Information recognition method, storage medium, and electronic device

Abstract

The application discloses an information identification method, a storage medium and an electronic device, and relates to big data, database, account management and other technologies in a cloud technology scenario. The method comprises the following steps: obtaining account behavior information corresponding to a target account to be identified; constructing a basic portrait feature of the target account based on historical behavior information in the account behavior information, and constructing a business vertical feature of the target account based on business behavior information in the account behavior information; performing aggregation processing on the basic portrait feature and the business vertical feature in at least two time dimensions to obtain at least two aggregated portrait features in different time dimensions; in the case that a target feature is obtained based on the aggregated portrait feature, inputting the target feature into an educational background identification model; and obtaining an output result of the educational background identification model, wherein the output result comprises account educational background information associated with the target account. The application solves the technical problem of low information identification accuracy.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to an information identification method, a storage medium, and an electronic device. Background Technology

[0002] In information recognition scenarios with complex behavioral characteristics, ensuring the accuracy of information recognition requires guaranteeing both the quantity and quality of reference information. However, with the increasing awareness of information security and the growing emphasis on account privacy, information silos can easily form, leading to lower accuracy in information recognition. Therefore, the problem of low accuracy in information recognition exists.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides an information identification method, storage medium, and electronic device to at least solve the technical problem of low accuracy in information identification.

[0005] According to one aspect of the embodiments of this application, an information identification method is provided, comprising: acquiring account behavior information corresponding to a target account to be identified; constructing basic profile features of the target account based on historical behavior information in the account behavior information, and constructing business vertical features of the target account based on business behavior information in the account behavior information, wherein the historical behavior information is behavior information performed by the target account within a preset time range, and the business behavior information is behavior information performed by the target account on a target object within a preset range; performing aggregation processing on the basic profile features and the business vertical features in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions; when a target feature is obtained based on the aggregated profile features, inputting the target feature into an educational qualification recognition model, wherein the information identification model is a neural network model trained using target sample features for identifying the educational qualification of an account, and the target sample features are sample features obtained after processing initial sample features using a metric learning algorithm; and acquiring the output result of the educational qualification recognition model, wherein the output result includes the educational qualification information of the account associated with the target account.

[0006] According to another aspect of the embodiments of this application, another information identification method is provided, including: obtaining account behavior information corresponding to each of multiple sample accounts; obtaining multiple initial sample features based on the account behavior information corresponding to each of the multiple sample accounts, wherein each of the multiple initial sample features corresponds one-to-one with each of the multiple sample accounts; processing the multiple initial sample features using a metric learning algorithm to obtain multiple target sample features; inputting the multiple target sample features into an initial educational qualification recognition model for training to obtain a trained educational qualification recognition model; and identifying the account educational qualification information associated with the account to be identified based on the trained educational qualification recognition model.

[0007] According to another aspect of the embodiments of this application, an information identification device is also provided, comprising: a first acquisition unit, configured to acquire account behavior information corresponding to a target account to be identified; a construction unit, configured to construct basic profile features of the target account based on historical behavior information in the account behavior information, and construct business vertical features of the target account based on business behavior information in the account behavior information, wherein the historical behavior information is behavior information performed by the target account within a preset time range, and the business behavior information is behavior information performed by the target account on a target object within a preset range; an aggregation unit, configured to perform aggregation processing on the basic profile features and the business vertical features in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions; a first input unit, configured to input the target features into an education level recognition model when the target features are acquired based on the aggregated profile features, wherein the information identification model is a neural network model trained using target sample features for identifying the education level of an account, and the target sample features are sample features obtained after processing the initial sample features using a metric learning algorithm; and a second acquisition unit, configured to acquire the output result of the education level recognition model, wherein the output result includes the education level information of the account associated with the target account.

[0008] As an optional solution, the aggregation unit includes: a first acquisition module, configured to acquire basic profile features and business vertical features within a first time period, wherein the preset time range includes the first time period, the basic profile features within the first time period are behavioral information performed by the target account within the first time period, and the business behavioral information is behavioral information performed by the target account on target objects within the preset range within the first time period; and a first aggregation module, configured to aggregate the basic profile features and business vertical features within the first time period to obtain aggregated profile features in the first time dimension, wherein the aggregated profile features in at least two different time dimensions include the first time... The system comprises: a first time dimension aggregated profile features; a second acquisition module, used to acquire basic profile features and business vertical features within a second time period, wherein the preset time range includes the second time period, the basic profile features within the second time period are the behavioral information performed by the target account within the second time period, and the business behavioral information is the behavioral information performed by the target account on target objects within the preset range within the second time period; and a second aggregation module, used to aggregate the basic profile features and business vertical features within the second time period to obtain aggregated profile features of the second time dimension, wherein the aggregated profile features of at least two different time dimensions include the aggregated profile features of the second time dimension.

[0009] As an optional solution, the first aggregation module includes an aggregation submodule, which is used to aggregate the basic profile features and business vertical features of the data in the first time period through an aggregation function. The aggregation function includes at least one of the following aggregation methods: summation, median, and standard deviation.

[0010] As an optional solution, the above-mentioned device includes at least one of the following: a first processing unit, configured to perform normalized numerical feature processing on the aggregated portrait features before inputting the target features into the academic qualification recognition model; and a second processing unit, configured to perform discretized non-numerical feature processing on the aggregated portrait features before inputting the target features into the academic qualification recognition model.

[0011] As an optional solution, the second processing unit includes at least one of the following: a first processing module for performing feature digitization processing on the features belonging to the categorical values ​​in the above-mentioned aggregated portrait features; a second processing module for replacing the category to which the categorical features in the above-mentioned aggregated portrait features belong with the frequency of occurrence of the categorical features; a third processing module for converting the high-dimensional sparse categorical variables in the above-mentioned aggregated portrait features into low-dimensional dense continuous variables; a fourth processing module for filling the missing values ​​of the continuous features in the above-mentioned aggregated portrait features by selecting the average value of the feature values ​​of all continuous features in the above-mentioned aggregated portrait features; or, by selecting the median value of the feature values ​​of all continuous features in the above-mentioned aggregated portrait features; a fifth processing module for filling the missing values ​​of the discrete features in the above-mentioned aggregated portrait features by selecting the most frequently occurring feature value of all discrete features in the above-mentioned aggregated portrait features; and a sixth processing module for summarizing multiple values ​​under the same category in the above-mentioned aggregated portrait features into the same information.

[0012] As an optional solution, the third processing module mentioned above includes: a processing submodule, which is used to embed features into the categorical variables representing the account behavior trajectory of the target account based on a DNN model, to obtain continuous variables representing the behavior patterns of the account behavior trajectory.

[0013] As an optional solution, the above-mentioned construction unit includes: a third acquisition module, used to acquire media information of a specific type associated with the account's educational information; and a construction module, used to acquire the behavior information performed by the target account on the media information of the specific type from the account behavior information, and to construct the business vertical features based on the behavior information performed by the target account on the media information of the specific type.

[0014] According to another aspect of the embodiments of this application, another information recognition device is also provided, including: a third acquisition unit, configured to acquire account behavior information corresponding to each of a plurality of sample accounts; a fourth acquisition unit, configured to acquire a plurality of initial sample features based on the account behavior information corresponding to each of the plurality of sample accounts, wherein each of the plurality of initial sample features corresponds one-to-one with each of the plurality of sample accounts; a third processing unit, configured to process the plurality of initial sample features using a metric learning algorithm to obtain a plurality of target sample features; a second input unit, configured to input the plurality of target sample features into an initial academic qualification recognition model for training to obtain a trained academic qualification recognition model; and a first recognition unit, configured to recognize the account academic qualification information associated with the account to be recognized based on the trained academic qualification recognition model.

[0015] As an optional scheme, the third processing unit includes: a partitioning module for partitioning the multiple initial sample features into a training set and a test set; a first training module for training the initial sample features in the training set using the metric learning algorithm to obtain a mapping matrix W and a kernel matrix M; a first calculation module for calculating the original distance of each initial sample feature in the training set based on the mapping matrix W and the kernel matrix M; a clustering module for clustering the initial sample features in the training set using the original distances to obtain K cluster centers, where K is a natural number; a second calculation module for calculating the first distance from each initial sample feature in the training set to the cluster centers, and sorting the initial sample features in the training set based on the feature similarity corresponding to the first distance to obtain a first sorting result; and an adding module for using the initial sample features in the training set as the target sample features and adding them to the target feature space according to the first sorting result.

[0016] As an optional solution, the second input unit includes: a second training module, used to input the features of the multiple target samples into an initial educational qualification recognition model for training until the training convergence condition is met; obtaining the current educational qualification recognition model and inputting the target feature space into the current educational qualification recognition model; using a first distance unit, clustering each target sample feature in the target feature space to obtain M first clustering results, where M is a natural number; using a second distance unit, clustering each target sample feature in the target feature space to obtain M second clustering results; calculating the M first clustering results to obtain a first co-correlation matrix; calculating the M second clustering results to obtain a second co-correlation matrix; based on the first co-correlation matrix and the second co-correlation matrix, obtaining the target clustering result for each target sample feature in the target feature space, wherein the target clustering result is used to indicate the probability that the target sample feature belongs to the feature corresponding to the target educational qualification; when the target clustering result meets the training convergence condition, determining the current educational qualification recognition model as the trained educational qualification recognition model.

[0017] As an optional approach, obtaining the target clustering result of each target sample feature in the target feature space based on the first co-correlation matrix and the second co-correlation matrix includes: determining whether the elements in the first distance matrix corresponding to the first co-correlation matrix need to be adjusted using a first threshold corresponding to the first co-correlation matrix; determining whether the elements in the second distance matrix corresponding to the second co-correlation matrix need to be adjusted using a second threshold corresponding to the second co-correlation matrix; if the elements in the first distance matrix and / or the elements in the second distance matrix need to be adjusted, and the current iteration number is less than or equal to the target threshold, determining whether the elements in the second distance matrix need to be adjusted using the second threshold, until the elements in the first distance matrix and / or the elements in the second distance matrix do not need to be adjusted, or the current iteration number is greater than the target threshold; and obtaining the target clustering result using a hierarchical clustering algorithm if the elements in the first distance matrix and the elements in the second distance matrix do not need to be adjusted.

