Object classification method and related apparatus

Abstract

The application discloses an object classification method and related device, and can be applied to the fields of cloud technology, cloud security, artificial intelligence, block chain, map Internet of Vehicles, etc. The method comprises the following steps: performing screening processing on target object samples in a target object set based on at least one screening path, to obtain target object samples meeting predetermined quality conditions; performing filtering processing on the target object samples through segmentation testing, to obtain training object samples corresponding to test results meeting predetermined living state matching; constructing portrait features of the training object samples; performing information extraction processing on the portrait features according to information extraction strategies corresponding to data characteristics of the portrait features, to obtain feature vectors corresponding to the training object samples; and training a state recognition model based on deep learning by using the feature vectors, to obtain a trained state recognition model for recognizing the living state of an object to be recognized. The application can effectively improve the living state recognition effect of an object.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an object classification method and related device. BACKGROUND

[0002] Object classification is a task of identifying the category of an object, and the object classification can have a task of identifying the life state category of the object. The life state category can be, for example, a marriage state category or a parenting state category. In the related technical solutions, there are a data rule based on manual experience to identify the life state, and a data mining method based on non-deep learning to identify the life state. Due to the limitation of the data rule and the low feature mining ability of the non-deep learning data mining method, the classification effect of the life state category of the target object is often poor. SUMMARY

[0003] The embodiments of the present application provide an object classification method and related device, which can effectively improve the classification effect of the life state category of the object. The related device can include an object classification device, a computer readable storage medium, an electronic device, and a computer program product or a computer program.

[0004] The embodiments of the present application provide the following technical solutions:

[0005] According to an embodiment of the present application, an object classification method includes: performing screening processing on object samples included in a target object set based on at least one screening path to obtain target object samples meeting predetermined quality conditions; performing filtering processing on the target object samples through segmentation testing to obtain training object samples corresponding to a test result of a predetermined life state match; constructing portrait features of the training object samples, and determining a corresponding information extraction strategy according to data characteristics of the portrait features; performing information extraction processing on the portrait features according to the information extraction strategy to obtain a feature vector corresponding to the training object samples; and training a state recognition model based on deep learning by using the feature vector to obtain a trained state recognition model, and performing classification processing on a to-be-recognized object by using the trained state recognition model to obtain a life state category of the to-be-recognized object.

[0006] According to an embodiment of the present application, an object classification device comprises: a screening module configured to perform screening processing on object samples included in a target object set based on at least one screening path, to obtain target object samples meeting predetermined quality conditions; a filtering module configured to perform filtering processing on the target object samples through segmentation testing, to obtain training object samples corresponding to test results meeting predetermined living state matching conditions; a construction module configured to construct portrait features of the training object samples, and determine corresponding information extraction strategies according to data characteristics of the portrait features; an extraction module configured to perform information extraction processing on the portrait features according to the information extraction strategies, to obtain feature vectors corresponding to the training object samples; and an identification module configured to train a state recognition model based on deep learning by using the feature vectors, to obtain a trained state recognition model, and perform classification processing on a to-be-identified object by using the trained state recognition model, to obtain a living state category of the to-be-identified object.

[0007] In some embodiments of the present application, the screening module comprises: a first recall unit configured to determine a reference object sample from object samples in the target object set, the reference object sample being an object sample matching a business experience rule; a second recall unit configured to perform object expansion processing according to the reference object sample, to determine an expansion object sample matching the reference object sample from the target object set; and a sample determination unit configured to determine the reference object sample and the expansion object sample as target object samples meeting predetermined quality conditions.

[0008] In some embodiments of the present application, the expansion object sample comprises a union of a first object sample and a second object sample; the second recall unit comprises: a classification subunit configured to perform classification processing on a specific object sample by using a target classifier, to obtain the first object sample matching the reference object sample in the specific object sample, the specific object sample comprising part of the reference object sample and object samples other than the reference object sample in the target object set, the target classifier being obtained by training with part of the reference object sample as positive samples and the specific object sample as negative samples; and an expansion subunit configured to perform similar crowd expansion processing based on the reference object sample, to obtain the second object sample matching the reference object sample in the target object set.

[0009] In some embodiments of the present application, the device further comprises a third recall unit configured to: perform object identification processing on a to-be-identified object sample based on a deep neural network to obtain a third object sample matching the reference object sample in the to-be-identified object sample, the to-be-identified object sample comprising an object sample other than the reference object sample in the target object set; perform clustering processing on the to-be-identified object sample to determine a fourth object sample matching the reference object sample in the to-be-identified object sample; and determine a union of the reference object sample, the extended object sample, the third object sample, and the fourth object sample as a target object sample meeting a predetermined quality condition.

[0010] In some embodiments of the present application, the filtering module comprises: a grouping unit configured to group the target object samples and set a corresponding control group for each object group obtained by grouping to form an experimental group pair; a stratification unit configured to perform flow stratification processing on each experimental group pair as online traffic according to a predetermined number of experimental layers to obtain stratification results of each experimental group pair at the predetermined experimental layers, the stratification results of the predetermined experimental layers being orthogonal; a testing unit configured to perform stratification control testing on each experimental group pair based on the predetermined experimental layers according to the stratification results of each experimental group pair at the predetermined experimental layers to obtain a testing result of each experimental group pair; and a matching unit configured to determine an experimental group pair corresponding to a testing result matching a predetermined life state and to determine a target object sample in an object group of the determined experimental group pair as the training object sample.

[0011] In some embodiments of the present application, the testing unit is configured to perform stratification control testing on each experimental group pair based on the predetermined experimental layers according to the stratification results of each experimental group pair at the predetermined experimental layers, a testing duration of the stratification control testing being greater than a predetermined duration.

[0012] In some embodiments of the present application, the construction module comprises: a data acquisition unit configured to acquire log data of a history of the training object sample participating in target business generation; and a feature extraction unit configured to perform feature extraction processing on the training object sample based on the log data to generate the portrait feature composed of a feature of an object attribute, a feature of a device attribute, and a feature of a network attribute.

[0013] In some embodiments of the present application, the extraction module comprises: a policy determination unit configured to determine an information extraction policy corresponding to a data characteristic of each sub-image feature in the image feature; an information extraction unit configured to perform information extraction processing on each sub-image feature based on the information extraction policy corresponding to each sub-image feature to obtain an extracted sub-feature vector; and a vector generation unit configured to generate a feature vector corresponding to the training object sample based on the extracted sub-feature vector.

[0014] In some embodiments of the present application, the extracted sub-feature vector comprises a first sub-feature vector and a second sub-feature vector; the information extraction unit comprises: a feature screening sub-unit configured to perform screening processing on each sub-image feature to obtain a sub-image feature meeting a business logic and a sub-image feature not meeting the business logic; a first extraction sub-unit configured to perform information extraction processing on the sub-image feature meeting the business logic based on the corresponding information extraction policy to obtain a first sub-feature vector corresponding to the sub-image feature meeting the target business logic; and a second extraction sub-unit configured to perform missing value marking processing on the sub-image feature not meeting the business logic to obtain a missing value marking vector as a second sub-feature vector corresponding to the sub-image feature not meeting the business logic.

[0015] In some embodiments of the present application, the vector generation unit is configured to perform splicing processing on the extracted sub-feature vector to generate the feature vector corresponding to the training object sample.

[0016] In some embodiments of the present application, the identification module comprises: a model acquisition unit configured to acquire a state identification model constructed based on a deep cross network; a model calling unit configured to perform life state identification based on the feature vector by using the state identification model to obtain a predicted life state corresponding to the training object sample; and a model optimization unit configured to perform parameter adjustment on the state identification model according to the predicted life state to generate a trained state identification model.

[0017] In some embodiments of the present application, the identification module comprises a state identification unit configured to: acquire a feature vector of the to-be-identified object; perform state identification based on the feature vector of the to-be-identified object by using the trained state identification model to obtain a life state identification result; and determine a life state category of the to-be-identified object according to the life state identification result, wherein the life state category comprises a marriage state category or a parenting state category.

[0018] According to another embodiment of the present application, a computer readable storage medium has a computer program stored thereon, and when the computer program is executed by a processor of a computer, the computer program causes the computer to perform the method according to the embodiments of the present application.

[0019] According to another embodiment of the present application, an electronic device comprises a memory storing a computer program, and a processor reading the computer program stored in the memory to execute the method described in the embodiments of the present application.

[0020] According to another embodiment of the present application, a computer program product or computer program comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to make the computer device execute the method provided in various optional implementations described in the embodiments of the present application.

[0021] In the embodiments of the present application, object samples included in a target object set are screened based on at least one screening path to obtain target object samples meeting predetermined quality conditions; the target object samples are filtered through segmentation testing to obtain training object samples corresponding to test results matching predetermined life states; portrait features of the training object samples are constructed, and corresponding information extraction strategies are determined according to data characteristics of the portrait features; information extraction processing is performed on the portrait features according to the information extraction strategies to obtain feature vectors corresponding to the training object samples; and a state recognition model based on deep learning is trained using the feature vectors to obtain a trained state recognition model, which classifies a to-be-recognized object to obtain a life state category of the to-be-recognized object.

[0022] In this way, through the overall process of screening processing based on at least one screening path, filtering processing through segmentation testing, and information extraction processing on portrait features according to information extraction strategies corresponding to data characteristics of the portrait features, a high-quality feature vector reflecting life state information can be obtained, and a trained state recognition model with significant performance can be obtained by training a state recognition model based on deep learning using the feature vector, which can accurately recognize and classify a to-be-recognized object, effectively improving the classification effect of the life state category of the object. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1 A schematic diagram of a system to which the embodiments of the present application can be applied is shown.

[0025] Figure 2 A flow chart of an object classification method according to an embodiment of the present application is shown.

[0026] Figure 3 A flow chart of an object classification method according to another embodiment of the present application is shown.

[0027] Figure 4 A framework diagram of a state recognition model according to an embodiment of the present application is shown.

[0028] Figure 5 A terminal interface diagram of a process of recognizing a life state of an object under a scenario by applying an embodiment of the present application is shown.

