a shallow sub-network, and performing embedding representation on the high-dimensional sparse features based on an embedding layer in the shallow sub-network to obtain corresponding low-dimensional dense embedding vectors;

inputting the low-dimensional dense embedded vector into a compression reward and punishment layer in the shallow sub-network, extracting the importance degree between sparse feature fields in the high-dimensional sparse feature based on the compression reward and punishment layer, and obtaining a corresponding importance degree weighting vector;

inputting the low-dimensional dense embedded vector and the importance weighting vector into a bilinear crossing layer in a shallow sub-network, and performing cross feature extraction on the low-dimensional dense embedded vector and the importance weighting vector based on the bilinear crossing layer to obtain a first cross vector corresponding to the low-dimensional dense embedded vector and a second cross vector corresponding to the importance weighting vector;

and splicing the first cross vector and the second cross vector based on the merging layer in the shallow sub-network to obtain a target cross vector.

Optionally, the estimation unit 1103 is specifically configured to:

and taking the product of the low-dimensional dense embedding vector and the feature set weight vector as an importance degree weight vector output by the embedding layer.

Optionally, the low-dimensional dense embedding vector and the importance weighting vector are in a matrix form; the estimation unit 1103 is specifically configured to:

based on a preset parameter matrix in a bilinear cross layer, carrying out cross feature extraction on different sparse features in the low-dimensional dense embedded vector to obtain a first cross vector; and

and performing cross feature extraction on different sparse features in the importance weighting vector based on a preset parameter matrix in the bilinear cross layer to obtain a second cross vector.

Optionally, the deep sub-network includes a plurality of hidden layers, and the pre-estimating unit 1103 is specifically configured to:

Optionally, the apparatus further comprises a training unit 1105:

the training unit 1105 is configured to train and obtain an attention estimation model by:

according to training samples in the training sample data set, performing loop iterative training on the attention prediction model, and outputting the trained attention prediction model when the training is finished; wherein, the following operations are executed in the one-time loop iteration training process:

selecting at least one training sample from a training sample data set, and inputting the at least one training sample into an attention estimation model;

based on the attention prediction model, respectively extracting the characteristics of high-dimensional sparse characteristics in at least one training sample to obtain each prediction probability value output by the attention prediction model;

Optionally, the apparatus further comprises:

the acquisition unit 1106 is configured to divide the first set time period into different time intervals according to different time dimensions before the feature extraction unit 1101 extracts account features corresponding to the target object based on historical behavior data of the target object for the target activity;

Optionally, when the portrait features include behavioral portrait features of the target object, the feature extraction unit 1101 is specifically configured to:

acquiring network connection track information related to login equipment used by a target object and application flow use behavior information;

determining a network behavior mode corresponding to a target object by analyzing the network connection track information;

and performing feature extraction by using the behavior information through a network behavior mode and application flow to obtain behavior portrait features corresponding to the target object.

Optionally, when the representation feature includes a device representation feature of a login device used by the target object, the feature extraction unit 1101 is specifically configured to:

and extracting the characteristics of basic attribute information related to the login equipment used by the target object to obtain the equipment portrait characteristics corresponding to the target object.

Optionally, when the portrait features include a feedback portrait feature of the target object moving to the target within a second set time period, the feature extraction unit 1101 is specifically configured to:

In the above embodiment, feature extraction and feature splicing are performed based on historical behavior data generated by a target object on a target activity and image features related to the target object, high-dimensional sparse features are constructed, prediction is performed based on the features, and the attention of the target object on the target activity is obtained through analysis. The method realizes the capture of high-dimensional sparse characteristic information, and can effectively improve the prediction accuracy of the target object on the attention degree of the target activity.

For convenience of description, the above parts are separately described as modules (or units) according to functional division. Of course, the functionality of the various modules (or units) may be implemented in the same one or more pieces of software or hardware when implementing the present application.

As will be appreciated by one skilled in the art, aspects of the present application may be embodied as a system, method or program product. Accordingly, various aspects of the present application may be embodied in the form of: an entirely hardware embodiment, an entirely software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects that may all generally be referred to herein as a "circuit," module "or" system.

The electronic equipment is based on the same inventive concept as the method embodiment, and the embodiment of the application also provides the electronic equipment. In one embodiment, the electronic device may be a server, such as server 120 shown in FIG. 1. In this embodiment, the electronic device may be configured as shown in fig. 12, and include a memory 1201, a communication module 1203, and one or more processors 1202.

A memory 1201 for storing computer programs executed by the processor 1202. The memory 1201 may mainly include a storage program area and a storage data area, where the storage program area may store an operating system, a program required for running an instant messaging function, and the like; the storage data area can store various instant messaging information, operation instruction sets and the like.

Memory 1201 may be a volatile memory (volatile memory), such as a random-access memory (RAM); the memory 1201 may also be a non-volatile memory (non-volatile memory), such as a ROM, a flash memory (flash memory), a hard disk (HDD) or a solid-state drive (SSD); or the memory 1201 is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited to such. The memory 1201 may be a combination of the above memories.