[0018] As an optional solution, the above-mentioned device further includes: a fifth acquisition unit, used to acquire account behavior information corresponding to each of the multiple sample accounts; a sixth acquisition unit, used to acquire multiple first sample features based on the account behavior information corresponding to each of the multiple first sample accounts, wherein each of the multiple first sample features corresponds one-to-one with each of the multiple first sample accounts; a fourth processing unit, used to process the multiple first sample features using the above-mentioned metric learning algorithm to obtain multiple second sample features; a third input unit, used to input the multiple second sample features into an initial gender recognition model for training to obtain a trained gender recognition model; and a second recognition unit, used to recognize the account gender information associated with the account to be recognized based on the trained gender recognition model.

[0019] According to another aspect of the embodiments of this 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 executes the computer instructions, causing the computer device to perform the information identification method described above.

[0020] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described information identification method through the computer program.

[0021] In this embodiment, account behavior information corresponding to the target account to be identified is obtained; basic profile features of the target account are constructed based on historical behavior information in the account behavior information, and business vertical features of the target account are constructed based on business behavior information in the account behavior information, wherein the historical behavior information is the behavior information performed by the target account within a preset time range, and the business behavior information is the behavior information performed by the target account on target objects within a preset range; the basic profile features and the business vertical features are aggregated in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions; when the target features are obtained based on the aggregated profile features, the target features are input into an education level recognition model, wherein the information recognition model is a neural network trained using target sample features to identify the education level of the account. The network model, wherein the target sample features are obtained by processing the initial sample features using a metric learning algorithm; the output of the above-mentioned academic qualification recognition model is obtained, wherein the output of the above-mentioned output includes the academic qualification information of the account associated with the target account, and the account behavior information is specifically divided into two categories: one is comprehensive but not vertically specific historical behavior information, and the other is incomplete but highly vertical business behavior information. The features of the two are aggregated using the time dimension. In this way, directional features belonging to the specific business scenario of academic qualification recognition are constructed, which fills the gap in reference information in the academic qualification recognition scenario. Then, the metric learning algorithm is used to process the initial sample features, and the information recognition model trained by the improved sample features is used to process the above-mentioned directional features, thereby achieving the technical effect of improving the accuracy of information recognition, and thus solving the technical problem of low information recognition accuracy.

[0022] In this embodiment, account behavior information corresponding to multiple sample accounts is obtained; multiple initial sample features are obtained based on the account behavior information corresponding to the multiple sample accounts, wherein each initial sample feature corresponds one-to-one with each sample account; the multiple initial sample features are processed using a metric learning algorithm to obtain multiple target sample features; the multiple target sample features are input into an initial education level recognition model for training to obtain a trained education level recognition model; the education level information associated with the account to be identified is identified based on the trained education level recognition model. By using a metric learning algorithm to process the initial sample features, the sample data with a small amount of information is expanded and extended, thereby achieving the technical objective of improving the training quality of the information recognition model on the basis of limited training resources, thus realizing the technical effect of improving the accuracy of information recognition, and solving the technical problem of low information recognition accuracy. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0024] Figure 1 This is a schematic diagram of an application environment for an optional information recognition method according to an embodiment of this application;

[0025] Figure 2 This is a schematic diagram of the flow of an optional information identification method according to an embodiment of this application;

[0026] Figure 3 This is a schematic diagram of an optional information recognition method according to an embodiment of this application;

[0027] Figure 4 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0028] Figure 5 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0029] Figure 6 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0030] Figure 7 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0031] Figure 8 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0032] Figure 9 This is a schematic diagram of the flow of another optional information identification method according to an embodiment of this application;

[0033] Figure 10 This is a schematic diagram of another optional information identification method according to an embodiment of this application;

[0034] Figure 11 This is a schematic diagram of an optional information identification device according to an embodiment of this application;

[0035] Figure 12 This is a schematic diagram of another optional information identification device according to an embodiment of this application;

[0036] Figure 13 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation

[0037] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0038] Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to achieve data computing, storage, processing, and sharing.

[0039] Cloud technology is a collective term for network technologies, information technologies, integration technologies, management platform technologies, and application technologies applied to the cloud computing business model. It can form resource pools, providing flexible and convenient on-demand access. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data at different levels will be processed separately, and various industry data will all require robust system support, which can only be achieved through cloud computing.

[0040] A database, simply put, can be viewed as an electronic filing cabinet—a place to store electronic files, where accounts can perform operations such as adding, querying, updating, and deleting data. A "database" is a collection of data stored together in a certain way, shareable with multiple accounts, with minimal redundancy, and independent of application programs.

[0041] A Database Management System (DBMS) is a computer software system designed to manage databases, generally possessing basic functions such as storage, retrieval, security, and backup. DBMSs can be classified according to the database model they support, such as relational or XML (Extensible Markup Language); or according to the type of computer they support, such as server clusters or mobile devices; or according to the query language used, such as SQL (Structured Query Language) or XQuery; or according to performance priorities, such as maximum scale or maximum operating speed; or other classification methods. Regardless of the classification method used, some DBMSs can cross categories, for example, supporting multiple query languages ​​simultaneously.

[0042] Big data refers to data sets that cannot be captured, managed, and processed within a certain timeframe using conventional software tools. It represents massive, rapidly growing, and diverse information assets that require new processing models to achieve stronger decision-making, insightful discovery, and process optimization capabilities. With the advent of the cloud era, big data has attracted increasing attention. Big data requires specialized technologies to effectively process large amounts of data within a tolerable timeframe. Technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet, and scalable storage systems.

[0043] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be connected directly or indirectly through wired or wireless communication methods, which is not limited herein.

[0044] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0045] According to one aspect of the embodiments of this application, an information identification method is provided. Optionally, as an optional implementation, the above-described information identification method may be applied to, but is not limited to, [examples of other methods]. Figure 1 The environment shown may include, but is not limited to, account device 102 and server 112. Account device 102 may include, but is not limited to, display 108, processor 106 and memory 1004. Server 112 includes database 114 and processing engine 116.

[0046] The specific process can be summarized in the following steps:

[0047] Step S102, the account device 102 obtains the account identifier of the target account 1002 to be identified;

[0048] Steps S104-S106: Send the account identifier to server 112 via network 110;

[0049] In steps S108-S110, server 112 obtains account behavior information corresponding to account identifier from database 114, and then calculates account education information corresponding to account behavior information through processing engine 116.

[0050] In steps S112-S114, the account education information is sent to the account device 102 via network 110. The account device 102 displays the account education information on the display 108 via processor 106 and stores the account education information in the memory 104.

[0051] remove Figure 1Beyond the examples shown, the above steps can be completed with server assistance. Specifically, the server can perform steps such as acquiring the candidate probe set of the target virtual model, obtaining obstruction levels, filtering probes, and rendering lighting, thereby reducing the server's processing load. The account device 102 includes, but is not limited to, handheld devices (such as mobile phones), laptops, desktop computers, and in-vehicle devices. This application does not limit the specific implementation of the account device 102.

[0052] Alternatively, as an optional implementation, such as Figure 2 As shown, the information recognition methods include:

[0053] S202, Obtain account behavior information corresponding to the target account to be identified;

[0054] S204, construct the basic profile features of the target account based on the historical behavior information in the account behavior information, and construct the business vertical features of the target account based on the business behavior information in the account behavior information. The historical behavior information is the behavior information that the target account has performed within a preset time range, and the business behavior information is the behavior information that the target account has performed on the target object within a preset range.

[0055] S206, perform aggregation processing on the basic profile features and business vertical features in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0056] S208, when the target features are obtained based on the aggregated portrait features, the target features are input into the education level recognition model. The information recognition model is a neural network model trained using the target sample features to identify the education level of the account. The target sample features are the sample features obtained after processing the initial sample features using a metric learning algorithm.

[0057] S210, Obtain the output of the education level recognition model, where the output includes the education level information of the account associated with the target account.

[0058] Optionally, in this embodiment, the above information recognition method can be applied, but is not limited to, in business scenarios involving the recognition of an account's highest educational qualification. For example, in scenarios where it is necessary to identify an account's highest educational qualification for targeted product recommendations, non-private basic behavioral information of the account is first obtained. However, since the behavioral characteristics associated with the educational qualification itself are quite complex, and the aforementioned non-private basic behavioral information is usually not directly related to the behavioral characteristics associated with the educational qualification itself, directly using the aforementioned non-private basic behavioral information for subsequent educational qualification recognition would lead to a problem of low accuracy in educational qualification recognition. In this embodiment, to improve the accuracy of educational qualification recognition, diversified feature extraction and feature processing are performed on the aforementioned non-private basic behavioral information to obtain relevant features with a higher correlation to the behavioral characteristics associated with the educational qualification itself. These relevant features are then used for subsequent educational qualification recognition, thereby improving the accuracy of educational qualification recognition.

[0059] Optionally, in this embodiment, the above information recognition method can be applied to other business scenarios, but is not limited to. For example, the above information recognition method can be applied to the business scenario of account gender recognition. First, the non-privacy basic behavioral information of the account is obtained, and then the non-privacy basic behavioral information is subjected to diversified feature extraction and feature processing to obtain relevant features with higher correlation to the business scenario of account gender recognition. Then, the relevant features are used for subsequent gender recognition (such as adjusting the education recognition model to a gender recognition model and then inputting the relevant features into the gender recognition model) to improve the accuracy of gender recognition.

[0060] Optionally, in this embodiment, account behavior information can be understood as, but is not limited to, the collected behavior information of an account within a preset time range or a preset space range, including historical behavior information and business behavior information. The historical behavior information is the behavior information of the target account within the preset time range (such as the account's purchase of virtual items through the target account), and the business behavior information is the behavior information of the target account on the target object within the preset range (such as the account's viewing of media resources through the target account), etc.