[0029] Figure 6 A block diagram of an object classification apparatus according to an embodiment of the present application is shown.

[0030] Figure 7 A block diagram of an electronic device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0032] Figure 1 A schematic diagram of a system 100 to which embodiments of the present application can be applied is shown. As shown in the figure, the system 100 can include a server 101 and a terminal 102. The server 101 and the terminal 102 can be directly or indirectly connected through wireless communication, which is not specially limited in the present application. Figure 1

[0033] The server 101 can be a stand-alone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud technology services such as big data and artificial intelligence platforms.

[0034] ​In an embodiment of the present example, the server 101 is a cloud server, and the server 101 can provide an artificial intelligence cloud service. For example, the server 101 can provide an object life state recognition service based on artificial intelligence. In an embodiment, the server 101 and the terminal 102 can be nodes in a blockchain network, and the data security can be ensured based on the blockchain network.

[0035] The terminal 102 can be any device, including but not limited to a mobile phone, a computer, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, a VR / AR device, a smart watch, a computer, and the like.

[0036] In an embodiment of the present example, the server 101 can perform the following operations. The server 101 can perform a screening process on object samples included in a target object set based on at least one screening path, to obtain target object samples that meet predetermined quality conditions. The server 101 can perform a filtering process on the target object samples through segmentation testing, to obtain training object samples corresponding to test results that match a predetermined life state. The server 101 can construct portrait features of the training object samples, and determine a corresponding information extraction strategy according to data characteristics of the portrait features. The server 101 can perform information extraction processing on the portrait features according to the information extraction strategy, to obtain a feature vector corresponding to the training object samples. The server 101 can train a state recognition model based on deep learning using the feature vector, to obtain a trained state recognition model. The server 101 can perform classification processing on a to-be-recognized object using the trained state recognition model, to obtain a life state category of the to-be-recognized object.

[0037] In an example, the to-be-recognized object can be an object corresponding to the terminal 102, for example, a user using the terminal 102. It can be understood that the terminal 102 can be multiple, and the user can participate in a target service (such as logging in to a target platform such as a game platform or a map vehicle network platform to perform a specific service) on different terminals 102 through unique information (such as a unique mobile phone number or a unique account) of the user. Through a specific client on the terminal 102, log data generated by the to-be-recognized object participating in the target service can be collected according to the agreed legal agreement and the unique information. The server 101 can obtain the log data, and can extract descriptive features of the to-be-recognized object from the log data, and then construct portrait features of the to-be-recognized object. Then, the server 101 can perform information extraction processing on the portrait features according to an information extraction strategy corresponding to data characteristics of the portrait features, to obtain a feature vector corresponding to the to-be-recognized object.

[0038] Figure 2 A flowchart of an object classification method according to an embodiment of the present application is schematically shown. The execution subject of the object classification method can be any terminal, such as Figure 1 the server 101 or the terminal 102 shown.

[0039] As Figure 2As shown, the object classification method can include steps S210 to S250.

[0040] Step S210, screening the object samples included in the target object set based on at least one screening path to obtain target object samples meeting predetermined quality conditions; step S220, filtering the target object samples through segmentation testing to obtain training object samples corresponding to test results matching predetermined life state; step S230, constructing portrait features of the training object samples and determining corresponding information extraction strategies according to data characteristics of the portrait features; step S240, performing information extraction processing on the portrait features according to the information extraction strategies to obtain feature vectors corresponding to the training object samples; and step S250, training a state recognition model based on deep learning using the feature vectors to obtain a trained state recognition model for classifying processing of a to-be-recognized object to obtain a life state category of the to-be-recognized object.

[0041] In this way, based on steps S210 to S250, the overall process of screening processing through at least one screening path, filtering processing through segmentation testing, and information extraction processing on the portrait features according to the information extraction strategies corresponding to the data characteristics of the portrait features can obtain high-quality feature vectors reflecting life state information. Further, training a state recognition model based on deep learning using the feature vectors can obtain a trained state recognition model with significant performance. The trained state recognition model can accurately recognize and classify the life state of a to-be-recognized object, effectively improving the classification effect of the life state category of the object.

[0042] The specific processes of each step performed when recognizing the life state of an object are described below.

[0043] Step S210, screening the object samples included in the target object set based on at least one screening path to obtain target object samples meeting predetermined quality conditions.

[0044] The target object set is a set of object samples, and an object sample is a sample used to describe a user. Each object sample can correspond to a unique identifier, a unique mobile phone number, or a unique account number, etc.

[0045] The predetermined quality condition is a condition describing the quality of the object sample itself, which can be the accuracy of the object sample itself, etc. The predetermined quality condition can be represented by business experience rules, classification results of target classifiers, etc. For example, object samples matching business experience rules can be target object samples meeting predetermined quality conditions.

[0046] The screening path is a path of the screening processing, and the screening path can include at least one, for example, 1 or 3, and the like. Each screening path can correspond to a type of recall manner (for example, one type can be a recall manner based on a business experience rule matching, and another type can be a recall manner based on supervised learning), and at least one recall manner can be corresponded under each screening path.

[0047] The screening processing based on at least one screening path can accurately obtain the target object sample meeting the predetermined quality condition, and further improve the overall life state recognition result of the object.

[0048] In an embodiment, referring to Figure 3 , in step S210, the target object sample is screened based on at least one screening path to obtain a target object sample meeting a predetermined quality condition, including:

[0049] In step S211, a reference object sample is determined from the object sample in the target object set, and the reference object sample refers to an object sample matched with a business experience rule. In step S212, object expansion processing is performed according to the reference object sample to determine an expansion object sample matched with the reference object sample from the target object set. In step S213, the reference object sample and the expansion object sample are determined as the target object sample meeting the predetermined quality condition.

[0050] In this embodiment, the screening processing is performed based on at least two screening paths, further accurately obtaining the target object sample meeting the predetermined quality condition, and obtaining the order of magnitude of the target object sample meeting the demand.

[0051] Specifically, in the first screening path, the screening processing is performed based on the matching manner of the business experience rule, and the object sample matched with the business experience rule in the object sample in the target object set is obtained, that is, the reference object sample is obtained.

[0052] The reference object sample can be specified by a user, and the reference object sample specified by the user is obtained when recalled. For the object sample, the user can set a statistical logic based on business experience, and can also count key features to circumscribe the reference object sample through manual review, and then the reference object sample can be matched with the business experience rule, and the quality of the reference object sample is high.

[0053] In the second screening path, the object expansion processing is performed based on the reference object sample, and further recalls the expansion object sample matched with the reference object sample from the target object set. The object expansion processing according to the reference object sample can include: classifying a specific object sample by using a target classifier, and performing similar crowd expansion processing based on the reference object sample, and the like, to obtain the object sample matched with the reference object sample.

[0054] Finally, all the benchmark object samples and the extended object samples recalled under the two screening paths, i.e., the target object samples meeting the predetermined quality conditions, are collected.

[0055] In an embodiment, the extended object samples include the union of the first object samples and the second object samples; refer to Figure 3 In step S212, object extension processing is performed according to the benchmark object samples to determine the extended object samples matching the benchmark object samples from the target object set, including:

[0056] In step S2121, the target classifier is used to classify the specific object samples to obtain the first object samples matching the benchmark object samples in the specific object samples, the specific object samples including the partial benchmark object samples and the object samples other than the benchmark object samples in the target object set, the target classifier being trained with the partial benchmark object samples as positive samples and the specific object samples as negative samples; in step S2122, similar population extension processing is performed based on the benchmark object samples to obtain the second object samples matching the benchmark object samples in the object samples in the target object set.

[0057] In this embodiment, the second screening path includes two recall modes, and the applicant finds that the life state recognition effect can be further improved in this way.

[0058] In step S2121, i.e., the first recall mode, for example, the target object set is W, the set of the benchmark object samples is D, D is a part of W, the set of the object samples other than the benchmark object samples in W is K, a part of the benchmark object samples in W is extracted to obtain W1, the combination of W1 and K, i.e., the specific object samples U (i.e., the specific object samples including the partial benchmark object samples and the object samples other than the benchmark object samples in the target object set), is obtained, then the target classifier is trained with W1 as positive samples and U as negative samples, and further, the trained target classifier can calculate the probability that the object samples belong to the positive samples.

[0059] The target classifier is used to classify the specific object samples, and the probability that the object samples in the specific object samples belong to the positive samples (i.e., the probability of matching the benchmark object samples) can be obtained, and further, when the probability corresponding to a certain object sample is higher than a predetermined threshold, the certain object sample can be determined as the first object sample matching the benchmark object samples.

[0060] The second recall manner in step S2122 is to perform similar crowd expansion processing based on the reference object sample. In one example, the similar crowd expansion model trained based on the reference object sample is used to filter the object samples in the target object set other than the reference object sample based on the similar crowd expansion model to obtain the second object sample matched with the reference object sample. In another example, the community division based on the social network graph of the object samples in the target object set is performed to obtain the second object sample matched with the reference object sample (for example, the object sample located in the same community as the reference object sample).

[0061] In step S2121, the Spy algorithm can be used, and the target classifier can be a Bayesian classifier. The target classifier can be trained based on the EM algorithm with W1 as the positive sample and U as the negative sample. In step S2122, the similar crowd expansion method of the Lookalike algorithm can be used. In this way, the second filtering path based on supervised learning is implemented.

[0062] In one embodiment, referring to Figure 3 The method further includes step S213 of performing filtering processing based on the third filtering path. Step S213 can include the following steps:

[0063] In step S2131, the object recognition processing is performed on the to-be-identified object sample based on the deep neural network to obtain the third object sample matched with the reference object sample in the to-be-identified object sample. The to-be-identified object sample includes the object sample other than the reference object sample in the target object set. In step S2132, the clustering processing is performed on the to-be-identified object sample to determine the fourth object sample matched with the reference object sample in the to-be-identified object sample. In step S2133, the union of the reference object sample, the expanded object sample, the third object sample, and the fourth object sample is determined as the target object sample meeting the predetermined quality condition.

