CN116702016A

CN116702016A - Object attribute identification method, device, equipment and storage medium

Info

Publication number: CN116702016A
Application number: CN202210167564.9A
Authority: CN
Inventors: 樊鹏
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-02-23
Filing date: 2022-02-23
Publication date: 2023-09-05

Abstract

The application discloses an object attribute identification method, device, equipment and storage medium, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic, internet of vehicles and the like, and the method comprises the following steps: acquiring characteristics of an object to be identified; performing object attribute identification processing on the object features to be identified based on the object attribute identification model to obtain target attribute tags corresponding to the object features to be identified; the training method of the object attribute identification model comprises the following steps: acquiring characteristics of a sample object; pre-training the characteristics of the sample object based on a preset network to obtain a first attribute identification model; obtaining an initial attribute identification model based on the first attribute identification model and the first output result; and updating a loss function of the initial attribute identification model based on the objective function, and training to obtain the object attribute identification model. The object attribute identification model improves the identification accuracy of the object attribute.

Description

Object attribute identification method, device, equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method, an apparatus, a device, and a storage medium for identifying object attributes.

Background

In the related art, a data rule is generally determined by using a manual experience, so that an object attribute tag is determined. The product operation is based on business experience, and a rule for identifying the life style of a user is set, for example, the user who is aged 25-35 years, always living in the first line city and in the specific technical field is considered to belong to a house, namely, a male and a female, and the probability is higher. The method for determining the data rules based on the manual experience not only uses a very limited number of rules, but also cannot capture the high-dimensional characteristic information of interactions between the rules, and most importantly cannot determine the optimal parameters of each rule, so that the accuracy of determining the object attributes is low.

Therefore, it is necessary to provide a method, an apparatus, a device and a storage medium for identifying object attributes, so as to improve the accuracy of identifying object attributes.

Disclosure of Invention

The application provides a method, a device, equipment and a storage medium for identifying object attributes, which can improve the accuracy of identifying object attributes.

In one aspect, the present application provides an object attribute identification method, the method including:

acquiring characteristics of an object to be identified;

performing object attribute identification processing on the object features to be identified based on an object attribute identification model to obtain target attribute tags corresponding to the object features to be identified;

The training method of the object attribute identification model comprises the following steps:

acquiring characteristics of a sample object; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features;

pre-training the sample object characteristics based on a preset network to obtain a first attribute identification model;

obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged;

updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model.

Another aspect provides an object attribute identification apparatus, the apparatus comprising:

the object feature acquisition module is used for acquiring the object feature to be identified;

the attribute tag identification module is used for carrying out object attribute identification processing on the object characteristics to be identified based on an object attribute identification model to obtain target attribute tags corresponding to the object characteristics to be identified;

The model training module is used for training the object attribute identification model;

wherein, the model training module includes:

the sample object feature acquisition sub-module is used for acquiring sample object features; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features;

the first attribute identification model determining submodule is used for pre-training the characteristics of the sample object based on a preset network to obtain a first attribute identification model;

the initial attribute identification model determining submodule is used for obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged;

and the object attribute identification model determining submodule is used for updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model.

Another aspect provides an object attribute identification device comprising a processor and a memory having stored therein at least one instruction or at least one program loaded and executed by the processor to implement an object attribute identification method as described above.

Another aspect provides a computer storage medium storing at least one instruction or at least one program loaded and executed by a processor to implement an object attribute identification method as described above.

Another aspect provides a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions so that the computer device executes to implement the object attribute identification method as described above.

The object attribute identification method, the device, the equipment and the storage medium provided by the application have the following technical effects:

the method comprises the steps of obtaining the characteristics of an object to be identified; performing object attribute identification processing on the object features to be identified based on an object attribute identification model to obtain target attribute tags corresponding to the object features to be identified; the training method of the object attribute identification model comprises the following steps: acquiring characteristics of a sample object; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features; pre-training the sample object characteristics based on a preset network to obtain a first attribute identification model; obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged; updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model. According to the application, based on the first attribute identification model and the first output result, an initial attribute identification model with higher accuracy is obtained; updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model, and continuing training to obtain an object attribute identification model capable of accurately identifying the object attribute; thereby improving the recognition accuracy of the object attribute.

Drawings

In order to more clearly illustrate the embodiments of the application or the technical solutions and advantages of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are only some embodiments of the application, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of an object attribute identification system according to an embodiment of the present application;

FIG. 2 is a schematic flow chart of an object attribute identification method according to an embodiment of the present application;

FIG. 3 is a flowchart of a method for obtaining characteristics of a sample object according to an embodiment of the present application;

FIG. 4 is an exemplary first method for determining an initial attribute identification model provided by an embodiment of the present application;

FIG. 5 is a second example of a method for determining an initial attribute identification model according to an embodiment of the present application;

FIG. 6 is a flow chart of a method for training an initial attribute identification model according to an embodiment of the present application;

FIG. 7 is a flowchart of a method for verifying an object attribute identification model according to an embodiment of the present application;

FIG. 8 is a flowchart of a method for determining characteristics of a target sample object according to a sorting result according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a DenseNet model according to an embodiment of the present application;

fig. 10 is a schematic diagram of a tag identification principle of a DenseNet model according to an embodiment of the present application;

FIG. 11 is a graphical representation of AUC versus user lifestyle predictions for different models provided by an embodiment of the present application;

FIG. 12 is a schematic diagram of a comparison of the effects of different models provided by embodiments of the present application on a user's lifestyle prediction;

fig. 13 is a schematic structural diagram of an object attribute identifying apparatus according to an embodiment of the present application.

Fig. 14 is a schematic structural diagram of a server according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

The embodiment of the application provides an object attribute identification method, an object attribute identification device, object attribute identification equipment and a storage medium. Specifically, the object attribute identifying method according to the embodiment of the present application may be executed by a computer device, where the computer device may be a terminal or a server. The embodiment of the application can be applied to various scenes such as data security, cloud technology, artificial intelligence, intelligent traffic and the like.

First, partial terms or terminology appearing in the course of describing the embodiments of the application are explained as follows:

artificial intelligence (Artificial Intelligence, AI) is the theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.

The intelligent transportation is a new generation information technology such as the Internet of things, space perception, cloud computing, mobile Internet and the like in the whole transportation field, and the theories and tools such as traffic science, system methods, artificial intelligence, knowledge mining and the like are comprehensively utilized, the comprehensive perception, deep fusion, active service and scientific decision making are taken as targets, and the related data of the transportation are deeply mined by constructing a real-time dynamic information service system to form a problem analysis model, so that the improvement of the industry resource allocation optimizing capability, public decision making capability, industry management capability and public service capability is realized, the transportation is promoted to be safer, more efficient, more convenient, more economical, more environment-friendly and more comfortable to operate and develop, and the transportation related industry is driven to be transformed and upgraded.

Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence.

Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.

Deep Learning (DL): is a branch of machine learning, an algorithm that attempts to abstract data at a high level using multiple processing layers, either comprising complex structures or consisting of multiple nonlinear transformations. Deep learning is the inherent rule and expression level of characteristic data of a learning sample object, and information obtained in the learning process is greatly helpful to interpretation of data such as characters, images, sounds and the like. The final goal of deep learning is to enable a machine to analyze learning capabilities like a person, and to recognize text, images, and sound data. Deep learning is a complex machine learning algorithm that achieves far greater results in terms of speech and image recognition than prior art.

Neural Networks (NN): a deep learning model imitating the structure and function of a biological neural network is disclosed in the fields of machine learning and cognitive science.

Characterization learning: in the field of machine learning, token learning (or feature learning) is a collection of techniques that transform raw data into a solution that can be efficiently developed by machine learning. Before the feature learning algorithm appears, a machine learning researcher needs to establish features from domain knowledge (domain knowledge) of the original data by using technologies such as manual feature engineering (manual feature learning), and then deploy a related machine learning algorithm. While manual feature engineering is effective for applying machine learning, it is also difficult, expensive, time consuming, and relies on strong expertise. Feature learning compensates for this by allowing the machine to learn not only the features of the data, but also to use those features to accomplish a specific task.