[0061] Optionally, in this embodiment, constructing the basic profile features of the target account based on the historical behavior information in the account behavior information can be understood as constructing the basic profile features of the target account based solely on the historical behavior information in the account behavior information, or it can be constructed based on the historical behavior information in the account behavior information and other account information. Among them, other account information can include, but is not limited to, basic account information, basic account information, account association information (such as relevant information of other accounts that have an association relationship with the target account), etc.

[0062] Similarly, in this embodiment, constructing the business vertical features of the target account based on the business behavior information in the account behavior information can be understood, but is not limited to, constructing the business vertical features of the target account based solely on the business behavior information in the account behavior information. It can also be understood, but is not limited to, constructing the business vertical features of the target account based on the other account information and the business behavior information in the account behavior information.

[0063] Optionally, in this embodiment, in order to aggregate features across different time spans, the basic profile features and business vertical features are aggregated in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0064] Optionally, in this embodiment, the information recognition model is a neural network model trained using target sample features to identify the educational level of an account. The target sample features are sample features obtained after processing the initial sample features using a metric learning algorithm. The object of metric learning is usually the distance between sample feature vectors. The purpose of metric learning can be, but is not limited to, reducing or limiting the distance between samples of the same type while increasing the distance between samples of different types through training and learning.

[0065] Optionally, in this embodiment, a seed account group is obtained through manual annotation and business logic, and seed account profile features are generated, including: basic account attributes (such as gender), etc.; abnormal accounts are filtered based on the profile, such as: filtering accounts that have used WeChat for more than 24 hours, etc.; then, based on a clustering metric learning framework, the reordering features of samples in the metric space are obtained to improve the information representation ability of the features; then, the metric learning framework method optimized by co-coefficient matrix is ​​used, combined with the analytic hierarchy process, to pool and weight the values ​​of the samples on different distance metric dimensions, and finally, the probability of the sample belonging to the positive example is fitted.

[0066] To further illustrate, alternatively, for example... Figure 3 As shown, the specific steps are as follows:

[0067] Step S302, Raw data preparation: Based on manual annotation and business experience, identify positive and negative training samples that are strongly correlated with the business, have normal data distribution, and reasonable account profiles;

[0068] Specifically, for example, seed accounts with labels are obtained based on manual annotation and business logic. A batch of seed accounts is roughly recalled based on rules, then filtered by manual screening, and finally verified based on business logic. The basic profile may include, but is not limited to, non-privacy behavioral data of the account within a specific app, such as whether the target application is installed, whether the target application's harassment blocking function is used, or the answering assistant function, etc. Abnormal account type indicators are calculated. In real business scenarios, there may be fake accounts and cases of computers controlling mobile phones. To eliminate the impact of non-genuine accounts on modeling and analysis, abnormal account detection indicators are set based on business experience, such as the account's traffic usage in the target application and the time distribution of traffic generation, etc. Abnormal seed accounts are filtered based on the distribution anomaly theorem. In addition, the "Raida criterion" can be used, but is not limited to, for outlier judgment criteria. The filtered normal seed accounts are stored offline in HDFS. The filtered clean data is stored in the Hadoop Distributed File System (HFDS) for easy access in subsequent processes.

[0069] Step S304, Offline feature processing: Construct portrait features of training samples, and based on the vertical characteristics of the features, combine the time dimension and different feature processing methods to produce high-dimensional feature vectors;

[0070] Specifically, for example, building basic profile features can be, but is not limited to, constructing rich account profiles based on historical account behavior data. These profiles include at least one of the following: basic account attributes, basic device attributes, network connection attributes, etc. Based on business characteristics, vertical business features can be constructed, including, but are not limited to, click-through rates and conversion rates for specific types of media resources. Combining the time dimension, profile features and business features across different time spans can be aggregated, such as calculating aggregated profiles for the account over the past six months, three months, one month, and one week. The aggregation method can be, but is not limited to, summation, median, and standard deviation, using any one or more of these three methods. Further feature processing can be performed using normalized numerical and discretized non-numerical methods, where the normalization method can be, but is not limited to, Gaussian normalization. Furthermore, the processed features are merged and stored offline in HDFS for quick access in subsequent processes. Finally, the feature processing logic is solidified, and offline automated calculations are performed periodically, with the offline calculation results pushed to the online storage engine.

[0071] Step S306, Clustering Metric Learning Framework: Based on training samples and feature vectors, and based on K-Means clustering and metric learning, obtain the re-ranking features of accounts belonging to positive examples;

[0072] Specifically, for example, read in the low-order and high-order feature results and concatenate them column by column; divide the features into a training set probFea and a test set galFea; further, for the training set probFea, use a metric learning algorithm to train and calculate the mapping matrix W and the kernel matrix M, where the calculation method can refer to the following formulas (1), (2) and (3):

[0073]

[0074] ∑ E W=J(W)∑ I W (2)

[0075]

[0076] Further utilizing the mapping matrix W and kernel matrix M mentioned above, the distance dist from each sample in the test set to the original test set is calculated. The calculation method can be found in the following formulas (4), (5) and (6):

[0077]

[0078]

[0079]

[0080] Furthermore, for example Figure 4 As shown, clustering is performed based on the K-Means algorithm to obtain K cluster centers. The specific steps are as follows:

[0081] Step S402: Cluster the samples to be clustered into 3 categories;

[0082] Step S404: Select 3 center points;

[0083] Step S406: For each sample, find the nearest centroid to complete one clustering operation;

[0084] Step S408: Determine whether the clustering of sample points before and after a clustering is the same. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0085] Step S410: Update the centroids based on the clustering results;

[0086] Step S412: For each sample, find the nearest centroid and complete the secondary clustering;

[0087] Step S414: Determine whether the clustering of sample points before and after the second clustering is the same. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0088] Step S416: Update the centroids based on the clustering results;

[0089] Step S411: For each sample, find the nearest centroid and complete multiple clustering operations;

[0090] Step S420: Determine whether the clustering of sample points is the same before and after multiple clustering. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0091] Step S422: Assuming the algorithm terminated in the previous step (step S420), the final clustering result is presented;

[0092] Further calculations are performed to determine the distance from each sample to the cluster center, and a re-ranking is then performed based on similarity matching. The re-ranking results are then added to a new feature space, for example... Figure 5 As shown, the specific steps are as follows:

[0093] Step S502, input data;

[0094] Step S504: Extract features;

[0095] Step S506: Use metric learning to train the data and obtain W and M;

[0096] Step S508: Use the K-means algorithm to obtain the cluster centers of the training set;

[0097] Step S510: Calculate the distance based on the cluster centers using W and M;

[0098] Step S512: Calculate the reordering matching matrix.

[0099] Step S308, Co-cooperation matrix framework: Use K-Means clustering to obtain the co-cooperation matrix, and use the co-cooperation matrix in the optimization of metric learning;

[0100] Specifically, for example, the low-order and high-order feature results are read in and concatenated column-wise (including reordering); then, K-Means is used to obtain M basis clustering results, and the co-coherence matrix A is calculated; the distance metric of the K-Means algorithm is replaced with Mahalanobis distance to obtain M new clustering results, and the co-coherence matrix B is calculated. Furthermore, the Mahalanobis distance between d-dimensional samples x and y can be defined, but is not limited to, as:

[0101]

[0102] Further utilizing the metric learning algorithm of the co-cooperation relation matrix, the distance matrix A and the threshold e are calculated. The calculation method can be found in the following formula (7):

[0103]

[0104] Furthermore, the final clustering results are obtained using the analytic hierarchy process (AHP). The calculation process is as follows: Figure 6 As shown, the specific steps are as follows:

[0105] Step S602, construct the judgment matrix;

[0106] Step S604: Calculate the single-layer weight subset;

[0107] Step S606: Check the consistency of the single layer. If it passes, proceed to step S608; otherwise, proceed to step S602.

[0108] Step S608: Calculate the single-layer weight subset;

[0109] Step S610: Check the consistency of the single layer. If it passes, proceed to step S612; otherwise, proceed to step S602.

[0110] Step S612, derive the indicator weights.

[0111] Further calculate the original distance and Mahalanobis distance of the samples, perform average pooling, and normalize the results. The normalized result is then approximated as the probability of belonging to a positive example. For example... Figure 7 As shown, first assume the input dataset is X = {x, x2, ..., xn}, the number of clusters in the clustering result is K, the maximum number of iterations is i, and the output clustering result is π*. The specific steps are as follows:

[0112] Step S702: Use the k-means algorithm to obtain M basis clustering results and calculate the co-coherence matrix;

[0113] Step S704: Obtain the distance matrix A and the threshold c using a metric learning algorithm based on the co-cooperation matrix;

[0114] Step S706: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0115] Step S708: Replace the k-means algorithm distance metric with Mahalanobis distance to obtain M basis clustering results, and calculate the co-coherence matrix;

[0116] Step S710: Obtain the distance matrix A and the threshold e using a metric learning algorithm based on the co-cooperation matrix;

[0117] Step S712: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0118] Step S714: Determine if the S matrix and D matrix have changed. If they have changed, proceed to step S716; otherwise, proceed to step S711.

[0119] Step S716: Determine if the number of iterations is greater than i. If yes, jump to step S711; otherwise, jump to step S708.

[0120] Step S711: Use the hierarchical clustering algorithm to obtain the final clustering result π*.

[0121] Optionally, in this embodiment, the account's educational background information may refer to, but is not limited to, the target account's current highest educational level, such as below undergraduate, undergraduate, or above undergraduate. In addition, the account's educational background information may also refer to, but is not limited to, the probability information of the target account's educational background, such as the probability that the target account is below undergraduate is 30%, the probability that the target account is undergraduate is 60%, and the probability that the target account is above undergraduate is 10%.