[0064] In this embodiment, the filtering processing is performed based on the third filtering path. In this embodiment, the third filtering path includes two recall manners. The applicant finds that in this way, the life state recognition effect can be further improved, and the order of magnitude of the target object sample meeting the demand can be further ensured.

[0065] In step S2131, the first recall manner, the object recognition processing is performed on the to-be-identified object sample based on the pre-trained deep neural network (DNN). The probability that the object sample in the to-be-identified sample belongs to a specific sample (i.e., the probability of matching with the reference object sample) can be determined. When the probability corresponding to a certain object sample is higher than a predetermined threshold, the certain object sample can be determined as the third object sample matched with the reference object sample.

[0066] The second recall manner, i.e., step S2132, is to perform clustering processing on the to-be-recognized object samples, i.e., to perform clustering processing on the to-be-recognized object samples based on a clustering model of unsupervised learning, and then, the fourth object sample (for example, the object sample located in the same object sample clustering cluster as the reference object sample) matching the reference object sample in the to-be-recognized object sample can be determined based on the clustering result. In one embodiment, the object samples in the to-be-recognized object samples are clustered based on a K-Means clustering model.

[0067] In step S220, the target object samples are filtered through the split test to obtain the training object samples corresponding to the test results matching the predetermined life state.

[0068] In the embodiment of the present example, the split test, also known as bucket test, can verify the quality of the target object samples in the online real traffic environment to obtain the test results of the target object samples. For example, the response of the target object samples to a specific advertisement is tested in the online real traffic environment to obtain the test results, i.e., the response of the target object samples to the specific advertisement.

[0069] The predetermined life state is, for example, married, married and not pregnant, etc., and the test result matching the predetermined life state is, for example, the test result matching the married state. According to the test result, the target object samples corresponding to the test results not matching the predetermined life state can be filtered, and the quality of the target object samples is further filtered, and the remaining target object samples are the training object samples corresponding to the test results matching the predetermined life state. Based on the training object samples, the state recognition model can be accurately trained, and the accuracy of the trained state recognition model is improved.

[0070] In one embodiment, referring to Figure 3 In step S220, the target object samples are filtered through the split test to obtain the training object samples corresponding to the test results matching the predetermined life state, including:

[0071] In step S221, the target object samples are grouped, and a corresponding control group is set for each object group obtained by grouping to form an experimental group pair; in step S222, each experimental group pair is respectively taken as online traffic to perform traffic stratification processing according to the number of predetermined experimental layers to obtain the stratification results of each experimental group pair in the predetermined experimental layers, and the stratification results between the predetermined experimental layers are orthogonal; in step S223, based on the stratification results of each experimental group pair in the predetermined experimental layers, each experimental group pair is respectively taken as online traffic to perform stratification control test based on the predetermined experimental layers to obtain the test results of each experimental group pair; in step S224, the experimental group pair corresponding to the test results matching the predetermined life state is determined, and the target object samples in the object group of the determined experimental group pair are taken as training object samples.

[0072] Grouping the target object samples, for example, 20 target object samples are divided into 5 object groups, each object group includes at least one target object sample, the number of groups can be set according to requirements. After grouping, set a corresponding control group for each object group, for example, set a control group A1 for the A object group, and then A and A1 form an experimental group pair.

[0073] The number of predetermined experimental layers is, for example, 2 layers, and each predetermined experimental layer performs an independent test experiment. Each experimental group pair is respectively treated as online traffic and subjected to flow layering processing according to the number of predetermined experimental layers, that is, the object samples in the object group and the control group in each experimental group pair are subjected to layering processing as user traffic, for example, the target object samples in the object group are divided into at least one sub-object group for the first layer of the predetermined experimental layer, and at least one sub-object group for the second layer of the predetermined experimental layer, and similarly, the object samples in the control group can be divided into at least one sub-control group for each predetermined experimental layer, and then the layering results of each experimental group pair in the predetermined experimental layer are obtained.

[0074] Further, the layering results between the predetermined experimental layers are orthogonal, for example, the target object samples in the sub-object group A corresponding to the first layer of the predetermined experimental layer are distributed in each sub-object group of the second layer of the predetermined experimental layer, and the target object samples in the sub-object group B corresponding to the first layer of the predetermined experimental layer are also distributed in each sub-object group of the second layer of the predetermined experimental layer.

[0075] Further, according to the layering results of each experimental group pair in the predetermined experimental layer, each experimental group pair can be subjected to layering control testing based on the test experiment set in the predetermined experimental layer, and the test results of each experimental group pair are obtained, for example, the response of the object samples in the control group and the object group to a specific advertisement. The control group can include at least one object sample, and the object sample in the control group can be an object sample that meets the quality requirements of the business, and the test result of the control group can be regarded as a standard test result.

[0076] Further, the experimental group pair corresponding to the test result matched with the predetermined life state can be determined, for example, the object group with a better response to a specific advertisement than the control group can be determined as the experimental group pair corresponding to the test result matched with the predetermined life state, and then the target object samples in the object group in the determined experimental group pair can be accurately used as training object samples.

[0077] In one embodiment, step S223, according to the layering results of each experimental group pair in the predetermined experimental layer, each experimental group pair is respectively subjected to layering control testing based on the predetermined experimental layer, and the test duration of the layering control testing is greater than the predetermined duration.

[0078] Each experimental group pair is pushed to the online experimental traffic pool to perform a stratified control test for a duration greater than a predetermined duration, for example, each experimental group pair is pushed to the online experimental traffic pool to continuously recover test results for more than a predetermined number of weeks (a number of weeks greater than a predetermined duration), so that the test for a duration greater than the predetermined duration can further exclude the abnormal influence of a specific time point on the experimental effect.

[0079] In step S230, portrait features of the training object samples are constructed, and an information extraction strategy is determined according to data characteristics of the portrait features.

[0080] The portrait features are features describing the training object samples, such as age, game duration, and device type. Through a legal agreement, log data generated by each training object sample in the history of participating in the target business can be obtained, and descriptive features of each training object sample can be extracted from the log data, and then the portrait features of each training object sample are constructed.

[0081] In one embodiment, in step S230, the portrait features of the training object samples are constructed, including: obtaining log data generated by each training object sample in the history of participating in the target business; and performing feature extraction processing on the training object samples based on the log data to obtain portrait features composed of features of all attributes in the log data.

[0082] In one embodiment, in step S230, the portrait features of the training object samples are constructed, including: obtaining log data generated by each training object sample in the history of participating in the target business; and performing feature extraction processing on the training object samples based on the log data to obtain portrait features composed of features of all attributes in the log data.

[0083] The log data generated by each training object sample in the history of participating in the target business can include log data generated when playing a game and log data generated when registering and logging into a business platform. The object attributes are attributes of the object samples, such as age. The device data are attributes of devices associated with the object samples, such as a model attribute of a device logged into by the object samples. The network attributes are attributes of networks associated with the object samples, such as a network point of interest (WiFi POI) attribute of the object samples.

[0084] From the log data, a descriptive feature (for example, 18 years old) corresponding to a feature keyword (for example, age) can be extracted based on the feature keyword. In the feature extraction processing, features of object attributes, features of device attributes, and features of network attributes are extracted, and the portrait features are composed of the features of the object attributes, the features of the device attributes, and the features of the network attributes. The applicant finds that the state recognition model can be efficiently and accurately trained.

[0085] In step S240, information extraction processing is performed on the image feature according to the information extraction strategy, to obtain the feature vector corresponding to the training object sample.

[0086] In the embodiment of the present example, the image feature can include at least one sub-image feature, each sub-image feature corresponding to its own data characteristic, and the data characteristic of each sub-image feature can be determined according to a preset strategy query table to determine the information extraction strategy corresponding to the data characteristic of each sub-image feature, and then the information extraction processing is performed on the image feature according to the information extraction strategy, to obtain the feature vector corresponding to the training object sample, and an accurate feature vector can be obtained, which can be used to accurately train the state recognition model.

[0087] In one embodiment, referring to Figure 3 , in step S240, information extraction processing is performed on the image feature according to the information extraction strategy, to obtain the feature vector corresponding to the training object sample, including:

[0088] In step S241, the information extraction strategy corresponding to the data characteristic of each sub-image feature in the image feature is determined; in step S242, information extraction processing is performed on each sub-image feature based on the information extraction strategy corresponding to each sub-image feature, to obtain the extracted sub-feature vector; and in step S243, the feature vector corresponding to the training object sample is generated based on the extracted sub-feature vector.

[0089] According to the preset strategy query table, the information extraction strategy corresponding to the data characteristic of each sub-image feature can be determined, and then based on the information extraction strategy corresponding to each sub-image feature, information extraction processing can be performed on each sub-image feature to obtain the extracted sub-feature vector.

[0090] For example, the information extraction strategy corresponding to the sub-image feature such as gender can be a One-Hot Encoding-based extraction strategy; the information extraction strategy corresponding to the sub-image feature such as a network interest point (WiFi POI) can be a Count Encoding-based extraction strategy, for example, a certain object sample went to the network interest point of "food-Chinese food-Cantonese food" for 3 times in a week; for some sub-image features with strong sparsity, the information extraction strategy corresponding thereto can be determined to be a Category Embedding-based extraction strategy, which can avoid overfitting of a state recognition model and improve model stability, and the Category Embedding-based introduction of a neural network can convert a high-dimensional sparse classification variable into a low-dimensional dense feature variable; for some sub-image features with multiple values, the information extraction strategy corresponding thereto can be determined to be a Category Embedding-based extraction strategy, which can integrate multiple values into the same information, for example, the multiple values of the system version feature of an Android mobile phone include "4.2", "4.4", and "5.0", which can be integrated into "low version Android system", so that greater positive benefits can be brought compared with directly One-Hot Encoding of the "Android system version" feature.

[0091] Thus, the sub-feature vector extracted according to the information extraction strategy corresponding to the data characteristics has excellent reliability, and based on the extracted sub-feature vector, a feature vector corresponding to a training object sample can be accurately generated. The feature vector corresponding to the training object sample can be obtained by concatenating or accumulating the extracted sub-feature vector.

