WO2023051085A1 - 对象识别方法、装置、设备、存储介质和程序产品 - Google Patents

对象识别方法、装置、设备、存储介质和程序产品 Download PDF

Info

Publication number
WO2023051085A1
WO2023051085A1 PCT/CN2022/113686 CN2022113686W WO2023051085A1 WO 2023051085 A1 WO2023051085 A1 WO 2023051085A1 CN 2022113686 W CN2022113686 W CN 2022113686W WO 2023051085 A1 WO2023051085 A1 WO 2023051085A1
Authority
WO
WIPO (PCT)
Prior art keywords
candidate
feature
extraction
training
information
Prior art date
Application number
PCT/CN2022/113686
Other languages
English (en)
French (fr)
Inventor
樊鹏
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2023051085A1 publication Critical patent/WO2023051085A1/zh
Priority to US18/335,569 priority Critical patent/US20230326185A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features

Definitions

  • the present application relates to the field of computer technology, in particular to an object recognition method, device, computer equipment, storage medium and program product.
  • the prediction result can be applied to a variety of scenarios, for example, it can generate relevant push information for the object based on the prediction result, and realize the effective transmission of information.
  • Different categories of objects are allocated corresponding shares of network resources based on the prediction results, and so on.
  • a method of object recognition comprising:
  • the set of candidate objects includes a plurality of candidate objects
  • Obtain multiple candidate object information of the candidate object in multiple dimensions perform feature extraction on each candidate object information, obtain candidate object features corresponding to each candidate object information, and fuse each candidate object feature to obtain the object extraction feature corresponding to the candidate object, based on object Extracting features for object category probability recognition, and obtaining the recognition probability that the candidate object belongs to the target object category;
  • the trained target object recognition model For each sub-object set formed by each sub-extraction feature set, based on the recognition probability corresponding to each candidate object in the sub-object set, select a representative object from the sub-object set; the candidate object information of the selected representative object is used For training the target object recognition model, the trained target object recognition model is used to identify whether the object belongs to the target object category.
  • An object recognition device comprising:
  • a candidate object set acquiring module configured to acquire a candidate object set; the candidate object set includes a plurality of candidate objects;
  • the probabilistic identification module is used to acquire multiple candidate object information of the candidate object in multiple dimensions, perform feature extraction on each candidate object information, obtain candidate object features corresponding to each candidate object information, and fuse each candidate object feature to obtain the candidate object corresponding Object extraction features, object category probability identification based on object extraction features, to obtain the identification probability that the candidate object belongs to the target object category;
  • the clustering module is used to cluster the object extraction features corresponding to the candidate objects, obtain the sub-extraction feature sets corresponding to each cluster category, and form the candidate objects corresponding to the object extraction features in the same sub-extraction feature set into a sub-object set ;
  • the representative object selection module is used to select representative objects from the sub-object sets based on the recognition probabilities corresponding to each candidate object in the sub-object sets for each sub-object set formed by each sub-extraction feature set;
  • the candidate object information representing the object is used to train the target object recognition model, and the trained target object recognition model is used to recognize whether the object belongs to the target object category.
  • a computer device including a processor and a memory; the memory is used to store computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor realizes the object recognition in the embodiment of the present application method.
  • a non-volatile computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the processor implements the object recognition method in the embodiment of the present application .
  • a computer program product comprising computer readable instructions stored in a computer readable storage medium; a processor of a computer device reads the computer readable instructions from the computer readable storage medium, and the processor executes the computer readable instructions
  • the readable instructions enable the computer device to implement the object recognition method in the embodiment of the present application.
  • Fig. 1 (a) is the application environment diagram of object recognition method in an embodiment
  • Fig. 1 (b) is a schematic diagram of the processing procedure of the object recognition method in one embodiment
  • Fig. 1 (c) is a schematic diagram of the processing procedure of the object recognition method in another embodiment
  • FIG. 2 is a schematic flow chart of an object recognition method in yet another embodiment
  • Fig. 3 is a schematic flow chart of an object recognition method in another embodiment
  • FIG. 4 is a processing architecture diagram of an object recognition method in an embodiment
  • FIG. 5 is a schematic diagram of an offline processing flow of an object recognition method in an embodiment
  • Fig. 6 is a schematic diagram of an online processing flow of an object recognition method in an embodiment
  • Fig. 7 is a model effect comparison diagram of using different models to identify user property status in one embodiment
  • Fig. 8 is a business effect comparison diagram of different models for object real estate state identification provided by the embodiment of the present application.
  • FIG. 9 is a structural block diagram of an object recognition device provided by an embodiment of the present application.
  • FIG. 10 is an internal structural diagram of a computer device provided by an embodiment of the present application.
  • an embodiment means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least some of the embodiments of the present application.
  • the occurrences of this phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is understood explicitly and implicitly by those skilled in the art that the embodiments described in this application can be combined with other embodiments.
  • One of the application areas is to group and classify objects based on deep learning models; Judging whether to push content related to buying a house to the subject; another example, based on the prediction result of the subject’s car purchase status based on the deep learning model, intelligently judge whether to push content related to car purchase to the subject; another example, based on the deep learning model to analyze the subject’s renting status Based on the prediction results, intelligently judge whether to push content related to renting to the object.
  • the real estate status of the object refers to whether the object currently owns real estate, that is, whether the object has bought a house; in the scenario where the real estate status of the object is predicted by the deep learning model, the greater the predicted probability output by the deep learning model, the greater the probability that the object owns the real estate , at this time, it is not necessary to push the content related to buying a house to the object.
  • the object’s car purchase status refers to whether the object currently owns a vehicle, that is, whether the object has already purchased a car; in the scenario where the object’s car purchase status is predicted by the deep learning model, the greater the prediction probability output by the deep learning model, the greater the probability that the object owns a vehicle. At this time, you can not push the content related to buying a car to the object. The lower the predicted probability output by the deep learning model is, the lower the probability that the object owns a vehicle. At this time, you can push the content related to buying a car to the object.
  • the renting status of the object refers to whether the object is currently renting a house; in the scene where the renting state of the object is predicted by the deep learning model, the greater the predicted probability output by the deep learning model, the greater the probability that the object has already rented a house.
  • Push the content related to renting The lower the prediction probability output by the deep learning model is, the lower the probability that the object has rented a house. At this time, you can push the content related to renting to the object.
  • the solution provided by the embodiment of this application involves the deep learning technology of artificial intelligence, which can be applied to scenarios such as cloud technology, cloud security, artificial intelligence and smart transportation;
  • the object information involved in this application including but not limited to object device information, object Behavioral information, etc.
  • data including but not limited to data for display, data for analysis, etc.
  • this application also provides the corresponding object Authorization entry, for the object to choose to authorize or choose to refuse.
  • the object recognition method provided by this application can be executed by a local computing device/local computing system, or by a distributed computing system; a distributed computing system is, for example, a server cluster composed of multiple physical servers, or a terminal device and a server. system.
  • Fig. 1(a) is an application environment diagram of the object recognition method in one embodiment.
  • the terminal 102 communicates with the server 104 through a communication network.
  • the terminal 102 can interact with the server 104 through the communication network;
  • the terminal 102 can be, but not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices, and the Internet of Things devices can be smart Speakers, smart TVs, smart air conditioners, smart car equipment, etc.
  • Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, and the like.
  • the server 104 may be implemented by an independent server or a server cluster or cloud server composed of multiple servers.
  • the server 104 may store the data involved in the object recognition method through a data storage system, and the data storage system may be integrated in the server 104 or set separately from the server 104 .
  • the object recognition method provided in this application may be executed cooperatively by the terminal 102 and the server 104 , may be executed solely by the terminal 102 , or may be executed solely by the server 104 .
  • the server 104 can obtain a set of candidate objects, the set of candidate objects includes multiple candidate objects, the server 104 obtains multiple candidate object information of the candidate objects in multiple dimensions, and performs feature extraction on each candidate object information , obtain the candidate object features corresponding to each candidate object information, fuse each candidate object feature to obtain the object extraction feature corresponding to the candidate object, perform object category probability identification based on the object extraction feature, and obtain the identification probability that the candidate object belongs to the target object category, and then, the server 104 cluster the object extraction features corresponding to the candidate objects to obtain the sub-extraction feature sets corresponding to each cluster category, and form the candidate objects corresponding to the object extraction features in the same sub-extraction feature set into a sub-object set, and then the server 104 for Each sub-object set composed of each sub-ex
  • the so-called target object category can be a category in different application scenarios.
  • it can be predicted whether the object belongs to the object category of the property ownership status, and it can also be predicted whether the object belongs to the property owner category.
  • the object category of the vehicle can also predict whether the object belongs to the object category of the renting state, and realize the accurate prediction of whether the object owns real estate, whether it owns a vehicle, and whether it rents a house, so as to accurately determine the potential needs of real estate information, car purchase information, and rental information.
  • Objects, pushing relevant information for such objects can not only improve the effectiveness of information transmission, but also avoid the occupation and waste of computer resources and network resources caused by generating a large amount of invalid relevant information and pushing them to unmatched objects, saving computer resources and Internet resources.
  • the object recognition method provided by this application mainly includes: performing feature extraction on multiple objects, for example, performing feature extraction on objects A, B, C, D, and E, and the extracted features are called Object extraction features, and then obtain the object extraction features of each object, for example, feature a of object A, feature b of object B, feature c of object C, feature d of object D, and feature e of object E; then, based on each object
  • the object extraction features of the object the probability that the identified object belongs to the target object category, the identified probability is called the recognition probability, and then the recognition probability P of each object belonging to the target object category is obtained.
  • the recognition probability of object A belonging to the target object category is P A
  • the recognition probability that object B belongs to the target object class is P B
  • the recognition probability that object C belongs to the target object class is P C
  • the recognition probability that object D belongs to the target object class is P D
  • the recognition probability that object E belongs to the target object class for P E is P E
  • the objects are clustered based on the object extraction features of each object to obtain multiple object sets; according to the recognition probability that the object belongs to the target object category, the corresponding representative objects are selected from each object set, and the representative objects are added to the training sample set , use the training sample set to build a target object recognition model, and use the target object recognition model to predict whether the object belongs to the target object category.
  • the object extraction features are obtained by performing feature extraction on the candidate object information. Since the candidate objects are clustered according to the object extraction features of the candidate objects, the candidate objects belonging to the same category can be divided into the same sub-object set. . Since the recognition probability that the candidate object belongs to the target object category is obtained according to the object extraction features of the candidate object, the representative objects selected from each sub-object set according to the recognition probability of the candidate object can belong to the sub-objects of the same category Select representative representative objects from the collection, and the selected representative objects can not only cover all categories of candidate objects, but also have the representative characteristics of each category, so that the target object recognition model trained by using the candidate object information of representative objects , can dig out the differences of different categories and the representative characteristics of each category, and improve the accuracy of predicting whether the object belongs to the target object category. In the information push scenario, it can improve the effectiveness of push information transmission, avoid pushing a large amount of relevant information to unmatched objects, and save computer resources.
  • the present application also provides an object recognition method, as shown in Figure 1(c), which mainly includes: obtaining a small number of seed objects based on manual labeling and business logic, and constructing object screening based on a small number of seed objects
  • the model through the object screening model, performs multiple rounds of iterative processing on other objects, and obtains the object extraction features of a large number of objects and the recognition probability belonging to the target object category; then, based on the object extraction features of the objects, a large number of objects are clustered to obtain Multiple object sets, and according to the recognition probability that the object belongs to the target object category, select representative objects from each object set, and add the representative objects to the training sample set; use the training sample set for model training to obtain the target object recognition model, use The target object recognition model predicts whether an object belongs to the target object category.
  • FIG. 2 it is a schematic flowchart of an object recognition method in an embodiment.
  • Fig. 1 (b) Fig. 1 (c) and Fig. 2 introduce the object recognition method that the present application provides, this method can be applied in the server 104 shown in Fig. 1 (a) such as computer equipment, mainly comprises the following steps:
  • Step S202 acquiring a set of candidate objects.
  • the object is an object with a category.
  • the category to which the object belongs can be "in the state of owning the real estate" or “not in the state of owning the real estate”; in the scenario of predicting whether the object owns a vehicle In , the category of the object can be "vehicle owned” or "not owned vehicle”.
  • Candidate objects are objects waiting to be selected for model training, and representative objects can be selected from the candidate object set for model training. For example, objects A, B, C, D, and E shown in FIG. 1 ; wherein, multiple candidate objects form a candidate object set, that is, the candidate object set includes multiple candidate objects, and multiple refers to at least two.
  • the computer device may acquire the candidate object set when responding to the sample selection instruction.
  • the set of candidate objects may be carried in the sample selection instruction, or may be pre-stored.
  • Step S204 obtaining multiple candidate object information of the candidate object in multiple dimensions, performing feature extraction on each candidate object information, obtaining candidate object features corresponding to each candidate object information, and merging each candidate object feature to obtain the object extraction feature corresponding to the candidate object , perform object class probability recognition based on object extraction features, and obtain the recognition probability that the candidate object belongs to the target object class.
  • the object information is information related to the object, such as at least one of object gender information, object device information, or information about a network to which the object device is connected.
  • the object device may include a smart watch, a mobile phone, or a laptop computer used by the object, and the device information may include at least one of a device resolution or a number of cores of a central processing unit (CPU) of the device.
  • the network connected to the target device can include WiFi (Wireless Fidelity) and mobile network; if the network connected to the target device is WiFi, then the information on the network connected to the target device can include connecting to WiFi every day At least one of the earliest time of , or the number of different WiFis connected to.
  • the object is a candidate object, and the information of the object may be referred to as candidate object information.
  • Candidate object features are features extracted from candidate object information; since candidate object information can be divided into multiple dimensions, such as the gender information of the candidate object, the device information of the candidate object, and the network information connected to the device of the candidate object, therefore , different candidate object features can be obtained by performing feature extraction on different candidate object information, for example, features obtained by performing feature extraction on candidate object gender information, and for example, features obtained by performing feature extraction on candidate object device information.
  • the object extraction feature is a feature obtained by fusing the candidate object features of multiple dimensions of the same candidate object. For example, for the object A belonging to the candidate object, the gender information, device information and network connected to the device of object A The information is extracted separately to obtain the candidate object features of multiple dimensions. The candidate object features of these dimensions are respectively recorded as a x , a y and a z . Then, a x , a y and a z can be fused to obtain a r , get the object extraction features of object A based on a r .
  • the way of fusing the features of candidate objects can be to sum a x , a y and a z to get the average value, or to sum a x , a y and a z , or to a x , a y and a z are weighted and summed to take the average value, and these features can also be further input into the fully connected layer for processing to obtain a fused feature.
  • the fused candidate object features can be candidate object features of all dimensions, and correspondingly, the obtained object extraction feature includes one feature value; for example, if the candidate object has p-dimensional candidate object features (a 1 , a 2 , a 3 ,..., a p ), then the p-dimensional candidate object features can be fused, and the obtained a r can be used as the feature value of the object extraction feature, and the number of feature values included in the object extraction feature for one.
  • the fused candidate object features can also be part-dimensional candidate object features, and correspondingly, the obtained object extraction features include more than one feature value; for example, if the candidate object has a q-dimensional candidate object feature (a 1 , a 2 , a 3 ,..., a q ), then only the candidate object features a 1 , a 2 and a 3 can be fused, and the obtained a r can be spliced with other unfused candidate object features to obtain
  • the feature value of the object extraction feature, the number of feature values included in the object extraction feature is at least one.
  • the computer device can obtain multiple candidate object information of the candidate object, and perform feature extraction on each candidate object information, obtain candidate object features corresponding to each candidate object information, and fuse each candidate object feature to obtain candidate object object extraction Features, as shown in Figure 1, the object extraction feature a of object A, the object extraction feature b of object B, the object extraction feature c of object C, the object extraction feature d of object D, the object extraction feature e of object E, and each The object extraction features include at least one feature value, respectively expressed as [a 1 , a 2 , a 3 , ..., a n ], [b 1 , b 2 , b 3 , ..., b n ], [c 1 , c 2 , c 3 ,..., c n ], [d 1 , d 2 , d 3 ,..., d n ], [e 1 , e 2 , e 3 ,..., e n ].
  • the category to which the object belongs may be referred to as an object category;
  • the target object category is the object category to be recognized by the object recognition model.
  • the object categories to be identified by the object recognition model are "in a state of owning a property” and “not in a state of owning a property”.
  • Real estate status is the target object category; for another example, in the scenario of predicting whether an object owns a vehicle, the object categories to be recognized by the object recognition model are "in a state of owning a vehicle” and “not in a state of owning a property”.
  • "in Owning a vehicle” or “not owning a property” are the target object categories.
  • the recognition probability that an object belongs to the target object class is how likely the object is to belong to the target object class.
  • the greater the recognition probability that the object belongs to the target object category the greater the possibility that the object belongs to the target object category; the smaller the recognition probability that the object belongs to the target object category, the smaller the possibility that the object belongs to the target object category.
  • the computer device obtains feature a of object A, feature b of object B, feature c of object C, feature d of object D, and feature e of object E. Since each feature describes the corresponding object, the computer device By analyzing feature a, feature b of object B, feature c of object C, feature d of object D, and feature e of object E, it can be determined that the recognition probability that object A belongs to the target object category is P A , and object B belongs to the target The recognition probability of the object class is P B , the recognition probability of object C belonging to the target object class is PC , the recognition probability of object D belonging to the target object class is PD , and the recognition probability of object E is PE of the target object class. For example, the computer device can input the object extraction features of each candidate object into the classification layer of the object screening model, and the classification layer can output the recognition probability that the candidate object belongs to the target object category.
  • Step S206 clustering the object extraction features corresponding to the candidate objects to obtain sub-extraction feature sets corresponding to each cluster category, and combining candidate objects corresponding to object extraction features in the same sub-extraction feature set to form a sub-object set.
  • Clustering the object extraction features corresponding to each candidate object can divide similar candidate objects together and separate dissimilar candidate objects; since the candidate objects are described by object extraction features, this clustering is the Clustering of object extraction features of objects, wherein the set of object extraction features corresponding to multiple candidate objects divided together can be regarded as a sub-extraction feature set, and the object extraction features in the same sub-extraction feature set correspond to Candidate objects of the candidate objects form a sub-object set.
  • object extraction features of objects A, E, and C are divided into the same sub-extraction feature set, and objects A, E, and C form a sub-object set. Multiple sub-object sets can be obtained after clustering, and the number of sub-object sets is consistent with the number of clustering categories.
  • the similarity between object extraction features corresponding to candidate objects belonging to the same sub-object set is greater than the similarity between object extraction features corresponding to candidate objects belonging to different sub-object sets; for example, objects A, E and C belong to The same sub-object set, objects B and D belong to the same sub-object set, then the similarity between the object extraction features of object A and object E is greater than the similarity between the object extraction features of object A and object B, and object The similarity between the respective object extraction features of E and object C is greater than the similarity between the respective object extraction features of object E and object D.
  • the above clustering can be realized by K-means algorithm (k-means algorithm) or density peak clustering algorithm (clustering by fast search and find of density peaks, DPC).
  • Step S208 for each sub-object set composed of each sub-extraction feature set, based on the recognition probabilities corresponding to each candidate object in the sub-object set, select a representative object from the sub-object set.
  • the representative object corresponding to the sub-object set is a representative candidate object in each candidate object of the sub-object set, and the representative object can represent the candidate objects in the sub-object set;
  • the number of representative objects may be one or more, and the number of selected representative objects is determined according to the set first preset condition.
  • the selected candidate object information representing the object is used to train the target object recognition model, and the trained target object recognition model is used to identify whether the object belongs to the target object category.
  • the computer device After the computer device obtains the sub-object set including objects A, C, and E, it uses object A as the representative object according to the recognition probability corresponding to each object; for another example, after obtaining the sub-object set including objects B and D, According to the recognition probability corresponding to each object, the object D is taken as the representative object. For another example, after obtaining the sub-object set including objects A, C, and E, objects A and E are used as representative objects according to the recognition probabilities corresponding to each object.
  • step S208 may specifically include the following steps: for each sub-object set composed of each sub-extracted feature set, the computer device respectively acquires candidate objects whose recognition probabilities satisfy the first preset condition from the sub-object sets, As the corresponding representative object in the sub-object collection.
  • the first preset condition includes at least one of the recognition probability being greater than the probability threshold, and the recognition probability being ranked before the first sorting threshold.
  • the first preset condition the recognition probability is greater than the probability threshold:
  • the computer device may use a candidate object whose recognition probability of the candidate object in the sub-object set is greater than the probability threshold as a representative object.
  • the representative object of each sub-object set is determined by means of a probability threshold to ensure the accuracy of the selection of the representative object, and obtain a training sample set for building the target object recognition model, thereby ensuring that the prediction of the target object recognition model is accurate sex.
  • the corresponding candidate objects can be selected from each sub-object set through the same probability threshold; however, in some scenarios, the recognition probability thresholds of all candidate objects included in some sub-object sets are less than the probability Threshold, at this time, the representative object of the sub-object set cannot be selected with this probability threshold, resulting in the omission of the sub-object set.
  • the computer device can set multiple levels of probability thresholds, such as high-level probability thresholds, medium-level probability thresholds, and low-level probability thresholds.
  • the probability threshold of the current level cannot be used to select the representative object in the sub-object set
  • the probability threshold lower than the current level can be used to select the representative object in the sub-object set again.
  • the representative object of the sub-object set including objects A, C and E cannot be selected by using the high-level probability threshold, the representative object of the sub-object set is selected again by using the medium-level probability threshold;
  • the probability threshold of the low-level probability threshold still cannot select the representative object of the sub-object set, the representative object of the sub-object set is selected by using the low-level probability threshold.
  • the sub-object can be The candidate object with the highest recognition probability in the set is used as the representative object of the sub-object set, and the recognition probability in the sub-object set can also be ranked in the top according to the order of recognition probability from large to small (such as the first 3, the top 5 ) candidates as representative objects.
  • the probability sorting is a sorting of recognition probabilities from large to small; the first sorting threshold may be 3, 5 or other numerical values.
  • the computer device sorts the recognition probabilities of each candidate object from large to small, and it can be obtained that the recognition probability of object A>the recognition probability of object E>the recognition probability of object C recognition probability. If the first sorting threshold is 2, object A and object E may be used as representative objects of the sub-object set.
  • the representative objects of each sub-object set are determined by probabilistic sorting to ensure the accuracy of the representative objects, and a training sample set for building the target object recognition model is obtained, thereby ensuring the prediction accuracy of the target object recognition model.
  • the representative objects may be added to the training sample set; the training sample set is used for model training, so as to train and obtain a target object recognition model for identifying the target object category.
  • the samples in the training sample set are samples used for model training.
  • the object information corresponding to the object in the training sample can be obtained as the object feature
  • the object category corresponding to the object in the training sample can be obtained as the object label.
  • Based on Object features and object labels are supervised for training to obtain a target object recognition model.
  • the target object recognition model can include support vector machines (support vector machines, SVM), convolutional neural network (Convolutional Neural Network, CNN), long short-term memory network (Long Short Term Memroy, LSTM), or Real-time Attention based Look-alike Model (RALM, similarity model based on real-time attention) and other models.
  • support vector machines support vector machines, SVM
  • convolutional neural network Convolutional Neural Network, CNN
  • long short-term memory network Long Short Term Memroy, LSTM
  • LSTM Long Short Term Memroy
  • RLM Real-time Attention based Look-alike Model
  • the Real-time Attention based Look-alike Model (RALM) model is a similarity-based look-alike model, which includes two parts: “object representation learning” and “look-alike model learning”.
  • the RALM model adopts a double-tower structure.
  • the input on the left is the Embedding (Embedding, embedded representation) of the seed object, and the input on the right is the Embedding of the target object.
  • the Embedding on both sides passes through a layer of FC (Fullconnection, fully connected layer ) to complete the mapping to the low-dimensional space. Since the Embedding of the target object on the right is obtained through object representation learning, in order to prevent overfitting, the first layer FC of the twin towers is shared.
  • the tower on the left can get the Embedding corresponding to each cluster, and then input the Embedding of each cluster and the Embedding of the target object to the Global Attention Unit (global attention unit) and the Local Attention Unit (local attention unit) respectively. ) to get Global Embedding (global embedding representation) and Local Embedding (local embedding representation).
  • RALM has two effects:
  • a global attention unit is utilized to learn the global representation of the seed object, which weights the representation of a single object and penalizes noisy objects, which is more robust than all object weights equally.
  • a local attention unit is utilized to learn a local representation of the seed object, which weights the relevance of the seed object to the target object. The local attention unit learns the representation of the seed object dynamically based on the target object. For different target objects, the learned representation of the seed object is different, which greatly improves the expressiveness of the seed object representation.
  • the RALM model involves a clustering process, which requires iteration and is time-consuming, and the number of cluster centers directly affects the clustering effect. In the scene of object real estate status prediction, it is better to choose 50-80 cluster centers online.
  • the evaluation index of the classification effect of the model can include AUC (Area under Curve, the area under the curve). The larger the AUC value, the more likely the current model is to rank positive samples in front of negative samples and obtain better classification results.
  • Parameter tuning refers to the grid optimization of the hyperparameters of the selected model, in order to expect the evaluation index AUC to be improved.
  • the object extraction features are obtained by extracting the features of the candidate object information. Since the candidate objects are clustered according to the object extraction features of the candidate objects, the candidate objects belonging to the same category can be divided into the same sub-object set middle. Since the recognition probability that the candidate object belongs to the target object category is obtained according to the object extraction features of the candidate object, the representative objects selected from each sub-object set according to the recognition probability of the candidate object can belong to the sub-objects of the same category Select representative representative objects from the collection. The selected representative objects can not only cover all categories of candidate objects, but also have the representative characteristics of each category.
  • the representative object can be used as a training sample, and more samples for model training can be mined from a large number of samples, so that the target object recognition obtained by using the candidate object information training of the representative object
  • the model can dig out the differences of different categories and the representative characteristics of each category, and then improve the accuracy of predicting whether the object belongs to the target object category.
  • the object extraction feature is obtained by feature extraction through an object screening model
  • the step of obtaining the object screening model includes: obtaining the seed object corresponding to the target object category; obtaining the seed object information corresponding to the seed object, and converting the seed object information to As the training feature in the training sample, the target object category is used as the label in the training sample to form the training sample; the model training is performed based on the training sample to obtain the object screening model.
  • these objects can be labeled manually or automatically.
  • the labeling results are more credible.
  • Annotated objects are regarded as objects with higher confidence, and the objects with higher confidence may be called seed objects.
  • the seed object can be an object with a high confidence level above, and the category is marked as belonging to the target object category; where the negative sample is not objects belonging to the target object category, and positive samples are objects belonging to the target object category.
  • the object information of the seed object may be referred to as seed object information, and the specific introduction of the object information may refer to the introduction of the object information corresponding to step S204.
  • the object screening model is used to predict the probability that the unlabeled object belongs to the target object category.
  • This process can be regarded as the process of labeling the unlabeled object;
  • the object screening model can be DeepFM (Deep Factorization Machine ) model or FM (Factorization Machine) model.
  • the object screening model is constructed by using the seed object with high confidence.
  • the seed object information can be used as the characteristics of the training sample, and the seed object belongs to the target object category as the label of the training sample.
  • a training sample is formed, and the training sample is used for model training to obtain an object screening model.
  • the seed object is used to construct the object screening model. Since the seed object has a high degree of confidence that the target object belongs to the category, when the object screening model predicts whether other objects belong to the target object category, the prediction accuracy can be improved.