[0122] It should be noted that account behavior information is specifically divided into two categories: one is comprehensive but lacks verticality in historical behavior information, and the other is incomplete but highly vertical business behavior information. The features of the two are aggregated using the time dimension. In this way, directional features specific to the business scenario of academic qualification recognition are constructed, which fills the gap in reference information in the academic qualification recognition scenario. Then, the initial sample features are processed using a metric learning algorithm, and the information recognition model trained with the improved sample features is used to process the above-mentioned directional features, thereby improving the accuracy of information recognition.

[0123] To further illustrate, optional examples include... Figure 8As shown, the process involves obtaining account behavior information 804 corresponding to the target account 802 to be identified; constructing basic profile features 806-1 of the target account 802 based on historical behavior information 04-1 in the account behavior information 804, and constructing business vertical features 806-2 of the target account 802 based on business behavior information 804-2 in the account behavior information 804; performing aggregation processing on the basic profile features 806-1 and business vertical features 806-2 in at least two time dimensions to obtain at least two aggregated profile features 808 in different time dimensions; and then obtaining... When the target features are obtained, the target features are input into the education level recognition model 810. The information recognition model 810 is a neural network model trained using the target sample features to identify the education level of the account. The target sample features are the sample features obtained after processing the initial sample features using a metric learning algorithm. The output results of the education level recognition model 810 are obtained. The output results include the account education information 812 associated with the target account 802, such as the probability that the target account is below undergraduate level is 30%, the probability that the target account is undergraduate level is 60%, and the probability that the target account is above undergraduate level is 10%.

[0124] The embodiments provided in this application obtain account behavior information corresponding to the target account to be identified; construct basic profile features of the target account based on historical behavior information in the account behavior information, and construct business vertical features of the target account based on business behavior information in the account behavior information, wherein historical behavior information refers to the behavior information performed by the target account within a preset time range, and business behavior information refers to the behavior information performed by the target account on target objects within a preset range; aggregate the basic profile features and business vertical features in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions; when the target features are obtained based on the aggregated profile features, the target features are input into an education level recognition model, wherein the information recognition model is trained using the target sample features and is used to identify the education level of the account. The neural network model uses target sample features obtained by processing initial sample features using a metric learning algorithm. It acquires the output of the educational qualification recognition model, which includes the educational qualification information associated with the target account. The account behavior information is specifically divided into two categories: comprehensive but insufficiently vertical historical behavior information, and incomplete but highly vertical business behavior information. The features of these two categories are aggregated using the time dimension. This method constructs directional features specific to the educational qualification recognition business scenario, filling the gaps in reference information in this scenario. The initial sample features are then processed using a metric learning algorithm, and the information recognition model trained with the refined sample features is used to further process these directional features, thereby improving the accuracy of information recognition.

[0125] As an optional approach, the basic profile features and business vertical features are aggregated in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions, including:

[0126] S1. Obtain the basic profile features and business vertical features within the first time period. The preset time range includes the first time period. The basic profile features within the first time period are the behavioral information performed by the target account within the first time period. The business behavioral information is the behavioral information performed by the target account on the target object within the preset range within the first time period. Aggregate the basic profile features and business vertical features within the first time period to obtain the aggregated profile features of the first time dimension. At least two aggregated profile features of different time dimensions include the aggregated profile features of the first time dimension.

[0127] S2, obtain the basic profile features and business vertical features within the second time period. The preset time range includes the second time period. The basic profile features within the second time period are the behavioral information performed by the target account within the second time period, and the business behavioral information are the behavioral information performed by the target account on target objects within the preset range within the second time period. Aggregate the basic profile features and business vertical features within the second time period to obtain aggregated profile features in the second time dimension. At least two aggregated profile features in different time dimensions include the aggregated profile features in the second time dimension.

[0128] Optionally, in this embodiment, "first" and "second" are illustrative examples and can be understood as multiple, but are not limited to, such as a first time period and a second time period, but are not limited to obtaining only the basic profile features and business vertical features of two time periods.

[0129] As an optional approach, the basic profile features and business-specific vertical features within the first time period are aggregated to obtain the aggregated profile features for the first time dimension, including:

[0130] The basic profile features and business vertical features within the first time period are aggregated using aggregation functions. The aggregation function can be summarized in at least one of the following ways: summation, median, or standard deviation.

[0131] Optionally, in this embodiment, the basic profile features and business vertical features of different time spans are aggregated by combining the time dimension, and the aggregated profile (aggregated profile features) of the account in the past six months, the past three months, the past one month and the past one week are calculated. The aggregation method selected is summation, median and standard deviation.

[0132] Optionally, in this embodiment, aggregate functions, also called group functions, can be used to perform statistics and calculations on data in a table, but are not limited to. They are generally used in conjunction with group by to count and calculate grouped data, such as count(col): to find the total number of rows in a specified column, max(col): to find the maximum value in a specified column, min(col): to find the minimum value in a specified column, sum(col): to find the sum of a specified column, avg(col): to find the average value in a specified column, etc.

[0133] As an alternative approach, before inputting the target features into the education level recognition model, the method includes at least one of the following:

[0134] S1, normalizes the aggregated image features into numerical features;

[0135] S2 performs discretization of the aggregated image features into non-numerical features.

[0136] Optionally, in this embodiment, the normalized numerical feature processing method may be, but is not limited to, selecting the Gaussian normalized feature processing method.

[0137] As an optional approach, the aggregated image features are discretized into non-numerical feature processing, including at least one of the following:

[0138] The features belonging to the category values ​​in the aggregated portrait features are processed by feature digitization;

[0139] Replace the category to which the classification feature belongs in the aggregated portrait feature with the frequency of the classification feature;

[0140] Transform high-dimensional sparse categorical variables in aggregated image features into low-dimensional dense continuous variables;

[0141] For missing values ​​of continuous features in aggregated portrait features, select the average value of all continuous features in the aggregated portrait features to fill the missing values ​​of continuous features; or, select the median value of all continuous features in the aggregated portrait features to fill the missing values ​​of continuous features.

[0142] For missing values ​​of discrete features in aggregated portrait features, select the feature value that appears most frequently among all discrete features in aggregated portrait features and fill the missing values ​​of continuous features.

[0143] Multiple values ​​of variables belonging to the same category in aggregated portrait features are grouped into the same information.

[0144] Optionally, in this embodiment, the features belonging to the category values ​​in the aggregated portrait features are processed by feature digitization, for example, by using one-hot encoding to process the features belonging to the category values ​​such as account gender.

[0145] Optionally, in this embodiment, the category to which the classification feature belongs in the aggregated portrait features is replaced with the frequency of the classification feature. For example, the Count Encoding method is used to replace the classification features with their occurrence count. For example, if 'Peking' appears 10 times in a certain category, then 'Peking' will be replaced with 10. Specifically, for the WiFiPOI features of the account, Count Encoding is used to identify the degree of interest of the account and this POI. For example, the account visited the POI "Food-Chinese Food-Cantonese Food" a total of 3 times in the current week.

[0146] Optionally, data analysis reveals that many category features exhibit strong sparsity. To avoid model overfitting and improve model stability, this embodiment transforms the high-dimensional sparse categorical variables in the aggregated portrait features into low-dimensional dense continuous variables. For example, it employs Category Embedding, introducing a neural network to convert the high-dimensional sparse categorical variables into low-dimensional dense embedding variables.

[0147] Optionally, for handling missing values ​​of features, in this embodiment, for missing values ​​of continuous features in aggregated portrait features, the average value of the feature values ​​of all continuous features in aggregated portrait features is selected to fill the missing values ​​of continuous features; or, the median value of the feature values ​​of all continuous features in aggregated portrait features is selected to fill the missing values ​​of continuous features; for missing values ​​of discrete features in aggregated portrait features, the feature value of the most frequently occurring discrete feature in aggregated portrait features is selected to fill the missing values ​​of continuous features, etc.; in addition, missing values ​​can also be handled by converting missing values ​​into Embedding representations.

[0148] Optionally, in this embodiment, multiple values ​​under variables belonging to the same category in the aggregated profile features are summarized into the same information. For example, Consolidation Encoding can be used to summarize multiple values ​​under certain category variables into the same information. For instance, if the target system version feature includes three values: "4.2", "4.4", and "5.0", these three values ​​can be summarized as "lower version of the target system" based on experience. Experiments have shown that Consolidation Encoding provides greater positive benefits than directly one-hot encoding the "target system version" feature.

[0149] As an alternative approach, the high-dimensional sparse categorical variables in the aggregated image features are transformed into low-dimensional dense continuous variables, including:

[0150] Based on the DNN model, feature embedding is performed on the categorical variables used to represent the account behavior trajectory of the target account to obtain continuous variables used to represent the behavior patterns of the account behavior trajectory.

[0151] Optionally, in this embodiment, Category features are input into the DNN model to train Embedding features.

[0152] Optionally, in this embodiment, the WiFi connection trajectory data of the account is embedded based on the MST-CNN deep learning network to capture the account's WiFi behavior pattern information.

[0153] Optionally, in this embodiment, based on the List-Embedding method, embedding is extracted from the traffic usage behavior sequences of accounts using different categories of applications under the same system. For example, the TrafficEmbedding of the target type of App is used to obtain low-dimensional dense account behavior features.

[0154] As an optional approach, business vertical characteristics of the target account can be constructed based on business behavior information within the account behavior information, including:

[0155] S1, obtain media information of a specific type associated with the account's educational background information;

[0156] S2: Obtain the behavioral information of the target account on media information of a specific type from the account behavior information, and construct business vertical features based on the behavioral information of the target account on media information of a specific type.

[0157] Optionally, in this embodiment, in order to improve the adaptability to the business scenario of academic qualification recognition, the media information of a specific type associated with the account's academic qualification information is first obtained, and then the behavior information of the target account on the media information of the specific type is obtained from the account behavior information. Based on the behavior information of the target account on the media information of the specific type, a business vertical feature is constructed.