[0092] In an embodiment, referring to Figure 3 , the extracted sub-feature vector includes a first sub-feature vector and a second sub-feature vector; in step S242, information extraction processing is performed on each sub-image feature based on the information extraction strategy corresponding to each sub-image feature, to obtain the extracted sub-feature vector, including:

[0093] In step S2421, each sub-image feature is subjected to screening processing to obtain a sub-image feature meeting the business logic and a sub-image feature not meeting the business logic; in step S2422, the sub-image feature meeting the business logic is subjected to information extraction processing based on the corresponding information extraction strategy, to obtain a first sub-feature vector corresponding to the sub-image feature meeting the target business logic; in step S2423, the sub-image feature not meeting the business logic is subjected to missing value marking processing, to obtain a missing value marking vector as a second sub-feature vector corresponding to the sub-image feature not meeting the business logic.

[0094] When screening each sub-image feature, the sub-image feature can be verified according to preset business logic to obtain a sub-image feature meeting the business logic and a sub-image feature not meeting the business logic. The preset business logic can be set according to actual business requirements.

[0095] The sub-image feature meeting the business logic is subjected to information extraction processing based on an information extraction strategy corresponding to a data characteristic of the sub-image feature to obtain a first sub-feature vector corresponding to the sub-image feature meeting the target business logic.

[0096] The sub-image feature not meeting the business logic is subjected to NaN embedding processing to obtain a NaN embedding vector as a second sub-feature vector corresponding to the sub-image feature not meeting the business logic. The applicant finds that, compared with methods such as “elimination” and “average value filling”, the NaN embedding processing method has the greatest positive effect on the state recognition model.

[0097] In an embodiment, based on the extracted sub-feature vectors, a feature vector corresponding to a training object sample is generated, including: splicing the extracted sub-feature vectors to generate the feature vector corresponding to the training object sample.

[0098] In this embodiment, based on splicing the extracted sub-feature vectors, the spliced vector is the generated feature vector corresponding to the training object sample, which can have excellent positive effects on the state recognition model.

[0099] In step S250, the state recognition model based on deep learning is trained using the feature vector to obtain a trained state recognition model, which is used to classify the to-be-recognized object to obtain a life state category of the to-be-recognized object.

[0100] In the embodiment of the present example, the feature vectors of all training object samples can be randomly divided into a training set and a test set, or can be divided according to the time windows to which the training object samples belong, with earlier time as the training set and later time as the validation set. In one example, the ratio of the training set to the validation set is 5:1.

[0101] Then, the feature vectors in the training set can be input into the state recognition model based on deep learning to perform parameter optimization on the state recognition model, and the validation set is used for verification. Finally, a parameter-optimized state recognition model, i.e., a trained state recognition model, is obtained, which can accurately recognize and classify the life state of the to-be-recognized object based on the trained state recognition model.

[0102] Further, after obtaining the post-practice state recognition model, the post-practice state recognition model can be tested online based on segmentation testing, the effect of the model is evaluated based on evaluation indexes such as advertisement click rate and advertisement conversion rate, and the parameters in the model are further adjusted according to the evaluation result. The recognition effect of the state recognition model can also be evaluated based on the evaluation index AUC, and it is determined whether the parameters are optimized according to the evaluation result.

[0103] Further, the trained post-practice state recognition model can be solidified. The trained post-practice state recognition model is solidified based on the TensorFlow Saver() method, and four files are generated: a checkpoint text file recording the path information list of the model file; a model.ckpt.data text file recording network weight information; and model.ckpt.index.data and.index are binary files saving variable weight information in the model. After solidifying the model, the client can use the solidified trained post-practice state recognition model based on the way of calling the service interface.

[0104] In an embodiment, the state recognition model based on deep learning is trained using the feature vector, comprising:

[0105] The state recognition model based on deep cross network is obtained, and the state recognition model is used to recognize the life state based on the feature vector to obtain the predicted life state corresponding to the training object sample. The state recognition model is adjusted according to the predicted life state to generate the trained post-practice state recognition model.

[0106] In this embodiment, the state recognition model based on deep learning is a state recognition model based on deep cross network (DCN), as shown in Figure 4 As shown in the state recognition model based on deep cross network (Deep&Cross Network), the feature vector is first input into the embedding and stacking layer 310, then the output of the embedding and stacking layer 310 is input into a cross network 320 and a deep network 330 parallel to the cross network, and then the outputs of the cross network 320 and the deep network 330 are combined by a combination output layer 340 to combine the outputs of the cross network 320 and the deep network 330, and then the prediction is performed to output the predicted life state P corresponding to the training object sample.

[0107] In the state recognition model based on deep cross network (Deep&Cross Network, DCN), the processing process of the feature vector is referred toFigure 4 A feature vector of a training sample can be composed of vector M and vector Z. In the embedding and stacking layer 310, the high-dimensional sparse vector Z (such as the vector obtained by one-hot encoding) can be embedded based on the feature embedding matrix Wz to generate a low-dimensional embedding vector Z1. Then, the embedding vector Z1 is superimposed with the dense vector M to generate a stacked vector x0. Then, the stacked vector x0 is input into the cross-network 320. In the cross-network 320, the stacked vector x0 can be processed by feature cross-processing from the first to the nth layer of the network, and finally the first output vector x is generated. n In the first layer, the formula can be used. Feature cross processing is performed to generate vector x1, and subsequent layers follow the formula. Feature cross processing is performed to generate vectors x for each layer. i , where w c b is the weight parameter. c This is the bias parameter. Simultaneously, the stacked vector x0 is input into the deep network 330. Within the deep network 330, the stacked vector x0 is sequentially propagated forward from layer 1 to layer n, ultimately generating the second output vector h. n In the first layer, the formula can be used. Propagate to generate vector h1, and subsequent layers follow the formula. Feature cross processing is performed to generate vectors h for each layer. i , where w h b is the weight parameter. h Here, x is the bias parameter, and ReLU is the activation function. Finally, in the combined output layer 340, the first output vector x is... n With the second output vector h n The combined vector x is obtained by concatenation s Then, based on the sigmoid function, the predicted living state corresponding to the training object sample is calculated. P represents the probability that a training sample belongs to a certain life state, and w s is the weight parameter, and b is the bias parameter.

[0108] Based on the predicted life state, the parameters in the state recognition model can be adjusted and updated. The adjusted parameters may include the feature embedding matrix Wz, w c b c w h b h w s Finally, based on the validation set, the state recognition model is validated. When the prediction accuracy of the state recognition model meets the requirements, the state recognition model with optimized parameters is obtained, which is the post-trained state recognition model.

[0109] The state recognition model based on the deep cross network (Deep&Cross Network, DCN) can effectively learn the bounded prediction cross features by explicitly applying feature cross on each layer without manual feature engineering or exhaustive search. Moreover, the network is simple and effective, the polynomial series of each layer is the highest and is determined by the layer depth, the network is composed of all cross terms, and the cross term coefficients are different. The cross network is memory efficient, saves memory, and is easy to implement. At the same time, compared with DNN, the parameter quantity of the cross entropy (LogLoss) is less than one order of magnitude. In the model, the ReLu function can be used as the activation function; and the Dropout can be added during the training process, that is, the neural network unit is temporarily discarded from the network according to a certain probability. Further, the state recognition model based on the deep cross network (Deep&Cross Network, DCN) can be trained to obtain a high-performance trained state recognition model for life state recognition, further improving the life state recognition effect.

[0110] In an embodiment, obtaining the trained state recognition model to perform life state recognition on a to-be-recognized object includes: obtaining a feature vector of the to-be-recognized object; using the trained state recognition model to perform state recognition based on the feature vector of the to-be-recognized object to obtain a life state recognition result; and determining a life state category of the to-be-recognized object according to the life state recognition result, wherein the life state category includes a marriage state category or a parenting state category.

[0111] In this embodiment, the trained state recognition model is used to perform state recognition based on the feature vector of the to-be-recognized object to obtain a life state recognition result, such as a married probability of 80%, an unmarried probability of 20%, or a parenting probability of 60%, an unparenting probability of 60%, etc. Further, the life state category of the to-be-recognized object can be determined based on the life state recognition result, and the life state category can include a marriage state category or a parenting state category, for example, the life state category of the to-be-recognized object can be determined as “married” according to the result “married probability of 80%” in the life state recognition result whose probability is greater than a predetermined threshold.

[0112] Obtaining the feature vector of the to-be-recognized object can include: obtaining log data generated by the to-be-recognized object participating in a target business through a predetermined legal protocol, extracting descriptive features of the to-be-recognized object from the log data, constructing portrait features to the to-be-recognized object, and then performing information extraction processing on the portrait features according to an information extraction strategy corresponding to the data characteristics of the portrait features to obtain a feature vector corresponding to the to-be-recognized object.

[0113] In an embodiment, obtaining the feature vector of the object to be identified can include: first, pulling log data generated by online real-time business of the object to be identified, and calculating general features (such as features of object attributes) of the object to be identified; calculating specific features (such as features of network attributes or device attributes) of the log data based on specific feature logic set by the business; then, splicing the general features and the specific features as real-time features on the line. Then, reading offline log data (for example, historical log data) of the object to be identified for feature extraction to obtain offline features, such as features of object attributes, network attributes, and device attributes. Then, based on computing engines such as Spark and TensorFlow, information extraction processing can be performed based on information extraction strategies corresponding to data characteristics of different features to generate feature vectors corresponding to each feature. Finally, based on the feature vectors corresponding to all real-time features on the line and offline features, a feature vector of the object to be identified can be generated, for example, splicing the feature vectors corresponding to all real-time features on the line and offline features into a long vector, and the long vector is the feature vector of the object to be identified.

[0114] According to the method described in the above embodiment, the following will be further described in detail by combining application scenarios. For example, Figure 5 A flowchart of applying the embodiment of the present application to identify the life state of an object in a scenario is shown. In this scenario, the meanings of related terms are the same as in the foregoing embodiments, and reference can be made to the descriptions in the foregoing embodiments. In this scenario, the foregoing embodiments of the present application are applied to identify the marriage state of an object.