  • the above-mentioned step of obtaining multiple candidate object information of a candidate object, and extracting the candidate object features corresponding to each candidate object information may specifically include: the computer device combines the multiple candidate object information of the candidate object in multiple dimensions It is input to the feature extraction layer of the object screening model for feature extraction, and the candidate object features corresponding to each candidate object information are obtained.
  • the above step of fusing the features of each candidate object to obtain the object extraction feature corresponding to the candidate object, performing object category probability identification based on the object extraction feature, and obtaining the recognition probability that the candidate object belongs to the target object category may specifically include: inputting each candidate object feature into the computer device Go to the classification layer of the object screening model, so that the classification layer fuses the features of each candidate object to obtain the object extraction feature; obtain the recognition probability that the output candidate object belongs to the target object category after the classification layer performs object category probability identification on the object extraction feature.
  • the object screening model may include a feature extraction layer and a classification layer.
  • the feature extraction layer is used to extract deep-level features used to describe the object
  • the feature extraction layer can be a convolutional layer, and the features extracted by the feature extraction layer can be called object features
  • the object information input to the feature extraction layer is the candidate object
  • the features extracted by the feature extraction layer can be called candidate object features, or deep features.
  • the classification layer is mainly to fuse the deep-level features extracted by the feature extraction layer to obtain the object extraction features, and determine the category of the object according to the object extraction features, and determine the recognition probability that the object belongs to the target object category.
  • the classification layer can be the whole connection layer.
  • the computer device can input the gender information of the object A, the device information and the network information connected to the device into the feature extraction layer of the object screening model to perform feature extraction, and then obtain the candidate object features corresponding to each candidate object information; then , the computer equipment inputs the features of each candidate object into the classification layer of the object screening model, so that the classification layer fuses the features of each candidate object to obtain the object extraction features of object A, and makes the classification layer perform object classification based on the object extraction features of object A Probabilistic recognition, to obtain the recognition probability that object A belongs to the target object category.
  • the object screening model built based on the seed object uses the object screening model built based on the seed object to predict whether these candidate objects belong to the target object category, and obtain more samples required for training the target object recognition model, and improve Predictive Accuracy of Object Recognition Models.
  • the above-mentioned object extraction features corresponding to the candidate objects are clustered to obtain a sub-extraction feature set corresponding to each cluster category, and the candidate objects corresponding to the object extraction features in the same sub-extraction feature set form a sub-object
  • the step of assembling specifically includes: obtaining the feature distance between the object extraction features corresponding to different candidate objects; for each object extraction feature, based on each feature distance, determining the number of object extraction features located in the area where each object extraction feature is located , based on the number, get the regional object density of each object extraction feature; select the cluster center based on the regional object density of each object extraction feature, and cluster the object extraction features based on the cluster center to obtain the sub-extraction features corresponding to each cluster category set; the candidate objects corresponding to the object extraction features in the same sub-extraction feature set form a sub-object set.
  • the feature distance is the distance between the object extraction features of different candidate objects, for example, the object extraction features of candidate objects A and B [a 1 , a 2 , a 3 ,..., a n ] and [b 1 , b 2 , b 3 ,..., b n ]
  • the characteristic distance between can be
  • the area where the object extraction feature is located is the area around the object extraction feature, and the area around the object extraction feature can be an area within a preset radius centered on the object extraction feature.
  • This embodiment calculates the area object density of the object extraction feature of each candidate object, and uses the calculation of the area object density of the object extraction feature of object A as an example to introduce:
  • the computer device After the computer device obtains the feature distance between the object extraction feature of object A and the object extraction features of other objects, when it is determined that the feature distance between the object extraction feature of object C and the object extraction feature of object A is less than the preset radius , it is considered that the object extraction feature of object C is located in the area of the preset radius centered on the object extraction feature of object A; in the above way, it can be determined whether the object extraction feature of other objects is located in the area centered on the object extraction feature of object A In the area of the preset radius; Then, when it is determined that the object extraction feature of object C and the object extraction feature of object D are located in the area of the preset radius centered on the object extraction feature of object A, it can be determined that the object is located in the object
  • the number of object extraction features in an area with a preset radius centered on the object extraction feature of A is 2, and this number is used as the area object density of the object extraction feature of object A.
  • the computer device can determine the area object density of the object extraction features of objects B, C, D, and E; before clustering, when the area object density of the object extraction features of both objects A and E Larger, the object extraction features of object A and object E can be used as cluster centers, and the object extraction features of objects B, C, and D can be clustered based on these two cluster centers; when the obtained In the clustering results, the object extraction features of object A, the object extraction features of object C, and the object extraction features of object D are divided together to form a corresponding sub-extraction feature set, the object extraction features of object B and the object extraction features of object E features are grouped together and form a corresponding sub-extraction feature set, then object A, object C and object D can be grouped together and form a corresponding sub-object set, and object B and object E can be divided together and form a corresponding sub-object set collection of objects.
  • the regional object density of each object extraction feature is determined, and the clustering center is selected according to the regional object density, which can improve the accuracy of clustering the object extraction features.
  • the corresponding candidate objects with relatively similar object extraction features can be classified into the same class, and the corresponding candidate objects with less similar object extraction features can be classified into different classes, so as to improve the accuracy of classifying the corresponding candidate objects.
  • the subsequent representative objects selected from the sub-object sets of each category can more comprehensively cover the candidate objects with large differences in object extraction characteristics, and reduce the situation that the selected representative objects have small differences and narrow coverage, thereby avoiding representative objects.
  • the problem of poor training effect of the target object recognition model caused by the one-sidedness of object selection improves the accuracy of the target object recognition model to identify object categories obtained through training.
  • the above-mentioned step of selecting a cluster center based on the regional object density of each object extraction feature, clustering the object extraction features based on the cluster center, and obtaining the sub-extraction feature set corresponding to each cluster category may include: In the feature of the area object density greater than the object extraction feature, determine the adjacent extraction feature of the object extraction feature; use the feature distance between the object extraction feature and the adjacent extraction feature as the target distance corresponding to the object extraction feature; based on the object The area object density of the extracted feature and the target distance corresponding to the object extracted feature are selected to select the cluster center.
  • the feature of the area object density greater than the area object density of the object extraction feature is: for the target object extraction feature, the area object density is greater than the area object density of the target object extraction feature of other object extraction features.
  • the neighboring extraction feature is an object extraction feature having the smallest feature distance from the target object extraction feature among other object extraction features having an area object density greater than that of the target object extraction feature.
  • the cluster center can be selected in combination with the area object density and the target distance of the object extraction feature; wherein, the calculation of the target distance of the object extraction feature of object C is used as an example to introduce:
  • the computer device can regard the object extraction feature of object A as the object extraction feature adjacent to the object C Object extraction features, and the feature distance
  • the computer device After the computer device obtains the target distance of the object extraction feature of each object in the above manner, the object extraction feature with larger area object density and target distance can be used as the cluster center.
  • the corresponding target distance is determined based on the object extraction features adjacent to the object extraction features of the candidate object, and the cluster center is selected in combination with the target distance and the area object density, which can improve the accuracy of object extraction feature clustering.
  • the representative objects selected in the sub-object collection of each category can cover more comprehensively the candidate objects with large differences in object extraction characteristics, thereby avoiding the problem of poor training effect of the target object recognition model caused by the one-sidedness of representative object selection, and improving training.
  • the obtained target object recognition model can identify the accuracy of the object category, thereby improving the accuracy of object recognition.
  • the above step of clustering the object extraction features based on the clustering center to obtain the sub-extraction feature sets corresponding to each cluster category may include: the computer device obtains the current object extraction features of the cluster category to be determined; Obtain the clustering center whose regional object density is greater than the regional object density of the current object extraction feature, as the candidate clustering center corresponding to the current object extraction feature; based on the distance between the current object extraction feature and the candidate clustering center, select the current cluster center from the candidate center set The adjacent clustering center corresponding to the object extraction feature, and the current object extraction feature is added to the sub-extraction feature set corresponding to the adjacent clustering center.
  • the computer device can form a set of candidate centers according to the candidate cluster centers corresponding to the current object extraction features, for example, the current object extraction features are the object extraction features of object F, and the cluster centers include the object extraction features of object G, object I
  • the area object density of the object extraction feature of object F is greater than the area object density of object extraction feature of object F
  • the area object density of object extraction feature of object G and the area object density of object extraction feature of object J are all smaller than the area object density of object extraction feature of object F
  • the object extraction features of object I and object K can be used as the candidate cluster centers corresponding to the object extraction features of object F, and form a set of candidate centers.
  • the adjacent clustering center is: in the set of candidate centers corresponding to the extracted features of the current object, the clustering center with the smallest feature distance to the extracted features of the current object.
  • the object extraction features of object I and the object extraction features of object K when the feature distance between the object extraction features of object F and the object extraction features of object I is smaller than the object extraction features of object F and
  • the computer device can use the object extraction features of object I as the adjacent clustering center of the object extraction features of object F, and divide the object extraction features of object F into the corresponding In the sub-extracted feature set.
  • the candidate cluster center of the current object extraction feature is determined based on the area object density, and the adjacent cluster center is determined based on the distance between the current object extraction feature and the candidate cluster center. Clustering center, and then divide the current object extraction feature into the corresponding sub-extraction feature set to improve the accuracy of clustering division.
  • the object recognition method provided by the present application further includes: based on the recognition probability that the candidate object belongs to the target object category, determining the influence weight of each candidate object on the training gradient change of the target object recognition model;
  • the influence weight of the training gradient change of the object recognition model is to select the candidate objects satisfying the second preset condition from the candidate object set, and add the candidate objects satisfying the second preset condition to the training sample set.
  • the target object recognition model can adopt the gradient descent training method, and the training samples used in each round of training will have an impact on the gradient change of the model.
  • the influence weight of the candidate object on the training gradient change of the target object recognition model reflects the influence degree of the candidate object on the gradient change in the training process of the target object recognition model, and the influence weight of the training gradient change is positively correlated with the recognition probability, that is, The greater the recognition probability that the candidate object belongs to the target object category, the greater the gradient change is when the candidate object is used to train the target object recognition model.
  • the second preset condition includes that the influence weight of the training gradient change is greater than the influence weight threshold, or at least one of the influence weight ranking is before the second sort threshold.
  • the two second preset conditions are introduced below:
  • the computer device determines the influence weight of each candidate object on the training gradient change of the object recognition model based on the recognition probability of each candidate object, and adds the candidate objects whose influence weight of the training gradient change is greater than the influence weight threshold in the training sample set.
  • the computer device determines the influence weight of each candidate object on the training gradient change of the object recognition model based on the recognition probability of each candidate object; according to the order of the influence weight of the training gradient change from large to small, the The candidate objects are sorted, and the candidate objects whose influence weight of the training gradient change is in the top ranks are added to the training sample set.
  • the influence weight of the corresponding training gradient change is determined, and the second preset condition is selected.
  • Candidate objects are used as training samples, and the dual strategies of "most representative” and “most discriminative” are used to select training samples.
  • the selected representative objects can not only cover all categories of candidates, but also have representatives of each category.
  • the unique characteristics make the target object recognition model trained by using the candidate object information of the representative object to mine the differences of different categories and the representative characteristics of each category, thereby improving the accuracy of predicting whether the object belongs to the target object category.
  • the step of training the target object recognition model includes: selecting similar first training objects and second training objects from the training sample set, forming the first training objects and the second training objects into an object group; combining the objects
  • the training object information corresponding to each training object in the group is input into the same feature extraction layer, and the training object characteristics corresponding to each training object are extracted; based on the training object characteristics, the object similarity between the training objects in the object group is obtained;
  • the model loss value is obtained based on the object similarity; the model loss value has a negative correlation with the object similarity; the object recognition model is trained based on the model loss value, and the target object recognition model is obtained.
  • the object information of the training object may be referred to as the training object information, and the introduction of the object information may refer to the content of the above step S204.
  • the model loss value is negatively correlated with the object similarity, for example, the larger the model loss value, the larger the object similarity, and the smaller the model loss value, the smaller the object similarity.
  • the target object recognition model is a RALM model
  • the RALM model adopts a double-tower structure
  • similar first training objects and second training objects can be selected in the training sample set to form an object group; then, the first training object
  • the object information of the second training object and the object information of the second training object are input into the same feature extraction layer to complete the mapping from the high-dimensional space to the low-dimensional space, and the features (such as Embedding features) extracted by the feature extraction layer are used as the training object features, based on The training object features of the first training object and the training object characteristics of the second training object, the object similarity between the first training object and the second training object is obtained; based on the object similarity, the RALM which is negatively correlated with the object similarity is obtained The loss value of the model; based on the model loss value, the object recognition model is trained, the construction of the RALM model is completed, and the target object recognition model is obtained.
  • similar training objects are input into the same feature extraction layer to obtain the corresponding training object features, and the model training is performed according to the model loss value obtained by the similarity between the training objects, so as to ensure the accuracy of the target object recognition model constructed. forecast accuracy.
  • the step of obtaining the training object information corresponding to the training object includes: the computer device obtains the object information category; obtains the training object in the time dimension set corresponding to the object information category and the information statistics angle; obtains each information statistics in the time dimension , the object information statistical value corresponding to the training object obtained based on the statistics of information statistics; the object information statistical value is the information statistical value corresponding to the object information category; in the time dimension set, the object information statistical value corresponding to each information statistical time dimension is calculated. Aggregation, using the aggregated object information as the training object information corresponding to the training object.
  • the object information used to describe the object has different categories, and the categories may be called object information categories, for example, the duration of playing games or the duration of browsing commodities.
  • the time dimension of information statistics is the time dimension of statistical object information, which represents the length of time corresponding to the statistical information; the time dimension set includes multiple time dimensions of information statistics, such as one day, one week, three months, and six months.
  • the angle of information statistics is the time unit of statistics, such as the unit of day, or the unit of week.
  • the statistical value of the object information corresponding to the training object is obtained according to the statistical angle of information statistics in each information statistical time dimension; for example, if the information statistical time dimension is one week, and the information statistical angle is one day, the object within one week is obtained After playing the game, you can determine the daily playing time of the object, and use the daily playing time of the object as the object information statistical value, and the object information statistical value corresponds to the object information category of the game playing time.
  • the statistical value of the object information under the information statistical angle can also be determined. For example, under the information statistical time dimension of the object’s game playing time for three months, it is possible to determine the number of times the object plays games every day. duration.
  • the computer device After the computer device obtains the time dimension set corresponding to the object information category of the training object and the angle of information statistics, it determines the statistical value of the object information corresponding to the training object obtained based on the statistics of the angle of information statistics in each time dimension of information statistics.
  • the object information statistics of the same object information category in different time dimensions are aggregated.
  • the aggregation method can be at least one of average value, variance, standard deviation, or summation, and then the aggregated object information is used as the corresponding training object.
  • the training object information is used for model training.
  • the training object information used for model training is obtained by aggregating the same object information category under different information statistics time dimensions, it is possible to avoid stuffing all the values of the same type of object information under different time dimensions into the
  • the collinearity caused by the model leads to poor model effect, improves the prediction effect of the model, improves the accuracy of object recognition, improves the effectiveness of information transmission, avoids pushing a large amount of information to unmatched objects, and saves computer resources.
  • the above-mentioned step of performing information aggregation on object information statistical values corresponding to each information statistical time dimension in the time dimension set, and using the aggregated object information as the training object information corresponding to the training object may include: computer equipment based on In the time dimension set, the statistical value of the object information corresponding to each information statistical time dimension is obtained to obtain the comprehensive information statistical value; the statistical value difference between the statistical value of each object information and the comprehensive information statistical value is determined; the statistical value corresponding to the time dimension set is obtained based on the statistical value difference Dispersion, the statistical value dispersion is used as the training object information corresponding to the training object.
  • the comprehensive information statistical value is obtained by statistically calculating the object information statistical value, for example, the average value of multiple object information statistical values.
  • the statistical value difference represents the difference between the statistical value of any object information and the comprehensive information statistical value, for example, it may be a difference or a ratio.
  • the statistical value dispersion represents the degree of dispersion of the statistical value of each object information, and is positively correlated with the statistical value difference.
  • the manner of obtaining the dispersion degree of the statistical value based on the difference of the statistical value may be: the computer device adds the differences of the statistical values, and uses the addition result as the dispersion degree of the statistical value.
  • the manner of obtaining the dispersion degree of the statistical value based on the difference of the statistical value may also be: summing the squares of the differences of the statistical values, and using the result of the summation of the squares as the dispersion degree of the statistical value.
  • the way to obtain the dispersion of statistical values based on statistical value differences can also be: sum the squares of the differences of statistical values, and perform square root processing on the results obtained by the square summation, and use the results obtained by the square root processing as statistical values Dispersion.
  • each information statistics time dimension has a corresponding object information statistics value; then, determine the statistical value difference between the object information statistics value and the comprehensive information statistics value of each information statistics time dimension , get the difference of 3 statistical values, sum the squares of the differences of the 3 statistical values, and process the square root of the result obtained by the sum of the squares, use the result of the square root processing as the dispersion of the statistical value, and use the statistical value Dispersion is used as the training object information corresponding to the training object.
  • the daily game playing time determined under the information statistical time dimension of the object's game playing time for three months is 3 hours
  • the duration of playing games is 3.5 hours
  • the daily duration of playing games determined under the information statistical time dimension of the duration of the object’s playing games in a week is 2.5 hours
  • the average of 3 hours, 3.5 hours and 2.5 hours can be taken as 3 hours
  • the statistical value of the comprehensive information and respectively determine the difference between the duration of 3 hours, 3.5 hours and 2.5 hours and the statistical value of the comprehensive information (that is, the statistical value difference) is 0 hour, 0.5 hour and 0.5 hour respectively; according to 0 hour, 0.5
  • the difference between hours and 0.5 hours can determine the statistical value dispersion of 3 hours, 3.5 hours and 2.5 hours, and use the statistical value dispersion as the training object information for model training.
  • the model can learn the influence of the magnitude of the object's behavior change on the object category, so according to the degree of dispersion of the statistical value of the object information, the The training object information for model training improves the prediction accuracy of the target object recognition model.
  • the embodiment of the present application also provides an object recognition method, which can be applied to scenarios such as cloud technology, cloud security, artificial intelligence, and smart transportation; this embodiment can be executed by a computer device, including the steps shown in Figure 3:
  • Step S302 acquiring the seed object corresponding to the target object category.
  • the category to which an object belongs may be referred to as an object category;
  • the target object category is the object category to be recognized by the object recognition model.
  • the object categories to be identified by the object recognition model are "owning a property" and “not owning a property”.
  • Real estate status is the target object category; for another example, in the scenario of predicting whether an object owns a vehicle, the object categories to be recognized by the object recognition model are "in a state of owning a vehicle” and “not in a state of owning a property”.
  • "in Owning a vehicle” and “not owning a property” are the target object categories.
  • step S304 the seed object information corresponding to the seed object is obtained, and the seed object information is used as the training feature in the training sample, and the target object category is used as the label in the training sample to form a training sample.
  • the object information of the seed object can be called seed object information, such as the gender information of the seed object, the device information of the candidate object, and the network information connected to the device of the candidate object.
  • the seed object information is used as a feature describing the seed object, and
  • the target object in the above step S302 is used as the label of the training sample; the training feature and the training label of the training sample are used to form the training sample.
  • Step S306 perform model training based on the training samples to obtain an object screening model.
  • the object screening model is mainly used to predict the probability that an unlabeled object belongs to the target object category, that is, to label the unlabeled object.
  • the object screening model can be a DeepFM model or an FM model; the object screening model can include a feature extraction layer and a classification layer.
  • the feature extraction layer mainly extracts deep-level features used to describe objects, and the classification layer mainly extracts the features from the feature extraction layer.
  • the deep-level features are fused to obtain object extraction features, and the category to which the object belongs is determined according to the object extraction features, and the recognition probability that the object belongs to the target object category is determined.
  • Step S308 acquiring a set of candidate objects.
  • the candidate objects included in the candidate object set are unmarked objects, that is, objects whose category is undetermined; when the computer device responds to the sample selection instruction, select a preset number of candidates from the unmarked multiple candidate objects objects, and form a set of candidate objects.
  • Step S310 input multiple candidate object information in multiple dimensions of the candidate object into the feature extraction layer of the object screening model for feature extraction, and obtain candidate object features corresponding to each candidate object information.
  • the feature of the candidate object is the feature obtained by feature extraction of the candidate object information, which can be divided into multiple dimensions; the feature extraction layer can be a convolutional layer, which is used to extract corresponding features from multiple candidate object information of the candidate object , to obtain the candidate object features of each candidate object information.
  • Step S312 input the features of each candidate object into the classification layer of the object screening model, so that the classification layer fuses the features of each candidate object to obtain object extraction features.
  • the object extraction feature is a feature obtained by fusing the candidate object features of some dimensions or all dimensions; the classification layer can be a fully connected layer, which is used to fuse the candidate object features to obtain the object extraction feature corresponding to the candidate object.
  • Step S314 obtaining the identification probability that the candidate object output by the classification layer after performing object category probability identification on the object extraction feature belongs to the target object category.
  • the recognition probability is the likelihood, eg probability, that the object belongs to the target object class.
  • the greater the recognition probability the more likely the object belongs to the target object category, and the greater the recognition probability, the less likely the object belongs to the target object category.
  • the above classification layer After the above classification layer obtains the object extraction features, it performs object category probability identification based on the object extraction features to obtain the identification probability that the candidate object belongs to the target object category.
  • Step S316 acquiring feature distances between object extraction features corresponding to different candidate objects.
  • the feature distance is the distance between the object extraction features of different candidate objects. For example, if two object extraction features [a 1 , a 2 , a 3 , ..., a n ] and [b 1 , b 2 , b 3 ,..., b n ], then the characteristic distance can be
  • the computer device can determine the characteristic distance between any two candidate objects according to the above formula.
  • the area where the object extraction feature is located is the area around the object extraction feature, and the area around the object extraction feature may be an area within a preset radius centered on the object extraction feature.
  • the computer device obtains the feature distance between the object extraction feature of object A and the object extraction features of other objects, when determining the object C
  • the feature distance between the object extraction feature and the object extraction feature of object A is less than the preset radius
  • it can be considered that the object extraction feature of object C is located in the area of the preset radius centered on the object extraction feature of object A; according to the above
  • it can be determined whether the object extraction features of other objects are located within the preset radius area centered on the object extraction features of object A; then, when it is determined that the object extraction features of object C and the object extraction features of object D are located in the area centered on object A
  • the object extraction feature of object A is within the area of the preset radius, the number of object extraction features located in the area
  • the area object densities of the object extraction features of the objects B, C, D, and E can be determined in the manner described above.
  • Step S320 among the features whose area object density is greater than the area object density of the object extraction feature, determine the adjacent extraction feature of the object extraction feature.
  • the feature of area object density greater than the area object density of the object extraction feature refers to other object extraction features whose area object density is greater than the area object density of the target object extraction feature for the target object extraction feature.
  • the adjacent extraction feature refers to the object extraction feature with the smallest feature distance between the target object extraction feature and the target object extraction feature among other object extraction features whose area object density is greater than that of the target object extraction feature.
  • the object extraction feature of object A when the area object density of the object extraction feature of object A and the area object density of the object extraction feature of object E are greater than the area object density of the object extraction feature of object C, and the object extraction feature of object A is the same as that of object C
  • between the extracted features is smaller than the feature distance
  • the object extracted features of object A can be used as the object extracted with object C
  • step S322 the feature distance between the object extraction feature and the adjacent extraction feature is used as the target distance corresponding to the object extraction feature.
  • the computer device can also use the feature distance
  • Step S324 based on the area object density of the object extraction feature and the target distance corresponding to the object extraction feature, the cluster center is selected.
  • the computer device After the computer device obtains the target distance of the object extraction feature of each object in the above manner, the object extraction feature with larger area object density and target distance can be used as the cluster center.
  • Step S326 acquiring the current object extraction features of the cluster category to be determined.
  • the computer device After the computer device selects the cluster center, it can determine the cluster category to which the object extraction features of other candidate objects in the candidate object set belong, and can select any object extraction feature from the object extraction features of other candidate objects as the current extraction feature.
  • Step S328 acquiring a cluster center whose regional object density is greater than the regional object density of the current object extraction feature, as a candidate cluster center corresponding to the current object extraction feature; the candidate cluster centers corresponding to the current object extraction feature form a candidate center set;
  • the candidate cluster centers corresponding to the current object extraction feature form the candidate center set.
  • the current object extraction feature is the object extraction feature of object F
  • the cluster centers include the object extraction feature of object G, the object extraction feature of object I, and the object K
  • the object extraction features of object extraction feature of object J and the object extraction features of object J, in the object extraction features of each object included in the cluster center, the area object density of the object extraction feature of object I and the area object density of the object extraction feature of object K are larger than object
  • the area object density of the object extraction feature of F, the area object density of the object extraction feature of object G and the area object density of the object extraction feature of object J are all smaller than the area object density of the object extraction feature of object F, then the object I can be
  • the object extraction features and the object extraction features of object K are used as candidate cluster centers corresponding to the object extraction features of object F, and form a candidate center set.
  • Step S330 based on the distance between the current object extraction feature and the candidate cluster center, select the adjacent cluster center corresponding to the current object extraction feature from the candidate center set, and add the current object extraction feature to the sub-extraction corresponding to the adjacent cluster center feature set.
  • the adjacent clustering center refers to the clustering center with the smallest feature distance to the current object extraction feature in the candidate center set corresponding to the current object extraction feature.
  • the object extraction feature of object I can be used as the adjacent clustering center of the object extraction feature of object F, and the object extraction feature of object F can be divided into object I In the corresponding sub-extraction feature set.