[0158] Alternatively, as an alternative implementation method, such as Figure 9 As shown, the information recognition methods include:

[0159] S902, obtain the account behavior information corresponding to each of the multiple sample accounts;

[0160] S904, based on the account behavior information corresponding to each of the multiple sample accounts, multiple initial sample features are obtained, wherein each initial sample feature in the multiple initial sample features corresponds one-to-one with each sample account in the multiple sample accounts;

[0161] S906 uses a metric learning algorithm to process multiple initial sample features to obtain multiple target sample features;

[0162] S908, input the features of multiple target samples into the initial academic qualification recognition model for training, and obtain the trained academic qualification recognition model;

[0163] S910 identifies the educational background information of accounts associated with the account to be identified based on a trained educational background recognition model.

[0164] Optionally, in this embodiment, the above information recognition method can be applied, but is not limited to, in business scenarios involving the recognition of an account's highest educational qualification. For example, in scenarios where it is necessary to identify an account's highest educational qualification for targeted product recommendations, non-private basic behavioral information of the account is first obtained. However, since the behavioral characteristics associated with the educational qualification itself are quite complex, and the aforementioned non-private basic behavioral information is usually not directly related to the behavioral characteristics associated with the educational qualification itself, directly using the aforementioned non-private basic behavioral information for subsequent educational qualification recognition would lead to a problem of low accuracy in educational qualification recognition. In this embodiment, to improve the accuracy of educational qualification recognition, diversified feature extraction and feature processing are performed on the aforementioned non-private basic behavioral information to obtain relevant features with a higher correlation to the behavioral characteristics associated with the educational qualification itself. These relevant features are then used for subsequent educational qualification recognition, thereby improving the accuracy of educational qualification recognition.

[0165] Optionally, in this embodiment, the above information recognition method can be applied to other business scenarios, but is not limited to. For example, the above information recognition method can be applied to the business scenario of account gender recognition. First, the non-privacy basic behavioral information of the account is obtained, and then the non-privacy basic behavioral information is subjected to diversified feature extraction and feature processing to obtain relevant features with higher correlation to the business scenario of account gender recognition. Then, the relevant features are used for subsequent gender recognition (such as adjusting the education recognition model to a gender recognition model and then inputting the relevant features into the gender recognition model) to improve the accuracy of gender recognition.

[0166] Optionally, in this embodiment, account behavior information can be understood as, but is not limited to, the collected behavior information of an account within a preset time range or a preset space range, including historical behavior information and business behavior information. The historical behavior information is the behavior information of the target account within the preset time range (such as the account's purchase of virtual items through the target account), and the business behavior information is the behavior information of the target account on the target object within the preset range (such as the account's viewing of media resources through the target account), etc.

[0167] Optionally, in this embodiment, constructing the basic profile features of the target account based on the historical behavior information in the account behavior information can be understood as constructing the basic profile features of the target account based solely on the historical behavior information in the account behavior information, or it can be constructed based on the historical behavior information in the account behavior information and other account information. Among them, other account information can include, but is not limited to, basic account information, basic account information, account association information (such as relevant information of other accounts that have an association relationship with the target account), etc.

[0168] Similarly, in this embodiment, constructing the business vertical features of the target account based on the business behavior information in the account behavior information can be understood, but is not limited to, constructing the business vertical features of the target account based solely on the business behavior information in the account behavior information. It can also be understood, but is not limited to, constructing the business vertical features of the target account based on the other account information and the business behavior information in the account behavior information.

[0169] Optionally, in this embodiment, in order to aggregate features across different time spans, the basic profile features and business vertical features are aggregated in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0170] Optionally, in this embodiment, the information recognition model is a neural network model trained using target sample features to identify the educational level of an account. The target sample features are sample features obtained after processing the initial sample features using a metric learning algorithm. The object of metric learning is usually the distance between sample feature vectors. The purpose of metric learning can be, but is not limited to, reducing or limiting the distance between samples of the same type while increasing the distance between samples of different types through training and learning.

[0171] Optionally, in this embodiment, a seed account group is obtained through manual annotation and business logic, and seed account profile features are generated, including: basic account attributes (such as gender), etc.; abnormal accounts are filtered based on the profile, such as: filtering accounts that have used WeChat for more than 24 hours, etc.; then, based on a clustering metric learning framework, the reordering features of samples in the metric space are obtained to improve the information representation ability of the features; then, the metric learning framework method optimized by co-coefficient matrix is ​​used, combined with the analytic hierarchy process, to pool and weight the values ​​of the samples on different distance metric dimensions, and finally, the probability of the sample belonging to the positive example is fitted.

[0172] To further illustrate, alternatively, for example... Figure 3 As shown, the specific steps are as follows:

[0173] Step S302, Raw data preparation: Based on manual annotation and business experience, identify positive and negative training samples that are strongly correlated with the business, have normal data distribution, and reasonable account profiles;

[0174] Specifically, for example, seed accounts with labels are obtained based on manual annotation and business logic. A batch of seed accounts is roughly recalled based on rules, then filtered by manual screening, and finally verified based on business logic. A basic profile of the seed accounts is obtained, which may include, but is not limited to, some non-privacy behavior data of the account in a specific app, such as whether the target application is installed, whether the target application's harassment blocking function is used, and the answering assistant function, etc. Anomaly account type indicators are calculated. In real business scenarios, there may be fake accounts and cases of computers controlling mobile phones. In order to eliminate the impact of non-real accounts on modeling and analysis, anomaly account detection indicators are set based on business experience, such as the account's traffic usage in the target application and the time distribution of traffic generation, etc. Anomaly seed accounts are filtered based on the distribution anomaly theorem. In addition, the "Raida criterion" can be used, but is not limited to, for outlier judgment criteria. The filtered normal seed accounts are stored offline in HDFS. The filtered clean data is stored in the distributed file system (The Hadoop Distributed File System, abbreviated as HFDS) for easy access in subsequent processes.

[0175] Step S304, Offline feature processing: Construct portrait features of training samples, and based on the vertical characteristics of the features, combine the time dimension and different feature processing methods to produce high-dimensional feature vectors;

[0176] Specifically, for example, building basic profile features can be, but is not limited to, constructing rich account profiles based on historical account behavior data. These profiles include at least one of the following: basic account attributes, basic device attributes, network connection attributes, etc. Based on business characteristics, vertical business features can be constructed, including, but are not limited to, click-through rates and conversion rates for specific types of media resources. Combining the time dimension, profile features and business features across different time spans can be aggregated, such as calculating aggregated profiles for the account over the past six months, three months, one month, and one week. The aggregation method can be, but is not limited to, summation, median, and standard deviation, using any one or more of these three methods. Further feature processing can be performed using normalized numerical and discretized non-numerical methods, where the normalization method can be, but is not limited to, Gaussian normalization. Furthermore, the processed features are merged and stored offline in HDFS for quick access in subsequent processes. Finally, the feature processing logic is solidified, and offline automated calculations are performed periodically, with the offline calculation results pushed to the online storage engine.

[0177] Step S306, Clustering Metric Learning Framework: Based on training samples and feature vectors, and based on K-Means clustering and metric learning, obtain the re-ranking features of accounts belonging to positive examples;

[0178] Specifically, for example, read in the low-order and high-order feature results and concatenate them by column; divide the features into a training set probFea and a test set galFea; further train the training set probFea using a metric learning algorithm, and calculate the mapping matrix W and the kernel matrix M, wherein the calculation method can refer to the above formulas (1), (2) and (3).

[0179] Further utilize the above mapping matrix W and kernel matrix M to calculate the distance dist from each sample in the test set to the original test set. The calculation method can refer to the above formulas (4), (5) and (6).

[0180] Furthermore, for example Figure 4 As shown, clustering is performed based on the K-Means algorithm to obtain K cluster centers. The specific steps are as follows:

[0181] Step S402: Cluster the samples to be clustered into 3 categories;

[0182] Step S404: Select 3 center points;

[0183] Step S406: For each sample, find the nearest centroid to complete one clustering operation;

[0184] Step S408: Determine whether the clustering of sample points is the same before and after a clustering. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0185] Step S410: Update the centroids based on the clustering results;

[0186] Step S412: For each sample, find the nearest centroid and complete the secondary clustering;

[0187] Step S414: Determine whether the clustering of sample points before and after the second clustering is the same. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0188] Step S416: Update the centroids based on the clustering results;

[0189] Step S411: For each sample, find the nearest centroid and complete multiple clustering operations;

[0190] Step S420: Determine whether the clustering of sample points is the same before and after multiple clustering. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0191] Step S422: Assuming the algorithm terminated in the previous step (step S420), the final clustering result is presented;

[0192] Further calculations are performed to determine the distance from each sample to the cluster center, and a re-ranking is performed based on similarity matching. The results of this re-ranking are then added to a new feature space, for example... Figure 5 As shown, the specific steps are as follows:

[0193] Step S502, input data;

[0194] Step S504: Extract features;

[0195] Step S506: Use metric learning to train the data and obtain W and M;

[0196] Step S508: Use the K-means algorithm to obtain the cluster centers of the training set;

[0197] Step S510: Calculate the distance based on the cluster centers using W and M;

[0198] Step S512: Calculate the reordering matching matrix.

[0199] Step S308, Co-cooperation matrix framework: Use K-Means clustering to obtain the co-cooperation matrix, and use the co-cooperation matrix in the optimization of metric learning;

[0200] Specifically, for example, the low-order and high-order feature results are read in and concatenated column-wise (including reordering); then, K-Means is used to obtain M basis clustering results, and the co-coherence matrix A is calculated; the distance metric of the K-Means algorithm is replaced with Mahalanobis distance to obtain M new clustering results, and the co-coherence matrix B is calculated. Furthermore, the Mahalanobis distance between d-dimensional samples x and y can be defined, but is not limited to, as:

[0201]

[0202] Further utilize the metric learning algorithm of the co-cooperation relation matrix to calculate the distance matrix A and the threshold e. The calculation method can refer to the above formula (7);

[0203] Furthermore, the final clustering results are obtained using the analytic hierarchy process (AHP). The calculation process is as follows: Figure 6 As shown, the specific steps are as follows:

[0204] Step S602, construct the judgment matrix;

[0205] Step S604: Calculate the single-layer weight subset;

[0206] Step S606: Check the consistency of the single layer. If it passes, proceed to step S608; otherwise, proceed to step S602.