Metric learning: the object is typically the distance of the feature vectors of the samples, and the purpose of metric learning is to reduce or limit the distance between like samples while increasing the distance between different classes of samples by training and learning. Metric learning (metric learning) study how to learn a distance function over a particular task, so that the distance function can help neighbor-based algorithms (kNN, k-means, etc.) achieve better performance.

User lifestyle: refers to the current long-term lifestyle of the user. The value is as follows: working maniman, living superman, house man and woman.

The k-nearest neighbor method (k-nearest neighbors algorigthm) is one of the most basic methods for classification and regression, and when there is no distribution about training data, the k-nearest neighbor method is the first most conceivable. The k-nearest neighbor method is input as a feature vector of the instance, and output as a class of the instance. The algorithm idea is that given a training data set, corresponding to each data point in the input space, if a new data point classification is to be determined, then the k nearest data points of the target data point are taken, then the number of each classification in the k data points is counted, and the classification with the largest number is taken as the classification of the target data point. The k-nearest neighbor method typically uses Euclidean distance to characterize the distance of two data points, the smaller the distance, the closer the two data points are.

Company a WiFi housekeeper: the WiFi housekeeper of the company A is free WiFi management software under the flag of the company A, supports billions of public WiFi hotspots, can be connected by one key without inputting a password, comprehensively evaluates the hotspots through the five-star WiFi standard of the company A in all directions, the connection speed, the network quality and the like, and ensures that bot viruses, risks and false WiFi are avoided.

AUC: is an important and common index for measuring the classification performance of a machine learning model, and can only be used for the case of two classifications. The AUC is known as Area under the ROC curve, which is the area between the ROC curve and the x-axis (FPR axis). AUC considers the ranking quality of the model predictions, reflecting the proportion of the model that ranks the positive cases before the negative cases (if auc=1, it is stated that the model 100% ranks all positive cases before the negative cases). The full name of the ROC is Receiver OperatingCharacteristic curve, which is a curve drawn by taking the false positive rate FPR as a horizontal axis and the true positive rate TPR as a vertical axis, and the closer the ROC curve is to the upper left corner, the better the performance of the model is shown. If the ROC curve of model a can completely "encase" the ROC curve of model B, then it can be concluded that model a's performance is better than model B's performance, but that the ROC curves of the two models tend to intersect, where AUC is needed for comparison performance.

It should be noted that the terms "first," "second," and the like in the description and the claims of the present application and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the application described herein may be implemented in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or server that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed or inherent to such process, method, article, or apparatus, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

The embodiment of the application establishes a set of accurate and effective automatic identification system with strong reusability for 'user life style prediction' based on a deep learning technology, wherein the user life style can comprise 'house men and women', 'working manikin', 'super life' and the like. In theory, the recognition probability of the corresponding scene can be recognized only by adjusting the characteristics of the sample object input into other scenes, such as 'technical field of the user', 'highest academy of the user', and the like. The identification of the life style of the user is helpful for improving the fine operation of the App flow.

It should be noted that, the training process and the actual prediction process of the object attribute identification model may be completed in the server or may be completed in the terminal. When the training process and the actual prediction process of the model are completed at the middle end of the server and the trained object attribute is required to be used for identifying the model, the object set to be identified can be input into the server, and after the actual prediction of the server is completed, the obtained prediction label information of the object to be identified is sent to the terminal for display. When the training process and the actual prediction process of the model are completed in the terminal and the trained object attribute identification model is required to be used, the object set to be identified can be input into the terminal, and the terminal displays the prediction label information of the object to be identified after the actual prediction of the terminal is completed. When the training process of the model is completed in the server and the actual prediction process of the model is completed in the terminal, and the trained object attribute is required to be used for identifying the model, the object set to be identified can be input into the terminal, and the terminal displays the prediction label information of the object to be identified after the actual prediction of the terminal is completed. Optionally, the trained model file (model file) in the server may be transplanted to the terminal, if the input object set to be identified needs to be predicted, the object set to be identified is input to the trained model file (model file), and the prediction label information of the object to be identified may be obtained through calculation.

Referring to fig. 1, fig. 1 is a schematic diagram of an object attribute identifying system according to an embodiment of the present application, and as shown in fig. 1, the object attribute identifying system may at least include a server 01 and a client 02.

Specifically, in the embodiment of the present application, the server 01 may include an independently operating server, or a distributed server, or a server cluster formed by a plurality of servers, and may also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network ), and basic cloud computing services such as big data and artificial intelligence platforms. The server 01 may include a network communication unit, a processor, a memory, and the like. Specifically, the server 01 may be configured to train an object attribute recognition model, and perform object attribute recognition processing on the object feature to be recognized according to the model, so as to obtain a target attribute tag corresponding to the object feature to be recognized.

Specifically, in the embodiment of the present application, the client 02 may include smart phones, desktop computers, tablet computers, notebook computers, digital assistants, intelligent wearable devices, intelligent sound boxes, vehicle terminals, intelligent televisions, and other types of entity devices, or may include software running in the entity devices, for example, web pages provided by some service providers to users, or may also provide applications provided by the service providers to users. Specifically, the client 02 may be configured to query, online, a target attribute tag corresponding to a feature of the object to be identified.

In the following description, fig. 2 is a schematic flow chart of an object attribute identifying method according to an embodiment of the present application, where the method operation steps described in the examples or the flowcharts are provided, but more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps recited in the embodiments is merely one way of performing the order of steps and does not represent a unique order of execution. When implemented in a real system or server product, the methods illustrated in the embodiments or figures may be performed sequentially or in parallel (e.g., in a parallel processor or multithreaded environment). As shown in fig. 2, the method may include:

s201: and obtaining the characteristics of the object to be identified.

In the embodiment of the application, the object feature to be identified can be a feature corresponding to the target object, and the object feature to be identified and the sample object feature are features of the same category; the characteristics of the object to be identified can be one or more, and the characteristic type can be determined according to the attribute label to be predicted; for example, if the attribute tag to be predicted is a life style of the user, the object feature to be identified is a feature associated with the life style of the user; if the attribute label to be predicted is the highest academy of the user, the object to be identified is characterized as the feature associated with the academy.

S203: and carrying out object attribute identification processing on the object feature to be identified based on the object attribute identification model to obtain a target attribute tag corresponding to the object feature to be identified.

In the embodiment of the application, the object attribute recognition model can be obtained through pre-training, or the object attribute recognition model can be generated in real time after the object characteristics to be recognized are obtained, and the target attribute label is predicted.

s301: acquiring characteristics of a sample object; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features;

in the embodiment of the application, the sample attribute tag can comprise a plurality of types, and can be specifically determined according to the sample object attribute, wherein the object attribute can comprise life style, highest school or technical field of the user. The life modes of the user can comprise three types of working manikins, living superpersons and home men and women, and the three life modes can be characterized by adopting the identification information as an attribute tag. For example, the sample attribute tags may be a number 1,2,3,1 that characterizes a work manikin, 2 that characterizes a super living person, 3 that characterizes a home male and female.

In the embodiment of the application, the sample object which is strongly related to the service, has normal data distribution and reasonable object association data can be found out from the original sample object based on the manual annotation and the service experience, the seed object in the sample object is screened out, and the abnormal seed object is filtered out, so that the sample object is obtained. For example, the object may be a user and the object attribute may be a user's lifestyle.