[0115] Referring to Figure 5 The process of identifying the life state of an object in this scenario can include steps S410 to S440.

[0116] Step S410: Target object sample circumscription, that is, object samples included in a target object set are filtered based on at least one filtering path to obtain target object samples meeting predetermined quality conditions.

[0117] In step S410, the object samples included in the target object set are filtered based on at least one filtering path to obtain target object samples meeting predetermined quality conditions, which can include: step S411, determining a reference object sample from the object samples in the target object set, the reference object sample referring to an object sample matching a business experience rule; step S412, performing object expansion processing based on the reference object sample to determine an expansion object sample matching the reference object sample from the target object set; and step S413, determining the reference object sample and the expansion object sample as target object samples meeting predetermined quality conditions.

[0118] In the first screening path, the matching based on the business experience rule is used for screening processing to determine the object samples in the target object set that match the business experience rule, i.e., to determine the reference object samples. The reference object samples can be specified by the user, and the reference object samples specified by the user can be obtained during recall. Specifically, for the object samples, the user can set statistical logic based on business experience to circumscribe the reference object samples at step S4111, and can count the key features to determine the reference object samples through manual review at step S4112. Then, the reference object samples can match the business experience rule, and the quality of the reference object samples is high.

[0119] In the second screening path, the object expansion processing is performed based on the reference object samples to further recall the expansion object samples that match the reference object samples from the target object set. The object expansion processing based on the reference object samples can include: using the target classifier to perform classification processing on the specific object samples, and using the reference object samples to perform similar crowd expansion processing to obtain the object samples that match the reference object samples.

[0120] The second screening path includes two recall methods. Specifically, at step S412, the object expansion processing is performed based on the reference object samples to recall the expansion object samples that match the reference object samples from the target object set, including: at step S4121, identifying high-quality object samples based on the Spy algorithm. Specifically, the target classifier is used to perform classification processing on the specific object samples to obtain the first object samples that match the reference object samples in the specific object samples. The specific object samples include part of the reference object samples and the object samples other than the reference object samples in the target object set. The target classifier is trained with the part of the reference object samples as positive samples and the specific object samples as negative samples. At step S4122, high-quality object samples are identified based on the Lookalike algorithm. Specifically, the similar crowd expansion processing is performed based on the reference object samples to obtain the second object samples that match the reference object samples in the object samples in the target object set. Then, the expansion object samples include the union of the first object samples and the second object samples.

[0121] Step S4121 is the first recall method. For example, the target object set is W, the set of reference object samples is D, D is a part of W, the set of object samples other than the reference object samples in W is K, a part of the reference object samples in W is extracted to obtain W1, the combination of W1 and K is the specific object samples U (i.e., the specific object samples include part of the reference object samples and the object samples other than the reference object samples in the target object set), then the target classifier is trained with W1 as positive samples and U as negative samples. Then, the trained target classifier can calculate the probability that the object samples belong to the positive samples.

[0122] The target classifier is used to classify the specific object samples, so that the probability (i.e., the probability of matching the reference object sample) that the object samples in the specific object samples belong to the positive samples can be obtained. Then, when the probability corresponding to a certain object sample is higher than a predetermined threshold, it can be determined that the certain object sample is the first object sample that matches the reference object sample.

[0123] The step S4122 is a second recall method, in one example, the similar crowd expansion model trained based on the reference object sample is used to filter the object samples in the target object set other than the reference object sample based on the similar crowd expansion model, so that the second object sample matching the reference object sample is obtained. In another example, the social network graph of the object samples in the target object set is used to divide the communities, so that the second object sample matching the reference object sample (e.g., the object sample in the same community as the reference object sample) is obtained.

[0124] The step S4121 can be implemented based on the Spy algorithm, and the target classifier can be a Bayesian classifier. The target classifier can be trained based on the EM algorithm with W1 as the positive sample and U as the negative sample. The step S4122 is implemented based on the Lookalike algorithm similar crowd expansion method. The second filtering path based on supervised learning is implemented.

[0125] Further, the step S413 of filtering processing combined with the third filtering path is further included. The third filtering path includes two recall methods. Specifically, the step S413 can include:

[0126] The step S4131 is to identify high-quality object samples based on a deep neural network. Specifically, the object recognition processing is performed on the to-be-identified object samples based on the deep neural network, so that the third object sample matching the reference object sample in the to-be-identified object samples is obtained. The to-be-identified object samples include the object samples other than the reference object sample in the target object set. The step S4132 is to identify high-quality object samples based on a clustering model. Specifically, the clustering processing is performed on the to-be-identified object samples, so that the fourth object sample matching the reference object sample in the to-be-identified object samples is determined. The step S4133 is to determine the union of the reference object sample, the expanded object sample, the third object sample, and the fourth object sample as the target object sample meeting the predetermined quality condition.

[0127] The step S4131, i.e., the first recall mode, can perform object recognition processing on the to-be-recognized object sample based on a pre-trained deep neural network (DNN) model, can determine the probability (i.e., the matching probability with the reference object sample) that the object sample in the to-be-recognized sample belongs to a specific sample, and can determine that the certain object sample is the third object sample matched with the reference object sample when the probability corresponding to the certain object sample is higher than a predetermined threshold.

[0128] The step S4132, i.e., the second recall mode, performs clustering processing on the to-be-recognized object sample. Specifically, the clustering processing is performed on the to-be-recognized object sample based on a clustering model (K-Means clustering model) of unsupervised learning, and then the fourth object sample (e.g., the object sample located in the same object sample clustering cluster as the reference object sample) matched with the reference object sample in the to-be-recognized object sample can be determined based on the clustering result.

[0129] The step S420, sample verification and storage: i.e., filtering processing is performed on the target object sample through segmentation testing to obtain the training object sample corresponding to the test result matched with the predetermined life state. After obtaining the training object sample, the data of the training object sample can be stored in a distributed file system (HDFS). The predetermined life state can be a predetermined marriage state.

[0130] In the step S420, filtering processing is performed on the target object sample through segmentation testing to obtain the training object sample corresponding to the test result matched with the predetermined life state, including:

[0131] The step S421, sample grouping and setting of a control group: specifically, the target object sample is grouped, and a corresponding control group is set for each object group obtained by grouping to form an experimental group pair; the step S422, flow layering processing: specifically, each experimental group pair is respectively taken as online traffic for flow layering processing according to the number of predetermined experimental layers to obtain the layering result of each experimental group pair in the predetermined experimental layers, and the layering results between the predetermined experimental layers are orthogonal; the step S423, layering control testing and recovery of experimental results: specifically, according to the layering result of each experimental group pair in the predetermined experimental layers, each experimental group pair is respectively taken as online traffic for layering control testing based on the predetermined experimental layers to obtain the test result of each experimental group pair; the step S424, selection of training object samples: specifically, the experimental group pair corresponding to the test result matched with the predetermined life state is determined, and the target object sample in the object group of the determined experimental group pair is taken as the training object sample.

[0132] The target object samples are grouped, for example, 20 target object samples are divided into 5 object groups, each object group includes at least one target object sample, and the number of groups can be set according to requirements. After grouping, a corresponding control group is set for each object group, for example, a control group A1 is set for the A object group, and A and A1 form an experimental group pair.

[0133] The number of predetermined experimental layers is, for example, 2 layers, and each predetermined experimental layer performs an independent test experiment. Each experimental group pair is respectively treated as online traffic and subjected to flow layering processing according to the number of predetermined experimental layers, that is, the object samples in the object group and the control group in each experimental group pair are subjected to layering processing as user traffic, for example, the target object samples in the object group are divided into at least one sub-object group for the first predetermined experimental layer and at least one sub-object group for the second predetermined experimental layer, and similarly, the object samples in the control group can be divided into at least one sub-control group for each predetermined experimental layer, and then the layering results of each experimental group pair in the predetermined experimental layers are obtained.

[0134] Further, the layering results between the predetermined experimental layers are orthogonal, for example, the target object samples in the sub-object group A corresponding to the first predetermined experimental layer are distributed in each sub-object group of the second predetermined experimental layer, and the target object samples in the sub-object group B corresponding to the first predetermined experimental layer are also distributed in each sub-object group of the second predetermined experimental layer.

[0135] Further, according to the layering results of each experimental group pair in the predetermined experimental layers, each experimental group pair can be subjected to layering control testing based on the test experiments set in the predetermined experimental layers, and the test results of each experimental group pair are obtained, for example, the response of the object samples in the control group and the object group to a specific advertisement. The control group can include at least one object sample, and the object sample in the control group can be an object sample that meets the quality requirements of the business, and the test result of the control group can be regarded as a standard test result.

[0136] Further, the experimental group pair corresponding to the test result matched with the predetermined life state can be determined, for example, the object group with a better response to a specific advertisement than the control group can be determined as the experimental group pair corresponding to the test result matched with the predetermined life state, and then the target object samples in the object group in the determined experimental group pair can be accurately used as training object samples.

[0137] In one embodiment, step S423, according to the layering results of each experimental group pair in the predetermined experimental layers, each experimental group pair is subjected to layering control testing based on the predetermined experimental layers, and the test duration of the layering control testing is greater than the predetermined duration.

[0138] Each experimental group pair is pushed to the online experimental traffic pool to perform a stratified control test for a duration greater than a predetermined duration, for example, each experimental group pair is pushed to the online experimental traffic pool to continuously recover test results for more than a predetermined number of weeks (a number of weeks greater than a predetermined duration), so that the test for a duration greater than a predetermined duration can further exclude abnormal influences on experimental effects at a specific time point.

[0139] At step S430, feature construction and model training are performed, including: at step S431, feature construction, specifically, portrait features of the training object samples are constructed, and information extraction strategies corresponding to the data characteristics of the portrait features are determined; at step S432, information extraction (which can include screening processing, missing value marking, splicing processing, etc.), specifically, information extraction processing is performed on the portrait features according to the information extraction strategies to obtain feature vectors corresponding to the training object samples; and at step S433, model training, specifically, the state recognition model based on deep learning is trained using the feature vectors to obtain a trained state recognition model.