[0207] Step S608: Calculate the single-layer weight subset;

[0208] Step S610: Check the consistency of the single layer. If it passes, proceed to step S612; otherwise, proceed to step S602.

[0209] Step S612, derive the indicator weights.

[0210] Further calculate the original distance and Mahalanobis distance of the samples, perform average pooling, and normalize the results. The normalized result is then approximated as the probability of belonging to a positive example. For example... Figure 7 As shown, first assume the input dataset is X = {x, x2, ..., xn}, the number of clusters in the clustering result is K, the maximum number of iterations is i, and the output clustering result is π*. The specific steps are as follows:

[0211] Step S702: Use the k-means algorithm to obtain M basis clustering results and calculate the co-coherence matrix;

[0212] Step S704: Obtain the distance matrix A and the threshold c using a metric learning algorithm based on the co-cooperation matrix;

[0213] Step S706: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0214] Step S708: Replace the k-means algorithm distance metric with Mahalanobis distance to obtain M basis clustering results, and calculate the co-coherence matrix;

[0215] Step S710: Obtain the distance matrix A and the threshold e using a metric learning algorithm based on the co-cooperation matrix;

[0216] Step S712: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0217] Step S714: Determine if the S matrix and D matrix have changed. If they have changed, proceed to step S716; otherwise, proceed to step S711.

[0218] Step S716: Determine if the number of iterations is greater than i. If yes, jump to step S711; otherwise, jump to step S708.

[0219] Step S711: Use the hierarchical clustering algorithm to obtain the final clustering result π*.

[0220] Optionally, in this embodiment, the account's educational background information may refer to, but is not limited to, the target account's current highest educational level, such as below undergraduate, undergraduate, or above undergraduate. In addition, the account's educational background information may also refer to, but is not limited to, the probability information of the target account's educational background, such as the probability that the target account is below undergraduate is 30%, the probability that the target account is undergraduate is 60%, and the probability that the target account is above undergraduate is 10%.

[0221] It should be noted that by using metric learning algorithms to process the features of the initial samples, the sample data with relatively small amounts of information can be expanded and extended. This improves the training quality of the information recognition model and thus enhances the accuracy of information recognition, given limited training resources.

[0222] To further illustrate, optional examples include... Figure 10 As shown, the specific steps are as follows:

[0223] Step S1002: Based on manual annotation and business logic, obtain seed accounts with Label information;

[0224] Step S1004: Obtain the basic profile of the seed account. Considering that the "highest education level tag of the account" is a long-term and stable requirement of the account, it is necessary to calculate the features of the account for multiple time periods and then use sum pooling to compress the features.

[0225] Step S1006: Calculate the evaluation indicators for abnormal account types, such as the account's traffic usage in the target product and the time distribution of traffic generation.

[0226] Step S1008: Based on the distribution anomaly theorem, filter out abnormal seed accounts, such as using the "Laida criterion" as an outlier judgment standard;

[0227] Step S1010: Determine whether the seed account quantity meets the standard. For example, if the minimum quantity of positive and negative samples is 100,000, proceed to step S1002; otherwise, proceed to step S1012.

[0228] Step S1012: Construct account features, such as basic profile features and business vertical type features;

[0229] Step S1014: Combine the time dimension to construct aggregated features;

[0230] Step S1016: Feature processing is performed using normalized numerical data and discretized non-numerical data methods;

[0231] Step S1011: Input the Category features into the DNN model to train the Embedding features;

[0232] Step S1020: Merge the processed features and store them offline in HDFS;

[0233] Step S1022: Solidify the feature processing logic, perform scheduled offline automated calculations, and push the offline calculation results to the online storage engine;

[0234] Step S1024: Read in the low-order and high-order feature results and concatenate them column by column;

[0235] Step S1026: Divide the features into a training set probFea and a test set galFea;

[0236] Step S1028: For the training set probFea, use the metric learning algorithm to train and calculate the mapping matrix W and the kernel matrix M;

[0237] Step S1030: Using the mapping matrix W and kernel matrix M mentioned above, calculate the distance dist from each sample in the test set to the original test set.

[0238] Step S1032: Perform clustering based on the K-Means algorithm to obtain K cluster centers;

[0239] Step S1034: Calculate the distance from each sample to the cluster center and re-rank them based on similarity matching;

[0240] Step S1036: Add the reordering result to the new feature space;

[0241] Step S1038: Read in the low-order and high-order feature results and concatenate them by column (including reordering results);

[0242] Step S1040: Use K-Means to obtain M basis clustering results and calculate the co-co-relationship matrix A;

[0243] Step S1042: Replace the distance metric of the K-Means algorithm with Mahalanobis distance, obtain M clustering results again, and calculate the co-co-relationship matrix B;

[0244] Step S1044: Calculate the distance matrix A and the threshold e using the metric learning algorithm of the co-cooperation relation matrix;

[0245] Step S1046: Obtain the final clustering results using the analytic hierarchy process (AHP).

[0246] Step S1048: Calculate the original distance and Mahalanobis distance of the samples, perform average pooling operation, normalize the result, and approximate the normalized result as the probability of belonging to the positive example.

[0247] The embodiments provided in this application obtain account behavior information corresponding to multiple sample accounts; based on the account behavior information corresponding to the multiple sample accounts, multiple initial sample features are obtained, wherein each initial sample feature corresponds one-to-one with each sample account; the multiple initial sample features are processed using a metric learning algorithm to obtain multiple target sample features; the multiple target sample features are input into an initial education level recognition model for training to obtain a trained education level recognition model; based on the trained education level recognition model, the account education information associated with the account to be identified is identified. By using a metric learning algorithm to process the initial sample features, the sample data with a small amount of information is expanded and extended, thereby achieving the technical objective of improving the training quality of the information recognition model on the basis of limited training resources, and thus realizing the technical effect of improving the accuracy of information recognition.

[0248] As an alternative approach, a metric learning algorithm is used to process multiple initial sample features to obtain multiple target sample features, including:

[0249] S1 divides multiple initial sample features into training and test sets;

[0250] S2, use the metric learning algorithm to train the initial sample features in the training set to obtain the mapping matrix W and the kernel matrix M;

[0251] S3, based on the mapping matrix W and the kernel matrix M, calculates the original distance of each initial sample feature in the training set in the training set;

[0252] S4, use the original distance to cluster the initial sample features in the training set to obtain K cluster centers, where K is a natural number;

[0253] S5, calculate the first distance from each initial sample feature in the training set to the cluster center, and sort the initial sample features in the training set based on the feature similarity corresponding to the first distance to obtain the first sorting result;

[0254] S6, take the initial sample features in the training set as the target sample features, and add them to the target feature space according to the first sorting result.

[0255] It should be noted that multiple initial sample features are divided into training and test sets; the initial sample features in the training set are trained using a metric learning algorithm to obtain a mapping matrix W and a kernel matrix M; based on the mapping matrix W and the kernel matrix M, the original distance of each initial sample feature in the training set is calculated; the initial sample features in the training set are clustered using the original distances to obtain K cluster centers, where K is a natural number; the first distance from each initial sample feature in the training set to the cluster center is calculated, and the initial sample features in the training set are sorted based on the feature similarity corresponding to the first distance to obtain a first sorting result; the initial sample features in the training set are used as target sample features and added to the target feature space according to the first sorting result.

[0256] To further illustrate, alternatively, low-order and high-order feature results can be read in and concatenated column by column; the features can be divided into a training set probFea and a test set galFea; further, for the training set probFea, a metric learning algorithm can be used for training, and the mapping matrix W and kernel matrix M can be calculated, wherein the calculation method can refer to the above formulas (1), (2) and (3):

[0257] Further utilizing the mapping matrix W and kernel matrix M mentioned above, the distance dist from each sample in the test set to the original test set is calculated. The calculation method can be found in formulas (4), (5), and (6) above:

[0258] Furthermore, for example Figure 4 As shown, clustering is performed based on the K-Means algorithm to obtain K cluster centers. The specific steps are as follows:

[0259] Step S402: Cluster the samples to be clustered into 3 categories;

[0260] Step S404: Select 3 center points;

[0261] Step S406: For each sample, find the nearest centroid to complete one clustering operation;

[0262] Step S408: Determine whether the clustering of sample points is the same before and after a clustering. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0263] Step S410: Update the centroids based on the clustering results;

[0264] Step S412: For each sample, find the nearest centroid and complete the secondary clustering;

[0265] Step S414: Determine whether the clustering of sample points before and after the second clustering is the same. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0266] Step S416: Update the centroids based on the clustering results;

[0267] Step S411: For each sample, find the nearest centroid and complete multiple clustering operations;

[0268] Step S420: Determine whether the clustering of sample points is the same before and after multiple clustering. If they are the same, terminate (step S422); if they are different, continue to the next step (step S420).

[0269] Step S422: Assuming the algorithm terminated in the previous step (step S420), the final clustering result is presented;

[0270] Further calculations are performed to determine the distance from each sample to the cluster center, and a re-ranking is then performed based on similarity matching. The re-ranking results are then added to a new feature space, for example... Figure 5 As shown, the specific steps are as follows:

[0271] Step S502, input data;

[0272] Step S504: Extract features;

[0273] Step S506: Use metric learning to train the data and obtain W and M;

[0274] Step S508: Use the K-means algorithm to obtain the cluster centers of the training set;

[0275] Step S510: Calculate the distance based on the cluster centers using W and M;

[0276] Step S512: Calculate the reordering matching matrix.