  • step S332 the candidate objects corresponding to the object extraction features in the sub-extraction feature set are formed into a sub-object set.
  • object extraction features of objects A, E, and C are divided into the same sub-extraction feature set, and the computer device can form objects A, E, and C into a sub-object set.
  • Step S334 from the sub-object set, obtain the candidate object whose recognition probability satisfies the first preset condition, as the corresponding representative object in the sub-object set.
  • the computer device After the computer device obtains the sub-object set including objects A, C and E, it uses object A as the representative object according to the recognition probability of each object; for another example, after the computer device obtains the sub-object set including objects B and D , according to the recognition probability of each object, the object D is taken as the representative object.
  • the samples in the training sample set are samples used for model training.
  • the object information corresponding to the object in the training sample can be obtained as an object feature
  • the object category corresponding to the object in the training sample can be obtained as an object label.
  • the influence weight of the training gradient change is the degree of gradient change during the training process of the object recognition model, and the influence weight of the training gradient change is positively correlated with the recognition probability, that is, the greater the recognition probability of the candidate object belonging to the target object category, the higher the probability of using the candidate object.
  • the object trains the object recognition model, and the gradient changes more.
  • the computer device can positively correlate the recognition probability of the candidate object belonging to the target object category and the influence weight of the training gradient change to obtain the influence weight of each candidate object on the training gradient change of the target object recognition model.
  • Step S340 based on the influence weight of each candidate object on the training gradient change of the target object recognition model, select the candidate object satisfying the second preset condition from the candidate object set, and add the candidate object satisfying the second preset condition to the training sample set.
  • the second preset condition includes that the influence weight of the training gradient change is greater than the influence weight threshold or at least one of the influence weight ranking is before the second sorting threshold.
  • the influence weight of each candidate object on the training gradient change of the object recognition model is determined based on the recognition probability of each candidate object, and the candidate objects whose influence weight of the training gradient change is greater than the influence weight threshold are added to the training sample set.
  • each candidate object For all candidate objects included in the candidate object set, determine the influence weight of each candidate object on the training gradient change of the object recognition model based on the recognition probability of each candidate object; according to the order of the influence weight of the training gradient change from large to small, each candidate The objects are sorted, and the candidate objects with the influence weight of the training gradient change in the top few are added to the training sample set.
  • Step S342 using the training sample set to perform model training to obtain a target object recognition model used to identify the category of the target object.
  • the computer equipment performs supervised training based on the object features and object labels of each training sample in the training sample set to obtain a target object recognition model;
  • the target object recognition model may include support vector machines (SVM), convolutional neural network (Convolutional Neural Network, CNN), long short-term memory network (Long Short Term Memroy, LSTM), or Real-time Attention based Look-alike Model (RALM, similarity model based on real-time attention) and other models.
  • SVM support vector machines
  • CNN convolutional neural network
  • LSTM Long Short Term Memroy
  • RALM Real-time Attention based Look-alike Model
  • the feature extraction is performed on the candidate object information to obtain the object extraction feature. Since each candidate object is clustered according to the object extraction feature of the candidate object, it can be ensured that the candidate objects in the same sub-object set are relatively similar; then , since the recognition probability of the candidate object belonging to the target object category is obtained according to the object extraction features of the candidate object, therefore, the representative objects selected from each sub-object set according to the recognition probability of the candidate object can represent each representative object to the greatest extent Other candidate objects in the sub-object set where the object is located; the representative object is used as the training sample for model training to improve the prediction accuracy of the target object recognition model and realize the effective transmission of information; and, using the seed object to build the object screening model, due to the seed The confidence that the object belongs to the target object category is high, so it can improve the accuracy of the object screening model to predict that other objects belong to the target object; The strategy selects training samples, improves the accuracy of model prediction, improves the accuracy of object recognition, improves the effectiveness of information transmission
  • FIG. 4 mainly includes 5 parts, among which, sample preparation and continuous learning These two parts of the framework can be seen as offline data preparation.
  • Figure 5 shows the process of building an offline model, which mainly corresponds to the four parts of sample preparation, continuous learning framework, offline feature processing, and offline model training in Figure 4;
  • Figure 6 shows the online calling process, which mainly corresponds to the online The model calls this section.
  • "in a state of owning a property” and "not in a state of owning a property” can be regarded as target object categories.
  • This application example can be executed by computer equipment, and mainly includes: obtaining seed object groups based on manual labeling and business logic, and generating seed object portrait features, including: object basic attributes (such as gender, etc.), device basic attributes (such as Device model, etc.), network connection attributes (such as the number of times you have connected to your home Wi-Fi in the past month). Then, filter abnormal objects based on portraits, for example: filter objects that use a specific application program for more than 24 hours. In most business scenarios, the number of positive and negative samples that can be directly obtained through manual labeling and business experience is very small (generally less than 10,000), which cannot meet the minimum sample requirements for training models. This application example is based on the continuous learning framework.
  • the online work includes: First, regularly pull the latest feature set of the object from the online storage engine. Then, the business side sets the unique feature calculation logic, exports the original data from the online real-time log, and completes the feature calculation based on the online calculation engine. Then, concatenate the features, input them into the model, and output the probability that the current object has already purchased a house.
  • This application example can improve the accuracy of the recognition of purchased house objects, especially when the number of seed objects is small, the continuous learning framework and deep learning model can still ensure high prediction accuracy and effect stability.
  • This application example provides a more accurate object grouping scheme for common business scenarios such as product operation and advertisement placement, improves the effectiveness of information transmission, avoids pushing a large amount of information to unmatched objects, and saves computer resources.
  • Offline feature processing Construct the image features of the training samples, and based on the vertical characteristics of the features, combine the time dimension and different feature processing methods to generate high-dimensional feature vectors.
  • Offline model training Based on training samples and feature vectors, find a model with good classification effect and stability.
  • the above offline data preparation includes two parts: sample preparation and continuous learning framework; among them, the sample preparation part mainly includes the following steps:
  • Step S502 acquiring the manually marked seed object.
  • the seed object with label information (the label information refers to whether the object is in the state of owning a property) is obtained.
  • a batch of seed objects is roughly recalled based on rules, then filtered based on manual screening, and finally verified based on business logic.
  • Step S504 acquiring the basic image of the seed object.
  • the basic portrait includes some non-private behavior data of the subject in the application, such as whether to install a specific application, whether to use the harassment blocking function provided by the application, whether to use the answering assistant function provided by the application, etc.
  • Step S506 obtaining the index value of the abnormal object type evaluation index under each seed object.
  • abnormal object type evaluation indicators will be set based on business experience; abnormal type evaluation indicators include but are not limited to the traffic usage of objects in the application program, the time distribution of traffic generation, etc.; computer equipment can According to the traffic usage of each seed object in the application and the time distribution of traffic generation, the index value corresponding to each seed object can be obtained.
  • Step S508 based on the distribution of index values of abnormal type evaluation indicators in various sub-objects, the seed objects whose index values are distributed in the normal interval are regarded as normal seed objects, and the seed objects whose index values are distributed in the abnormal interval are regarded as abnormal seed objects, and Filter exception seed objects.
  • the "Raida criterion" can be used to judge the abnormal interval. Specifically, assuming that a set of test data only contains random errors, the standard deviation is obtained by calculating and processing it, and an interval is determined according to a certain probability. The error in the interval is not a random error but a gross error, and the data containing this error should be eliminated.
  • Step S510 storing normal seed objects.
  • the storage method can be offline storage
  • the server used for storage can be HDFS (The Hadoop Distributed File System, Hadoop Distributed File System), which is convenient for quick access in subsequent processes.
  • the continuous learning framework is to increase high-quality positive and negative samples. This part mainly includes the following steps:
  • Step S512 use the above-mentioned small number of normal seed objects to perform model training to obtain an object screening model
  • the object screening model obtained at this time can be denoted as M0;
  • the object screening and recognition model can be a DeepFM model.
  • Step S514 input the unlabeled samples into the object screening model to perform feature extraction and predict the probability that the samples belong to the target object category, so as to label the samples.
  • Step S516 storing the probability that each sample belongs to the target object category and the object extraction features of each sample output by the object screening model.
  • Step S518, selecting samples with higher representativeness and samples with higher discrimination from each sample
  • the selection process of samples with high representativeness mainly includes: clustering each sample to obtain each sample cluster, and for any sample cluster, the sample with a higher probability of the sample belonging to the target object category in any sample cluster is taken as a representative sample;
  • the selection process of samples with higher discrimination mainly includes: Based on the idea of the maximum gradient length (Expected Gradient Length), among all samples, the samples with a higher probability of the sample belonging to the target object category are taken as samples with higher discrimination.
  • the maximum gradient length Exected Gradient Length
  • Step S520 if the object screening model needs to be adjusted, the object screening model can be fine-tuned by using samples with high representativeness and high recognition ability until the adjusted object screening model reaches the performance threshold.
  • Step S522 determining the total number of samples with high representativeness and samples with high discrimination.
  • Step S524 judge whether the total quantity is greater than or equal to the quantity threshold; if so, proceed to step S526, otherwise return to step S514, and continue to iterate to obtain more samples with higher representativeness and higher discrimination.
  • the number threshold is determined according to the number of training samples required for training the target object recognition model.
  • the above-mentioned offline feature processing part mainly includes the following steps:
  • Step S526, constructing the portrait feature and business vertical type feature of the object.
  • the basic portrait features are mainly constructed based on the historical behavior data of the object, including: basic attributes of the object, basic attributes of the device, network connection attributes, etc.; basic attributes of the object are not limited to: gender, place of origin, city of residence, etc.; basic attributes of the device include but Not limited to: mobile phone resolution, API_Level (interface level), number of CPU cores, etc.; network connection attributes include but not limited to: the number of connected Wi-Fi, the earliest time to connect to Wi-Fi every day, etc.
  • vertical type characteristics including the click rate and conversion rate of the object for a specific type of advertisement.
  • specific types of advertisements include house purchase advertisements, house rental advertisements, decoration advertisements, and the like.
  • Step S528, aggregating similar profile features or business vertical type features in different time dimensions.
  • the time dimension can be nearly half a year/nearly 3 months/nearly 1 month/nearly 1 week; for the same type of feature, the feature values under different time dimensions can be aggregated to obtain the aggregated feature value; the aggregation can be Sum, median, standard deviation and many other ways.
  • Step S530 perform normalization processing or discretization processing on the aggregated features, which includes normalizing numerical features and discretizing non-numeric features; where the normalization method can be Gaussian normalization Discretization; discretization includes the following methods:
  • One-Hot Encoding For example, for characteristics such as object gender.
  • Count Encoding frequency encoding: For example, for the WiFi POI (point of interest) feature of an object, Count Encoding will be used to identify the object and the degree of interest of this POI. For example, the subject went to the POI "Gourmet-Chinese Cuisine-Cantonese Cuisine" three times in a week.
  • Category Embedding (category embedding): If many category features have strong sparsity, in order to avoid model overfitting and improve model stability, a neural network is introduced to convert high-dimensional sparse classification variables into low-dimensional dense Embedding variable.
  • Embedding For the missing value processing of features, methods such as “elimination”, “average filling” and “missing mark” can be used. Among them, converting missing values into Embedding expression has the greatest effect on the model Positive earnings.
  • the above iii is mainly to input the Category feature (category feature) to the DNN model (Deep Neural Networks, deep neural network model) to train the Embedding feature.
  • the DNN model Deep Neural Networks, deep neural network model
  • 1WiFi Trajectory Embedding (WiFi trajectory embedding representation). Based on the MST-CNN (Masked Self-Supervised Transformer-Convolution neural network) deep learning network, embedding the object's WiFi connection trajectory data to capture the object's Wi-Fi behavior pattern information.
  • MST-CNN Mask Self-Supervised Transformer-Convolution neural network
  • 2App Traffic Embedding (APP traffic embedding representation). Based on the List-Embedding method, the Embedding extraction is performed on the traffic usage behavior sequence of the object using different types of App (application, application program), such as using the Traffic Embedding of the social type App, to obtain low-dimensional and dense object behavior characteristics.
  • App application, application program
  • Step S532 combining and storing the normalized numerical features and the discretized non-numerical features.
  • the storage method can be offline storage in HDFS, which is convenient for quick access in subsequent processes.
  • the computer device can also solidify the feature processing logic of the above steps S526 to S532, perform regular offline automatic calculation, and upload the features obtained from the offline calculation to the online storage engine for storage.
  • the computer device may also perform cleaning, filtering, and verification on the object's portrait features and business vertical type features. Since features are generated through multiple strategic logic paths, the data quality of features is difficult to guarantee, so feature data quality monitoring is required.
  • the specific method is as follows:
  • Quality standards include but are not limited to: the duration of using a specific App is less than 16 hours a day, etc.;
  • the preset strategy verify the existing features, and eliminate invalid and abnormal features. For example, the duration of use of a certain app by the subject cannot be longer than 24 hours, etc.;
  • the spliced vectors may be (0.2,0.1,1,...,-1,...,0).
  • the above offline model training part mainly includes the following steps:
  • Step S534 randomly divide the training sample set to obtain a training set and a test set.
  • the ratio of the training set to the verification set can be 5:1.
  • Step S536 based on preset parameters, train multiple models in parallel, and use the model with the best prediction effect among the multiple models after parallel training as the target object recognition model.
  • the models used in parallel training include but are not limited to: support vector machines (support vector machines, SVM), convolutional neural network (Convolutional Neural Network, CNN), long short-term memory network (Long Short Term Memroy, LSTM), Real-time Attention based Look-alike Model (RALM), etc.
  • Step S538, optimize the parameters of the target object recognition model to improve the prediction effect of the target object recognition model.
  • the evaluation index with the best prediction effect can be AUC; parameter tuning refers to the grid optimization of the hyperparameters of the selected model, in order to expect the evaluation index AUC to be improved.
  • Step S540 obtain the target object recognition model after parameter optimization, and verify the prediction effect of the target object recognition model after parameter optimization and the stability of the prediction effect on multiple verification sets, and judge the target object after parameter optimization Identify whether the prediction effect of the model is up to standard, if yes, go to step S542, otherwise return to step S538.
  • the evaluation index of the prediction effect of the target object recognition model may be AUC.
  • AUC has the following advantages: a)
  • the AUC indicator itself has nothing to do with the absolute value of the model prediction score, and only focuses on the ranking effect, which is closer to the needs of actual business; b)
  • the calculation method of AUC also considers the classification ability of the learner for positive and negative examples. In the case of unbalanced samples Under this condition, it is still possible to make a reasonable evaluation of the classifier.
  • Step S542 solidifying the target object recognition model after parameter optimization; the solidified target object recognition model can be used for advertising push, such as the push of real estate advertisements or educational advertisements; based on TensorFlow (an end-to-end open source machine
  • the Saver() method of the learning platform solidifies the trained model, and generates 4 files in total: a) checkpoint text file, which records the path information list of the model file; b) model.ckpt.data, which records the network weight information; c) model.ckpt.index.data and .index are binary files that store variable weight information in the model.
  • model training process from step S534 to step S542 above can also be solidified, and offline training, verification, alarm, and solidification can be performed at regular intervals.
  • the purpose of calling this part of the above-mentioned online model is to use the offline solidified model to complete online real-time prediction, which mainly includes the following steps:
  • Step S602 acquiring the target object recognition model after the parameters are optimized.
  • this step includes the following parts:
  • the trained model is solidified based on the Saver() method of TensorFlow;
  • the client pulls the latest model file from the cloud based on the method of calling the service interface.
  • Step S604 acquiring the portrait feature and business vertical type feature of the object to be identified.
  • portrait features and business vertical type features can be divided into real-time features and offline features; It mainly reflects the real-time behavior of the object in the application that day, and the granularity can be hourly, for example, the total time that the object uses a specific application in the day; the offline feature refers to the feature whose stability is higher than the threshold, and the offline feature is less likely Changes over time, such as the gender of the subject, the equipment the subject uses, etc.
  • Offline features also known as general features
  • basic attribute features such as the gender of the object
  • device attribute features such as the brand of the device used by the object
  • Wi-Fi connected to the device that day such as the number of Fi.
  • Computer equipment can be based on Spark and TensorFlow computing engines, and based on different feature processing rules, the offline features obtained by timing calculation of the data of the object to be recognized; feature processing rules include: One-Hot Encoding, Count Encoding, Category Embedding, NaN Embedding, Consolidation Encoding, WiFi Trajectory Embedding, App Traffic Embedding.
  • the computer device can also obtain the real-time data of the object from the real-time log of the object, and calculate the real-time data of the object according to the preset characteristic processing rules to obtain the real-time characteristics of the object, such as the total time the object uses a specific application program on the day and the first opening time of the day, etc. .
  • Step S606 splicing the portrait features of the object to be identified and the vertical type features of the business.
  • the features obtained after splicing can be called splicing features, which belong to high-dimensional vectors, such as (1,0,2,1.1,41,...,3,1,14).
  • Step S608 input the splicing feature into the target object recognition model, so that the target object recognition model predicts the probability that the object to be recognized belongs to the target object category based on the splicing feature;
  • Step S610 acquiring the probability that the object to be recognized output by the target object recognition model belongs to the category of the target object.
  • Step S612 if the probability that the object to be recognized belongs to the target object category is greater than or equal to the probability threshold, then determine that the object to be recognized belongs to the target object category.
  • the target object recognition model of this application example is the RALM model, and the prediction effect of the offline experiment evaluation RALM model mainly includes:
  • AUC Average under Curve
  • Evaluation indicators include: advertisement click-through rate, object real-name registration rate.
  • Figure 7 is a comparison of model effects of using different models to predict the state of the object’s real estate.
  • the RALM model is the model used in this application example. It can be seen that:
  • the RALM scheme has an average increase of 29.24% compared with the scheme based on artificial strong rules or non-deep learning scheme;
  • the RALM scheme has an average increase of 28.18% compared with the artificial strong rule-based scheme or the non-deep learning scheme;
  • Figure 8 is a comparison chart of the business effect of using different models to predict the state of the object’s real estate.
  • the RALM model is the model used in this application example. It can be seen that:
  • the RALM scheme has an average increase of 441.93% compared with the scheme based on artificial strong rules or the scheme without deep learning;
  • the RALM scheme has an average increase of 309.54% compared with the artificial strong rule-based scheme or the non-deep learning scheme.
  • This application example has strong reusability.
  • the object category of the corresponding scene can be predicted without modifying other content, saving the processing time of computer equipment.
  • changing the object category of the positive sample such as "group recognition of the object's car purchase status"
  • the server accumulates the corresponding log data, and finally uses the same method of feature splicing, feature processing, and model training to produce results.
  • steps in the flow charts of FIG. 2 to FIG. 6 are shown sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in FIGS. 2 to 6 may include multiple steps or stages. These steps or stages are not necessarily executed at the same time, but may be executed at different times. The steps or stages The order of execution is not necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least a part of steps or stages in other steps.
  • an object recognition device including:
  • a candidate object set acquisition module 902 configured to acquire a candidate object set; the candidate object set includes a plurality of candidate objects;
  • the probabilistic identification module 904 is configured to obtain multiple candidate object information of a candidate object in multiple dimensions, perform feature extraction on each candidate object information, obtain candidate object features corresponding to each candidate object information, and fuse each candidate object feature to obtain candidate object correspondence
  • the object extraction features of the object are extracted based on the object category probability identification, and the identification probability that the candidate object belongs to the target object category is obtained;
  • the clustering module 906 is used to cluster the object extraction features corresponding to the candidate objects, obtain the sub-extraction feature sets corresponding to each cluster category, and form the candidate objects corresponding to the object extraction features in the same sub-extraction feature set into a sub-object gather;
  • the representative object selection module 908 is configured to select representative objects from the sub-object sets based on the recognition probabilities corresponding to each candidate object in each sub-object set composed of sub-extracted feature sets.
  • the object extraction feature is obtained by performing feature extraction through an object screening model; the device further includes: an object screening model building module for obtaining a seed object corresponding to the target object category; obtaining seed object information corresponding to the seed object, The seed object information is used as the training feature in the training sample, and the target object category is used as the label in the training sample to form a training sample; model training is performed based on the training sample to obtain an object screening model.
  • the probability identification module 904 is further configured to input multiple candidate object information of candidate objects in multiple dimensions into the feature extraction layer of the object screening model for feature extraction, and obtain candidate objects corresponding to each candidate object information feature; each candidate object feature is input to the classification layer of the object screening model, so that the classification layer fuses each candidate object feature to obtain the object extraction feature; the output candidate object belongs to the target object after the classification layer performs object category probability identification on the object extraction feature Class recognition probabilities.
  • the clustering module 906 is also used to obtain the feature distance between the object extraction features corresponding to different candidate objects; for each object extraction feature, based on each feature distance, determine the area where each object extraction feature is located Based on the number of object extraction features in the object, the regional object density of each object extraction feature is obtained; the cluster center is selected based on the regional object density of each object extraction feature, and the object extraction features are clustered based on the cluster center to obtain each cluster.
  • the sub-extraction feature set corresponding to the class category; the candidate objects corresponding to the object extraction features in the same sub-extraction feature set form a sub-object set.
  • the clustering module 906 is also used to determine the adjacent extraction features of the object extraction features in the features whose area object density is greater than the area object density of the object extraction features;
  • the feature distance is used as the target distance corresponding to the object extraction feature;
  • the cluster center is selected based on the area object density of the object extraction feature and the target distance corresponding to the object extraction feature.
  • the clustering module 906 is also used to determine the object extraction feature with the smallest feature distance between the object extraction feature and the object extraction feature among the features whose area object density is greater than the object extraction feature; The object extraction feature with the smallest feature distance between them is used as the adjacent extraction feature of the object extraction feature.
  • the clustering module 906 is also used to obtain the current object extraction feature of the cluster category to be determined; to obtain the cluster center whose regional object density is greater than the current object extraction feature, as the current object extraction feature corresponding
  • the candidate cluster centers corresponding to the current object extraction features form the candidate center set; based on the distance between the current object extraction features and the candidate cluster centers, select the adjacent cluster centers corresponding to the current object extraction features from the candidate center set , adding the extracted feature of the current object to the sub-extracted feature set corresponding to the adjacent cluster center.
  • the clustering module 906 is also used to determine the cluster center with the smallest feature distance between the extracted features of the current object and the cluster center with the smallest feature distance between the extracted features of the current object in the set of candidate centers; The center is used as the adjacent clustering center corresponding to the current object extraction feature.
  • the device further includes an area determination module, configured to, when one of the objects is used as the target object, obtain the feature distance between the object extraction features of other objects and the object extraction features of the target object; when one of the other objects When the corresponding feature distance is smaller than the preset radius, it is determined that the object extraction feature of one of the other objects is located in the area where the object extraction feature of the target object is located.
  • an area determination module configured to, when one of the objects is used as the target object, obtain the feature distance between the object extraction features of other objects and the object extraction features of the target object; when one of the other objects When the corresponding feature distance is smaller than the preset radius, it is determined that the object extraction feature of one of the other objects is located in the area where the object extraction feature of the target object is located.
  • the representative object selection module 908 is further configured to, for each sub-object set composed of each sub-extraction feature set, obtain candidate objects whose recognition probabilities satisfy the first preset condition from the sub-object sets, as The corresponding representative object in the sub-object set; the first preset condition includes at least one of a recognition probability greater than a probability threshold or a probability ranking before the first sorting threshold.
  • the device further includes a second set processing module, configured to determine the influence weight of each candidate object on the training gradient change of the target object recognition model based on the recognition probability that the candidate object belongs to the target object category, and the influence weight of the training gradient change The weight is positively correlated with the recognition probability; based on the influence weight of each candidate object on the training gradient change of the target object recognition model, the candidate object that meets the second preset condition is selected from the candidate object set, and the candidate that satisfies the second preset condition The object is added to the training sample set; the second preset condition includes at least one of the influence weight of the training gradient change being greater than the influence weight threshold or the influence weight ranking being before the second sorting threshold.
  • the object category recognition probability is performed through a target object recognition model
  • the device further includes a target object recognition model building module, which is used to select similar first training objects and second training objects from the training sample set, and The first training object and the second training object form the object group; the training object information corresponding to each training object in the object group is input in the same feature extraction layer, and the training object characteristics corresponding to each training object are extracted; , get the object similarity between the training objects in the object group; get the model loss value based on the object similarity; the model loss value and the object similarity have a negative correlation; train the object recognition model based on the model loss value, and get the target object Identify the model.
  • the device further includes an object information processing module, which is used to acquire the object information category; acquire the time dimension set and information statistics angle corresponding to the training object in the object information category, and the time dimension set includes multiple information statistics time dimensions; In each information statistical time dimension, the statistical value of the object information corresponding to the training object obtained based on the statistical angle of information statistics; the statistical value of the object information is the statistical value of the information corresponding to the object information category; The statistical value of the object information is aggregated, and the aggregated object information is used as the training object information corresponding to the training object.
  • an object information processing module which is used to acquire the object information category; acquire the time dimension set and information statistics angle corresponding to the training object in the object information category, and the time dimension set includes multiple information statistics time dimensions; In each information statistical time dimension, the statistical value of the object information corresponding to the training object obtained based on the statistical angle of information statistics; the statistical value of the object information is the statistical value of the information corresponding to the object information category; The statistical value of the object information is aggregated, and the aggregate
  • the object information processing module is further configured to obtain the comprehensive information statistical value based on the object information statistical value corresponding to each information statistical time dimension in the time dimension set; determine the statistical value of each object information statistical value and the comprehensive information statistical value value difference; the statistical value dispersion corresponding to the time dimension set is obtained based on the statistical value difference, and the statistical value dispersion is used as the training object information corresponding to the training object; the statistical value dispersion is positively correlated with the statistical value difference.
  • the candidate object information is subjected to feature extraction to obtain the object extraction feature, and since each candidate object is clustered according to the object extraction feature of the candidate object, objects belonging to the same category can be divided into the same sub-object set; Since the recognition probability that the candidate object belongs to the target object category is obtained according to the object extraction features of the candidate object, the representative objects selected from each sub-object set according to the recognition probability of the candidate object can belong to the sub-objects of the same category Selecting representative objects from the collection improves the accuracy of object recognition, improves the effectiveness of information transmission, avoids pushing a large amount of information to unmatched objects, and saves computer resources.
  • Each module in the above-mentioned object recognition device can be fully or partially realized by software, hardware and a combination thereof.
  • the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 10 .
  • the computer device includes a processor and a memory connected through a system bus.
  • the computer device may also include a network interface.
  • the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions and a database.
  • the internal memory provides an environment for the execution of the operating system and computer readable instructions in the non-volatile storage medium.
  • the database of the computer device is used to store data for object recognition.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer readable instructions are executed by the processor, an object recognition method is implemented.
  • FIG. 10 is only a block diagram of a part of the structure related to the solution of this application, and does not constitute a limitation to the computer equipment on which the solution of this application is applied.
  • the specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.
  • a computer device including a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor implements the Methods.
  • a non-volatile computer-readable storage medium stores computer-readable instructions.
  • the processor implements the present invention. Apply the method in the examples.
  • a computer program product comprising computer readable instructions stored in a computer readable storage medium; a processor of a computer device reads the computer readable instructions from the computer readable storage medium Read instructions, the processor executes the computer-readable instructions, so that the computer device executes the method in the embodiment of the present application.
  • Nonvolatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, or optical memory, among others.
  • Volatile memory can include Random Access Memory (RAM) or external cache memory.
  • RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