[0140] In step S431, the portrait features of the training object samples are constructed, including: obtaining log data generated by the training object samples in history when participating in the target business; and performing feature extraction processing on the training object samples based on the log data to generate portrait features composed of features of object attributes, features of device attributes, and features of network attributes.

[0141] The log data generated by the training object samples in history when participating in the target business can include log data generated when playing games and log data generated when registering and logging in to a business platform, etc. The object attributes are attributes of the object samples themselves, such as age and the like. The device data are attributes of devices associated with the object samples, such as a model attribute of a device logged in by the object samples, etc. The network attributes are attributes of networks associated with the object samples, such as a network point of interest (WiFi POI) attribute of the object samples, etc.

[0142] In step S432, the information extraction processing is performed on the portrait features according to the information extraction strategies to obtain the feature vectors corresponding to the training object samples, including: determining information extraction strategies corresponding to data characteristics of each sub-portrait feature in the portrait features; performing information extraction processing on each sub-portrait feature based on the information extraction strategies corresponding to each sub-portrait feature to obtain extracted sub-feature vectors; and generating the feature vectors corresponding to the training object samples based on the extracted sub-feature vectors.

[0143] The information extraction strategy corresponding to each sub-image feature is determined according to the preset strategy query table, and then, based on the information extraction strategy corresponding to each sub-image feature, the sub-image feature meeting the business logic is subjected to information extraction processing based on the corresponding information extraction strategy, to obtain a first sub-feature vector corresponding to the sub-image feature meeting the target business logic.

[0144] The information extraction strategy corresponding to each sub-image feature is determined according to the preset strategy query table, and then, based on the information extraction strategy corresponding to each sub-image feature, the sub-image feature meeting the business logic is subjected to information extraction processing based on the corresponding information extraction strategy, to obtain a first sub-feature vector corresponding to the sub-image feature meeting the target business logic.

[0145] For example, the information extraction strategy corresponding to a sub-image feature such as gender can be an extraction strategy based on One-Hot Encoding; the information extraction strategy corresponding to a sub-image feature such as a network interest point (WiFi POI) can be an extraction strategy based on Count Encoding, such as that a certain object sample went to the network interest point “Food-Chinese food-Cantonese food” for 3 times in a week; for a sub-image feature with strong sparsity, the information extraction strategy corresponding thereto can be determined to be an extraction strategy based on Category Embedding, which can avoid overfitting of a state recognition model and improve the stability of the model, and Category Embedding based on a neural network can convert a high-dimensional sparse classification variable into a low-dimensional dense feature variable; for a sub-image feature with multiple values, the information extraction strategy corresponding thereto can be determined to be an extraction strategy based on Category Embedding, and Category Embedding based on the multiple values can be summarized into the same information, for example, the multiple values of the system version feature of an Android mobile phone include “4.2”, “4.4” and “5.0”, which can be summarized as “low version Android system”, so that compared with directly One-Hot encoding the “Android system version” feature, greater positive benefits can be brought.

[0146] Further, based on the extracted sub-feature vector, a feature vector corresponding to the training object sample is generated, including: splicing the extracted sub-feature vector to generate a feature vector corresponding to the training object sample.

[0147] In step S433, the state recognition model based on deep learning is trained using the feature vector, including: obtaining a state recognition model constructed based on a deep cross network; using the state recognition model to recognize the life state based on the feature vector to obtain the predicted life state corresponding to the training object sample; adjusting the parameters of the state recognition model according to the predicted life state to generate the trained state recognition model.

[0148] The feature vectors of all training object samples can be randomly divided into a training set and a test set, or can be divided according to the time window to which the training object sample belongs, with the earlier time as the training set and the later time as the verification set. In an example, the ratio of the training set and the verification set is 5:1.

[0149] Then, the feature vectors in the training set are input into the state recognition model constructed based on the deep cross network, the parameters of the state recognition model are optimized, and the verification set is used for verification. Finally, the state recognition model after parameter optimization, i.e., the trained state recognition model, is obtained. Based on the trained state recognition model, the life state of the to-be-recognized object can be accurately recognized.

[0150] After obtaining the trained state recognition model, the trained state recognition model can be tested online based on segmentation testing, the effect of the model is evaluated based on evaluation indexes such as advertisement click rate and advertisement conversion rate, and the parameters in the model are further adjusted according to the evaluation result. The recognition effect of the state recognition model can also be evaluated based on the evaluation index AUC, and whether the parameters are optimized is determined according to the evaluation result.

[0151] Further, after the trained state recognition model is trained, the model is solidified. The Saver() method based on TensorFlow is used to solidify the trained state recognition model, and four files are generated: a checkpoint text file recording the path information list of the model file; a model.ckpt.data text file recording network weight information; and model.ckpt.index.data and.index are binary files saving variable weight information in the model. After the model is solidified, the client can use the solidified trained state recognition model based on the calling service interface.

[0152] In step S440, the trained state recognition model is called online to classify and process the to-be-recognized object, and the life state category of the to-be-recognized object is obtained.

[0153] In step S440, the trained state recognition model is used to classify the to-be-identified object, including: in step S441, obtaining the feature vector of the to-be-identified object; in step S442, identifying the marital status, specifically, using the trained state recognition model to generate state recognition based on the feature vector of the to-be-identified object to obtain a life state recognition result; and in step S443, determining the marital status, specifically, determining the life state category of the to-be-identified object according to the life state recognition result, and the life state category refers to the marital status category.

[0154] The feature vector of the to-be-identified object can include: obtaining log data generated by the to-be-identified object participating in the target service through a predetermined legal protocol, extracting descriptive features of the to-be-identified object from the log data, and then constructing portrait features of the to-be-identified object, and then performing information extraction processing on the portrait features according to information extraction strategies corresponding to the data characteristics of the portrait features to obtain the feature vector corresponding to the to-be-identified object.

[0155] Specifically, obtaining the feature vector of the to-be-identified object can include: first, pulling log data generated by the to-be-identified object in online real-time services to calculate general features (such as object attribute features) of the to-be-identified object; calculating specific features (such as network attribute or device attribute features) of the log data based on specific feature logic set by the service; then, splicing the general features and the specific features as real-time features on the line. Then, read the offline log data (such as historical log data) of the to-be-identified object to extract features, and obtain offline features such as object attribute, network attribute, and device attribute features. Then, based on Spark and TensorFlow computing engines, information extraction processing can be performed based on information extraction strategies corresponding to the data characteristics of different features to generate a feature vector corresponding to each feature. Finally, based on the feature vectors corresponding to all real-time features on the line and offline features, a long vector is obtained, which is the feature vector of the to-be-identified object.

[0156] The trained state recognition model is used to generate state recognition based on the feature vector of the to-be-identified object to obtain a life state recognition result, such as a married probability of 80%. Then, the marital status of the to-be-identified object can be determined based on the life state recognition result, for example, a married probability of 80% is greater than a certain threshold (such as 70%) to determine that the to-be-identified object is married.

[0157] In this way, based on steps S410 to S440, the overall process of screening processing through at least one screening path, filtering processing through segmentation testing, and information extraction processing of portrait features according to information extraction strategies corresponding to data characteristics of the portrait features, a high-quality feature vector reflecting life state information can be obtained. Further, using the feature vector to train a state recognition model based on deep learning can obtain a trained state recognition model with significant performance. The trained state recognition model can accurately recognize and classify the marriage state of the object to be recognized, effectively improving the classification effect of the life state category (marriage state category in this scenario) of the object.

[0158] To better implement the object classification method provided by the embodiments of the present application, the embodiments of the present application further provide an object classification device based on the above object classification method. The meanings of the terms are the same as those in the above object classification method, and specific implementation details can be referred to the description in the method embodiments. Figure 6 A block diagram of an object classification device according to an embodiment of the present application is shown.

[0159] As shown in Figure 6 The object classification device 500 can include a screening module 510, a filtering module 520, a construction module 530, an extraction module 540, and an identification module 550.

[0160] The screening module 510 can be configured to perform screening processing on object samples included in a target object set based on at least one screening path, to obtain target object samples meeting predetermined quality conditions. The filtering module 520 can be configured to perform filtering processing on the target object samples through segmentation testing, to obtain training object samples corresponding to test results of predetermined life state matching. The construction module 530 can be configured to construct portrait features of the training object samples, and determine corresponding information extraction strategies according to data characteristics of the portrait features. The extraction module 540 can perform information extraction processing on the portrait features according to the information extraction strategies, to obtain feature vectors corresponding to the training object samples. The identification module 550 can be configured to train a state recognition model based on deep learning using the feature vectors, to obtain a trained state recognition model, and perform classification processing on an object to be recognized using the trained state recognition model, to obtain a life state category of the object to be recognized.

[0161] In some embodiments of the present application, the screening module comprises: a first recall unit configured to determine a reference object sample from the object samples in the target object set, the reference object sample being an object sample matching a business experience rule; a second recall unit configured to perform object expansion processing based on the reference object sample to determine an expanded object sample matching the reference object sample from the target object set; and a sample determination unit configured to determine the reference object sample and the expanded object sample as target object samples meeting a predetermined quality condition.

[0162] In some embodiments of the present application, the expanded object sample comprises a union of a first object sample and a second object sample; and the second recall unit comprises: a classification subunit configured to perform classification processing on a specific object sample using a target classifier to obtain the first object sample matching the reference object sample in the specific object sample, the specific object sample comprising part of the reference object sample and an object sample other than the reference object sample in the target object set, the target classifier being trained using part of the reference object sample as positive samples and the specific object sample as negative samples; and an expansion subunit configured to perform similar crowd expansion processing based on the reference object sample to obtain the second object sample matching the reference object sample in the object samples in the target object set.