The processor 1202 may include one or more Central Processing Units (CPUs), a digital processing unit, or the like. A processor 1202, configured to implement any of the above-described activity attention prediction methods when invoking a computer program stored in the memory 1201.

The communication module 1203 is used for communicating with the terminal device and other servers.

In the embodiment of the present application, the specific connection medium between the memory 1201, the communication module 1203 and the processor 1202 is not limited. In the embodiment of the present application, the memory 1201 and the processor 1202 are connected through the bus 1204 in fig. 12, the bus 1204 is represented by a thick line in fig. 12, and the connection manner between other components is only schematically illustrated and is not limited. The bus 1204 may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown in FIG. 12, but this is not intended to represent only one bus or type of bus.

The memory 1201 stores therein a computer storage medium, and the computer storage medium stores therein computer-executable instructions for implementing the activity attention prediction method according to the embodiment of the present application. The processor 1202 is configured to perform the activity attention prediction method described above, as shown in fig. 2.

In another embodiment, the electronic device may also be other electronic devices, such as the terminal device 110 shown in fig. 1. In this embodiment, the structure of the electronic device may be as shown in fig. 13, including: a communication component 1310, a memory 1320, a display unit 1330, a camera 1340, a sensor 1350, audio circuitry 1360, a bluetooth module 1370, a processor 1380, and so forth.

The communication component 1310 is for communicating with a server. In some embodiments, a circuit wireless fidelity module may be included, the Wifi module belongs to a short-distance wireless transmission technology, and the electronic device may help the user to send and receive information through the Wifi module.

Memory 1320 may be used to store software programs and data. The processor 1380 performs various functions of the terminal device 110 and data processing by executing software programs or data stored in the memory 1320. The memory 1320 may include high speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device. Memory 1320 stores an operating system that enables terminal device 110 to operate. The memory 1320 in the present application may store an operating system and various application programs, and may also store codes for executing the activity attention prediction method according to the embodiment of the present application.

The display unit 1330 may also be used to display information input by or provided to the user and a Graphical User Interface (GUI) of various menus of the terminal apparatus 110. Specifically, display unit 1330 may include a display screen 1332 provided on the front face of terminal device 110. The display 1332 may be configured in the form of a liquid crystal display, a light emitting diode, or the like. The display unit 1330 can be used to display a presentation interface related to the target activity in the embodiment of the present application, and the like.

The display unit 1330 may also be configured to receive input numeric or character information and generate signal input related to user settings and function control of the terminal device 110, and specifically, the display unit 1330 may include a touch screen 1331 disposed on the front surface of the terminal device 110 and configured to collect touch operations by a user thereon or nearby, such as clicking a button, dragging a scroll box, and the like.

Touch screen 1331 may cover display screen 1332, or touch screen 1331 and display screen 1332 may be integrated to implement the input and output functions of terminal device 110, and after integration, the touch screen may be referred to as a touch display screen for short. The display unit 1330 may display the application programs and the corresponding operation steps.

The camera 1340 may be used to capture still images, and the user may upload images captured by the camera 1340 to the review via the client. The number of the cameras 1340 may be one or more. The object generates an optical image through the lens and projects the optical image to the photosensitive element. The photosensitive element may be a Charge Coupled Device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The light sensing elements convert the light signals into electrical signals, which are then passed to a processor 1380 for conversion into digital image signals.

The terminal device may further comprise at least one sensor 1350, such as an acceleration sensor 1351, a distance sensor 1352, a fingerprint sensor 1353, a temperature sensor 1354. The terminal device may also be configured with other sensors such as a gyroscope, barometer, hygrometer, thermometer, infrared sensor, light sensor, motion sensor, and the like.

The audio circuit 1360, speaker 1361, microphone 1362 may provide an audio interface between the user and the terminal device 110. The audio circuit 1360 may transmit the electrical signal converted from the received audio data to the speaker 1361, and the electrical signal is converted into a sound signal by the speaker 1361 and output. Terminal device 110 may also be configured with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 1362 converts the collected sound signal into an electrical signal, which is received by the audio circuit 1360 and converted into audio data, and then outputs the audio data to the communication module 1310 to be transmitted to, for example, another terminal device 110, or outputs the audio data to the memory 1320 for further processing.

The bluetooth module 1370 is used for information interaction with other bluetooth devices having bluetooth modules through a bluetooth protocol. For example, the terminal device may establish a bluetooth connection with a wearable electronic device (e.g., a smart watch) that is also equipped with a bluetooth module through the bluetooth module 1370, so as to perform data interaction.

The processor 1380 is a control center of the terminal device, connects various parts of the entire terminal device using various interfaces and lines, performs various functions of the terminal device and processes data by operating or executing software programs stored in the memory 1320 and calling data stored in the memory 1320. In some embodiments, processor 1380 may include one or more processing units; the processor 1380 may also integrate an application processor, which primarily handles operating systems, user interfaces, application programs, etc., and a baseband processor, which primarily handles wireless communications. It will be appreciated that the baseband processor may not be integrated into the processor 1380. The processor 1380 in the present application may run an operating system, an application, a user interface display, and a touch response, as well as the activity attention prediction method of the embodiments of the present application. Additionally, a processor 1380 is coupled to the display unit 1330.