一种对象识别方法，包括：对候选对象在多个维度的多个候选对象信息进行提取得到候选对象特征，对融合各候选对象特征得到的对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率(S204)；对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合(S206)；对于按各子提取特征集合组成的各个子对象集合，分别基于子对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象(S208)；选取得到的代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。

Description

对象识别方法、装置、设备、存储介质和程序产品
本申请要求于2021年09月30日提交中国专利局，申请号为202111161862.9，申请名称为“对象识别方法、装置、设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，特别是涉及一种对象识别方法、装置、计算机设备、存储介质和程序产品。
背景技术
在人工智能时代，通过人工智能的方式从海量的数据中挖掘出有价值的信息，以实现信息的有效传递，是当前研究热点之一；例如，通过挖掘不同对象的信息，利用挖掘的信息构建并训练人工智能模型，使用该人工智能模型预测对象所属的类别，该预测结果可以应用于多种场景，比如可以基于该预测结果为对象生成相关的推送信息，实现信息的有效传递，又比如可以基于该预测结果为不同类别的对象分配相应份额的网络资源，等等。
训练样本的选取对于人工智能模型的训练效果起到非常重要的作用，如果训练样本选取不合适或不准确，则会影响模型训练效果，导致后续的模型预测准确性较低。
发明内容
一种对象识别方法，包括：
获取候选对象集合；候选对象集合包括多个候选对象；
获取候选对象在多个维度的多个候选对象信息，对各候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率；
对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合；
对于按各子提取特征集合组成的各个子对象集合，分别基于子对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象；选取得到的所述代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
一种对象识别装置，包括：
候选对象集合获取模块，用于获取候选对象集合；候选对象集合包括多个候选对象；
概率识别模块，用于获取候选对象在多个维度的多个候选对象信息，对各候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率；
聚类模块，用于对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合；
代表对象选取模块，用于对于按各子提取特征集合组成的各个子对象集合，分别基于子 对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象；选取得到的所述代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
一种计算机设备，包括处理器、存储器；所述存储器用于存储计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述处理器实现本申请实施例中的对象识别方法。
一种非易失性的计算机可读存储介质，计算机可读存储介质存储有计算机可读指令，该计算机可读指令被处理器执行时，使得该处理器实现本申请实施例中的对象识别方法。
一种计算机程序产品，包括计算机可读指令，该计算机可读指令存储在计算机可读存储介质中；计算机设备的处理器从计算机可读存储介质读取该计算机可读指令，处理器执行该计算机可读指令，使得该计算机设备实现本申请实施例中的对象识别方法。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1(a)为一个实施例中对象识别方法的应用环境图；
图1(b)为一个实施例中对象识别方法的处理过程示意图；
图1(c)为另一个实施例中对象识别方法的处理过程示意图；
图2为又一个实施例中对象识别方法的流程示意图；
图3为再一个实施例中对象识别方法的流程示意图；
图4为一个实施例中对象识别方法的处理架构图；
图5为一个实施例中对象识别方法的离线处理流程示意图；
图6为一个实施例中对象识别方法的在线处理流程示意图；
图7为一个实施例中用不同模型进行用户房产状态识别的模型效果比对图；
图8为本申请实施例提供的一种用不同模型进行对象房产状态识别的业务效果比对图；
图9为本申请实施例提供的一种对象识别装置的结构框图；
图10为本申请实施例提供的一种计算机设备的内部结构图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
在本申请中提及“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本申请的至少一些实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是，本申请所描述的实施例可以与其它实施例相结合。
随着深度学习技术研究和进步，深度学习技术在许多领域展开研究和应用，其中一个应用领域是基于深度学习模型对对象进行分群分类；例如，基于深度学习模型对对象房产状态的预测结果，智能判断是否向对象推送与买房有关的内容；又例如，基于深度学习模型对对象购车状态的预测结果，智能判断是否向对象推送与买车有关的内容；再例如，基于深度学 习模型对对象租房状态的预测结果，智能判断是否向对象推送与租房有关的内容。
其中，对象房产状态是指对象当前是否拥有房产，即对象是否已买房；在通过深度学习模型预测对象房产状态的场景中，深度学习模型输出的预测概率越大，该对象拥有房产的概率越大，此时可以不向该对象推送与买房有关的内容，深度学习模型输出的预测概率越小，该对象拥有房产的概率越小，此时可以向该对象推送与买房有关的内容。
对象购车状态是指对象当前是否已拥有车辆，即对象是否已购车；在通过深度学习模型预测对象购车状态的场景中，深度学习模型输出的预测概率越大，该对象拥有车辆的概率越大，此时可以不向该对象推送与买车有关的内容，深度学习模型输出的预测概率越小，该对象拥有车辆的概率越小，此时可以向该对象推送与买车有关的内容。
对象租房状态是指对象当前是否已经租房；在通过深度学习模型预测对象租房状态的场景中，深度学习模型输出的预测概率越大，该对象已经租房的概率越大，此时可以不向该对象推送与租房有关的内容，深度学习模型输出的预测概率越小，该对象已经租房的概率越小，此时可以向该对象推送与租房有关的内容。
本申请实施例提供的方案涉及人工智能的深度学习的技术，可以应用于云技术、云安全、人工智能和智慧交通等场景；本申请所涉及的对象信息(包括但不限于对象设备信息、对象行为信息等)和数据(包括但不限于用于展示的数据、分析的数据等)，均为经对象授权或者经过各方充分授权的信息和数据；对应的，本申请还提供有相应的对象授权入口，供对象选择授权或者选择拒绝。
本申请提供的对象识别方法，可以由本地计算设备/本地计算系统执行，也可以由分布式计算系统执行；分布式计算系统例如是多个物理服务器构成的服务器集群、或者终端设备和服务器构成的系统。
图1(a)为一个实施例中对象识别方法的应用环境图。其中，终端102与服务器104通过通信网络进行通信。其中，终端102可以通过通信网络与服务器104进行交互；终端102可以但不限于是各种台式计算机、笔记本电脑、智能手机、平板电脑、物联网设备和便携式可穿戴设备，物联网设备可为智能音箱、智能电视、智能空调、智能车载设备等。便携式可穿戴设备可为智能手表、智能手环、头戴设备等。服务器104可以用独立的服务器或者是多个服务器组成的服务器集群或云服务器来实现。服务器104可以通过数据存储系统存储对象识别方法所涉及的数据，该数据存储系统可以集成在服务器104中，也可以与服务器104分开设置。
本申请提供的对象识别方法，可以由终端102和服务器104协同执行，可以由终端102单独执行，也可以由服务器104单独执行。以服务器104单独执行为例，服务器104可以获取候选对象集合，该候选对象集合包括多个候选对象，服务器104获取候选对象在多个维度的多个候选对象信息，对各候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率，随后，服务器104对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合，继而服务器104对于按各子提取特征集合组成的各个子对象集合，分别基于子对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象，选取得到的代表对象的候选对象信息用于 训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
本申请提供的对象识别方法中，所谓的目标对象类别可以是不同应用场景下的类别，例如，通过该对象识别方法，可以预测对象是否属于拥有房产状态的对象类别，还可以预测对象是否属于拥有车辆的对象类别，还可以预测对象是否属于租房状态的对象类别，实现对对象是否拥有房产、是否拥有车辆、是否租房的精准预测，从而准确确定对房产信息、购车信息、租房信息有潜在需要的对象，为这类对象推送相关信息，不仅可以提高信息传递的有效性，还可以避免生成大量无效相关信息推送至不匹配的对象带来的计算机资源与网络资源的占用与浪费，节约计算机资源与网络资源。
如图1(b)所示，本申请提供的对象识别方法主要包括：对多个对象进行特征提取，例如对对象A、B、C、D和E进行特征提取，将提取得到的特征称为对象提取特征，进而得到各对象的对象提取特征，例如，对象A的特征a、对象B的特征b、对象C的特征c、对象D的特征d和对象E的特征e；接着，基于各对象的对象提取特征，识别对象属于目标对象类别的概率，将识别到的概率称为识别概率，进而得到各对象属于目标对象类别的识别概率P，如对象A属于目标对象类别的识别概率为P A，对象B属于目标对象类别的识别概率为P B，对象C属于目标对象类别的识别概率为P C，对象D属于目标对象类别的识别概率为P D，以及对象E属于目标对象类别的识别概率为P E。并且，基于各对象的对象提取特征对对象进行聚类，得到多个对象集合；按照对象属于目标对象类别的识别概率，分别从每一对象集合选取对应的代表对象，将代表对象加入训练样本集合，利用该训练样本集合构建目标对象识别模型，利用该目标对象识别模型预测对象是否属于目标对象类别。
上述对象识别方法，通过对候选对象信息进行特征提取得到对象提取特征，由于根据候选对象的对象提取特征对各候选对象进行聚类，因此可以将属于同一类别的候选对象划分在同一个子对象集合中。由于候选对象属于目标对象类别的识别概率是根据候选对象的对象提取特征得到的，因此，根据候选对象的识别概率分别从各子对象集合中选取得到的代表对象，可以从属于同一类别的子对象集合中选取出具有代表性的代表对象，选出的代表对象既能覆盖所有类别的候选对象，又具备每个类别的代表性特点，使得利用代表对象的候选对象信息训练得到的目标对象识别模型，能够挖掘出不同类别的差异性与每个类别的代表性特点，进行提高预测对象是否属于目标对象类别的准确度。在信息推送场景下，可以提高推送信息传递的有效性，避免将大量相关信息推送至不匹配的对象，节约计算机资源。
本申请还提供一种对象识别方法，如图1(c)所示，该方法主要包括：基于人工标注和业务的逻辑方式，获取少量的种子对象，在少量的种子对象基础上，构建对象筛选模型，经过对象筛选模型，对其他对象进行多轮迭代处理，得到大量对象的对象提取特征和属于目标对像类别的识别概率；接着，基于对象的对象提取特征，对大量对象进行聚类，得到多个对象集合，并根据对象属于目标对像类别的识别概率，从各对象集合中选取代表对象，将代表对象加入训练样本集合中；利用训练样本集合进行模型训练，得到目标对象识别模型，利用该目标对象识别模型预测对象是否属于目标对象类别。
如图2所示，为一个实施例中对象识别方法的流程示意图。以下结合图1(b)、图1(c)和图2介绍本申请提供的对象识别方法，该方法可以应用于计算机设备例如图1(a)所示的服务器104中，主要包括如下步骤：
步骤S202，获取候选对象集合。
其中，对象是具有所属类别的对象，在预测对象是否拥有房产的场景中，对象所属的类别可以是“处于拥有房产状态”或是“不处于拥有房产状态”；在预测对象是否拥有车辆的场景中，对象所属的类别可以是“处于拥有车辆状态”或是“不处于拥有车辆状态”。候选对象是等待被挑选以用于模型训练的对象，可以从候选对象集合中，挑选出有代表性的对象，用于进行模型训练。例如图1所示的对象A、B、C、D和E；其中，多个候选对象形成候选对象集合，也即候选对象集合包括多个候选对象，多个是指至少两个。
本申请实施例中，计算机设备可以响应于样本选取指令时，获取候选对象集合。该候选对象集合可以是样本选取指令中携带的，也可以是预先存储的。
步骤S204，获取候选对象在多个维度的多个候选对象信息，对各候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率。
其中，对象信息是与对象相关的信息，例如对象性别信息、对象设备信息、或对象设备所连接的网络的信息的至少一种。对象设备可以包括对象使用的智能手表、手机或笔记本电脑等，设备信息可以包括设备分辨率，或设备的中心处理器(central processing unit，CPU)的内核数量的至少一种。如果对象设备具有上网功能，那么该对象设备所连接的网络可以包括WiFi(Wireless Fidelity)、移动网络；如果对象设备连接的网络是WiFi，那么与对象设备所连接的网络的信息可以包括每天连接WiFi的最早时间，或连接到的不同WiFi的数量的至少一种。当对象待被挑选以用于模型训练时，该对象为候选对象，那么可以将该对象的信息称为候选对象信息。
候选对象特征是对候选对象信息进行特征提取得到的特征；由于候选对象信息可以分为多个维度，例如候选对象的性别信息、候选对象的设备信息和候选对象的设备所连接的网络信息，因此，对不同候选对象信息进行特征提取，可以得到不同的候选对象特征，例如，对候选对象的性别信息进行特征提取得到的特征，又例如，对候选对象的设备信息进行特征提取得到的特征。
对象提取特征是对同一候选对象的多个维度的候选对象特征进行融合后得到的特征，示例性地，针对属于候选对象的对象A，对对象A的性别信息、设备信息和设备所连接的网络信息分别进行特征提取，得到多个维度的候选对象特征，这些维度的候选对象特征分别记为a x、a y和a z，接着，可以对a x、a y和a z进行融合，得到a r，基于a r得到对象A的对象提取特征。
其中，融合候选对象特征的方式可以是对a x、a y和a z进行求和取平均值，也可以是对a x、a y和a z进行求和，还可以是对a x、a y和a z进行加权求和后取平均值，还可以是将这些特征进一步输入到全连接层中进行处理，得到一个融合后的特征。
被融合的候选对象特征可以是全部维度的候选对象特征，对应地，所得到的对象提取特征包括的特征值的数量为一个；示例性地，若候选对象具有p维的候选对象特征(a 1，a 2，a 3，…，a p)，那么可以对这p维的候选对象特征进行融合，将得到的a r作为对象提取特征的特征值，该对象提取特征所包括的特征值的数量为一个。
被融合的候选对象特征也可以是部分维度的候选对象特征，对应地，所得到的对象提取特征包括的特征值的数量大于一个；示例性地，若候选对象具有q维的候选对象特征(a 1， a 2，a 3，…，a q)，那么可以仅对候选对象特征a 1，a 2和a 3进行融合，将得到的a r和其他未进行融合的候选对象特征进行拼接，得到对象提取特征的特征值，该对象提取特征所包括的特征值的数量为至少一个。
本步骤中，计算机设备可以获取候选对象的多个候选对象信息，并对各个候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，并融合各个候选对象特征得到候选对象的对象提取特征，如图1所示的对象A的对象提取特征a、对象B的对象提取特征b、对象C的对象提取特征c、对象D的对象提取特征d、对象E的对象提取特征e，并且各对象提取特征包括至少一个特征值，分别表示为[a 1，a 2，a 3，…，a n]、[b 1，b 2，b 3，…，b n]、[c 1，c 2，c 3，…，c n]、[d 1，d 2，d 3，…，d n]、[e 1，e 2，e 3，…，e n]。
其中，对象所属的类别可以称为对象类别；目标对象类别是对象识别模型所要识别的对象类别。例如，在预测对象是否拥有房产的场景中，对象识别模型所要识别的对象类别是“处于拥有房产状态”和“不处于拥有房产状态”，此时，“处于拥有房产状态”或“不处于拥有房产状态”为目标对象类别；又例如，在预测对象是否拥有车辆的场景中，对象识别模型所要识别的对象类别是“处于拥有车辆状态”和“不处于拥有房产状态”，此时，“处于拥有车辆状态”或“不处于拥有房产状态”为目标对象类别。
对象属于目标对象类别的识别概率是对象属于目标对象类别的可能性大小。对象属于目标对象类别的识别概率越大，该对象属于目标对象类别的可能性越大；对象属于目标对象类别的识别概率越小，该对象属于目标对象类别的可能性越小。
示例性地，计算机设备得到对象A的特征a、对象B的特征b、对象C的特征c、对象D的特征d和对象E的特征e，由于各特征是描述对应对象的，因此，计算机设备分别对特征a、对象B的特征b、对象C的特征c、对象D的特征d和对象E的特征e进行分析，可以确定对象A属于目标对象类别的识别概率为P A，对象B属于目标对象类别的识别概率为P B，对象C属于目标对象类别的识别概率为P C，对象D属于目标对象类别的识别概率为P D，以及对象E属于目标对象类别的识别概率为P E。例如，计算机设备可以将各个候选对象的对象提取特征分别输入到对象筛选模型的分类层中，分类层可以输出候选对象属于目标对象类别的识别概率。
步骤S206，对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合。
对各候选对象对应的对象提取特征进行聚类，可以将相似的候选对象划分在一起，将不相似的候选对象分开；由于候选对象是通过对象提取特征描述，因此，该聚类是对各候选对象的对象提取特征的聚类，其中，被划分在一起的多个候选对象对应的对象提取特征形成的集合，可以看成是一个子提取特征集合，同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合，示例性地，对象A、E和C的对象提取特征被划分至同一个子提取特征集合中，对象A、E和C组成子对象集合。聚类后可以得到多个子对象集合，子对象集合的数量与聚类类别的数量一致。
另外，属于同一子对象集合的候选对象对应的对象提取特征间的相似度，大于属于不同子对象集合的候选对象对应的对象提取特征间的相似度；示例性地，对象A、E和C属于同一子对象集合，对象B和D属于同一个子对象集合，那么对象A和对象E各自的对象提 取特征之间的相似度，大于对象A和对象B各自的对象提取特征之间的相似度，对象E和对象C各自的对象提取特征之间的相似度，大于对象E和对象D各自的对象提取特征之间的相似度。
上述聚类可以通过K均值算法(k-means算法)或密度峰值聚类算法(clustering by fast search and find of density peaks，DPC)实现。
步骤S208，对于按各子提取特征集合组成的各个子对象集合，分别基于子对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象。
其中，与子对象集合对应的代表对象是该子对象集合的各候选对象中具代表性的候选对象，该代表对象能代表该子对象集合中的候选对象；从一个子对象集合中选取得到的代表对象的数量可以是一个也可以是多个，选取得到的代表对象的数量根据设定的第一预设条件确定。选取得到的代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
例如，计算机设备在得到包括对象A、C和E的子对象集合后，根据各对象对应的识别概率，将对象A作为代表对象；又如，在得到包括对象B和D的子对象集合后，根据各对象对应的识别概率，将对象D作为代表对象。再如，在得到包括对象A、C和E的子对象集合后，根据各对象对应的识别概率，将对象A和E作为代表对象。
在一些实施例中，步骤S208可以具体包括如下步骤：对于按各子提取特征集合组成的各个子对象集合，计算机设备分别从子对象集合中，获取识别概率满足第一预设条件的候选对象，作为子对象集合中对应的代表对象。其中，第一预设条件包括识别概率大于概率阈值，与，识别概率排序在第一排序阈值之前这两者中的至少一个，以下具体介绍这两个第一预设条件：
(1)第一预设条件：识别概率大于概率阈值：
在该第一预设条件下，计算机设备可以将子对象集合中候选对象的识别概率大于该概率阈值的候选对象作为代表对象。
该实施例中，通过概率阈值的方式确定各子对象集合的代表对象，保证代表对象的选取的准确性，得到用于构建目标对象识别模型的训练样本集合，从而证目标对象识别模型的预测准确性。
在一些场景中，可以通过相同的概率阈值，分别从各子对象集合中选取对应的候选对象；但是，在一些场景中，某些子对象集合包括的全部候选对象的识别概率阈值均小于该概率阈值，此时无法用该概率阈值选取该子对象集合的代表对象，导致该子对象集合被遗漏。
为避免无法为某些子对象集合选取代表对象导致遗漏的问题，在一些实施例中，计算机设备可以将设置多个等级的概率阈值，例如高等级的概率阈值、中等级的概率阈值和低等级的概率阈值，当按照当前等级的概率阈值不能选取到子对象集合中的代表对象时，可以利用低于当前等级的概率阈值，再次选取该子对象集合中的代表对象。
示例性地，当利用高等级的概率阈值不能选取到包括对象A、C和E的子对象集合的代表对象时，利用中等级的概率阈值再次选取该子对象集合的代表对象；当利用中等级的概率阈值仍无法选取到该子对象集合的代表对象时，利用低等级的概率阈值选取该子对象集合的代表对象。
在上述场景中，若利用设定的各等级的概率阈值都无法选取到包括对象A、C和E的 子对象集合的代表对象，那么为保证该子对象集合不被遗漏，可以将该子对象集合中识别概率最大的候选对象作为该子对象集合的代表对象，也可以按照识别概率由大到小的顺序，将该子对象集合中识别概率排在前几个(如前3个、前5个)的候选对象作为代表对象。
(2)第一预设条件：概率排序在第一排序阈值之前：
其中，概率排序是识别概率从大到小的排序；第一排序阈值可以是3、5或其他数值。
示例性地，针对包括对象A、C和E的子对象集合，计算机设备按照各候选对象的识别概率从大到小的排序，可以得到对象A的识别概率>对象E的识别概率>对象C的识别概率。若第一排序阈值为2，那么可以将对象A和对象E作为该子对象集合的代表对象。
该实施例中，通过概率排序的方式确定各子对象集合的代表对象，保证代表对象的准确性，得到用于构建目标对象识别模型的训练样本集合，从而保证目标对象识别模型的预测准确性。
在一些实施例中，计算机设备在得到上述代表对象后，可以将代表对象加入到训练样本集合中；该训练样本集合用于进行模型训练，以训练得到识别目标对象类别的目标对象识别模型。
训练样本集合中的样本是用于进行模型训练的样本，例如在训练时，可以获取训练样本中对象对应的对象信息，作为对象特征，获取训练样本中对象对应的对象类别，作为对象标签，基于对象特征以及对象标签进行有监督的训练，得到目标对象识别模型。
其中，目标对象识别模型可以包括支持向量机(support vector machines,SVM)、卷积神经网络(Convolutional Neural Network,CNN)、长短时记忆网络(Long Short Term Memroy,LSTM)、或者Real-time Attention based Look-alike Model(RALM，基于实时注意的相似性模型)等模型。
Real-time Attention based Look-alike Model(RALM)模型是一个基于相似性的look-alike模型，包含“对象表示学习”和“look-alike模型学习”两部分。
1)RALM模型采用双塔的结构，左侧的输入是种子对象的Embedding(Embedding，嵌入表示)，右侧的输入是目标对象的Embedding，两侧的Embedding经过一层FC(Fullconnection，全连接层)完成到低维空间的映射。由于右侧的目标对象的Embedding是经过对象表征学习得到的，为了防止过拟合，双塔的第一层FC是共享的。在经过FC层之后，左侧的塔可以得到各个簇对应的Embedding，然后将各个簇的Embedding和目标对象的Embedding分别输入给Global Attention Unit(全局注意力单元)和Local Attention Unit(局部注意力单元)就可以得到Global Embedding(全局嵌入表示)和Local Embedding(局部嵌入表示)。
2)RALM模型的迭代训练阶段。在反向传播的过程中，因为User Embedding(对象嵌入表示)的值可能会发生改变，所以为了保证种子对象簇的Embedding和User Embedding保持同步，在每一轮迭代之后，都必须重新进行聚类操作。
3)RALM相比其它模型，有两个效果：
a)优化了“对象表示学习”的效果。针对多域(Multi-Fields)对象兴趣表示学习问题，RALM引入了“注意力融合层(Attention Merge Layer)”的深度兴趣网络，它解决了由强相关特征和弱相关特征分别带来的过拟合和噪音问题
b)提高了种子对象在表示学习的鲁棒性和适应性。利用全局注意单元来学习种子对象 的全局表示，全局注意单元对单个对象的表示进行加权，并且惩罚噪音对象，这比所有对象权重一样更具有鲁棒性。利用局部注意单元来学习种子对象的局部表示，它对种子对象与目标对象的相关性进行加权。局部注意单元动态地基于目标对象来学习种子对象的表示，对于不同的目标对象，学习到的种子对象表示也不一样，这极大地提升了种子对象表示的表达能力。
4)训练RALM：RALM模型涉及聚类过程，聚类过程需要迭代且比较耗时，聚类中心数直接影响聚类效果。在对象房产状态预测的场景，线上的聚类中心数选择50-80较优。
本步骤中，可以并行训练多个模型，从中选出分类效果最好的模型，对效果最好的模型，进行参数调优。模型分类效果的评价指标可以包括AUC(Area under Curve，曲线下的面积)，AUC值越大，当前模型越有可能将正样本排在负样本前面，得到更好的分类结果。参数调优是指对选择模型的超参数进行网格寻优，以期待评价指标AUC能获得提升。
上述对象识别方法中，通过对候选对象信息进行特征提取得到对象提取特征，由于根据候选对象的对象提取特征对各候选对象进行聚类，因此可以将属于同一类别的候选对象划分在同一个子对象集合中。由于候选对象属于目标对象类别的识别概率是根据候选对象的对象提取特征得到的，因此，根据候选对象的识别概率分别从各子对象集合中选取得到的代表对象，可以从属于同一类别的子对象集合中选取出具有代表性的代表对象，选出的代表对象既能覆盖所有类别的候选对象，又具备每个类别的代表性特点，
在后续进行目标对象识别模型的训练时，可以将代表对象作为训练样本，从海量的样本中挖掘出更多的用于模型训练的样本，使得利用代表对象的候选对象信息训练得到的目标对象识别模型，能够挖掘出不同类别的差异性与每个类别的代表性特点，进而提高预测对象是否属于目标对象类别的准确度。
在一些实施例中，对象提取特征是通过对象筛选模型进行特征提取得到的，得到对象筛选模型的步骤包括：获取目标对象类别对应的种子对象；获取种子对象对应的种子对象信息，将种子对象信息作为训练样本中的训练特征，将目标对象类别作为训练样本中的标签，组成训练样本；基于训练样本进行模型训练，得到对象筛选模型。
其中，针对类别未标注的对象，可以通过人工标注或自动标注的方式对这些对象进行类别标注，当人工标注和自动标注的准确性较高时，标注结果较为可信，此时可以将类别已标注的对象作为置信度较高的对象，该置信度较高的对象可以称为种子对象。
在一些场景中，负样本较为容易获取，正样本较难获取，此时，该种子对象可以是上述置信度较高的对象中，类别标注为属于目标对象类别的对象；其中，负样本为不属于目标对象类别的对象，正样本为属于目标对象类别的对象。