[0163] In some embodiments of the present application, the device further comprises a third recall unit configured to: perform object recognition processing on a to-be-recognized object sample based on a deep neural network to obtain a third object sample matching the reference object sample in the to-be-recognized object sample, the to-be-recognized object sample comprising an object sample other than the reference object sample in the target object set; perform clustering processing on the to-be-recognized object sample to determine a fourth object sample matching the reference object sample in the to-be-recognized object sample; and determine a union of the reference object sample, the expanded object sample, the third object sample, and the fourth object sample as target object samples meeting a predetermined quality condition.

[0164] In some embodiments of the present application, the filtering module comprises: a grouping unit configured to group the target object samples and set a corresponding control group for each object group obtained by grouping, to form an experimental group pair; a stratification unit configured to perform flow stratification processing on each of the experimental group pairs as online traffic according to a predetermined number of experimental layers, to obtain stratification results of each of the experimental group pairs at the predetermined experimental layers, the stratification results of the predetermined experimental layers being orthogonal; a testing unit configured to perform stratification control testing on each of the experimental group pairs based on the predetermined experimental layers according to the stratification results of each of the experimental group pairs at the predetermined experimental layers, to obtain a testing result of each of the experimental group pairs; and a matching unit configured to determine an experimental group pair corresponding to a testing result matched with a predetermined life state, and to take a target object sample in an object group of the determined experimental group pair as the training object sample.

[0165] In some embodiments of the present application, the testing unit is configured to perform stratification control testing on each of the experimental group pairs based on the predetermined experimental layers according to the stratification results of each of the experimental group pairs at the predetermined experimental layers, a testing duration of the stratification control testing being greater than a predetermined duration.

[0166] In some embodiments of the present application, the construction module comprises: a data acquisition unit configured to acquire log data generated by a target business in which a history of the training object sample is involved; and a feature extraction unit configured to perform feature extraction processing on the training object sample based on the log data, to generate the portrait feature composed of a feature of an object attribute, a feature of a device attribute, and a feature of a network attribute.

[0167] In some embodiments of the present application, the extraction module comprises: a strategy determination unit configured to determine an information extraction strategy corresponding to a data characteristic of each sub-portrait feature in the portrait feature; an information extraction unit configured to perform information extraction processing on each of the sub-portrait features based on the information extraction strategy corresponding to each of the sub-portrait features, to obtain an extracted sub-feature vector; and a vector generation unit configured to generate a feature vector corresponding to the training object sample based on the extracted sub-feature vector.

[0168] In some embodiments of the present application, the extracted sub-feature vectors include a first sub-feature vector and a second sub-feature vector; the information extraction unit includes: a feature screening sub-unit configured to perform screening processing on each of the sub-image features to obtain sub-image features that meet the business logic and sub-image features that do not meet the business logic; a first extraction sub-unit configured to perform information extraction processing on the sub-image features that meet the business logic based on the corresponding information extraction strategy to obtain a first sub-feature vector corresponding to the sub-image features that meet the target business logic; and a second extraction sub-unit configured to perform missing value marking processing on the sub-image features that do not meet the business logic to obtain a missing value marking vector as a second sub-feature vector corresponding to the sub-image features that do not meet the business logic.

[0169] In some embodiments of the present application, the vector generation unit is configured to perform splicing processing on the extracted sub-feature vectors to generate a feature vector corresponding to the training object sample.

[0170] In some embodiments of the present application, the identification module includes: a model acquisition unit configured to acquire a state identification model constructed based on a deep cross network; a model calling unit configured to perform life state identification based on the feature vector using the state identification model to obtain a predicted life state corresponding to the training object sample; and a model optimization unit configured to perform parameter adjustment on the state identification model according to the predicted life state to generate the trained state identification model.

[0171] In some embodiments of the present application, the identification module includes a state identification unit configured to: acquire a feature vector of the to-be-identified object; perform state identification based on the feature vector of the to-be-identified object using the trained state identification model to obtain a life state identification result; and determine a life state category of the to-be-identified object according to the life state identification result, wherein the life state category includes a marriage state category or a parenting state category.

[0172] In this way, based on the object classification device 500, the overall process of performing screening processing through at least one screening path, then performing filtering processing through segmentation testing, and performing information extraction processing on the image features according to the information extraction strategy corresponding to the data characteristics of the image features can be implemented, and a high-quality feature vector reflecting life state information can be obtained. Further, the feature vector is used to train the state identification model based on deep learning, and a trained state identification model with significant performance can be obtained. The trained state identification model can accurately identify the marriage state of the to-be-identified object, and effectively improves the marriage state identification effect of the object.

[0173] It should be noted that although several modules or units of the device for action execution are mentioned in the foregoing detailed description, such division is not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into modules or units embodied by multiple modules or units.

[0174] In addition, the embodiments of the present application also provide an electronic device, which can be a terminal or a server, for example. Figure 7 As shown in the figure, it shows the structure schematic diagram of the electronic device related to the embodiments of the present application, in particular:

[0175] The electronic device can include a processor 601 with one or more processing cores, a memory 602 with one or more computer readable storage media, a power supply 603, an input unit 604, and the like. Those skilled in the art can understand that the structure of the electronic device shown in the figure does not constitute a limitation on the electronic device, and can include more or less components than the figure, or combine certain components, or different component arrangements. Among them: Figure 7

[0176] The processor 601 is the control center of the electronic device, which connects various parts of the computer device through various interfaces and lines, executes the software programs and / or modules stored in the memory 602 and calls the data stored in the memory 602, performs various functions and processes data of the computer device, and thus overall detects the electronic device. Optionally, the processor 601 can include one or more processing cores; preferably, the processor 601 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user pages and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 601.

[0177] ​The memory 602 can be used to store software programs and modules, and the processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the computer device, etc. In addition, the memory 602 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 602 can also include a memory controller to provide access for the processor 601 to the memory 602.

[0178] The electronic device also includes a power supply 603 for powering the various components. Preferably, the power supply 603 can be logically connected to the processor 601 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 603 can also include one or more direct current or alternating current power supplies, a recharging system, a power supply fault detection circuit, a power supply converter or inverter, a power supply status indicator, and the like.

[0179] The electronic device can also include an input unit 604, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0180] Although not shown, the electronic device can also include a display unit and the like, which will not be described here. Specifically, in the present embodiment, the processor 601 in the electronic device will load one or more executable files corresponding to the processes of the computer programs into the memory 602 according to the following instructions, and run the computer programs stored in the memory 602 by the processor 601, so as to realize various functions, such as the processor 601 can execute:

[0181] The object samples included in the target object set are screened based on at least one screening path to obtain target object samples meeting predetermined quality conditions; the target object samples are filtered through segmentation testing to obtain training object samples corresponding to test results matching a predetermined life state; portrait features of the training object samples are constructed, and an information extraction strategy is determined according to data characteristics of the portrait features; information extraction processing is performed on the portrait features according to the information extraction strategy to obtain a feature vector corresponding to the training object samples; and the feature vector is used to train a state recognition model based on deep learning to obtain a trained state recognition model, which classifies a to-be-recognized object to obtain a life state category of the to-be-recognized object.

[0182] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a computer program, or by a computer program controlling related hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0183] To this end, the embodiments of the present application further provide a computer readable storage medium, which stores a computer program capable of being loaded by a processor to execute the steps in any of the methods provided by the embodiments of the present application.

[0184] The computer readable storage medium can include a Read Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, etc.

[0185] Since the computer program stored in the computer readable storage medium can execute the steps in any of the methods provided by the embodiments of the present application, the beneficial effects of the methods provided by the embodiments of the present application can be achieved, which are described in detail in the above embodiments and will not be described here.

[0186] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementation manners of the above embodiments of the present application.

[0187] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. The present application is intended to cover any variations, uses or adaptations of the application following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice in the art to which the application pertains.

[0188] It should be understood that the application is not limited to the embodiments described above and shown in the drawings, but various modifications and changes can be made without departing from the scope of the application.

Claims

1. A method of object classification, characterized by, The method comprises the following steps: The object samples included in the target object set are filtered based on at least one filtering path to obtain target object samples meeting predetermined quality conditions, which comprises: determining a reference object sample from the object samples in the target object set, the reference object sample being an object sample matching a business experience rule; performing object expansion processing based on the reference object sample to determine an expanded object sample matching the reference object sample from the target object set; and determining the reference object sample and the expanded object sample as target object samples meeting predetermined quality conditions; The target object samples are filtered through a segmentation test to obtain training object samples corresponding to test results matching a predetermined life state; An image feature of the training object samples is constructed, which comprises: obtaining log data of a training object sample participating in a target business in the past; performing feature extraction processing on the training object sample based on the log data to obtain an image feature composed of features of all attributes in the log data; An information extraction strategy is determined according to the data characteristics of the image feature; Information extraction processing is performed on the image feature according to the information extraction strategy to obtain a feature vector corresponding to the training object sample; The feature vector is used to train a state recognition model based on deep learning to obtain a trained state recognition model, which is used to classify a to-be-recognized object to obtain a life state category of the to-be-recognized object.

2. The method of claim 1, wherein, The expanded object sample comprises a union of a first object sample and a second object sample; the object expansion processing based on the reference object sample to determine an expanded object sample matching the reference object sample from the target object set comprises: A target classifier is used to classify a specific object sample to obtain the first object sample matching the reference object sample in the specific object sample, the specific object sample comprising part of the reference object samples and object samples other than the reference object samples in the target object set, and the target classifier is trained with part of the reference object samples as positive samples and the specific object sample as negative samples; Similar crowd expansion processing is performed based on the reference object sample to obtain the second object sample matching the reference object sample in the object samples in the target object set.

3. The method of claim 2, wherein, The method further comprises: A deep neural network is used to perform object recognition processing on a to-be-recognized object sample to obtain a third object sample matching the reference object sample in the to-be-recognized object sample, the to-be-recognized object sample comprising object samples other than the reference object samples in the target object set; The to-be-recognized object sample is clustered to determine a fourth object sample matching the reference object sample in the to-be-recognized object sample; A union of the reference object sample, the expanded object sample, the third object sample and the fourth object sample is determined as a target object sample meeting predetermined quality conditions.