[0277] As an optional approach, multiple target sample features are input into the initial educational qualification recognition model for training, resulting in a trained educational qualification recognition model, including:

[0278] S1, Input the features of multiple target samples into the initial education level recognition model for training until the training convergence condition is met:

[0279] S2, obtain the current academic qualification recognition model and input the target feature space into the current academic qualification recognition model;

[0280] S3, using the first distance unit, cluster each target sample feature in the target feature space to obtain M first clustering results, where M is a natural number; using the second distance unit, cluster each target sample feature in the target feature space to obtain M second clustering results;

[0281] S4. Calculate the first co-colour relation matrix by performing calculations on the M first clustering results; calculate the second co-colour relation matrix by performing calculations on the M second clustering results.

[0282] S5. Based on the first co-co-relation matrix and the second co-co-relation matrix, obtain the target clustering result for each target sample feature in the target feature space, wherein the target clustering result is used to indicate the probability that the target sample feature belongs to the feature corresponding to the target education level.

[0283] S6. If the target clustering result meets the training convergence condition, the current academic qualification recognition model is determined as the trained academic qualification recognition model.

[0284] Optionally, in this embodiment, the first distance unit may be, but is not limited to, the distance unit of the original distance, and the second distance unit may be, but is not limited to, the distance unit of Mahalanobis distance;

[0285] To illustrate further, we can optionally read in low-order and high-order feature results, concatenate them column-wise (including reordering results), then use K-Means to obtain M basis clustering results, calculate the co-coherence matrix A, and replace the distance metric of the K-Means algorithm with Mahalanobis distance to obtain M clustering results again, and calculate the co-coherence matrix B.

[0286] As an optional approach, based on the first and second co-co ...

[0287] S1, use the first threshold corresponding to the first co-co ...

[0288] S2, if the elements in the first distance matrix and / or the elements in the second distance matrix need to be adjusted and the current iteration number is less than or equal to the target threshold, use the second threshold to determine whether the elements in the second distance matrix need to be adjusted, until the elements in the first distance matrix and / or the elements in the second distance matrix do not need to be adjusted, or the current iteration number is greater than the target threshold;

[0289] S3. Without needing to adjust the elements in the first distance matrix and the elements in the second distance matrix, the hierarchical clustering algorithm is used to obtain the target clustering result.

[0290] It should be noted that the first threshold corresponding to the first co-co ...

[0291] To further illustrate, calculate the original distance and Mahalanobis distance of the samples, perform average pooling, normalize the results, and approximate the normalized result as the probability of belonging to the positive example. For example... Figure 7 As shown, first assume the input dataset is X = {x, x2, ..., xn}, the number of clusters in the clustering result is K, the maximum number of iterations is i, and the output clustering result is π*. The specific steps are as follows:

[0292] Step S702: Use the k-means algorithm to obtain M basis clustering results and calculate the co-coherence matrix;

[0293] Step S704: Obtain the distance matrix A and the threshold c using a metric learning algorithm based on the co-cooperation matrix;

[0294] Step S706: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0295] Step S708: Replace the k-means algorithm distance metric with Mahalanobis distance to obtain M basis clustering results, and calculate the co-coherence matrix;

[0296] Step S710: Obtain the distance matrix A and the threshold e using a metric learning algorithm based on the co-cooperation matrix;

[0297] Step S712: For elements in the co-operation matrix that are greater than or equal to c, set the corresponding element in the S matrix to 1; for elements that are less than c, set the corresponding element in the D matrix to 1.

[0298] Step S714: Determine if the S matrix and D matrix have changed. If they have changed, proceed to step S716; otherwise, proceed to step S711.

[0299] Step S716: Determine if the number of iterations is greater than i. If yes, jump to step S711; otherwise, jump to step S708.

[0300] Step S711: Use the hierarchical clustering algorithm to obtain the final clustering result π*.

[0301] As an alternative approach, the method also includes:

[0302] S1, obtain the account behavior information corresponding to each of the multiple sample accounts;

[0303] S2, based on the account behavior information corresponding to each of the multiple first sample accounts, obtain multiple first sample features, wherein each first sample feature in the multiple first sample features corresponds one-to-one with each sample account in the multiple first sample accounts;

[0304] S3, using a metric learning algorithm to process multiple first sample features to obtain multiple second sample features;

[0305] S4. Input multiple second sample features into the initial gender recognition model for training to obtain a trained gender recognition model.

[0306] S5 identifies the gender information of the accounts associated with the accounts to be identified based on the trained gender recognition model.

[0307] Optionally, in this embodiment, the education recognition model has strong reusability: First, change the account type of the positive sample, such as the prediction of "account gender label", then the server accumulates the corresponding log data, and finally use the same feature concatenation, feature processing and model training methods to produce the result, which is the account education information replaced by the account gender information.

[0308] The embodiments provided in this application obtain account behavior information corresponding to multiple sample accounts; based on the account behavior information corresponding to multiple first sample accounts, multiple first sample features are obtained, wherein each first sample feature corresponds one-to-one with each sample account in the multiple first sample accounts; the multiple first sample features are processed using a metric learning algorithm to obtain multiple second sample features; the multiple second sample features are input into an initial gender recognition model for training to obtain a trained gender recognition model; based on the trained gender recognition model, the gender information of the account to be identified is recognized, thereby achieving the goal of recognizing the corresponding scene's recognition information even when the input samples from other scenes need to be adjusted, thus realizing the technical effect of improving the reusability of information recognition.

[0309] It is understood that in the specific implementation of this application, data such as account information are involved. When the above embodiments of this application are applied to specific products or technologies, account permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0310] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0311] According to another aspect of the embodiments of this application, an information identification apparatus for implementing the above-described information identification method is also provided. For example... Figure 11 As shown, the device includes:

[0312] The first acquisition unit 1102 is used to acquire account behavior information corresponding to the target account to be identified;

[0313] The construction unit 1104 is used to construct the basic profile features of the target account based on the historical behavior information in the account behavior information, and to construct the business vertical features of the target account based on the business behavior information in the account behavior information. The historical behavior information is the behavior information that the target account has performed within a preset time range, and the business behavior information is the behavior information that the target account has performed on the target object within a preset range.

[0314] Aggregation unit 1106 is used to perform aggregation processing on basic profile features and business vertical features in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0315] The first input unit 1108 is used to input the target features into the academic qualification recognition model when the target features are obtained based on the aggregated portrait features. The information recognition model is a neural network model trained using the target sample features to identify the academic qualification of the account. The target sample features are sample features obtained by processing the initial sample features using a metric learning algorithm.

[0316] The second acquisition unit 1110 is used to acquire the output results of the academic qualification recognition model, wherein the output results include the academic qualification information of the account associated with the target account.

[0317] Optionally, in this embodiment, the aforementioned information recognition device can be applied, but is not limited to, in business scenarios involving the recognition of an account's highest educational qualification. For example, in scenarios where it is necessary to identify an account's highest educational qualification for targeted product recommendations, non-private basic behavioral information of the account is first obtained. However, since the behavioral characteristics associated with the educational qualification itself are quite complex, and the aforementioned non-private basic behavioral information is usually not directly related to the behavioral characteristics associated with the educational qualification itself, directly using the aforementioned non-private basic behavioral information for subsequent educational qualification recognition would lead to a problem of low accuracy in educational qualification recognition. In this embodiment, to improve the accuracy of educational qualification recognition, diversified feature extraction and feature processing are performed on the aforementioned non-private basic behavioral information to obtain relevant features with a higher degree of correlation with the behavioral characteristics associated with the educational qualification itself. These relevant features are then used for subsequent educational qualification recognition, thereby improving the accuracy of educational qualification recognition.

[0318] Optionally, in this embodiment, the information recognition device can be applied to other business scenarios, but is not limited to. For example, the information recognition device can be applied to the business scenario of account gender recognition. First, the non-privacy basic behavioral information of the account is obtained, and then the non-privacy basic behavioral information is subjected to diversified feature extraction and feature processing to obtain relevant features with higher correlation to the business scenario of account gender recognition. Then, the relevant features are used for subsequent gender recognition (such as adjusting the education recognition model to a gender recognition model and then inputting the relevant features into the gender recognition model) to improve the accuracy of gender recognition.

[0319] Optionally, in this embodiment, account behavior information can be understood as, but is not limited to, the collected behavior information of an account within a preset time range or a preset space range, including historical behavior information and business behavior information. The historical behavior information is the behavior information of the target account within the preset time range (such as the account's purchase of virtual items through the target account), and the business behavior information is the behavior information of the target account on the target object within the preset range (such as the account's viewing of media resources through the target account), etc.

[0320] Optionally, in this embodiment, constructing the basic profile features of the target account based on the historical behavior information in the account behavior information can be understood as constructing the basic profile features of the target account based solely on the historical behavior information in the account behavior information, or it can be constructed based on the historical behavior information in the account behavior information and other account information. Among them, other account information can include, but is not limited to, basic account information, basic account information, account association information (such as relevant information of other accounts that have an association relationship with the target account), etc.

[0321] Similarly, in this embodiment, constructing the business vertical features of the target account based on the business behavior information in the account behavior information can be understood, but is not limited to, constructing the business vertical features of the target account based solely on the business behavior information in the account behavior information. It can also be understood, but is not limited to, constructing the business vertical features of the target account based on the other account information and the business behavior information in the account behavior information.

[0322] Optionally, in this embodiment, in order to aggregate features across different time spans, the basic profile features and business vertical features are aggregated in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0323] Optionally, in this embodiment, the information recognition model is a neural network model trained using target sample features to identify the educational level of an account. The target sample features are sample features obtained after processing the initial sample features using a metric learning algorithm. The object of metric learning is usually the distance between sample feature vectors. The purpose of metric learning can be, but is not limited to, reducing or limiting the distance between samples of the same type while increasing the distance between samples of different types through training and learning.