该种子对象的对象信息可以称为种子对象信息，对象信息的具体介绍可以参照对应于步骤S204的对象信息的介绍。
其中，对象筛选模型用于预测类别未标注的对象属于目标对象类别的概率，这一处理过程可以看作是对类别未标注的对象进行标注的过程；该对象筛选模型可以是DeepFM(Deep Factorization Machine)模型或FM(Factorization Machine)模型。
本步骤中，利用置信度较高的种子对象构建对象筛选模型，当种子对象为正样本时，可以将种子对象信息作为训练样本的特征，将种子对象属于目标对象类别作为该训练样本的标签，组成训练样本，并利用该训练样本进行模型训练，得到对象筛选模型。
该实施例中，利用种子对象构建对象筛选模型，由于种子对象属于目标对象类别的置信度较高，因此，在对象筛选模型预测其他对象是否属于目标对象类别的时候，可以提高预测准确性。
在一些实施例中，上述获取候选对象的多个候选对象信息，提取得到各个候选对象信息对应的候选对象特征的步骤，具体可以包括：计算机设备将候选对象在多个维度的多个候选对象信息输入到对象筛选模型的特征提取层中进行特征提取，得到各个候选对象信息对应的候选对象特征。
上述融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率的步骤，具体可以包括：计算机设备将各个候选对象特征输入到对象筛选模型的分类层，以使分类层融合各个候选对象特征得到对象提取特征；获取分类层对对象提取特征进行对象类别概率识别后输出的候选对象属于目标对象类别的识别概率。
其中，对象筛选模型可以包括特征提取层和分类层。其中，特征提取层用于提取用于描述对象的深层次特征，特征提取层可以是卷积层，特征提取层提取到的特征可以称为对象特征；输入到特征提取层的对象信息是候选对象信息时，特征提取层提取到的特征可以称为候选对象特征，也可以称为深度特征。分类层主要是对特征提取层提取到的深层次的特征进行融合，得到对象提取特征，并根据对象提取特征确定对象所属的类别，确定该对象属于目标对象类别的识别概率，分类层可以是全连接层。
示例性地，计算机设备可以将对象A的性别信息、设备信息和设备所连接的网络信息输入到对象筛选模型的特征提取层中进行特征提取，进而得到各个候选对象信息对应的候选对象特征；接着，计算机设备将各个候选对象特征输入到对象筛选模型的分类层中，以使分类层对各候选对象特征进行融合得到对象A的对象提取特征，使分类层基于对象A的对象提取特征进行对象类别概率识别，得到对象A属于目标对象类别的识别概率。
本实施例中，针对类别未标注的候选对象，利用基于种子对象构建的对象筛选模型对这些候选对象是否属于目标对象类别进行预测，得到更多目标对象识别模型进行训练时所需的样本，提高目标对象识别模型的预测准确性。
在一些实施例中，上述对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合的步骤，具体包括：获取对应于不同候选对象的对象提取特征之间的特征距离；对于各个对象提取特征，基于各特征距离，确定位于各个对象提取特征所在的区域内的对象提取特征的数量，基于数量，得到各个对象提取特征的区域对象密度；基于各个对象提取特征的区域对象密度选取聚类中心，基于聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合；同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合。
其中，特征距离是不同候选对象的对象提取特征之间的距离，例如，候选对象A和B的对象提取特征[a 1，a 2，a 3，…，a n]和[b 1，b 2，b 3，…，b n]之间的特征距离，可以是
Figure PCTCN2022113686-appb-000001
对象提取特征所在的区域是对象提取特征周围的区域，对象提取特征周围的区域可以是 以对象提取特征为中心的预设半径内的区域。
本实施例计算各个候选对象的对象提取特征的区域对象密度，以计算对象A的对象提取特征的区域对象密度为例进行介绍：
计算机设备在得到对象A的对象提取特征与其他各个对象的对象提取特征之间的特征距离后，当确定对象C的对象提取特征与对象A的对象提取特征之间的特征距离小于该预设半径时，认为对象C的对象提取特征位于以对象A的对象提取特征为中心的预设半径的区域内；按照上述方式，可以确定其他对象的对象提取特征是否位于以对象A的对象提取特征为中心的预设半径的区域内；接着，当确定对象C的对象提取特征和对象D的对象提取特征均位于以对象A的对象提取特征为中心的预设半径的区域内时，可以确定位于以对象A的对象提取特征为中心的预设半径的区域内的对象提取特征的数量为2，将该数量作为对象A的对象提取特征的区域对象密度。
同样地，按照上述方式，计算机设备可以确定对象B、C、D和E的对象提取特征的区域对象密度；在进行聚类前，当对象A和对象E二者的对象提取特征的区域对象密度较大，可以将对象A的对象提取特征和对象E的对象提取特征作为聚类中心，并基于这两个聚类中心对对象B、C和D的对象提取特征进行聚类；当在得到的聚类结果中，对象A的对象提取特征、对象C的对象提取特征和对象D的对象提取特征被划分至一起并形成对应的子提取特征集合，对象B的对象提取特征和对象E的对象提取特征被划分至一起并形成对应的子提取特征集合，那么可以将对象A、对象C和对象D划分至一起并形成对应的子对象集合，将对象B和对象E划分至一起并形成对应的子对象集合。
该实施例中，基于不同候选对象的对象提取特征之间的特征距离，确定各个对象提取特征的区域对象密度，并按照区域对象密度选取聚类中心，可以提升将对象提取特征进行聚类的准确性，进而可以将相应的对象提取特征较为相似的候选对象分到同一类，而将相应的对象提取特征不那么相似的候选对象分到不同的类，提升将相应的候选对象进行分类的准确性，后续从每个类别的子对象集合中选取的代表对象，能够较为全面覆盖对象提取特征差异性较大的候选对象，减少选取的代表对象差异性较小、覆盖面较窄的情况，从而避免代表对象选取的片面性带来的目标对象识别模型训练效果差的问题，提升训练得到的目标对象识别模型识别对象类别的准确性。
在一些实施例中，上述基于各个对象提取特征的区域对象密度选取聚类中心，基于聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合的步骤，可以包括：在区域对象密度大于对象提取特征的区域对象密度的特征中，确定对象提取特征的邻近提取特征；将对象提取特征与邻近提取特征之间的特征距离，作为对象提取特征对应的目标距离；基于对象提取特征的区域对象密度以及对象提取特征对应的目标距离，选取聚类中心。
区域对象密度大于对象提取特征的区域对象密度的特征是：针对目标对象提取特征而言，区域对象密度大于该目标对象提取特征的区域对象密度的其他对象提取特征。邻近提取特征是：在区域对象密度大于该目标对象提取特征的区域对象密度的其他对象提取特征中，与该目标对象提取特征之间的特征距离最小的对象提取特征。
本申请实施例可以结合对象提取特征的区域对象密度和目标距离，选取聚类中心；其中，以计算对象C的对象提取特征的目标距离为例介绍：
当对象A的对象提取特征的区域对象密度和对象E的对象提取特征的区域对象密度大 于对象C的对象提取特征的区域对象密度，且对象A的对象提取特征与对象C的对象提取特征之间的特征距离|AC|小于对象E的对象提取特征与对象C的对象提取特征之间的特征距离|EC|时，计算机设备可以将对象A的对象提取特征作为与对象C的对象提取特征邻近的对象提取特征，并将对象A的对象提取特征与对象C的对象提取特征之间的特征距离|AC|作为对象C的目标距离。
计算机设备在按照上述方式得到各对象的对象提取特征的目标距离后，可以将具有较大区域对象密度和目标距离的对象提取特征作为聚类中心。
上述实施例中，基于与候选对象的对象提取特征邻近的对象提取特征确定对应的目标距离，并结合目标距离和区域对象密度选取聚类中心，可以提高对象提取特征聚类的准确性，后续从每个类别的子对象集合中选取的代表对象，能够较为全面覆盖对象提取特征差异性较大的候选对象，从而避免代表对象选取的片面性带来的目标对象识别模型训练效果差的问题，提升训练得到的目标对象识别模型识别对象类别的准确性进而提升对象识别的准确度。
在一些实施例中，上述基于聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合的步骤，可以包括：计算机设备获取待确定聚类类别的当前对象提取特征；获取区域对象密度大于当前对象提取特征的区域对象密度的聚类中心，作为当前对象提取特征对应的候选聚类中心；基于当前对象提取特征与候选聚类中心的距离，从候选中心集合中选取当前对象提取特征对应的邻近聚类中心，将当前对象提取特征加入到将邻近聚类中心所对应的子提取特征集合中。
其中，计算机设备可以根据当前对象提取特征对应的候选聚类中心，组成候选中心集合，例如，当前对象提取特征是对象F的对象提取特征，且聚类中心包括对象G的对象提取特征、对象I的对象提取特征、对象K的对象提取特征和对象J的对象提取特征，在聚类中心包括的各对象的对象提取特征中，对象I的对象提取特征的区域对象密度和对象K的对象提取特征的区域对象密度均大于对象F的对象提取特征的区域对象密度，对象G的对象提取特征的区域对象密度和对象J的对象提取特征的区域对象密度均小于对象F的对象提取特征的区域对象密度，那么可以将对象I的对象提取特征和对象K的对象提取特征作为与对象F的对象提取特征对应的候选聚类中心，并组成候选中心集合。
邻近聚类中心是：在当前对象提取特征对应的候选中心集合中，与当前对象提取特征之间的特征距离最小的聚类中心。
在对象I的对象提取特征和对象K的对象提取特征这两个候选聚类中心中，当对象F的对象提取特征与对象I的对象提取特征之间的特征距离小于对象F的对象提取特征与对象K的对象提取特之间的特征距离时，计算机设备可以将对象I的对象提取特征作为对象F的对象提取特征的邻近聚类中心，并将对象F的对象提取特征划分至对象I对应的子提取特征集合中。
该实施例中，针对待确定聚类类别的当前对象提取特征，基于区域对象密度确定该当前对象提取特征的候选聚类中心，并基于该当前对象提取特征与候选聚类中心的距离确定邻近的聚类中心，进而将当前对象提取特征划分至对应的子提取特征集合中，提高聚类划分的准确性。
在一些实施例中，本申请提供的对象识别方法还包括：基于候选对象属于目标对象类别的识别概率，确定各个候选对象对于目标对象识别模型的训练梯度变化的影响权重；基于各 个候选对象对于目标对象识别模型的训练梯度变化的影响权重，从候选对象集合中选取满足第二预设条件的候选对象，将满足第二预设条件的候选对象加入到训练样本集合中。
其中，目标对象识别模型可采用梯度下降的训练方式，每轮训练使用的训练样本将对模型的梯度变化产生影响。候选对象对于目标对象识别模型的训练梯度变化的影响权重，反应了候选对象对目标对象识别模型训练过程中的梯度变化的影响程度，训练梯度变化的影响权重与识别概率成正相关关系，也即，候选对象属于目标对象类别的识别概率越大，那么利用该候选对象对目标对象识别模型进行训练，梯度变化也越大。
第二预设条件包括训练梯度变化的影响权重大于影响权重阈值，或者影响权重排序在第二排序阈值之前的至少一个，以下介绍这两个第二预设条件：
(1)第二预设条件：训练梯度变化的影响权重大于影响权重阈值：
针对候选对象集合包括的全部候选对象，计算机设备基于各候选对象的识别概率确定各候选对象对对象识别模型的训练梯度变化的影响权重，将训练梯度变化的影响权重大于影响权重阈值的候选对象加入训练样本集合中。
(2)第二预设条件：训练梯度变化的影响权重排序在第二排序阈值之前：
针对候选对象集合包括的全部候选对象，计算机设备基于各候选对象的识别概率确定各候选对象对对象识别模型的训练梯度变化的影响权重；按照训练梯度变化的影响权重由大到小的顺序，对各候选对象进行排序，并将训练梯度变化的影响权重处于前几名的候选对象加入到训练样本集合中。
该实施例中，在从各子对象集合中选取代表对象的基础上，结合各候选对象属于目标对象类别的识别概率确定对应的训练梯度变化的影响权重，并选取出满足第二预设条件的候选对象作为训练样本，实现“最具代表性”和“最具识别力”的双策略进行训练样本的选取，选出的代表对象既能覆盖所有类别的候选对象，又具备每个类别的代表性特点，使得利用代表对象的候选对象信息训练得到的目标对象识别模型，能够挖掘出不同类别的差异性与每个类别的代表性特点，进而提高预测对象是否属于目标对象类别的准确度。
在一些实施例中，训练目标对象识别模型的步骤，包括：从训练样本集合中选取相似的第一训练对象以及第二训练对象，将第一训练对象以及第二训练对象组成对象组；将对象组中各个训练对象对应的训练对象信息输入到同一个特征提取层中，提取得到各个训练对象分别对应的训练对象特征；基于训练对象特征，得到对象组中的训练对象之间的对象相似度；基于对象相似度得到模型损失值；模型损失值与对象相似度成负相关关系；基于模型损失值对对象识别模型进行训练，得到目标对象识别模型。
其中，训练对象的对象信息可以称为训练对象信息，关于对象信息的介绍可以参照上述步骤S204的内容。模型损失值与对象相似度成负相关关系，例如，模型损失值越大，对象相似度越大，模型损失值越小，对象相似度越小。
当目标对象识别模型是RALM模型时，由于该RALM模型采用双塔结构，可以在训练样本集合中选取相似的第一训练对象和第二训练对象，并形成对象组；接着，将第一训练对象的对象信息和第二训练对象的对象信息输入同一特征提取层，以完成高维空间到低维空间的映射，并将该特征提取层提取到的特征(如Embedding特征)作为训练对象特征，基于第一训练对象的训练对象特征和第二训练对象的训练对象特征，得到第一训练对象和第二训练对象之间的对象相似度；基于对象相似度得到与对象相似度成负相关关系的RALM模型 的损失值；基于模型损失值对对象识别模型进行训练，完成对RALM模型的构建，得到目标对象识别模型。
上述实施例中，将相似的训练对象输入同一特征提取层中得到对应的训练对象特征，并根据训练对象之间的相似度得到的模型损失值进行模型训练，保证构建得到的目标对象识别模型的预测准确性。
在一些实施例中，得到训练对象对应的训练对象信息的步骤，包括：计算机设备获取对象信息类别；获取训练对象在对象信息类别对应的时间维度集合以及信息统计角度；获取各个信息统计时间维度中，基于信息统计角度统计得到的训练对象对应的对象信息统计值；对象信息统计值为对象信息类别对应的信息统计值；对时间维度集合中，各个信息统计时间维度对应的对象信息统计值进行信息聚合，将聚合得到对象信息作为训练对象对应的训练对象信息。
其中，用于描述对象的对象信息具有不同的类别，该类别可以称为对象信息类别，例如玩游戏的时长或者浏览商品的时长。信息统计时间维度是统计对象信息的时间维度，代表的是统计的信息所对应的时间长度；时间维度集合包括多个信息统计时间维度，例如一天、一周、三个月和六个月等。信息统计角度是统计的时间单位，如以天为单位，又如以一周为单位。
其中，训练对象对应的对象信息统计值是根据各个信息统计时间维度中，基于信息统计角度统计得到的；示例性地，若信息统计时间维度是一周，信息统计角度是一天，在得到一周内对象玩游戏的时长后，则可以确定每天该对象玩游戏的时长，将该对象每天玩游戏的时长作为对象信息统计值，且该对象信息统计值对应玩游戏时长这一对象信息类别。
同样地，针对其他信息统计时间维度，也可以确定该信息统计角度下的对象信息统计值，例如在该对象三个月玩游戏的时长这一信息统计时间维度下，可以确定对象每天玩游戏的时长。
上述方式中，计算机设备在得到训练对象在对象信息类别对应的时间维度集合以及信息统计角度后，确定各个信息统计时间维度中，基于信息统计角度统计得到的训练对象对应的对象信息统计值，将在不同时间维度下同一对象信息类别的对象信息统计值进行聚合，聚合方式可以是求平均值、求方差、求标准差或者求和等的至少一种，接着将聚合得到对象信息作为训练对象对应的训练对象信息以用于模型训练。
该实施例中，由于用于模型训练的训练对象信息是对不同信息统计时间维度下的同一对象信息类别进行聚合得到的，因此，可以避免将同类对象信息在不同时间维度下的值全部塞入模型引起的共线性，导致模型效果差的情况，提高模型预测效果，提高了对象识别的准确度，提高信息传递的有效性，避免将大量信息推送至不匹配的对象，节约计算机资源。
在一些实施例中，上述对时间维度集合中，各个信息统计时间维度对应的对象信息统计值进行信息聚合，将聚合得到对象信息作为训练对象对应的训练对象信息的步骤，可以包括：计算机设备基于时间维度集合中，各个信息统计时间维度对应的对象信息统计值得到综合信息统计值；确定各个对象信息统计值与综合信息统计值的统计值差异；基于统计值差异得到时间维度集合对应的统计值离散度，将统计值离散度作为训练对象对应的训练对象信息。
其中，综合信息统计值是对对象信息统计值进行统计得到的，例如多个对象信息统计值的平均值。统计值差异代表任一对象信息统计值与综合信息统计值之间的差异，例如可以是 差值或者是比例。统计值离散度表征各个对象信息统计值的离散程度，与统计值差异成正相关关系。基于统计值差异获取统计值离散度的方式可以是：计算机设备对各统计值差异进行相加，将相加结果作为得到统计值离散度。基于统计值差异获取统计值离散度的方式还可以是：对各统计值差异进行平方求和，并将平方求和的结果作为统计值离散度。基于统计值差异获取统计值离散度的方式还可以是：对各统计值差异进行平方求和，并对平方求和得到的结果进行开方根处理，将开方根处理得到的结果作为统计值离散度。
举个例子，假设有3个信息统计时间维度，则每个信息统计时间维度有对应的对象信息统计值；接着，确定各信息统计时间维度的对象信息统计值与综合信息统计值的统计值差异，得到3个统计值差异，对3个统计值差异进行平方求和，并对平方求和得到的结果进行开方根处理，将开方根处理得到的结果作为统计值离散度，将统计值离散度作为训练对象对应的训练对象信息。
示例性地，若在对象三个月玩游戏的时长这一信息统计时间维度下确定的每天玩游戏的时长是3小时，在对象一个月玩游戏的时长这一信息统计时间维度下确定的每天玩游戏的时长是3.5小时，在对象一周玩游戏的时长这一信息统计时间维度下确定的每天玩游戏的时长是2.5小时，那么可以将3小时、3.5小时和2.5小时的平均值3小时作为综合信息统计值，并分别确定3小时、3.5小时和2.5小时这几个时长与综合信息统计值的差值(即统计值差异)分别为0小时、0.5小时和0.5小时；根据0小时、0.5小时和0.5小时这几个差值，可以确定3小时、3.5小时和2.5小时呈现出的统计值离散度，将统计值离散度作为训练对象信息进行模型训练。
上述实施例中，由于离散程度能够代表一个对象在一段时间的行为变化的幅度，因此使得模型能够学习到对象行为变化的幅度对对象类别的影响，因此根据对象信息统计值的离散程度得到用于模型训练的训练对象信息，提高目标对象识别模型的预测准确性，。
本申请实施例还提供一种对象识别方法，可以应用于云技术、云安全、人工智能、智慧交通等场景；本实施例可以由计算机设备执行，包括图3所示的步骤：
步骤S302，获取目标对象类别对应的种子对象。
对象所属类别可以称为对象类别；目标对象类别是对象识别模型所要识别的对象类别。例如，在预测对象是否拥有房产的场景中，对象识别模型所要识别的对象类别是“处于拥有房产状态”和“不处于拥有房产状态”，此时，“处于拥有房产状态”和“不处于拥有房产状态”为目标对象类别；又例如，在预测对象是否拥有车辆的场景中，对象识别模型所要识别的对象类别是“处于拥有车辆状态”和“不处于拥有房产状态”，此时，“处于拥有车辆状态”和“不处于拥有房产状态”为目标对象类别。针对类别未标注的对象，可以通过人工标注或自动标注的方式对这些对象进行类别标注，若人工标注和自动标注的准确性较高，标注结果较为可信，此时可以将类别已标注的对象作为置信度较高的对象，该置信度较高的对象可以称为种子对象。
步骤S304，获取种子对象对应的种子对象信息，将种子对象信息作为训练样本中的训练特征，将目标对象类别作为训练样本中的标签，组成训练样本。
该种子对象的对象信息可以称为种子对象信息，例如种子对象的性别信息、候选对象的设备信息和候选对象的设备所连接的网络信息，将该种子对象信息作为描述该种子对象的特征，得到训练样本的训练特征，将上述步骤S302的目标对象作为训练样本的标签；将训练 样本的训练特征和训练标签组成训练样本。
步骤S306，基于训练样本进行模型训练，得到对象筛选模型。
该对象筛选模型主要用于预测类别未标注的对象属于目标对象类别的概率，也即对类别未标注的对象进行标注。该对象筛选模型可以是DeepFM模型或FM模型；对象筛选模型可以包括特征提取层和分类层，特征提取层主要是提取用于描述对象的深层次特征，分类层主要是对特征提取层提取到的深层次的特征进行融合得到对象提取特征，并根据对象提取特征确定该对象所属的类别，确定该对象属于目标对象类别的识别概率。
步骤S308，获取候选对象集合。
该候选对象集合包括的候选对象是未经标注的对象，也即所属类别未定的对象；计算机设备可以响应于样本选取指令时，从未经标注的多个候选对象中任选预设数量的候选对象，并形成候选对象集合。
步骤S310，将候选对象在多个维度的多个候选对象信息输入到对象筛选模型的特征提取层中进行特征提取，得到各个候选对象信息对应的候选对象特征。
候选对象特征是对候选对象信息进行特征提取得到的特征，可以分为多个维度；该特征提取层可以是卷积层，用于从候选对象的多个候选对象信息中分别提取出对应的特征，得到各个候选对象信息的候选对象特征。
步骤S312，将各个候选对象特征输入到对象筛选模型的分类层，以使分类层融合各个候选对象特征得到对象提取特征。
对象提取特征是对部分维度或者全部维度的候选对象特征进行融合后得到的特征；该分类层可以是全连接层，用于对各候选对象特征进行融合得到候选对象对应的对象提取特征。
步骤S314，获取分类层对对象提取特征进行对象类别概率识别后输出的候选对象属于目标对象类别的识别概率。
识别概率是对象属于目标对象类别的可能性大小，例如是概率。识别概率越大，对象属于目标对象类别的可能性越大，识别概率越大，对象属于目标对象类别的可能性越小。
上述分类层得到对象提取特征后，基于对对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率。
步骤S316，获取对应于不同候选对象的对象提取特征之间的特征距离。
特征距离是不同候选对象的对象提取特征之间的距离，例如，若计算对象A和B这两个对象提取特征[a 1，a 2，a 3，…，a n]和[b 1，b 2，b 3，…，b n]之间的距离，那么可以该特征距离可以是
Figure PCTCN2022113686-appb-000002
针对候选对象集合中的任两个候选对象，计算机设备可以按照上述公式确定任两个候选对象之间的特征距离。
步骤S318，对于各个对象提取特征，基于各特征距离，确定位于各个对象提取特征所在的区域内的对象提取特征的数量，基于数量，得到各个对象提取特征的区域对象密度。
对象提取特征所在的区域是对象提取特征周围的区域，对象提取特征周围的区域可以是以对象提取特征为中心的预设半径内的区域。示例性地，在计算对象A的对象提取特征的区域对象密度的场景中，计算机设备在得到对象A的对象提取特征与其他各个对象的对象提取特征之间的特征距离后，当确定对象C的对象提取特征与对象A的对象提取特征之间 的特征距离小于该预设半径时，可以认为对象C的对象提取特征位于以对象A的对象提取特征为中心的预设半径的区域内；按照上述方式，可以确定其他对象的对象提取特征是否位于以对象A的对象提取特征为中心的预设半径的区域内；接着，当确定对象C的对象提取特征和对象D的对象提取特征位于以对象A的对象提取特征为中心的预设半径的区域内时，位于以对象A的对象提取特征为中心的预设半径的区域内的对象提取特征的数量为2，将该数量作为对象A的对象提取特征的区域对象密度。
同样地，按照上述方式可以确定对象B、C、D和E的对象提取特征的区域对象密度。
步骤S320，在区域对象密度大于对象提取特征的区域对象密度的特征中，确定对象提取特征的邻近提取特征。
区域对象密度大于对象提取特征的区域对象密度的特征是指针对目标对象提取特征而言，区域对象密度大于该目标对象提取特征的区域对象密度的其他对象提取特征。
邻近提取特征是指区域对象密度大于该目标对象提取特征的区域对象密度的其他对象提取特征中，与该目标对象提取特征之间的特征距离最小的对象提取特征。
示例性地，当对象A的对象提取特征的区域对象密度和对象E的对象提取特征的区域对象密度大于对象C的对象提取特征的区域对象密度，且对象A的对象提取特征与对象C的对象提取特征之间的特征距离|AC|小于对象E的对象提取特征与对象C的对象提取特征之间的特征距离|EC|时，可以将对象A的对象提取特征，作为与对象C的对象提取特征邻近的对象提取特征，也即将对象A的对象提取特征作为对象C的邻近提取特征。
步骤S322，将对象提取特征与邻近提取特征之间的特征距离，作为对象提取特征对应的目标距离。
示例性地，以计算对象C的对象提取特征的目标距离为例介绍：计算机设备还可以将对象A的对象提取特征与对象C的对象提取特征之间的特征距离|AC|作为对象C的目标距离。
步骤S324，基于对象提取特征的区域对象密度以及对象提取特征对应的目标距离，选取聚类中心。
计算机设备在按照上述方式得到各对象的对象提取特征的目标距离后，可以将具有较大区域对象密度和目标距离的对象提取特征作为聚类中心。
步骤S326，获取待确定聚类类别的当前对象提取特征。
计算机设备在选取聚类中心后，可以确定候选对象集合中其他候选对象的对象提取特征所属的聚类类别，可以从其他候选对象的对象提取特征中选取任一个对象提取特征作为当前提取特征。
步骤S328，获取区域对象密度大于当前对象提取特征的区域对象密度的聚类中心，作为当前对象提取特征对应的候选聚类中心；当前对象提取特征对应的候选聚类中心组成候选中心集合；
当前对象提取特征对应的候选聚类中心组成候选中心集合，例如，当前对象提取特征是对象F的对象提取特征，且聚类中心包括对象G的对象提取特征、对象I的对象提取特征、对象K的对象提取特征和对象J的对象提取特征，在聚类中心包括的各对象的对象提取特征中，对象I的对象提取特征的区域对象密度和对象K的对象提取特征的区域对象密度均大于对象F的对象提取特征的区域对象密度，对象G的对象提取特征的区域对象密度和对 象J的对象提取特征的区域对象密度均小于对象F的对象提取特征的区域对象密度，那么可以将对象I的对象提取特征和对象K的对象提取特征作为与对象F的对象提取特征对应的候选聚类中心，并组成候选中心集合。
步骤S330，基于当前对象提取特征与候选聚类中心的距离，从候选中心集合中选取当前对象提取特征对应的邻近聚类中心，将当前对象提取特征加入到将邻近聚类中心所对应的子提取特征集合中。
邻近聚类中心是指当前对象提取特征对应的候选中心集合中，与当前对象提取特征之间的特征距离最小的聚类中心。
示例性地，在对象I的对象提取特征和对象K的对象提取特征这两个候选聚类中心中，若对象F的对象提取特征与对象I的对象提取特征之间的特征距离小于对象F的对象提取特征与对象K的对象提取特之间的特征距离，那么可以将对象I的对象提取特征作为对象F的对象提取特征的邻近聚类中心，并将对象F的对象提取特征划分至对象I对应的子提取特征集合中。
步骤S332，将子提取特征集合中的对象提取特征对应的候选对象组成子对象集合。
例如，对象A、E和C的对象提取特征被划分至同一个子提取特征集合中，计算机设备可以将对象A、E和C组成子对象集合。
步骤S334，从子对象集合中，获取识别概率满足第一预设条件的候选对象，作为子对象集合中对应的代表对象。
例如，计算机设备在得到包括对象A、C和E的子对象集合后，根据各对象的识别概率，将对象A作为代表对象；又如，计算机设备在得到包括对象B和D的子对象集合后，根据各对象的识别概率，将对象D作为代表对象。
步骤S336，将代表对象加入到训练样本集合中。
训练样本集合中的样本是用于进行模型训练的样本，例如在训练时，可以获取训练样本中对象对应的对象信息，作为对象特征，获取训练样本中对象对应的对象类别，作为对象标签。
步骤S338，基于候选对象属于目标对象类别的识别概率，确定各个候选对象对于目标对象识别模型的训练梯度变化的影响权重。
训练梯度变化的影响权重是对象识别模型训练过程中的梯度变化程度，训练梯度变化的影响权重与识别概率成正相关关系，也即，候选对象属于目标对象类别的识别概率越大，那么利用该候选对象对对象识别模型进行训练，梯度变化也越大。
计算机设备可以将候选对象属于目标对象类别的识别概率以及识别概率与训练梯度变化的影响权重成正相关关系，得到各候选对象对于目标对象识别模型的训练梯度变化的影响权重。
步骤S340，基于各个候选对象对于目标对象识别模型的训练梯度变化的影响权重，从候选对象集合中选取满足第二预设条件的候选对象，将满足第二预设条件的候选对象加入到训练样本集合中。