4. The method of claim 1, wherein, The filtering processing of the target object samples through a segmentation test to obtain training object samples corresponding to test results matching a predetermined life state comprises: grouping the target object samples, and setting a corresponding control group for each object group obtained by grouping, to form an experimental group pair; performing flow layering processing on each of the experimental group pairs as online traffic according to a predetermined number of experimental layers, to obtain layering results of each of the experimental group pairs at the predetermined experimental layers, and the layering results between the predetermined experimental layers being orthogonal; performing layering control test on each of the experimental group pairs based on the predetermined experimental layers according to the layering results of each of the experimental group pairs at the predetermined experimental layers, to obtain a test result of each of the experimental group pairs; determining an experimental group pair corresponding to the test result matched with a predetermined life state, and taking a target object sample in an object group of the determined experimental group pair as the training object sample.

5. The method of claim 4, wherein, The performing layering control test on each of the experimental group pairs based on the predetermined experimental layers according to the layering results of each of the experimental group pairs at the predetermined experimental layers includes: The performing layering control test on each of the experimental group pairs based on the predetermined experimental layers according to the layering results of each of the experimental group pairs at the predetermined experimental layers includes that a test duration of the layering control test is greater than a predetermined duration.

6. The method of claim 1, wherein, The feature extraction processing on the training object sample based on the log data to obtain portrait features composed of features of all attributes in the log data includes: The feature extraction processing on the training object sample based on the log data to obtain portrait features composed of features of all attributes in the log data includes that the portrait features are composed of features of object attributes, features of device attributes, and features of network attributes.

7. The method of claim 1, wherein, The information extraction processing on the portrait features according to the information extraction strategy to obtain a feature vector corresponding to the training object sample includes: determining an information extraction strategy corresponding to a data characteristic of each sub-portrait feature in the portrait features; performing information extraction processing on each of the sub-portrait features based on the information extraction strategy corresponding to each of the sub-portrait features to obtain a sub-feature vector extracted; generating the feature vector corresponding to the training object sample based on the sub-feature vector extracted.

8. The method of claim 7, wherein, The sub-feature vector extracted includes a first sub-feature vector and a second sub-feature vector; the information extraction processing on each of the sub-portrait features based on the information extraction strategy corresponding to each of the sub-portrait features to obtain a sub-feature vector extracted includes: performing screening processing on each of the sub-portrait features to obtain a sub-portrait feature conforming to a business logic and a sub-portrait feature not conforming to the business logic; performing information extraction processing on the sub-portrait feature conforming to the business logic based on the corresponding information extraction strategy to obtain a first sub-feature vector corresponding to the sub-portrait feature conforming to the business logic; performing missing value marking processing on the sub-portrait feature not conforming to the business logic to obtain a missing value marking vector as a second sub-feature vector corresponding to the sub-portrait feature not conforming to the business logic.

9. The method of claim 7, wherein, The generating the feature vector corresponding to the training object sample based on the sub-feature vector extracted includes: performing splicing processing on the sub-feature vector extracted to generate the feature vector corresponding to the training object sample.

10. The method of claim 1, wherein, The training of the state recognition model based on deep learning by using the feature vector comprises: An state recognition model based on a deep cross network is obtained; The state recognition model is used to recognize the life state based on the feature vector, and a predicted life state corresponding to the training object sample is obtained; The state recognition model is adjusted in parameters according to the predicted life state, and a trained state recognition model is generated.

11. The method according to any one of claims 1 to 10, characterized in that, The trained state recognition model is used to classify the to-be-recognized object, comprising: The feature vector of the to-be-recognized object is obtained; The trained state recognition model is used to generate state recognition based on the feature vector of the to-be-recognized object, and a life state recognition result is obtained; The life state category of the to-be-recognized object is determined according to the life state recognition result, and the life state category comprises a marriage state category or a parenting state category.

12. An object classification apparatus characterized by comprising: Comprising: A screening module is configured to screen object samples included in a target object set based on at least one screening path to obtain target object samples meeting predetermined quality conditions; A filtering module is configured to filter the target object samples through segmentation testing to obtain training object samples corresponding to test results matching predetermined life states; A construction module is configured to construct portrait features of the training object samples and determine corresponding information extraction strategies according to data characteristics of the portrait features; An extraction module is configured to extract information from the portrait features according to the information extraction strategies to obtain feature vectors corresponding to the training object samples; An identification module is configured to train a state recognition model based on deep learning by using the feature vectors to obtain a trained state recognition model for classifying to-be-recognized objects to obtain life state categories of the to-be-recognized objects; The screening module comprises: A first recall unit is configured to determine a reference object sample from object samples in the target object set, wherein the reference object sample refers to an object sample matching a business experience rule; A second recall unit is configured to perform object expansion processing based on the reference object sample to determine an expansion object sample matching the reference object sample from the target object set; A sample determination unit is configured to determine the reference object sample and the expansion object sample as target object samples meeting predetermined quality conditions; The construction module comprises: A data acquisition unit is configured to acquire log data generated by a target business in the past of the training object sample; A feature extraction unit is configured to extract features of the training object sample based on the log data to obtain portrait features composed of all attributes in the log data.

13. The apparatus of claim 12, wherein, The expansion object sample comprises a union of a first object sample and a second object sample; The second recall unit comprises: The classification subunit is configured to perform classification processing on a specific object sample by using a target classifier, to obtain the first object sample matched with the reference object sample in the specific object sample, wherein the specific object sample includes part of the reference object sample and an object sample other than the reference object sample in the target object set, and the target classifier is trained by taking part of the reference object sample as a positive sample and the specific object sample as a negative sample. The expansion subunit is configured to perform similar population expansion processing based on the reference object sample, to obtain the second object sample matched with the reference object sample in the object sample in the target object set.

14. The apparatus of claim 13, wherein, The device further comprises: The third recall unit is configured to perform object identification processing on a to-be-identified object sample based on a deep neural network, to obtain the third object sample matched with the reference object sample in the to-be-identified object sample, wherein the to-be-identified object sample includes an object sample other than the reference object sample in the target object set; perform clustering processing on the to-be-identified object sample, to determine the fourth object sample matched with the reference object sample in the to-be-identified object sample; and determine the union of the reference object sample, the expanded object sample, the third object sample, and the fourth object sample as the target object sample meeting a predetermined quality condition.

15. The apparatus of claim 12, wherein, The filtering module comprises: The grouping unit is configured to group the target object samples, and set a corresponding control group for each object group obtained by grouping, to form an experimental group pair. The layering unit is configured to perform flow layering processing on each experimental group pair as online traffic according to a predetermined number of experimental layers, to obtain a layering result of each experimental group pair at the predetermined experimental layers, and the layering results between the predetermined experimental layers are orthogonal. The test unit is configured to perform layering control test on each experimental group pair based on the predetermined experimental layers according to the layering result of each experimental group pair at the predetermined experimental layers, to obtain a test result of each experimental group pair. The matching unit is configured to determine an experimental group pair corresponding to a test result matched with a predetermined life state, and take a target object sample in an object group in the determined experimental group pair as the training object sample.

16. The apparatus of claim 15, wherein, The test unit is configured to perform layering control test on each experimental group pair based on the predetermined experimental layers according to the layering result of each experimental group pair at the predetermined experimental layers, and a test duration of the layering control test is greater than a predetermined duration.

17. The apparatus of claim 12, wherein, The feature extraction unit is specifically configured to perform feature extraction processing on the training object sample based on the log data, to generate the portrait feature composed of a feature of an object attribute, a feature of a device attribute, and a feature of a network attribute.

18. The apparatus of claim 12, wherein, The extraction module comprises: The policy determination unit is configured to determine an information extraction policy corresponding to a data characteristic of each sub-portrait feature in the portrait feature. The information extraction unit is configured to perform information extraction processing on each sub-portrait feature based on the information extraction policy corresponding to each sub-portrait feature, to obtain an extracted sub-feature vector. The vector generation unit is configured to generate a feature vector corresponding to the training object sample based on the extracted sub-feature vectors.

19. The apparatus of claim 18, wherein, The extracted sub-feature vectors include a first sub-feature vector and a second sub-feature vector. The information extraction unit includes: The feature screening sub-unit is configured to perform screening processing on each of the sub-image features to obtain sub-image features that conform to the business logic and sub-image features that do not conform to the business logic. The first extraction sub-unit is configured to perform information extraction processing on the sub-image features that conform to the business logic based on a corresponding information extraction strategy to obtain a first sub-feature vector corresponding to the sub-image features that conform to the business logic. The second extraction sub-unit is configured to perform missing value marking processing on the sub-image features that do not conform to the business logic to obtain a missing value marking vector as a second sub-feature vector corresponding to the sub-image features that do not conform to the business logic.

20. The apparatus of claim 18, wherein, The vector generation unit is configured to perform splicing processing on the extracted sub-feature vectors to generate a feature vector corresponding to the training object sample.

21. The apparatus of claim 12, wherein, The recognition module includes: The model acquisition unit is configured to acquire a state recognition model constructed based on a deep cross network. The model calling unit is configured to use the state recognition model to perform life state recognition based on the feature vector to obtain a predicted life state corresponding to the training object sample. The model optimization unit is configured to perform parameter adjustment on the state recognition model according to the predicted life state to generate a trained state recognition model.

22. The apparatus of any one of claims 12 to 21, wherein, The recognition module includes: The state recognition unit is configured to acquire a feature vector of the to-be-recognized object, use the trained state recognition model to perform life state recognition based on the feature vector of the to-be-recognized object to obtain a life state recognition result, and determine a life state category of the to-be-recognized object according to the life state recognition result, where the life state category includes a marriage state category or a parenting state category.

23. A computer-readable storage medium, characterized in that, The computer program is stored on the computer and causes the computer to execute the method of any one of claims 1 to 11 when executed by a processor of the computer.

24. An electronic device, comprising: The computer program is stored on the computer and causes the computer to execute the method of any one of claims 1 to 11 when executed by a processor of the computer. The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 11. The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 11.

25. A computer program product, characterised in that, ​

Images