In an embodiment of the present application, as shown in fig. 3, the acquiring the sample object feature may include:

s30101: acquiring a seed object with Label (Label) information from an original sample object based on recall rules or business logic;

in the embodiment of the application, the recall rule comprises that the terminal equipment used by the object is provided with the application in the application white list, and the business logic comprises the activity track of the object. For example, a batch of candidate seed objects are recalled roughly based on recall rules, then the candidate seed objects are filtered based on manual screening, and finally the candidate seed objects are verified based on business logic to obtain seed objects with tag information. For example, the tag information may be a lifestyle tag of the subject, and may include 1,2,3; wherein 1 represents a working maniman, 2 represents a living superman, and 3 represents a house man and a house woman; the label information can also be the highest academic label of the object, and can comprise the family, the master and the doctor; the label information can also be a technical field label of the object, and can comprise teachers, doctors, public officers and the like. When the tag information is a life style tag (super life) of the object, the recall rule may install an application in an application white list for the terminal device used by the object, for example, the application white list includes a food shopping App, a clothing shopping App, a life payment App, a child care App, a financial App, and the like. For example, a financial class App is installed on the basis of terminal equipment used by the object to roughly recall a batch of candidate seed objects; the business logic may include the activity trajectory of the object in the target area, such as a supermarket, shopping mall, etc. where the object must often appear around the address. If the candidate seed object never passes through a living supermarket or the like, the candidate seed object needs to be removed so as to obtain the seed object with the tag information, and further sample object characteristics are obtained.

S30103: acquiring basic association data of the seed object;

in the embodiment of the application, the basic associated data includes non-privacy behavior data of the seed object in the target application set, for example, the target application set may be an a company series App, for example, the basic associated data includes some non-privacy behavior data of the seed object in the a company series App, for example, the non-privacy behavior data includes behavior data such as whether an a company WiFi manager is installed, whether an a company mobile phone manager harassment interception function is used, whether a time of using a specific App by the object reaches a certain time threshold, and the like. The basic association data of the seed object is not all basic association data characteristics of the seed object, but only contains part of basic association data characteristics of the seed object, the basic association data of the seed object is used for evaluating the quality of the seed object, the basic association data of the seed object is generally not subjected to cross processing of a time dimension, and whether the basic association data of the seed object meets specific abnormal index conditions is generally judged after the basic association data of the seed object is obtained so as to verify whether the value of the seed object in certain association data dimensions is an abnormal value. For example, if the tag information is a candidate seed object a for a house man or woman, and the time for browsing a certain official website per day exceeds 8 hours, it may be in a house office state instead of a house man or woman, and thus the candidate seed object a may be determined as an abnormal seed object. The above-mentioned basic associated data are obtained to obtain the consent of the user, and the corresponding data are used within the scope of meeting the laws and regulations.

S30105: calculating an abnormal object type evaluation index according to the basic association data of the seed object;

in the embodiment of the application, the abnormal object type evaluation index is used for evaluating whether an abnormal seed object exists in the seed objects. In a real service scene, false objects exist, and a mobile phone is controlled by a computer, in order to eliminate the influence of non-real objects on modeling analysis, abnormal object type evaluation indexes are set based on service experience, for example, the abnormal object type evaluation indexes include but are not limited to: traffic usage of the object in the target application set (company a series product), time distribution of traffic generation, and the like.

S30107: and filtering the abnormal seed object from the seed object according to the abnormal object type evaluation index and the distribution abnormal theorem to obtain an updated seed object with label information, and obtaining a sample object.

In the embodiment of the present application, the distribution anomaly theorem may include a rad criterion, and the rad criterion may be used as a criterion for anomaly values. The Laida criterion is also called as 3 sigma criterion, a group of detection data is assumed to contain only random errors, standard deviation is obtained by calculating the detection data, a section is determined according to a certain probability, and the error exceeding the section is considered to be not random error but coarse error, and the data containing the error should be removed. Sigma represents the standard deviation and μ represents the mean in the normal distribution. x=μ is the symmetry axis of the image, 3σ principle: the probability of the numerical distribution in (μ - σ, μ+σ) is 0.6827; the probability of the numerical distribution in (μ -2σ, μ+2σ) is 0.9544; the probability of the numerical distribution in (μ -3σ, μ+3σ) is 0.9974; it is considered that the value of Y is almost entirely concentrated in the (μ -3σ, μ+3σ) interval, and the possibility of exceeding this range is only less than 0.3%. When the abnormal seed object is filtered, the abnormal seed object can be directly deleted from the candidate sample objects.

S30109: based on the sample object, sample object features are acquired.

In the embodiment of the application, the sample object characteristics corresponding to the sample object can be obtained, so that the model training of the next step is performed.

In the embodiment of the application, the basic associated data characteristic of the sample object can be constructed based on the characteristic of the sample object, and the high-dimensional characteristic vector can be produced based on the service vertical type characteristic of the sample object by combining the time dimension and different characteristic processing methods. For example, the feature processing may be offline feature processing or online feature processing.

Optionally, a plurality of sample object features may be grouped into a sample object feature set prior to model training; and carrying out feature vector processing on the sample object feature set to obtain feature vectors of each sample object feature in the sample object feature set, and carrying out model training based on the feature vectors.

The feature vector processing is performed on the sample object feature set to obtain a feature vector of each sample object feature in the sample object feature set, including:

and constructing basic associated data features of each sample object feature in the sample object feature set.

In an embodiment of the present application, the constructing a basic associated data feature of each sample object feature in the sample object feature set includes:

s1: and constructing basic associated data characteristics of each sample object characteristic in the sample object characteristic set based on the object historical behavior data of each sample object characteristic in the sample object characteristic set, wherein the basic associated data characteristics comprise object basic attributes, equipment basic attributes and network connection attributes.

And constructing basic associated data features of each sample object feature in the sample object feature set based on the object historical behavior data of each sample object feature in the sample object feature set. Wherein the base associated data features may include: object basic properties, device basic properties, network connection properties, and the like. For example, the subject base attributes may include gender, age (1-10 years, 11-20 years, etc.), and the like. For example, the device base attributes may include a device brand, a device model, etc. of the terminal device used by the object. For example, the network connection attributes may include the number of times the subject connects Wi-Fi per day, wi-Fi coverage, etc.

S2: constructing a business vertical type feature of each sample object feature in the sample object feature set based on the business characteristics of each sample object feature in the sample object feature set;

In the embodiment of the application, the service characteristics may include that the terminal device used by the object is provided with an application in an application white list, and when the tag information is a life style tag (super life) of the object, the application white list includes a food shopping App, a clothing shopping App, a life payment App, a child care App, a financial App and the like. The service vertical type feature may include click rate, conversion rate, etc. of the object on the advertisement of the specific type, and if the advertisement of the specific type belongs to the advertisement of the electronic commerce type, the service vertical type feature further includes commodity purchase rate, etc. The App is obtained to obtain the consent of the user, and the corresponding data use is within the scope of meeting the laws and regulations.

S3: combining preset time dimensions to aggregate basic associated data features and business vertical type features of each sample object feature in the sample object feature set so as to aggregate aggregated basic associated data features and aggregated business vertical type features of different time dimensions;

in an embodiment of the application, the aggregated underlying associated data features of the subject are calculated for the last half year/last 3 months/last 1 month/last 1 week. For example, the method of aggregation may be selected from summation, median, standard deviation, and the like. For example, the vertical type feature of the service of the specific App is feature a, wherein the preset time dimension includes four time dimensions of about 1 day, about 1 week, about 1 month, about 3 months, and the feature a and the four time dimensions are cross-processed to become four features: a1, A2, A3 and A4, and then aggregating the features (A1, A2, A3 and A4) corresponding to the four different time dimensions to obtain an aggregated business vertical type feature A'. The basic associated data features and the business vertical type features are basic features of sample object features, and in order to enrich the features of the basic associated data features and the business vertical type features, which can be combined with the time dimension, the time dimension needs to be combined, and the aggregated basic associated data features and the aggregated business vertical type features with different time dimensions are aggregated.

S4: performing feature vector processing on the basic associated data features before and/or after aggregation and the business vertical type features to obtain first feature information of each sample object feature in the sample object feature set;

in the embodiment of the present application, the feature vector processing for the pre-aggregation and/or post-aggregation basic associated data features and the service vertical type features includes:

normalizing the numerical value type characteristics in the basic associated data characteristics and the business vertical type characteristics before and/or after aggregation; and discretizing the non-numerical type features in the basic associated data features and the business vertical type features before and/or after aggregation.