其中，第二预设条件包括训练梯度变化的影响权重大于影响权重阈值或者影响权重排序在第二排序阈值之前的至少一个，以下介绍这两个第二预设条件：
(1)第二预设条件：训练梯度变化的影响权重大于影响权重阈值：
针对候选对象集合包括的全部候选对象，基于各候选对象的识别概率确定各候选对象对对象识别模型的训练梯度变化的影响权重，将训练梯度变化的影响权重大于影响权重阈值的候选对象加入训练样本集合中。
(2)第二预设条件：影响权重排序在第二排序阈值之前：
针对候选对象集合包括的全部候选对象，基于各候选对象的识别概率确定各候选对象对对象识别模型的训练梯度变化的影响权重；按照训练梯度变化的影响权重由大到小的顺序，对各候选对象进行排序，并将训练梯度变化的影响权重处于前几名的候选对象加入到训练样本集合中。
步骤S342，利用训练样本集合进行模型训练，得到用于识别目标对象类别的目标对象识别模型。
计算机设备基于训练样本集合中各训练样本的对象特征以及对象标签进行有监督的训练，得到目标对象识别模型；该目标对象识别模型可以包括支持向量机(support vector machines,SVM)、卷积神经网络(Convolutional Neural Network,CNN)、长短时记忆网络(Long Short Term Memroy,LSTM)、或者Real-time Attention based Look-alike Model(RALM，基于实时注意的相似性模型)等模型。
上述实施例中，对候选对象信息进行特征提取得到对象提取特征，由于根据候选对象的对象提取特征对各候选对象进行聚类，因此可以保证同一子对象集合中的候选对象是较为相似的；接着，由于候选对象属于目标对象类别的识别概率是根据候选对象的对象提取特征得到的，因此，根据候选对象的识别概率分别从各子对象集合中选取得到的代表对象，可以最大程度地代表各代表对象所在的子对象集合中的其他候选对象；将代表对象作为模型训练的训练样本，提高目标对象识别模型的预测准确性，实现信息的有效传递；并且，利用种子对象构建对象筛选模型，由于种子对象属于目标对象类别的置信度较高，因此，可以提高对象筛选模型对其他对象属于目标对象进行预测的准确性；另外，融合“最具代表性”和“最具识别力”的这两个策略进行训练样本的选取，提高模型预测的准确性，提高了对象识别的准确度，提高信息传递的有效性，避免将大量信息推送至不匹配的对象，节约计算机资源。
为了更好地理解上述方法，以下结合图4至图6详细阐述一个本申请对象识别方法的应用实例；图4示出的技术架构图中，主要包括5个部分，其中，样本准备和持续学习框架这两个部分可以看成是离线数据准备。图5示出离线模型构建的过程，主要对应于图4的样本准备、持续学习框架、离线特征处理和离线模型训练这4个部分；图6示出在线调用过程，主要对应于图4的在线模型调用这个部分。在该应用实例中，可以将“处于拥有房产状态”和“未处于拥有房产状态”视为目标对象类别。
本应用实例可以由计算机设备执行，主要包括：通过基于人工标注和业务的逻辑的方式，获取种子对象群，生成种子对象画像特征，包括：对象基础属性(比如性别等)、设备基础属性(比如设备机型等)、网络连接属性(比如近1个月连接家庭Wi-Fi的次数)。然后，基于画像对异常对象进行过滤，比如：过滤使用特定应用程序的时长超过24小时的对象等。由于在大多数业务场景，能够通过人工标注和业务经验直接获取到的正负样本数量非常少(一般低于1万)，达不到训练模型的最低样本要求。本应用实例基于持续学习框架，在少量初始种子对象的基础上，在DeepFM模型上经过多轮迭代，结合“最具代表性策略”和“最具识别力策略”双策略融合方式，获得更多模型训练所需的正负样本。接着，为捕捉对 象在不同时间窗口的画像信息，结合时间维度，选择不同的“池化操作”，生成经过卷积处理后的特征向量。然后，将特征工程的结果，输入到多个机器学习模型进行Baseline训练，基于模型评价指标AUC初筛出最好的模型，对最佳模型进行最优参数寻优，获得最终的效果最佳的模型。最后，固化模型训练流程，定时离线训练、验证、告警、固化。在线工作包括：首先，定时从线上存储引擎，拉取对象最新的特征集合。接着，业务方设置特有特征计算逻辑，从线上实时日志，导出原数据，基于线上计算引擎，完成特征计算。然后，拼接特征、输入到模型、输出当前对象当前已购房的概率。
本应用实例可以提升对已购房对象识别的准确率，尤其是当种子对象量级较少的场景，持续学习框架和深度学习模型仍能保证较高的预测准确率和效果稳定性。本应用实例对于产品运营、广告投放等常见业务场景，提供了更为精准的对象分群方案，提高信息传递的有效性，避免将大量信息推送至不匹配的对象，节约计算机资源。
以下介绍更为具体的内容：
1)离线数据准备：
a)基于人工标注和业务经验，找出与业务强相关、数据分布正常、对象画像合理的正负训练样本；
b)基于持续学习框架，在a)的基础上，产出更多的高质量正负样本；
2)离线特征处理：构建训练样本的画像特征，并基于特征的垂直特性，结合时间维度、不同特征处理方法，产出高维特征向量。
3)离线模型训练：基于训练样本和特征向量，寻找分类效果佳且稳定的模型。
4)在线模型调用：基于离线训练模型和线上实时特征，对线上对象是否已购房的标签进行实时预测。
更具体地，上述离线数据准备包括样本准备和持续学习框架这两个部分；其中，样本准备这个部分主要包括以下步骤：
步骤S502，获取人工标注得到的种子对象。
具体来说，基于人工标注、业务逻辑，获取带有label信息(该label信息是指对象是否处于拥有房产状态的标签)的种子对象。基于规则粗召回一批种子对象，然后基于人工筛查的方式进行过滤，最后基于业务逻辑进行验证。
步骤S504，获取种子对象的基础画像。
其中，基础画像包括对象在应用程序的一些非隐私行为数据，比如是否安装某个特定应用程序、是否使用该应用程序提供的骚扰拦截功能、是否使用该应用程序提供的接听助理功能等。
步骤S506，获取异常对象类型评价指标在各个种子对象下的指标值。
在真实业务场景，会存在虚假对象、电脑操控手机的情况。为了剔除非真实对象对建模分析的影响，会基于业务经验设置异常对象类型评价指标；异常类型评价指标包括但不限于对象在应用程序的流量使用情况、流量产生的时间分布等；计算机设备可以根据每个种子对象在应用程序的流量使用情况和流量产生的时间分布，可以得到每个种子对象对应的指标值。
步骤S508，基于异常类型评价指标在各种子对象的指标值的分布，将指标值分布在正常区间的种子对象作为正常种子对象，将指标值分布在异常区间的种子对象作为异常种子对 象，并过滤异常种子对象。
其中，可以使用“拉依达准则”进行异常区间的判断，具体来说，假设一组检测数据只含有随机误差，对其进行计算处理得到标准偏差，按一定概率确定一个区间，认为凡超过这个区间的误差，就不属于随机误差而是粗大误差，含有该误差的数据应予以剔除。
步骤S510，存储正常种子对象。
具体来说，存储方式可以是离线存储，存储所用的服务器可以是HDFS(The Hadoop Distributed File System，Hadoop分布式文件系统)，便于后续流程的快速访问。
其中，持续学习框架是为了增加高质量的正负样本，该部分主要包括以下步骤：
步骤S512，利用上述少量的正常种子对象进行模型训练，得到对象筛选模型，此时得到的对象筛选模型可以记为M0；该对象筛选识别模型可以是DeepFM模型。
步骤S514，将未标注的样本，输入到对象筛选模型中进行特征提取并预测样本属于目标对象类别的概率，以对样本进行标注。
步骤S516，保存各样本属于目标对象类别的概率和对象筛选模型输出的各样本的对象提取特征。
步骤S518，从各样本中选择代表性较高的样本和识别力较高的样本；
其中，代表性较高的样本的选择过程主要包括：对各样本进行聚类得到各样本簇中，针对任一样本簇，将该任一样本簇中样本属于目标对象类别概率较高的样本作为代表性较高的样本；
识别力较高的样本的选择过程主要包括：基于最大梯度长度(Expected Gradient Length)的思想，在全部样本中，将样本属于目标对象类别概率较高的样本作为识别力较高的样本。
步骤S520，若对象筛选模型还需调整，则可以利用代表性较高的样本和识别力较高的样本对对象筛选模型进行微调，直至调整后的对象筛选模型达到性能阈值。
步骤S522，确定代表性较高的样本和识别力较高的样本的总数量。
步骤S524，判断总数量是否大于或等于数量阈值；若是则进入步骤S526，若否则返回步骤S514，继续迭代得到更多代表性较高的样本和识别力较高的样本。其中，数量阈值是根据进行针对目标对象识别模型的训练时所需的训练样本数量确定的。
上述离线特征处理这个部分，主要包括以下步骤：
步骤S526，构建对象的画像特征和业务垂直类型特征。
其中，基础画像特征主要是基于对象历史行为数据构建的，包括：对象基础属性、设备基础属性、网络连接属性等；对象基础属性但不限于：性别、籍贯、居住城市等；设备基础属性包括但不限于：手机分辨率、API_Level(接口等级)、CPU内核数等；网络连接属性包括但不限于：连接Wi-Fi的个数、每天连接Wi-Fi的最早时间等。
基于业务特性，构建业务垂直类型特征：垂直类型特征，包括对象对特定类型广告的点击率、转化率等。在一些场景，特定类型广告为：购房类广告、租房类广告、装修类广告等。
步骤S528，对不同时间维度下的同类画像特征或业务垂直类型特征进行聚合。其中，时间维度可以是近半年/近3个月/近1个月/近1周；针对同一类特征，可以将不同时间维度下的特征值进行聚合，得到聚合后的特征值；聚合可以是求和、中位数、标准差等多种方式。
步骤S530，对聚合后的特征进行归一化处理或者离散化处理，该处理包括对数值型特 征进行归一化以及对非数值型特征进行离散化；其中，归一化方法可以是高斯归一化；离散化处理包括以下方法：
i.One-Hot Encoding(独热编码)：例如对于对象性别等特征。
ii.Count Encoding(频数编码)：例如对于对象的WiFi POI(point of interest，兴趣点)特征，会用Count Encoding来标识对象和这个POI的兴趣程度。比如对象当周去了“美食-中国菜-粤菜”这个POI共3次。
iii.Category Embedding(类别嵌入)：若许多类目特征都存在较强的稀疏性，为了避免模型过拟合和提高模型稳定性，引入神经网络将高维稀疏分类变量转换为低维稠密的Embedding变量。
iv.NaN Embedding：对于特征的缺失值处理，可以使用“剔除”、“平均值填充”和“缺失标记”等方法，其中，将缺失值转为Embedding表达的方式，对模型的效果具有最大的正向收益。
v.Consolidation Encoding(合并编码)：某些类目变量下的多个取值，可以将其归纳成同一个信息。比如安卓手机的系统版本特征的多个取值里包括“4.2”、“4.4”和“5.0”三个，基于经验可以将这三个值归纳为“低版本安卓系统”。其中，Consolidation Encoding处理方式，比直接将“安卓系统版本”特征one-hot能带来更大的正向收益。
上述iii主要是将Category特征(类别特征)输入到DNN模型(Deep Neural Networks,深度神经网络模型)，训练Embedding特征，具体来说，可以包括如下内容：
①WiFi Trajectory Embedding(WiFi轨迹嵌入表示)。基于MST-CNN(Masked Self-Supervised Transformer-Convolution neural networ)深度学习网络，对对象的WiFi连接轨迹数据进行Embedding，捕捉对象Wi-Fi行为模式信息。
②App Traffic Embedding(APP流量嵌入表示)。基于List-Embedding方式，对对象使用不同类目App(application，应用程序)的流量使用行为序列进行Embedding提取，比如使用社交类型App的Traffic Embedding，获得低维稠密的对象行为特征。
步骤S532，将归一化后的数值型特征和离散化后的非数值型特征合并和存储。存储方式可以是离线存储在HDFS中，便于后续流程的快速访问。
其中，计算机设备还可以固化上述步骤S526至S532的特征处理逻辑，定时离线自动化计算，将离线计算得到的特征上传到线上存储引擎进行存储。
另外，在步骤S528之前，计算机设备还可以对对象的画像特征和业务垂直类型特征进行清洗、过滤、验证。由于特征通过多个策略逻辑通路产生，特征的数据质量难以保证，故需要进行特征数据质量监控。具体做法如下：
a)基于业务经验，制定特征质量标准，对计算好的特征进行清洗、过滤、验证。质量标准包括但不限于：每天使用特定App的时长低于16小时等；
b)按照预设策略，验证已有特征，对无效、异常特征进行剔除。比如对象使用某个App的时长不能大于24小时等；
c)拼接符合业务要求的特征，对不符合要求的特征进行缺失标记，最后实现入模向量拼接，拼接后的向量可能是(0.2,0.1,1,…,-1,…,0)。
上述离线模型训练这个部分，主要包括以下步骤：
步骤S534，对训练样本集合进行随机划分，得到训练集和测试集。
其中，可以按照样本所属的时间窗口进行划分，将时间较早的训练样本作为训练集(比如5月份的样本作为训练集)，时间较晚的训练样本作为验证集(比如6月份的样本作为验证集)；其中，训练集和验证集的比例可以是5:1。
步骤S536，基于预设参数，并行训练多个模型，将并行训练后的多个模型中预测效果最好的模型作为目标对象识别模型。并行训练所用的模型包括但不限于：支持向量机(support vector machines,SVM)、卷积神经网络(Convolutional Neural Network,CNN)、长短时记忆网络(Long Short Term Memroy,LSTM)、Real-time Attention based Look-alike Model(RALM)等。
步骤S538，对目标对象识别模型进行参数寻优以提升目标对象识别模型的预测效果。预测效果最好的评价指标可以是AUC；参数调优是指对选择模型的超参数进行网格寻优，以期待评价指标AUC能获得提升。
步骤S540，获取参数寻优后的目标对象识别模型，并在多份验证集上验证该参数寻优后的目标对象识别模型的预测效果以及预测效果的稳定性，判断参数寻优后的目标对象识别模型的预测效果是否达标，若是则进入步骤S542，若否则返回步骤S538。
其中，目标对象识别模型的预测效果的评价指标可以是AUC。AUC有如下优势：a)
AUC指标本身和模型预测score绝对值无关，只关注排序效果，更加贴近实际业务的需要；b)AUC的计算方法同时考虑了学习器对于正例和负例的分类能力，在样本不平衡的情况下，依然能够对分类器做出合理的评价。
步骤S542，对参数寻优后的目标对象识别模型进行固化；该固化后的目标对象识别模型可以用于广告推送，例如房产类广告或教育类广告的推送；基于TensorFlow(一个端到端开源机器学习平台)的Saver()方法固化训练好的模型，共产生4个文件：a)checkpoint文本文件，记录了模型文件的路径信息列表；b)model.ckpt.data，记录网络权重信息；c)model.ckpt.index.data和.index是二进制文件，保存模型中的变量权重信息。
其中，还可以固化上述步骤S534至步骤S542的模型训练流程，定时离线训练、验证、告警、固化。
更具体地，上述在线模型调用这个部分的目的是使用离线固化的模型完成线上实时预测，主要包括以下步骤：
步骤S602，获取上述参数寻优后的目标对象识别模型。
具体来说，本步骤包括如下部分：
a)目标对象识别模型在线下训练好后，基于TensorFlow的Saver()方法固化训练好的模型；
b)将目标对象识别模型的模型文件存储在云端；
c)客户端基于调用服务接口的方式，从云端拉取最新的模型文件。
步骤S604，获取待识别对象的画像特征和业务垂直类型特征。
其中，按照特征稳定性，可以将画像特征和业务垂直类型特征分为实时特征和离线特征；实时特征是指稳定性低于阈值的特征，该实时特征可以随着时间的变化而变化，实时特征主要反映对象当天在应用程序内产生的实时行为，粒度可以是小时级，例如是当天内对象使用特定应用程序的总时长；离线特征是指稳定性高于阈值的特征，该离线特征较不可能随着时间的变化而变化，例如对象的性别、对象使用的设备等。
计算机设备可以对待识别对象的数据进行定时计算得到的离线特征(也可以称为通用特征)，如对象的性别等基础属性特征、对象所用设备的品牌等设备属性特征和设备当天所连接的Wi-Fi个数等网络属性特征。
计算机设备可以基于Spark和TensorFlow计算引擎，基于不同的特征处理规则，对待识别对象的数据进行定时计算得到的离线特征；特征处理规则包括：One-Hot Encoding，Count Encoding，Category Embedding，NaN Embedding，Consolidation Encoding，WiFi Trajectory Embedding，App Traffic Embedding。
计算机设备还可以从对象实时日志中获取对象实时数据，并按照预设的特征处理规则对对象实时数据进行计算，得到对象实时特征，例如对象当天使用特定应用程序的总时长和当天首次打开时间等。
步骤S606，拼接待识别对象的画像特征和和业务垂直类型特征。
其中，拼接后得到的特征可以称为拼接特征，属于高维度的向量，例如(1,0,2,1.1,41,…,3,1,14)。
步骤S608，将拼接特征输入到目标对象识别模型中，以使目标对象识别模型基于拼接特征预测待识别对象属于目标对象类别的概率；
步骤S610，获取目标对象识别模型输出的待识别对象属于目标对象类别的概率。
步骤S612，若待识别对象属于目标对象类别的概率大于或等于概率阈值，则确定待识别对象属于目标对象类别。
本应用实例的目标对象识别模型是RALM模型，离线实验的评估RALM模型的预测效果主要包括：
1)数学指标评估：
a)AUC(Area under Curve)：AUC值越大，当前分类算法越有可能将正样本排在负样本前面，得到更好的分类结果；
2)线上实验评估：
a)基于A/B Test的线上流量，对模型的效果进行评估；
b)评估的指标有：广告点击率、对象实名登记率。
图7为使用不同模型进行对象房产状态预测的模型效果比对图，其中RALM模型是本应用实例所用的模型，可以看出：
a)从线下AUC效果来看，RALM方案相比基于人工强规则的方案或非深度学习的方案，平均提高29.24％；
b)从线上AUC效果来看，RALM方案相比基于人工强规则的方案或非深度学习的方案，平均提高28.18％；
图8为使用不同模型进行对象房产状态预测的业务效果比对图，其中RALM模型是本应用实例所用的模型，可以看出：
a)从广告点击率来看，RALM方案相比基于人工强规则的方案或非深度学习的方案，平均提高441.93％；
b)从对象实名登记率来看，RALM方案相比基于人工强规则的方案或非深度学习的方案，平均提高309.54％。
本应用实例具备很强的复用性，更换正例样本所属对象类别，即可预测相应场景的对 象类别，无需修改其他内容，节约计算机设备的处理时间，具体地，更换正例样本所属对象类别，比如“对象购车状态的群体识别”，然后服务端累计对应日志数据，最后使用相同的特征拼接、特征处理、模型训练的方法产出结果。
应该理解的是，虽然图2至图6的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，图2至图6中的至少一部分步骤可以包括多个步骤或者多个阶段，这些步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。
在一些实施例中，如图9所示，提供了一种对象识别装置，包括：
候选对象集合获取模块902，用于获取候选对象集合；候选对象集合包括多个候选对象；
概率识别模块904，用于获取候选对象在多个维度的多个候选对象信息，对各候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到候选对象对应的对象提取特征，基于对象提取特征进行对象类别概率识别，得到候选对象属于目标对象类别的识别概率；
聚类模块906，用于对候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合；
代表对象选取模块908，用于对于按各子提取特征集合组成的各个子对象集合，分别基于子对象集合中各个候选对象对应的识别概率，从子对象集合中选取得到代表对象。
在一些实施例中，对象提取特征是通过对象筛选模型进行特征提取得到的；装置还包括：对象筛选模型构建模块，用于获取目标对象类别对应的种子对象；获取种子对象对应的种子对象信息，将种子对象信息作为训练样本中的训练特征，将目标对象类别作为训练样本中的标签，组成训练样本；基于训练样本进行模型训练，得到对象筛选模型。
在一些实施例中，概率识别模块904，还用于将候选对象在多个维度的多个候选对象信息输入到对象筛选模型的特征提取层中进行特征提取，得到各个候选对象信息对应的候选对象特征；将各个候选对象特征输入到对象筛选模型的分类层，以使分类层融合各个候选对象特征得到对象提取特征；获取分类层对对象提取特征进行对象类别概率识别后输出的候选对象属于目标对象类别的识别概率。
在一些实施例中，聚类模块906，还用于获取对应于不同候选对象的对象提取特征之间的特征距离；对于各个对象提取特征，基于各特征距离，确定位于各个对象提取特征所在的区域内的对象提取特征的数量，基于数量，得到各个对象提取特征的区域对象密度；基于各个对象提取特征的区域对象密度选取聚类中心，基于聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合；同一个子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合。
在一些实施例中，聚类模块906，还用于在区域对象密度大于对象提取特征的区域对象密度的特征中，确定对象提取特征的邻近提取特征；将对象提取特征与邻近提取特征之间的特征距离，作为对象提取特征对应的目标距离；基于对象提取特征的区域对象密度以及对象 提取特征对应的目标距离，选取聚类中心。
在一些实施例中，聚类模块906，还用于在区域对象密度大于对象提取特征的区域对象密度的特征中，确定与对象提取特征间的特征距离最小的对象提取特征；将与对象提取特征间的特征距离最小的对象提取特征，作为对象提取特征的邻近提取特征。
在一些实施例中，聚类模块906，还用于获取待确定聚类类别的当前对象提取特征；获取区域对象密度大于当前对象提取特征的区域对象密度的聚类中心，作为当前对象提取特征对应的候选聚类中心；当前对象提取特征对应的候选聚类中心组成候选中心集合；基于当前对象提取特征与候选聚类中心的距离，从候选中心集合中选取当前对象提取特征对应的邻近聚类中心，将当前对象提取特征加入到将邻近聚类中心所对应的子提取特征集合中。
在一些实施例中，聚类模块906，还用于在候选中心集合中，确定与当前对象提取特征间的特征距离最小的聚类中心；将与当前对象提取特征间的特征距离最小的聚类中心，作为当前对象提取特征对应的邻近聚类中心。
在一些实施例中，装置还包括区域确定模块，用于，在将其中一个对象作为目标对象时，获取其他对象的对象提取特征与目标对象的对象提取特征间的特征距离；当其中一个其他对象对应的特征距离小于预设半径时，确定其中一个其他对象的对象提取特征位于目标对象的对象提取特征所在的区域内。
在一些实施例中，代表对象选取模块908，还用于对于按各子提取特征集合组成的各个子对象集合，分别从子对象集合中，获取识别概率满足第一预设条件的候选对象，作为子对象集合中对应的代表对象；第一预设条件包括识别概率大于概率阈值或者概率排序在第一排序阈值之前的至少一个。
在一些实施例中，装置还包括第二集合处理模块，用于基于候选对象属于目标对象类别的识别概率，确定各个候选对象对于目标对象识别模型的训练梯度变化的影响权重，训练梯度变化的影响权重与识别概率成正相关关系；基于各个候选对象对于目标对象识别模型的训练梯度变化的影响权重，从候选对象集合中选取满足第二预设条件的候选对象，将满足第二预设条件的候选对象加入到训练样本集合中；第二预设条件包括训练梯度变化的影响权重大于影响权重阈值或者影响权重排序在第二排序阈值之前的至少一个。
在一些实施例中，对象类别识别概率是通过目标对象识别模型进行的，装置还包括目标对象识别模型构建模块，用于从训练样本集合中选取相似的第一训练对象以及第二训练对象，将第一训练对象以及第二训练对象组成对象组；将对象组中各个训练对象对应的训练对象信息输入到同一个特征提取层中，提取得到各个训练对象分别对应的训练对象特征；基于训练对象特征，得到对象组中的训练对象之间的对象相似度；基于对象相似度得到模型损失值；模型损失值与对象相似度成负相关关系；基于模型损失值对对象识别模型进行训练，得到目标对象识别模型。
在一些实施例中，装置还包括对象信息处理模块，用于获取对象信息类别；获取训练对象在对象信息类别对应的时间维度集合以及信息统计角度，时间维度集合包括多个信息统计时间维度；获取各个信息统计时间维度中，基于信息统计角度统计得到的训练对象对应的对象信息统计值；对象信息统计值为对象信息类别对应的信息统计值；对时间维度集合中，各个信息统计时间维度对应的对象信息统计值进行信息聚合，将聚合得到对象信息作为训练对象对应的训练对象信息。
在一些实施例中，对象信息处理模块，还用于基于时间维度集合中，各个信息统计时间维度对应的对象信息统计值得到综合信息统计值；确定各个对象信息统计值与综合信息统计值的统计值差异；基于统计值差异得到时间维度集合对应的统计值离散度，将统计值离散度作为训练对象对应的训练对象信息；统计值离散度与统计值差异成正相关关系。
上述对象识别装置中，对候选对象信息进行特征提取得到对象提取特征，由于根据候选对象的对象提取特征对各候选对象进行聚类，因此可以将属于同一类别的对象划分在同一个子对象集合中；由于候选对象属于目标对象类别的识别概率是根据候选对象的对象提取特征得到的，因此，根据候选对象的识别概率分别从各子对象集合中选取得到的代表对象，可以从属于同一类别的子对象集合中选取出具有代表性的对象，提高了对象识别的准确度，提高信息传递的有效性，避免将大量信息推送至不匹配的对象，节约计算机资源。
关于对象识别装置的具体限定可以参见上文中对于对象识别方法的限定。上述对象识别装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图10所示。该计算机设备包括通过系统总线连接的处理器、存储器。该计算机设备还可以包括网络接口。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储对象识别的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种对象识别方法。
本领域技术人员可以理解，图10中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
在一些实施例中，提供了一种计算机设备，包括存储器、处理器，该存储器存储有计算机可读指令，该计算机可读指令被该处理器执行时，使得该处理器实现本申请实施例中的方法。
在一些实施例中，提供了一种非易失性的计算机可读存储介质，计算机可读存储介质存储有计算机可读指令，该计算机可读指令被处理器执行时，使得该处理器实现本申请实施例中的方法。
在一些实施例中，提供了一种计算机程序产品，包括计算机可读指令，该计算机可读指令存储在计算机可读存储介质中；计算机设备的处理器从计算机可读存储介质读取该计算机可读指令，处理器执行该计算机可读指令，使得该计算机设备执行本申请实施例中的方法。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，上述的计算机可读指令可存储于一非易失性的计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和易失性存储器中的至少一种。非易失性存储器可包括只读存储器 (Read-Only Memory，ROM)、磁带、软盘、闪存或光存储器等。易失性存储器可包括随机存取存储器(Random Access Memory，RAM)或外部高速缓冲存储器。作为说明而非局限，RAM可以是多种形式，比如静态随机存取存储器(Static Random Access Memory，SRAM)或动态随机存取存储器(Dynamic Random Access Memory，DRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上的实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (20)