In some possible embodiments, the various aspects of the activity attention prediction method provided by the present application may also be implemented in the form of a program product including program code for causing an electronic device to perform the steps of the activity attention prediction method according to various exemplary embodiments of the present application described above in this specification when the program product is run on the electronic device, for example, the electronic device may perform the steps as shown in fig. 2.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a ROM, an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

However, the program product of the present application is not so limited, and in the present application, the readable storage medium may be any tangible medium that can contain or store a program for use by or in connection with a command execution system, apparatus, or device.

A readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A readable signal medium may be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with a command execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

Program code for carrying out operations of the present application may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user computing device, partly on the user equipment, as a stand-alone software package, partly on the user computing device and partly on a remote computing device, or entirely on the remote computing device or server. In the case of remote computing devices, the remote computing device may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computing device (e.g., through the internet using an internet service provider).

While the preferred embodiments of the present application 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 alterations and modifications as fall within the scope of the application.

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

Claims

1. A method for activity attention prediction, the method comprising:

and determining the attention of the target object to the target activity based on the pre-estimated probability value.

2. The method of claim 1, wherein the predicting the interest level of the target object for the target activity based on the high-dimensional sparse features to obtain a predicted probability value for the target object comprises:

inputting the high-dimensional sparse features into a trained attention degree estimation model, extracting features of the high-dimensional sparse features based on the attention degree estimation model, and obtaining and outputting estimation probability values;

3. The method of claim 2, wherein the pre-estimation of interest model comprises a shallow sub-network and a deep sub-network;

inputting the high-dimensional sparse features into a trained attention prediction model, extracting features of the high-dimensional sparse features based on the attention prediction model, and obtaining and outputting prediction probability values, wherein the method comprises the following steps:

4. The method of claim 3, wherein the shallow sub-network comprises an embedding layer, a punishment-compression layer, a bilinear cross-over layer, and a merging layer;

the inputting the high-dimensional sparse feature into a shallow sub-network in the attention prediction model, performing cross feature extraction on the high-dimensional sparse feature based on the shallow sub-network, and acquiring a corresponding target cross vector includes:

inputting the low-dimensional dense embedded vector into a compression reward punishment layer in the shallow sub-network, and extracting the importance degree between sparse feature fields in the high-dimensional sparse feature based on the compression reward punishment layer to obtain a corresponding importance degree weighting vector;

5. The method of claim 4, wherein the inputting the low-dimensional dense embedded vector into a compression punishment layer in the shallow sub-network, and extracting importance degrees between sparse feature fields in the high-dimensional sparse feature based on the compression punishment layer to obtain a corresponding importance degree weighting vector comprises:

6. The method of claim 4, wherein the low-dimensional dense embedding vector and the importance weighting vector are in the form of a matrix; the performing cross feature extraction on the low-dimensional dense embedded vector and the importance degree weighting vector based on the bilinear cross layer to obtain a first cross vector corresponding to the low-dimensional dense embedded vector and a second cross vector corresponding to the importance degree weighting vector, including:

7. The method of claim 3, wherein the deep sub-network comprises a plurality of hidden layers, and the inputting the target cross vector into the deep sub-network and the mapping the target cross vector based on the deep sub-network to obtain the predicted probability values output by the predictive models of interest comprises:

8. The method of claim 2, wherein the attention prediction model is obtained by training:

respectively extracting the features of the high-dimensional sparse features in the at least one training sample based on the attention prediction model to obtain each predicted probability value output by the attention prediction model;

9. The method of claim 1, further comprising, prior to the extracting account features corresponding to a target object based on historical behavior data of the target object for target activities,:

10. The method of any one of claims 1 to 9, wherein when the image feature comprises a behavioral image feature of the target object, the extracting the image feature corresponding to the target object based on the image data related to the target object comprises:

11. The method of any of claims 1 to 9, wherein when the image features include device image features of a registration device used by the target object, the extracting image features corresponding to the target object based on image data associated with the target object comprises:

12. The method of any one of claims 1 to 9, wherein when the image feature comprises a feedback image feature of the target object on the target activity within a second set time period, the extracting the image feature corresponding to the target object based on the image data related to the target object comprises:

acquiring feedback behavior data generated by the target object aiming at the target activity within a second set time period;

13. An activity attention prediction device comprising:

the characteristic extraction unit is used for extracting account characteristics corresponding to a target object based on historical behavior data of the target object for target activities and extracting portrait characteristics corresponding to the target object based on portrait data related to the target object;

the characteristic splicing unit is used for performing characteristic splicing on the account characteristic and the portrait characteristic to obtain a high-dimensional sparse characteristic corresponding to the target object;

14. An electronic device, characterized in that it comprises a processor and a memory, wherein the memory has stored a program code which, when executed by the processor, causes the processor to carry out the steps of the method according to any one of claims 1 to 12.

15. Computer-readable storage medium, characterized in that it comprises program code means for causing an electronic device to carry out the steps of the method as claimed in any one of claims 1 to 12 when said program product is run on said electronic device.