Specifically, the numerical type features in the basic associated data features and the business vertical type features before and/or after aggregation are normalized to map the original feature values of the associated features to a certain range (such as [0,1] or [ -1,1 ]), so as to eliminate adverse effects caused by singular sample object feature data. For example, discretizing non-numeric ones of the pre-and/or post-aggregate base associated data features and business vertical type features may translate the associated features into features that are more easily identifiable by the model. For example, age group is a numeric feature and gender is a non-numeric feature.

For example, the normalization method may select gaussian normalization.

For example, the discretization processing may include the following methods, and specifically, the discretization processing may be performed by selecting a corresponding method according to the characteristics of the feature itself:

(1) One-Hot Encoding (One-Hot Encoding). For example, for the characteristics of the object sex, the discretization processing is performed by using One-HotEncoding.

(2) Frequency Encoding (Count Encoding). For example, for a WiFi map point of interest (Point of Interest, POI) feature of an object, count Encoding would be used to identify the object and the interest level of this POI. For example, the subject has consumed 3 times the POI of "food-Chinese dish-Yue dish".

(3) Category embedment (Category Embedding). According to data analysis, many categories of features have strong sparsity, and in order to avoid model overfitting and improve model stability, a neural network is introduced for Category Embedding to convert high-dimensional sparse classification variables into low-dimensional dense Embedding (Embedding) variables.

(4) Deletion value embedded expression (NaN Embedding). For the processing of the missing values of the features, the methods such as 'reject', 'average filling', 'missing mark' and the like are adopted in the experiment, and experimental results show that the mode of converting the missing values into the missing values by using NaN missing is adopted, so that the method has the greatest forward benefit on the effect of the model.

(5) Merging codes (Consolidation Encoding). Multiple values under certain category variables can be generalized to the same information. For example, three values of the system version characteristics of the android mobile phone include "4.2", "4.4" and "5.0", and these three values can be summarized as "low-version android system" based on experience. Experiments prove that the Consolidation Encoding processing mode can bring more forward benefits than the direct android system version feature one-hot.

S5: processing category characteristics in the first characteristic information of each sample object characteristic in the sample object characteristic set to obtain second characteristic information of each sample object characteristic in the sample object characteristic set;

in the embodiment of the application, the second characteristic information comprises an embedded characteristic, and the embedded characteristic comprises a WiFi track embedded characteristic and an App flow embedded characteristic.

Specifically, category (Category) features in the first feature information are input into a DNN model for training to obtain second feature information, wherein the second feature information comprises embedded features. The task of deep learning is to map high-dimensional raw data (such as images and sentences) to low-dimensional manifolds, so that the high-dimensional raw data becomes separable after being mapped to the low-dimensional manifolds, and the mapping is called Embedding. The Embedding feature can be understood as extracting the low-dimensional dense weight feature of the middle full-connection layer of the neural network, replacing the original feature expression, and the general dimension is the power of 2 to N, such as 128. The step of training the Embedding feature may include:

(1) WiFi track embedded (WiFi Trajectory Embedding) features are acquired. Based on the MST-CNN deep learning network, the WiFi connection track data of the object is subjected to Embedding, and template (Pattern) information of Wi-Fi behaviors of the object is captured. Wherein MST (multiple spanning tree) is a Multiple Spanning Tree (MST) standard of IEEE (Institute of Electrical and Electronics Engineers ) derived from the implementation of Cisco (Cisco) specific Multiple Instance Spanning Tree Protocol (MISTP). Convolutional neural networks (Convolutional Neural Networks, CNN) are a type of feedforward neural network (Feedforward Neural Networks) that contains convolutional calculations and has a deep structure, and are one of the representative algorithms of deep learning. By performing the eimbedding, discrete variables can be converted into continuous vectors, which also serve to visualize the relationships between different discrete variables. For example, pattern information may be used to represent different expressions of different types of objects existing on Wi-Fi connection, where the expressions are often implicitly represented in the Embedding, and Pattern information of Wi-Fi behavior of the object may be captured by Embedding WiFi connection track data of the object, for example, pattern information is of a male-in-home type, social type, or the like.

(2) An App traffic embedded (App Traffic Embedding) feature is obtained. Based on a List-Embedding mode, extracting the Embedding by using a behavior sequence for the flow of different categories of apps in a target application set (such as a company A series) of the object, such as Traffic Embedding of a social type App of the company A series, so as to obtain low-dimensional dense object behavior characteristics. The List-Embedding is a method of Embedding, and is different from other Embedding methods in that the List-Embedding splices the input of a plurality of values into a List (List).

S6: and combining the first characteristic information and the second characteristic information of each sample object characteristic in the sample object characteristic set to obtain a characteristic vector of each sample object characteristic in the sample object characteristic set.

In an embodiment of the present application, the merging the first feature information and the second feature information of each sample object feature in the sample object feature set to obtain a feature vector of each sample object feature in the sample object feature set includes:

and splicing the low-dimensional features and the high-dimensional features in the first feature information and the second feature information in columns to obtain feature vectors of each sample object feature in the sample object feature set.

After the first feature information and the second feature information of each sample object feature in the sample object feature set are combined to obtain feature vectors of each sample object feature in the sample object feature set, the obtained feature vectors of each sample object feature in the sample object feature set can be stored in a Hadoop distributed file system (The Hadoop Distributed File System, HDFS) in an offline manner, so that the quick access of subsequent processes is facilitated. The HDFS is used for storing files, and large-capacity, high-speed and reliable storage and access of data can be realized through a large-scale distributed server cluster. When merging, the first feature information and the low-dimensional feature and the high-dimensional feature in the second feature information of each sample object feature in the sample object feature set obtained after processing can be spliced according to columns so as to obtain feature vectors of each sample object feature in the sample object feature set.

In some embodiments, after step S6, the method may further include:

s7: and the solidification feature processing logic is used for timing offline automatic calculation and pushing (Push) an offline calculation result to an online storage engine. Specifically, the code of the feature processing logic is cured and the code of the feature processing logic is run at regular time.

S303: pre-training the sample object characteristics based on a preset network to obtain a first attribute identification model;

in the embodiment of the present application, the preset network may be DenseNet; denseNet (Dense Convolutional Network) is a convolutional neural network with dense connections. In this network there is a direct connection between any two layers, that is, the input to each layer of the network is the union of the outputs of all the previous layers, and the feature map learned by that layer is also passed directly to all the subsequent layers as input. The feature is characterized in that dense connection of all the front layers and the rear layers is realized, and feature reuse is realized through connection of features on a dimension (channel). These features allow DenseNet to achieve better performance than other network architectures with less parameters and computational costs. In a conventional convolutional neural network, if there are L layers, there are L connections; however, in DenseNet, there are L (L+1)/2 connections. In brief, the input of each layer comes from the outputs of all the preceding layers. The structure of the DenseNet model is shown in FIG. 9, where the input of H1 is x0 (input), the input of H2 is x0 and x1 (x 1 is the output of H1), the input of H3 is x0, x1 and x2 (x 2 is the output of H2), and the input of H4 is x0, x1, x2 and x3 (x 3 is the output of H3).

Compared with other network structures, denseNet proposes a more aggressive dense connection mechanism: i.e. all layers are interconnected, i.e. each layer will accept all its preceding layers as its additional input. It follows that in DenseNet, each layer is connected (concat) together with all previous layers in the channel dimension as input to the next layer.

As shown in fig. 10, the DenseNet includes 3 Dense blocks and each Block has the same number of layers. Before entering the first Dense Block, a Conv (output channel 16) is first passed, keeping the size (size) of the feature-map unchanged. Dense Block is internally densely connected. Between the two Dense blocks is a Transition Layer (Transition Layer). After the last Dense Block is an average pooling layer (avg pool) plus a Softmax Classifier (Classifier). The size of the feature-map of the 3 Dense blocks was 32×32, 16×16, and 8×8, respectively.