  1. 一种对象识别方法，由计算机设备执行，所述方法包括：
    获取候选对象集合；所述候选对象集合包括多个候选对象；
    获取所述候选对象在多个维度的多个候选对象信息，对各所述候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到所述候选对象对应的对象提取特征，基于所述对象提取特征进行对象类别概率识别，得到所述候选对象属于目标对象类别的识别概率；
    对所述候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个所述子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合；
    对于按各所述子提取特征集合组成的各个子对象集合，分别基于所述子对象集合中各个所述候选对象对应的识别概率，从所述子对象集合中选取得到代表对象；选取得到的所述代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
  2. 根据权利要求1所述的方法，其特征在于，所述对象提取特征是通过对象筛选模型进行特征提取得到的，所述对象筛选模型的确定步骤包括：
    获取所述目标对象类别对应的种子对象；
    获取所述种子对象对应的种子对象信息，将所述种子对象信息作为训练样本中的训练特征，将所述目标对象类别作为所述训练样本中的标签，组成所述训练样本；
    基于所述训练样本进行模型训练，得到所述对象筛选模型。
  3. 根据权利要求2所述的方法，其特征在于，所述获取所述候选对象在多个维度的多个候选对象信息，对各所述候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，包括：
    将所述候选对象在多个维度的多个多个候选对象信息，输入到所述对象筛选模型的特征提取层中进行特征提取，得到各个候选对象信息对应的候选对象特征；
    所述融合各个候选对象特征得到所述候选对象对应的对象提取特征，基于所述对象提取特征进行对象类别概率识别，得到所述候选对象属于目标对象类别的识别概率，包括：
    将各个候选对象特征输入到所述对象筛选模型的分类层，以使所述分类层融合各个候选对象特征得到对象提取特征；
    获取所述分类层对所述对象提取特征进行对象类别概率识别后输出的所述候选对象属于目标对象类别的识别概率。
  4. 根据权利要求1所述的方法，其特征在于，所述对所述候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个所述子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合，包括：
    获取对应于不同候选对象的对象提取特征之间的特征距离；
    对于各个对象提取特征，基于各特征距离，确定位于所述各个对象提取特征所在的区域内的对象提取特征的数量，基于所述数量，得到所述各个对象提取特征的区域对象密度；
    基于所述各个对象提取特征的区域对象密度选取聚类中心，基于所述聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合；同一个所述子提取特征集合中的对象提取特征对应的候选对象组成一个所述子对象集合。
  5. 根据权利要求4所述的方法，其特征在于，所述基于所述各个对象提取特征的区域对象密度选取聚类中心，包括：
    在区域对象密度大于所述对象提取特征的区域对象密度的特征中，确定所述对象提取特征的邻近提取特征；
    将所述对象提取特征与所述邻近提取特征之间的特征距离，作为所述对象提取特征对应的目标距离；
    基于所述对象提取特征的区域对象密度以及所述对象提取特征对应的目标距离，选取聚类中心。
  6. 根据权利要求5所述的方法，其特征在于，所述在区域对象密度大于所述对象提取特征的区域对象密度的特征中，确定所述对象提取特征的邻近提取特征，包括：
    在区域对象密度大于所述对象提取特征的区域对象密度的特征中，确定与所述对象提取特征间的特征距离最小的对象提取特征；
    将与所述对象提取特征间的特征距离最小的对象提取特征，作为所述对象提取特征的邻近提取特征。
  7. 根据权利要求4所述的方法，其特征在于，所述基于所述聚类中心对对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，包括：
    获取待确定聚类类别的当前对象提取特征；
    获取区域对象密度大于所述当前对象提取特征的区域对象密度的聚类中心，作为所述当前对象提取特征对应的候选聚类中心；所述当前对象提取特征对应的候选聚类中心组成候选中心集合；
    基于所述当前对象提取特征与候选聚类中心的距离，从所述候选中心集合中选取所述当前对象提取特征对应的邻近聚类中心，将所述当前对象提取特征加入到将所述邻近聚类中心所对应的子提取特征集合中。
  8. 根据权利要求7所述的方法，其特征在于，所述从所述候选中心集合中选取所述当前对象提取特征对应的邻近聚类中心，包括：
    在所述候选中心集合中，确定与当前对象提取特征间的特征距离最小的聚类中心；
    将与所述当前对象提取特征间的特征距离最小的聚类中心，作为所述当前对象提取特征对应的邻近聚类中心。
  9. 根据权利要求4所述的方法，其特征在于，确定位于各个对象提取特征所在的区域内的对象提取特征的步骤包括：
    对于任一对象提取特征，获取与任一其它对象提取特征间的特征距离；
    当所述特征距离小于预设半径时，确定所述其它对象提取特征位于所述对象提取特征所在的区域内。
  10. 根据权利要求1所述的方法，其特征在于，所述对于按各所述子提取特征集合组成的各个子对象集合，分别基于所述子对象集合中各个所述候选对象对应的识别概率，从所述子对象集合中选取得到代表对象，包括：
    对于按各所述子提取特征集合组成的各个子对象集合，分别从所述子对象集合中，获取识别概率满足第一预设条件的候选对象，作为所述子对象集合中对应的代表对象；
    所述第一预设条件包括识别概率大于概率阈值或者概率排序在第一排序阈值之前的至 少一个。
  11. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    基于候选对象属于目标对象类别的识别概率，确定各个候选对象对于目标对象识别模型的训练梯度变化的影响权重，所述训练梯度变化的影响权重与所述识别概率成正相关关系；
    基于各个候选对象对于目标对象识别模型的训练梯度变化的影响权重，从所述候选对象集合中选取满足第二预设条件的候选对象，将满足第二预设条件的候选对象加入到训练样本集合中；
    所述第二预设条件包括训练梯度变化的影响权重大于影响权重阈值与训练梯度变化的影响权重排序在第二排序阈值之前中的至少一个。
  12. 根据权利要求1至11任一项所述的方法，其特征在于，所述目标对象识别模型的训练步骤，包括：
    将选取的所述代表对象加入目标对象识别模型的训练样本集合中；
    根据所述目标对象识别模型的训练样本集合，对目标对象识别模型进行模型训练；
    通过训练好的目标对象识别模型，预测对象是否属于目标对象类别。
  13. 根据权利要求12所述的方法，其特征在于，得到所述目标对象识别模型的步骤，包括：
    从训练样本集合中选取相似的第一训练对象以及第二训练对象，将所述第一训练对象以及所述第二训练对象组成对象组；
    将所述对象组中各个训练对象对应的训练对象信息输入到同一个特征提取层中，提取得到各个训练对象分别对应的训练对象特征；
    基于所述训练对象特征，得到所述对象组中的训练对象之间的对象相似度；
    基于所述对象相似度得到模型损失值；所述模型损失值与所述对象相似度成负相关关系；
    基于所述模型损失值对对象识别模型进行训练，得到所述目标对象识别模型。
  14. 根据权利要求13所述的方法，其特征在于，得到训练对象对应的训练对象信息的步骤，包括：
    获取对象信息类别；
    获取所述训练对象在所述对象信息类别对应的时间维度集合以及信息统计角度，所述时间维度集合包括多个信息统计时间维度；
    获取各个所述信息统计时间维度中，基于所述信息统计角度统计得到的所述训练对象对应的对象信息统计值；所述对象信息统计值为所述对象信息类别对应的信息统计值；
    对所述时间维度集合中，各个所述信息统计时间维度对应的对象信息统计值进行信息聚合，将聚合得到对象信息作为所述训练对象对应的训练对象信息。
  15. 根据权利要求14所述的方法，其特征在于，所述对所述时间维度集合中，各个所述信息统计时间维度对应的对象信息统计值进行信息聚合，将聚合得到对象信息作为所述训练对象对应的训练对象信息，包括：
    基于所述时间维度集合中，各个所述信息统计时间维度对应的对象信息统计值得到综合信息统计值；
    确定各个所述对象信息统计值与所述综合信息统计值的统计值差异；
    基于所述统计值差异得到所述时间维度集合对应的统计值离散度，将所述统计值离散度 作为所述训练对象对应的训练对象信息；所述统计值离散度与所述统计值差异成正相关关系。
  16. 一种对象识别装置，其特征在于，所述装置包括：
    候选对象集合获取模块，用于获取候选对象集合；所述候选对象集合包括多个候选对象；
    概率识别模块，用于获取所述候选对象在多个维度的多个候选对象信息，对各所述候选对象信息进行特征提取，得到各个候选对象信息对应的候选对象特征，融合各个候选对象特征得到所述候选对象对应的对象提取特征，基于所述对象提取特征进行对象类别概率识别，得到所述候选对象属于目标对象类别的识别概率；
    聚类模块，用于对所述候选对象对应的对象提取特征进行聚类，得到各个聚类类别对应的子提取特征集合，将同一个所述子提取特征集合中的对象提取特征对应的候选对象组成一个子对象集合；
    代表对象选取模块，用于对于按各所述子提取特征集合组成的各个子对象集合，分别基于所述子对象集合中各个所述候选对象对应的识别概率，从所述子对象集合中选取得到代表对象；选取得到的所述代表对象的候选对象信息用于训练目标对象识别模型，训练好的目标对象识别模型用于识别对象是否属于目标对象类别。
  17. 根据权利要求16所述的装置，其特征在于，所述对象提取特征是通过对象筛选模型进行特征提取得到的；所述装置还包括：
    对象筛选模型构建模块，用于获取所述目标对象类别对应的种子对象；获取所述种子对象对应的种子对象信息，将所述种子对象信息作为训练样本中的训练特征，将所述目标对象类别作为所述训练样本中的标签，组成所述训练样本；基于所述训练样本进行模型训练，得到所述对象筛选模型。
  18. 一种计算机设备，其特征在于，包括处理器、存储器；所述存储器用于存储计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述处理器实现如权利要求1至15任一项所述的方法。
  19. 一种非易失性的计算机可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时，使得所述处理器实现如权利要求1至15任一项所述的方法。
  20. 一种计算机程序产品，包括计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时，使得所述处理器实现如权利要求1至15任一项所述的方法。
PCT/CN2022/113686 2021-09-30 2022-08-19 对象识别方法、装置、设备、存储介质和程序产品 WO2023051085A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/335,569 US20230326185A1 (en) 2021-09-30 2023-06-15 Object recognition method and apparatus, device, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111161862.9A CN115937556A (zh) 2021-09-30 2021-09-30 对象识别方法、装置、设备和存储介质
CN202111161862.9 2021-09-30

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/335,569 Continuation US20230326185A1 (en) 2021-09-30 2023-06-15 Object recognition method and apparatus, device, and storage medium

Publications (1)

Publication Number Publication Date
WO2023051085A1 true WO2023051085A1 (zh) 2023-04-06

Family

ID=85781273

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/113686 WO2023051085A1 (zh) 2021-09-30 2022-08-19 对象识别方法、装置、设备、存储介质和程序产品

Country Status (3)

Country Link
US (1) US20230326185A1 (zh)
CN (1) CN115937556A (zh)
WO (1) WO2023051085A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116595978A (zh) * 2023-07-14 2023-08-15 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140140583A1 (en) * 2012-08-22 2014-05-22 Canon Kabushiki Kaisha Image recognition apparatus and image recognition method for identifying object
CN110335248A (zh) * 2019-05-31 2019-10-15 上海联影智能医疗科技有限公司 医学图像病灶检测方法、装置、计算机设备和存储介质
CN110968802A (zh) * 2019-12-04 2020-04-07 上海风秩科技有限公司 一种用户特征的分析方法、分析装置及可读存储介质
CN111368926A (zh) * 2020-03-06 2020-07-03 腾讯科技（深圳）有限公司 图像筛选方法、装置和计算机可读存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140140583A1 (en) * 2012-08-22 2014-05-22 Canon Kabushiki Kaisha Image recognition apparatus and image recognition method for identifying object
CN110335248A (zh) * 2019-05-31 2019-10-15 上海联影智能医疗科技有限公司 医学图像病灶检测方法、装置、计算机设备和存储介质
CN110968802A (zh) * 2019-12-04 2020-04-07 上海风秩科技有限公司 一种用户特征的分析方法、分析装置及可读存储介质
CN111368926A (zh) * 2020-03-06 2020-07-03 腾讯科技（深圳）有限公司 图像筛选方法、装置和计算机可读存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116595978A (zh) * 2023-07-14 2023-08-15 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备
CN116595978B (zh) * 2023-07-14 2023-11-14 腾讯科技（深圳）有限公司 对象类别识别方法、装置、存储介质及计算机设备

Also Published As

Publication number Publication date
CN115937556A (zh) 2023-04-07
US20230326185A1 (en) 2023-10-12

Similar Documents

Publication Publication Date Title
WO2020249125A1 (zh) 用于自动训练机器学习模型的方法和系统
Chiroma et al. Progress on artificial neural networks for big data analytics: a survey
Verdhan Supervised learning with python
KR20200039852A (ko) 기업 경영 현황 분석 예측 모델링을 위한 기계학습 알고리즘 제공 방법
WO2023051085A1 (zh) 对象识别方法、装置、设备、存储介质和程序产品
Shuai et al. Relationship analysis of short-term origin–destination prediction performance and spatiotemporal characteristics in urban rail transit
Wang et al. Metro traffic flow prediction via knowledge graph and spatiotemporal graph neural network
Liu et al. Explainable spatially explicit geospatial artificial intelligence in urban analytics
Zhou et al. Learning with self-attention for rental market spatial dynamics in the Atlanta metropolitan area
Xu et al. Tourist Attraction Recommendation Method and Data Management Based on Big Data Analysis
CN115796331A (zh) 基于多模态城市知识图谱的城市资源预测方法及系统
CN116861226A (zh) 一种数据处理的方法以及相关装置
Yang et al. Precise marketing strategy optimization of E-commerce platform based on KNN clustering
Su et al. [Retracted] Advertising Popularity Feature Collaborative Recommendation Algorithm Based on Attention‐LSTM Model
Shan [Retracted] Multisensor Cross‐Media Data Mining Method Assisted by Expert System
Feng et al. Implementation of Short Video Click-Through Rate Estimation Model Based on Cross-Media Collaborative Filtering Neural Network
CN117009883B (zh) 对象分类模型构建方法、对象分类方法、装置和设备
Reyes et al. Data Stream Processing Method for Clustering of Trajectories
Karamanou et al. Explainable Graph Neural Networks: An Application to Open Statistics Knowledge Graphs for Estimating House Prices
Nohuddin Predictive trend mining for social network analysis
Tran et al. Improving Traffic Load Prediction with Multi-modality: A Case Study of Brisbane
Tang et al. Low-Cost and High-Performance Abnormal Trajectory Detection Based on the GRU Model with Deep Spatiotemporal Sequence Analysis in Cloud Computing
Tampakis et al. Offline Trajectory Analytics
Rahim-Taleqani Next Location Prediction Model: A Geohashed Based Recurrent Neural Network
Ren Entity Identification Based on Human Mobility Data