In this embodiment, the DenseNet network includes the following key point features:

a) Pooling Layer (Pooling Layer). When the size of the feature-map is different, dense Connectivity cannot be used. The DenseNet divides the network into multiple layers of densely connected Dense Blocks, downsampling operations (BN+1X1Conv+2×2avg pool) are performed between the Blocks (called Transition Layers), where BN (Batch Normalization) is the bulk normalization layer, conv (Convolutional layer) is the convolution layer, and avg pool is the average pooling layer.

b) A dense connection layer (Dense Connectivity). The blocks are densely connected by adopting Dense Connectivity. From between any layer to all subsequent layers. Between layers is bn+relu+3×3conv. Wherein ReLU (Rectified Liner Uints) is a sparse activation function layer.

c) Growth Rate (Growth Rate). If k feature-maps are generated per layer, the number of feature-maps input by the L-th layer is k0+k (L-1), where k0 is the channel input. A DenseNet differs from other network structures in that the layer (layer) of the DenseNet can be very narrow, such as k=12. K is referred to herein as the Growth Rate of the network.

d) Bottleneck layer (Bottleneck Layers). The 1×1conv can be used as a bottleneck layer before 3×3conv, and the dimension of the input feature-map is reduced, so that the calculation efficiency is improved. This structure was introduced in DenseNet-B, and 1X 1conv outputs a feature-map of 4 k.

In this example, denseNet has two prominent contributions compared to other models:

a) Optimizing nature integrates identity mapping (identity mapping), deep supervised learning (deep supervision), and the nature of varying depths. The gradient vanishing problem is relieved, the propagation of the characteristics is enhanced, the reuse of the characteristics is encouraged, and the number of parameters is greatly reduced.

b) Each layer has gradient from original input to loss function, thus realizing hidden depth supervision and being beneficial to training deep network; dense connectivity plays a regularized role, reducing overfitting on small-scale training sets.

In this embodiment, the whole training sample may be divided into a plurality of Batch (Batch/Batch of samples), and the Batch size of the training DenseNet (the size of each Batch of samples) may be set to 32; learning Rate (Learning Rate) was initially 0.1 divided by 5 at 50% and 75% epoch; when a complete data set passes through the neural network once and returns once, this process is called epoch once. (that is, all training samples are propagated forward and backward in the neural network once), and one Epoch is the process of training all training samples once. However, when the number of samples of one Epoch (i.e., all training samples) may be too large (for a computer), it is necessary to divide it into a plurality of small blocks, i.e., into a plurality of latches for training.

After the model prediction is completed, information recommendation can be performed on users with different labels according to the prediction result of the model, for example, advertisements with target types can be pushed to users with specific labels, and the advertisement exposure success rate and the advertisement click rate are counted, so that the prediction accuracy of the model is evaluated.

Evaluation of offline experiments:

1) Mathematical index evaluation:

a) AUC (Area under Curve): the larger the AUC value, the more likely the current classification algorithm will rank positive samples in front of negative samples, resulting in better classification results.

2) On-line experimental evaluation:

a) The effect of the model was evaluated based on the on-line flow of the A/B Test.

b) The evaluation index includes: advertisement exposure success rate, advertisement click rate.

S305: obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged;

in an embodiment of the present application, as shown in fig. 4, the obtaining an initial attribute identification model based on the first attribute identification model and the first output result includes:

s3051: acquiring a feature matrix corresponding to the sample object features when the training of the first attribute identification model converges;

in the embodiment of the application, when the training of the first attribute identification model converges, a feature matrix corresponding to the features of the sample object is obtained, and the feature matrix can be a weight matrix of the feature layer.

S3053: singular value decomposition is carried out on the feature matrix to obtain feature layer parameters;

In the embodiment of the application, singular value decomposition can be carried out on the feature matrix, and the feature layer parameters are determined through the decomposed matrix; the singular value decomposition (Singular Value Decomposition, abbreviated as SVD) is an algorithm widely applied in the field of machine learning, and can be used for characteristic decomposition in a dimension reduction algorithm, a recommendation system, natural language processing and other fields.

In the embodiment of the application, singular value decomposition processing can be carried out on the feature matrix after each training convergence of the subsequent model.

S3055: and obtaining an initial attribute identification model based on the first attribute identification model, the characteristic layer parameters and the first output result.

In the embodiment of the application, the relevance of the network full-connection layer can be removed based on a singular value decomposition algorithm, the distinguishing capability of the DenseNet feature layer is improved, and the orthogonality of the feature layer is maintained.

In an embodiment of the present application, the obtaining an initial attribute identification model based on the first attribute identification model, the feature layer parameter, and the first output result includes:

inputting the first output result into the first attribute identification model to perform constraint training by taking the characteristic layer parameters as constraint conditions to obtain a second attribute identification model;

In the embodiment of the application, the characteristic layer parameters after singular value decomposition processing can be fixed, and other layer parameters are updated until the first attribute identification model is re-converged; this training process is defined as constraint training. And obtaining a second attribute identification model through constraint training.

Canceling the constraint on the characteristic layer parameters, and inputting a second output result into the second attribute identification model to perform relaxation training to obtain an initial attribute identification model; and the second output result is the output result of the target pooling layer corresponding to the second attribute identification model when training is converged.

In the embodiment of the application, the model training process of canceling the constraint on the characteristic layer parameters is relaxation training. The target pooling layer may be a fifth pooling layer in the model-corresponding network. And when the training of the second attribute identification model converges, taking the output result of the target pooling layer corresponding to the model as an input characteristic again, and carrying out the model training of the next round, thereby obtaining the initial attribute identification model.

In the embodiment of the application, after constraint training, the constraint of the feature layer parameters can be canceled, and training of the second attribute identification model is continued until the model converges, so as to obtain an initial attribute identification model. The accuracy of the initial attribute recognition model can be improved through constraint training and relaxation training.

In some embodiments, as shown in fig. 5, the inputting the second output result into the second attribute identification model for relaxation training to obtain an initial attribute identification model includes:

s501: inputting the second output result into the second attribute identification model for relaxation training to obtain a third attribute identification model;

s503: acquiring an updated feature matrix corresponding to the sample object features when the training of the third attribute identification model converges;

in the embodiment of the application, after the relaxation training is finished, the corresponding updated feature matrix when the model converges can be continuously acquired, and the subsequent repeated training steps are carried out.

S505: singular value decomposition processing is carried out on the updated feature matrix to obtain updated feature layer parameters;

in the embodiment of the application, singular value decomposition is carried out on the updated feature matrix to obtain the decomposed updated matrix, thereby obtaining the corresponding updated feature layer parameters.

S507: repeating the constraint training and the relaxation training based on the third attribute identification model, the updated feature layer parameters and a third output result to obtain an initial attribute identification model; and the third output result is the output result of the target pooling layer corresponding to the third attribute identification model when training is converged.

In the embodiment of the application, after one round of constraint training and relaxation training is performed, the model obtained by training can be utilized to continue the next round of constraint training and relaxation training, and the stability of the initial attribute recognition model can be improved through repeated training for a plurality of times.

In the embodiment of the present application, as shown in fig. 6, the repeating the constraint training and the relaxation training based on the third attribute identification model, the updated feature layer parameter, and the third output result to obtain an initial attribute identification model includes:

s5071: based on the third attribute identification model, the updated feature layer parameters and a third output result, performing constraint training and relaxation training to obtain an updated identification model; and taking the third attribute identification model as a current identification model;

in the embodiment of the application, constraint training and relaxation training are the same as the steps, and the subsequent training adopts the model of the previous training and updates the characteristic parameters.

S5073: acquiring a first loss value when the current recognition model converges and a second loss value when the updated recognition model converges;

in the embodiment of the application, the model corresponds to a loss value when training converges each time, and the loss value is usually smaller than a preset value, and the preset value can be set according to actual conditions.

S5075: calculating a target value according to the first loss value and the second loss value; the target value is the absolute value or the ratio of the difference value of the first loss value and the second loss value;

s5077: and if the target value is smaller than or equal to a preset threshold value, determining the updated identification model as the initial attribute identification model.

In the embodiment of the application, the training end point of the model can be judged according to the magnitude of the target value.

In an embodiment of the present application, the method further includes:

s5079: if the target value is larger than the preset threshold value, the updated identification model is taken as the current identification model again;

s50711: performing constraint training and relaxation training based on the current recognition model to obtain a target model, and re-using the target model as an updated recognition model;

s50713: repeating the steps of obtaining the first loss value when the current recognition model converges and the second loss value when the updated recognition model converges to the step of performing constraint training and relaxation training based on the current recognition model to obtain a target model, and taking the target model as the updated recognition model again.

In an embodiment of the present application, the repeating steps include: acquiring a first loss value when the current recognition model converges and a second loss value when the updated recognition model converges; calculating a target value according to the first loss value and the second loss value; if the target value is larger than the preset threshold value, the updated identification model is taken as the current identification model again; and performing constraint training and relaxation training based on the current recognition model to obtain a target model, and re-using the target model as an updated recognition model.

In the embodiment of the application, the stability of the model can be determined according to the target value corresponding to the loss value, and if the target value is smaller, the model with better stability in two continuous training processes can be proved, and the training can be finished. If the target value is larger, the stability of the model is poorer, and the training needs to be continuously repeated.

In the model training process of some embodiments, constraint training and relaxation training may be repeated for a plurality of times, so as to obtain a loss value when each training converges, and stability of the model may be determined through a plurality of loss values, so as to determine whether to end the training process of the model. For example, the plurality of loss values may be grouped, each group including two, the loss value of each group is calculated to correspond to the target value, and model training is ended when the target values corresponding to the loss values of each group are all less than or equal to a preset threshold. And the last model is used as the initial attribute identification model.

S307: updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model.

In the embodiment of the application, an original loss function of the initial attribute identification model can be replaced by an objective function, model training can be continued, wherein the objective function can be AM-Softmax (additive margin Softmax), and feature and weight normalization is carried out on a regression target of Softmax loss. Softmax is used in the multi-classification process, which maps the output of multiple neurons into (0, 1) intervals, which can be understood as probabilities, thereby performing multi-classification; unlike Softmax, AMSoftmax belongs to a strategy of narrowing the intra-class distance and increasing the inter-class distance. The Softmax can only divide the boundary between the categories, while the AMSoftmax can reduce the inter-category distance to increase the inter-category distance, reduce the interval of the category to the range of the target area, and generate the inter-category distance with the size of an edge (margin).

In the embodiment of the application, the AM-softMax loss function is used for measurement learning, so that the recognition capability and recognition accuracy of the model can be improved.

In an embodiment of the present application, as shown in fig. 7, the updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model, and training the sample object feature according to the updated model to obtain the object attribute identification model includes:

s3071: updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model;

in the embodiment of the application, the original loss function of the initial attribute identification model can be replaced by the objective function to obtain the updated model.

In the embodiment of the application, after model training is completed, the model can be verified, so that the recognition accuracy of the model is improved.

S3073: training the sample object feature set according to the updated model to obtain an attribute identification model to be verified; the sample object feature set comprises at least two sample object features;

in the embodiment of the application, the sample object feature set is trained according to the updated model to obtain the attribute identification model to be verified.

S3075: acquiring a feature set of a test object; the test object features in the test object feature set are marked with test attribute tags;

in an embodiment of the present application, a test object feature set may include a plurality of test object features.

S3077: determining target sample object features from the sample object feature set based on the attribute identification model to be verified;

in an embodiment of the present application, as shown in fig. 8, the determining, based on the attribute identification model to be verified, a target sample object feature from the sample object feature set includes:

s30771: determining the comprehensive distance between each sample object feature and the target test object feature based on the attribute identification model to be verified; the target test object feature is any feature in the test object feature set;

in the embodiment of the application, the comprehensive distance can represent the similarity between the characteristics of the sample object and the characteristics of the target test object.

In an embodiment of the present application, the determining, based on the attribute identification model to be verified, a comprehensive distance between each sample object feature and a target test object feature includes:

calculating the original distance between the target test object feature and each sample object feature based on the attribute identification model to be verified;

In the embodiment of the present application, the original distance may be a euclidean distance, and the euclidean distance between the target test object feature and each sample object feature may be calculated.

Determining a preset number of candidate sample object features according to the original distance corresponding to each sample object feature;

in the embodiment of the application, candidate sample object characteristics can be determined according to the original distance; and determining an attribute tag corresponding to the characteristic of the target test object based on the K nearest neighbor method.

In the embodiment of the application, the preset number can be K, and the sample object features can be reordered according to the K neighbor method, so that the recognition capability of the model is improved.

The K-nearest neighbor method (K-nearest neighbors algorigthm) is one of the most basic methods for classification and regression, and when there is no distribution about training data, the K-nearest neighbor method is the first to be considered. The K neighbor method is input as the feature vector of the instance, and output as the class of the instance. The algorithm idea is that given a training data set, corresponding to each data point of the input space, a new data point classification is to be judged, then the K nearest data points of the target data point are taken, then the number of each classification in the K data points is counted, and the classification with the largest number is taken as the classification of the target data point. The K-nearest neighbor method typically uses euclidean distance to characterize the distance of two data points, the smaller the distance, the closer the two data points are.

In some embodiments, the K-nearest neighbor set can also be expanded to a more robust set, and then the distance calculated. Specifically, features in the collection may be added based on the current collection. For example, the radius of the circle is extended outward by the center of the current set, thereby determining more candidate sample object features.

In some embodiments, after determining the candidate sample object features, the test tag of the test object feature may be determined directly according to the K-nearest neighbor method; and determining the duty ratio of each candidate sample label in the K labels according to the candidate sample label corresponding to each characteristic in the K candidate sample object characteristics, determining the candidate sample label with the highest duty ratio as a test label, and matching the obtained test label with the labeled label, so that the model is verified according to the matching result.

Constructing a target sample feature set according to the preset number of candidate sample object features;

calculating a jaccard distance between the test object feature set and the target sample feature set;

and determining the comprehensive distance between each candidate sample object feature and the target test object feature based on the original distance corresponding to each candidate sample object feature and the Jacquard distance.

In the embodiment of the present application, the Jaccard Distance (Jaccard Distance) is an index for measuring the difference between two sets. The original distance and the Jacquard distance can be combined to determine the comprehensive distance corresponding to each candidate sample object feature, weights corresponding to the original distance and the Jacquard distance can be respectively set before the comprehensive distance is determined, and the sum of the weight distances is calculated according to the weights corresponding to the original distance and the Jacquard distance to obtain the comprehensive distance.

S30773: sorting the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature;

in the embodiment of the application, the sorting can be performed according to the comprehensive distance from big to small or from small to big.

S30775: and determining the characteristics of the target sample object according to the sequencing result.

In some embodiments, the sample object feature with the smallest combined distance may be determined as the target sample object feature according to the ranking result.

In the embodiment of the application, the probability value of the sample attribute label corresponding to each sample object characteristic as the target sample attribute label can be determined according to the comprehensive distance; and sequencing according to the probability value corresponding to each sample object characteristic, and determining the sample attribute label with the maximum probability value as the target sample attribute label.

S3079: acquiring a target sample attribute tag corresponding to the target sample object characteristic;

s30711: and if the target sample attribute tag is matched with the test attribute tag, determining the attribute identification model to be verified as the object attribute identification model.

In the embodiment of the application, matching the target sample attribute tag with the test attribute tag means that the target sample attribute tag and the test attribute tag are the same tag. If the attribute identification model to be verified is matched with the object attribute identification model, determining that the attribute identification model to be verified passes verification, wherein the attribute identification model to be verified is the object attribute identification model. If the attribute identification model to be verified is not matched with the attribute identification model to be verified, the training steps are repeated for training.

In some embodiments, after determining a preset number of candidate sample object features according to the original distance corresponding to each sample object feature, the method may further include:

determining candidate attribute labels corresponding to the characteristics of each candidate sample object;

dividing the characteristics of a preset number of candidate sample objects into a target number of candidate characteristic sets according to the candidate attribute labels; candidate sample objects in each candidate feature set correspond to the same candidate attribute tag;

in the embodiment of the application, the candidate sample object features can be classified according to the attribute tags, and the features of the same attribute tag form a set.

Calculating a first distance between each candidate feature in the candidate feature set and the test sample;

in the embodiment of the present application, the first distance may be a euclidean distance.

Calculating a second distance between each candidate feature set and the test sample set, wherein the second distance is a Jaccard distance;

the test samples in the test sample set are the same test attribute labels;

determining a composite distance between the test sample and each candidate feature set based on the first distance and the second distance;

determining a target candidate feature set according to the comprehensive distance corresponding to each candidate feature set;

determining a candidate attribute label corresponding to the target candidate feature set as an output label of the test sample;

and if the output label is consistent with the test attribute label, determining that the model verification is passed, and determining the attribute identification model to be verified as the object attribute identification model.

And if the output label is inconsistent with the test attribute label, training the attribute identification model to be verified repeatedly.

In the embodiment of the application, after model training is finished, the recognition accuracy of the model can be improved by verifying the model.

In the embodiment of the application, the attribute types corresponding to the sample characteristics can be replaced and applied to different scenes; for example, the method can be used for predicting the highest academic label of the user, the server collects corresponding log data, and the corresponding prediction model is obtained by using the same feature splicing, feature processing and model training methods, so that the method is applied.

In a specific embodiment, referring to fig. 11 and 12, fig. 11 shows a graph of AUC contrast of different models versus user lifestyle predictions, and fig. 12 shows a graph of effect contrast of different models versus user lifestyle predictions.

Among them, from the model effect comparison analysis of fig. 11, it is known that: a) From the aspect of off-line AUC effect, the solution of this embodiment is improved by 33.54% on average compared with other solutions; b) From the on-line AUC effect, the solution of this embodiment is improved by 31.42% on average compared to other solutions.

Among them, from the business effect comparison analysis of fig. 12, it can be seen that: a) Compared with other technical schemes, the scheme of the embodiment improves the average click rate of advertisements by 387.48%; b) From the aspect of advertisement conversion rate, compared with other technical schemes, the scheme of the embodiment is averagely improved by 387.72 percent.

The object attribute identification model provided by the embodiment of the application is obviously superior to other models through analysis of the model effect and the service effect, and the accuracy of object attribute identification is effectively improved through the object attribute identification model constructed by the training method provided by the embodiment of the application.

As can be seen from the technical solutions provided by the above embodiments of the present application, the embodiments of the present application obtain features of an object to be identified; performing object attribute identification processing on the object features to be identified based on an object attribute identification model to obtain target attribute tags corresponding to the object features to be identified; the training method of the object attribute identification model comprises the following steps: acquiring characteristics of a sample object; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features; pre-training the sample object characteristics based on a preset network to obtain a first attribute identification model; obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged; updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model. According to the application, based on the first attribute identification model and the first output result, an initial attribute identification model with higher accuracy is obtained; updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model, and continuing training to obtain an object attribute identification model; thereby improving the recognition accuracy of the object attribute recognition model.

The embodiment of the application can be realized by combining a Cloud technology or a blockchain network technology, wherein the Cloud technology refers to a hosting technology for integrating serial resources such as hardware, software, network and the like in a wide area network or a local area network to realize calculation, storage, processing and sharing of data. The cloud technology is based on the general names of network technology, information technology, integration technology, management platform technology, application technology and the like applied by the cloud computing business mode, can form a resource pool, and is flexible and convenient as required. Background services of technical network systems require a large amount of computing and storage resources, such as video websites, picture websites and more portal websites, so cloud technologies need to be supported by cloud computing.

It should be noted that cloud computing is a computing mode, which distributes computing tasks on a resource pool formed by a large number of computers, so that various application systems can acquire computing power, storage space and information service as required. The network that provides the resources is referred to as the "cloud". Resources in the cloud are infinitely expandable in the sense of users, and can be acquired at any time, used as needed, expanded at any time and paid for use as needed. As a basic capability provider of cloud computing, a cloud computing resource pool platform, referred to as a cloud platform for short, is generally called an infrastructure as a service (Infrastructure as a Service, iaaS), and multiple types of virtual resources are deployed in the resource pool for external clients to select for use. The cloud computing resource pool mainly comprises: computing devices (which may be virtualized machines, including operating systems), storage devices, and network devices.

The embodiment of the application also provides an object attribute identification device, as shown in fig. 13, which comprises:

an object feature to be identified acquisition module 1310, configured to acquire an object feature to be identified;

the attribute tag recognition module 1320 is configured to perform object attribute recognition processing on the object feature to be recognized based on an object attribute recognition model, so as to obtain a target attribute tag corresponding to the object feature to be recognized;

a model training module 1330 for training the object attribute identification model;

wherein the model training module 1330 includes:

a sample object feature acquisition submodule 13310 for acquiring sample object features; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features;

a first attribute identification model determination submodule 13320, configured to pre-train the sample object features based on a preset network, to obtain a first attribute identification model;

an initial attribute identification model determination submodule 13330, configured to obtain an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged;

An object attribute identification model determination submodule 13340 is configured to update a loss function of the initial attribute identification model based on an objective function to obtain an updated model, and train the sample object features according to the updated model to obtain the object attribute identification model.

In some embodiments, the initial attribute identification model determination submodule may include:

the feature matrix acquisition unit is used for acquiring a feature matrix corresponding to the sample object features when the training of the first attribute identification model converges;

the characteristic layer parameter determining unit is used for carrying out singular value decomposition treatment on the characteristic matrix to obtain characteristic layer parameters;

and the initial attribute identification model determining unit is used for obtaining an initial attribute identification model based on the first attribute identification model, the characteristic layer parameters and the first output result.

In some embodiments, the initial attribute identification model determining unit may include:

the first input subunit is used for inputting the first output result into the first attribute identification model to perform constraint training by taking the characteristic layer parameters as constraint conditions, so as to obtain a second attribute identification model;

The second input subunit is used for canceling the constraint on the characteristic layer parameters, inputting a second output result into the second attribute identification model for relaxation training, and obtaining an initial attribute identification model; and the second output result is the output result of the target pooling layer corresponding to the second attribute identification model when training is converged.

In some embodiments, the second input subunit may include:

a third attribute identification model determining subunit, configured to input the second output result into the second attribute identification model to perform relaxation training, so as to obtain a third attribute identification model;

the updated feature matrix obtaining subunit is used for obtaining an updated feature matrix corresponding to the sample object feature when the training of the third attribute identification model converges;

the updated feature layer parameter determination subunit is used for carrying out singular value decomposition processing on the updated feature matrix to obtain updated feature layer parameters;

the repeating subunit is configured to repeat the constraint training and the relaxation training based on the third attribute identification model, the updated feature layer parameter, and the third output result, to obtain an initial attribute identification model; and the third output result is the output result of the target pooling layer corresponding to the third attribute identification model when training is converged.

In some embodiments, the object attribute identification model determination submodule may include:

the model updating unit is used for updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model;

the attribute identification model training unit to be verified is used for training the sample object feature set according to the updated model to obtain an attribute identification model to be verified; the sample object feature set comprises at least two sample object features;

the test object feature set acquisition unit is used for acquiring the test object feature set; the test object features in the test object feature set are marked with test attribute tags;

a target sample object feature determining unit configured to determine a target sample object feature from the sample object feature set based on the attribute identification model to be verified;

the target sample attribute tag acquisition unit is used for acquiring a target sample attribute tag corresponding to the target sample object characteristic;

and the model determining unit is used for determining the attribute identification model to be verified as the object attribute identification model if the target sample attribute label is matched with the test attribute label.

In some embodiments, the target sample object feature determination unit may include:

The comprehensive distance determining subunit is used for determining the comprehensive distance between each sample object characteristic and the target test object characteristic based on the attribute identification model to be verified; the target test object feature is any feature in the test object feature set;

a sorting subunit, configured to sort the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature;

and the target sample object feature determining subunit is used for determining the target sample object feature according to the sequencing result.

In some embodiments, the integrated distance determination subunit may include:

the original distance calculating subunit is used for calculating the original distance between the target test object characteristic and each sample object characteristic based on the attribute identification model to be verified;

the candidate sample object feature determining subunit is used for determining a preset number of candidate sample object features according to the original distance corresponding to each sample object feature;

a target sample feature set construction subunit, configured to construct a target sample feature set according to the preset number of candidate sample object features;

a jaccard distance computing subunit, configured to compute a jaccard distance between the test object feature set and the target sample feature set;

And the distance determining subunit is used for determining the comprehensive distance between each candidate sample object feature and the target test object feature based on the original distance corresponding to each candidate sample object feature and the Jacquard distance.

The device and method embodiments in the device embodiments described are based on the same inventive concept.

The embodiment of the application provides object attribute identification equipment, which comprises a processor and a memory, wherein at least one instruction or at least one section of program is stored in the memory, and the at least one instruction or the at least one section of program is loaded and executed by the processor to realize the object attribute identification method provided by the embodiment of the method.

Embodiments of the present application also provide a computer storage medium that may be provided in a terminal to store at least one instruction or at least one program related to implementing an object attribute identification method in a method embodiment, where the at least one instruction or at least one program is loaded and executed by the processor to implement the object attribute identification method provided in the method embodiment.

Embodiments of the present application also provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes to implement the object attribute identification method provided by the above-mentioned method embodiment.

Alternatively, in an embodiment of the present application, the storage medium may be located on at least one network server of a plurality of network servers of the computer network. Alternatively, in the present embodiment, the storage medium may include, but is not limited to: a U-disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a removable hard disk, a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The memory according to the embodiments of the present application may be used to store software programs and modules, and the processor executes the software programs and modules stored in the memory, thereby performing various functional applications and data processing. The memory may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, application programs required for functions, and the like; the storage data area may store data created according to the use of the device, etc. In addition, the memory 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. Accordingly, the memory may also include a memory controller to provide access to the memory by the processor.

The object attribute identification method provided by the embodiment of the application can be executed in a mobile terminal, a computer terminal, a server or similar computing devices. Taking the operation on a server as an example, fig. 14 is a block diagram of a hardware structure of a server of an object attribute identifying method according to an embodiment of the present application. As shown in fig. 14, the server 1400 may vary considerably in configuration or performance and may include one or more central processing units (Central Processing Units, CPU) 1410 (the central processing unit 1410 may include, but is not limited to, a microprocessor MCU or a processing device such as a programmable logic device FPGA), a memory 1430 for storing data, one or more storage mediums 1420 (e.g., one or more mass storage devices) storing applications 1423 or data 1422. Wherein the memory 1430 and the storage medium 1420 may be transitory or persistent storage. The program stored on the storage medium 1420 may include one or more modules, each of which may include a series of instruction operations on a server. Still further, the central processor 1410 may be configured to communicate with a storage medium 1420, and execute a series of instruction operations in the storage medium 1420 on the server 1400. The server 1400 may also include one or more power supplies 1460, one or more wired or wireless network interfaces 1450, one or more input/output interfaces 1440, and/or one or more operating systems 1421, such as Windows Server, mac OS XTM, unixTM, linuxTM, freeBSDTM, and the like.

Input-output interface 1440 may be used to receive or transmit data via a network. The specific example of the network described above may include a wireless network provided by a communication provider of the server 1400. In one example, input/output interface 1440 includes a network adapter (Network Interface Controller, NIC) that may connect to other network devices through a base station to communicate with the internet. In one example, the input-output interface 1440 may be a Radio Frequency (RF) module for communicating with the internet wirelessly.

It will be appreciated by those of ordinary skill in the art that the configuration shown in fig. 14 is merely illustrative and is not intended to limit the configuration of the electronic device described above. For example, server 1400 may also include more or fewer components than shown in fig. 14, or have a different configuration than shown in fig. 14.

As can be seen from the above embodiments of the object attribute identifying method, apparatus, device or storage medium provided by the present application, the present application obtains features of an object to be identified; performing object attribute identification processing on the object features to be identified based on an object attribute identification model to obtain target attribute tags corresponding to the object features to be identified; the training method of the object attribute identification model comprises the following steps: acquiring characteristics of a sample object; the sample object features are marked with sample attribute tags, and the sample attribute tags represent sample object attributes corresponding to the sample object features; pre-training the sample object characteristics based on a preset network to obtain a first attribute identification model; obtaining an initial attribute identification model based on the first attribute identification model and a first output result; the first output result is an output result of a target pooling layer corresponding to the first attribute identification model when training is converged; updating the loss function of the initial attribute identification model based on an objective function to obtain an updated model, and training the sample object characteristics according to the updated model to obtain the object attribute identification model. According to the application, based on the first attribute identification model and the first output result, an initial attribute identification model with higher accuracy is obtained; updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model, and continuing training to obtain an object attribute identification model capable of accurately identifying the object attribute; thereby improving the recognition accuracy of the object attribute.

It should be noted that: the sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the foregoing description has been directed to specific embodiments of this specification. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for the apparatus, device, storage medium embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and references to the parts of the description of the method embodiments are only required.

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

The foregoing description of the preferred embodiments of the application is not intended to limit the application to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the application are intended to be included within the scope of the application.

Claims

1. An object attribute identification method, the method comprising:

acquiring characteristics of an object to be identified;

2. The method of claim 1, wherein the obtaining an initial attribute identification model based on the first attribute identification model and the first output result comprises:

acquiring a feature matrix corresponding to the sample object features when the training of the first attribute identification model converges;

singular value decomposition is carried out on the feature matrix to obtain feature layer parameters;

and obtaining an initial attribute identification model based on the first attribute identification model, the characteristic layer parameters and the first output result.

3. The method of claim 2, wherein the obtaining an initial attribute identification model based on the first attribute identification model, the feature layer parameters, and the first output result comprises:

4. A method according to claim 3, wherein said inputting the second output result into the second attribute identification model for relaxation training to obtain an initial attribute identification model comprises:

inputting the second output result into the second attribute identification model for relaxation training to obtain a third attribute identification model;

acquiring an updated feature matrix corresponding to the sample object features when the training of the third attribute identification model converges;

singular value decomposition processing is carried out on the updated feature matrix to obtain updated feature layer parameters;

repeating the constraint training and the relaxation training based on the third attribute identification model, the updated feature layer parameters and a third output result to obtain an initial attribute identification model; and the third output result is the output result of the target pooling layer corresponding to the third attribute identification model when training is converged.

5. The method according to any one of claims 1-4, wherein updating the loss function of the initial attribute identification model based on the objective function results in an updated model, and training the sample object features according to the updated model results in the object attribute identification model, comprising:

updating the loss function of the initial attribute identification model based on the objective function to obtain an updated model;

training the sample object feature set according to the updated model to obtain an attribute identification model to be verified; the sample object feature set comprises at least two sample object features;

acquiring a feature set of a test object; the test object features in the test object feature set are marked with test attribute tags;

determining target sample object features from the sample object feature set based on the attribute identification model to be verified;

acquiring a target sample attribute tag corresponding to the target sample object characteristic;

and if the target sample attribute tag is matched with the test attribute tag, determining the attribute identification model to be verified as the object attribute identification model.

6. The method of claim 5, wherein the determining target sample object features from the sample object feature set based on the attribute identification model to be verified comprises:

Determining the comprehensive distance between each sample object feature and the target test object feature based on the attribute identification model to be verified; the target test object feature is any feature in the test object feature set;

sorting the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature;

and determining the characteristics of the target sample object according to the sequencing result.

7. The method of claim 6, wherein determining the integrated distance of each sample object feature from a target test object feature based on the attribute identification model to be verified comprises:

8. An object property recognition apparatus, the apparatus comprising:

wherein, the model training module includes:

9. An object property recognition device, characterized in that the device comprises a processor and a memory, in which at least one instruction or at least one program is stored, which at least one instruction or at least one program is loaded and executed by the processor to implement the object property recognition method according to any one of claims 1-7.

10. A computer storage medium having stored therein at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by a processor to implement the object property identification method of any one of claims 1-7.

11. A computer program product comprising computer instructions which, when executed by a processor, implement the object property recognition method of any one of claims 1-7.