[0324] Optionally, in this embodiment, a seed account group is obtained through manual annotation and business logic to generate seed account profile features, including: basic account attributes (such as gender), etc.; abnormal accounts are filtered based on the profile, such as: filtering accounts that have used WeChat for more than 24 hours, etc.; then, based on a clustering metric learning framework, the reordering features of samples in the metric space are obtained to improve the information representation ability of the features; then, the metric learning framework method optimized by co-coefficient matrix is ​​used, combined with the analytic hierarchy process, to pool and weight the values ​​of the samples on different distance metric dimensions, and finally, the probability of a sample belonging to a positive example is fitted.

[0325] Optionally, in this embodiment, the account's educational background information may refer to, but is not limited to, the target account's current highest educational level, such as below undergraduate, undergraduate, or above undergraduate. In addition, the account's educational background information may also refer to, but is not limited to, the probability information of the target account's educational background, such as the probability that the target account is below undergraduate is 30%, the probability that the target account is undergraduate is 60%, and the probability that the target account is above undergraduate is 10%.

[0326] It should be noted that account behavior information is specifically divided into two categories: one is comprehensive but lacks verticality in historical behavior information, and the other is incomplete but highly vertical business behavior information. The features of the two are aggregated using the time dimension. In this way, directional features specific to the business scenario of academic qualification recognition are constructed, which fills the gap in reference information in the academic qualification recognition scenario. Then, the initial sample features are processed using a metric learning algorithm, and the information recognition model trained with the improved sample features is used to process the above-mentioned directional features, thereby improving the accuracy of information recognition.

[0327] For specific embodiments, please refer to the examples shown in the above information recognition device; these will not be repeated here.

[0328] The embodiments provided in this application obtain account behavior information corresponding to the target account to be identified; construct basic profile features of the target account based on historical behavior information in the account behavior information, and construct business vertical features of the target account based on business behavior information in the account behavior information, wherein historical behavior information refers to the behavior information performed by the target account within a preset time range, and business behavior information refers to the behavior information performed by the target account on target objects within a preset range; aggregate the basic profile features and business vertical features in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions; when the target features are obtained based on the aggregated profile features, the target features are input into an education level recognition model, wherein the information recognition model is trained using the target sample features and is used to identify the education level of the account. The neural network model uses target sample features obtained by processing initial sample features using a metric learning algorithm. It acquires the output of the educational qualification recognition model, which includes the educational qualification information associated with the target account. The account behavior information is specifically divided into two categories: comprehensive but insufficiently vertical historical behavior information, and incomplete but highly vertical business behavior information. The features of these two categories are aggregated using the time dimension. This method constructs directional features specific to the educational qualification recognition business scenario, filling the gaps in reference information in this scenario. The initial sample features are then processed using a metric learning algorithm, and the information recognition model trained with the refined sample features is used to further process these directional features, thereby improving the accuracy of information recognition.

[0329] As an optional solution, aggregation unit 1106 includes:

[0330] The first acquisition module is used to acquire basic profile features and business vertical features within a first time period. The preset time range includes the first time period. The basic profile features within the first time period are the behavioral information performed by the target account within the first time period, and the business behavioral information are the behavioral information performed by the target account on target objects within a preset range within the first time period. The first aggregation module is used to aggregate the basic profile features and business vertical features within the first time period to obtain aggregated profile features in the first time dimension. At least two aggregated profile features in different time dimensions include the aggregated profile features in the first time dimension.

[0331] The second acquisition module is used to acquire basic profile features and business vertical features within the second time period. The preset time range includes the second time period. The basic profile features within the second time period are the behavioral information performed by the target account within the second time period, and the business behavioral information are the behavioral information performed by the target account on target objects within the preset range within the second time period. The second aggregation module is used to aggregate the basic profile features and business vertical features within the second time period to obtain aggregated profile features in the second time dimension. At least two aggregated profile features in different time dimensions include aggregated profile features in the second time dimension.

[0332] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0333] As an optional solution, the first aggregation module includes:

[0334] The aggregation submodule is used to aggregate the basic profile features and business vertical features of the data in the first time period through aggregation functions. The aggregation function can be summarized in at least one of the following ways: summation, median, and standard deviation.

[0335] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0336] As an optional solution, the device includes at least one of the following:

[0337] The first processing unit is used to perform normalized numerical feature processing on the aggregated portrait features before inputting the target features into the education recognition model.

[0338] The second processing unit is used to perform discretization non-numerical feature processing on the aggregated portrait features before inputting the target features into the education recognition model.

[0339] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0340] As an optional solution, the second processing unit includes at least one of the following:

[0341] The first processing module is used to digitize the features that belong to the category values ​​in the aggregated portrait features;

[0342] The second processing module is used to replace the category to which the classification feature in the aggregated portrait features belongs with the frequency of the classification feature.

[0343] The third processing module is used to transform the high-dimensional sparse classification variables in the aggregated portrait features into low-dimensional dense continuous variables.

[0344] The fourth processing module is used to fill the missing values ​​of continuous features in the aggregated portrait features by selecting the average value of all continuous features in the aggregated portrait features; or, by selecting the median value of all continuous features in the aggregated portrait features.

[0345] The fifth processing module is used to fill in the missing values ​​of continuous features by selecting the most frequently occurring feature value among all discrete features in the aggregated portrait features.

[0346] The sixth processing module is used to summarize multiple values ​​of variables belonging to the same category in the aggregated portrait features into the same information.

[0347] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0348] As an optional solution, the third processing module includes:

[0349] The processing submodule is used to embed features from the categorical variables representing the account behavior trajectory of the target account based on the DNN model, so as to obtain continuous variables representing the behavior patterns of the account behavior trajectory.

[0350] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0351] As an optional solution, building unit 1104 includes:

[0352] The third acquisition module is used to acquire media information of a specific type associated with the account's educational background information;

[0353] The module is used to obtain the behavior information of the target account on media information of a specific type from the account behavior information, and to build business vertical features based on the behavior information of the target account on media information of a specific type.

[0354] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0355] According to another aspect of the embodiments of this application, an information identification apparatus for implementing the above-described information identification method is also provided. For example... Figure 12 As shown, the device includes:

[0356] The third acquisition unit 1202 is used to acquire account behavior information corresponding to each of the multiple sample accounts;

[0357] The fourth acquisition unit 1204 is used to acquire multiple initial sample features based on the account behavior information corresponding to each of the multiple sample accounts, wherein each initial sample feature in the multiple initial sample features corresponds one-to-one with each sample account in the multiple sample accounts.

[0358] The third processing unit 1206 is used to process multiple initial sample features using a metric learning algorithm to obtain multiple target sample features;

[0359] The second input unit 1208 is used to input the features of multiple target samples into the initial academic qualification recognition model for training, so as to obtain the trained academic qualification recognition model.

[0360] The first identification unit 1210 is used to identify the account education information associated with the account to be identified based on the trained education identification model.

[0361] Optionally, in this embodiment, the above information recognition method can be applied, but is not limited to, in business scenarios involving the recognition of an account's highest educational qualification. For example, in scenarios where it is necessary to identify an account's highest educational qualification for targeted product recommendations, non-private basic behavioral information of the account is first obtained. However, since the behavioral characteristics associated with the educational qualification itself are quite complex, and the aforementioned non-private basic behavioral information is usually not directly related to the behavioral characteristics associated with the educational qualification itself, directly using the aforementioned non-private basic behavioral information for subsequent educational qualification recognition would lead to a problem of low accuracy in educational qualification recognition. In this embodiment, to improve the accuracy of educational qualification recognition, diversified feature extraction and feature processing are performed on the aforementioned non-private basic behavioral information to obtain relevant features with a higher correlation to the behavioral characteristics associated with the educational qualification itself. These relevant features are then used for subsequent educational qualification recognition, thereby improving the accuracy of educational qualification recognition.

[0362] Optionally, in this embodiment, the above information recognition method can be applied to other business scenarios, but is not limited to. For example, the above information recognition method can be applied to the business scenario of account gender recognition. First, the non-privacy basic behavioral information of the account is obtained, and then the non-privacy basic behavioral information is subjected to diversified feature extraction and feature processing to obtain relevant features with higher correlation to the business scenario of account gender recognition. Then, the relevant features are used for subsequent gender recognition (such as adjusting the education recognition model to a gender recognition model and then inputting the relevant features into the gender recognition model) to improve the accuracy of gender recognition.

[0363] Optionally, in this embodiment, account behavior information can be understood as, but is not limited to, the collected behavior information of an account within a preset time range or a preset space range, including historical behavior information and business behavior information. The historical behavior information is the behavior information of the target account within the preset time range (such as the account's purchase of virtual items through the target account), and the business behavior information is the behavior information of the target account on the target object within the preset range (such as the account's viewing of media resources through the target account), etc.

[0364] Optionally, in this embodiment, constructing the basic profile features of the target account based on the historical behavior information in the account behavior information can be understood as constructing the basic profile features of the target account based solely on the historical behavior information in the account behavior information, or it can be constructed based on the historical behavior information in the account behavior information and other account information. Among them, other account information can include, but is not limited to, basic account information, basic account information, account association information (such as relevant information of other accounts that have an association relationship with the target account), etc.

[0365] Similarly, in this embodiment, constructing the business vertical features of the target account based on the business behavior information in the account behavior information can be understood, but is not limited to, constructing the business vertical features of the target account based solely on the business behavior information in the account behavior information. It can also be understood, but is not limited to, constructing the business vertical features of the target account based on the other account information and the business behavior information in the account behavior information.

[0366] Optionally, in this embodiment, in order to aggregate features across different time spans, the basic profile features and business vertical features are aggregated in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0367] Optionally, in this embodiment, the information recognition model is a neural network model trained using target sample features to identify the educational level of an account. The target sample features are sample features obtained after processing the initial sample features using a metric learning algorithm. The object of metric learning is usually the distance between sample feature vectors. The purpose of metric learning can be, but is not limited to, reducing or limiting the distance between samples of the same type while increasing the distance between samples of different types through training and learning.

[0368] Optionally, in this embodiment, a seed account group is obtained through manual annotation and business logic to generate seed account profile features, including: basic account attributes (such as gender), etc.; abnormal accounts are filtered based on the profile, such as: filtering accounts that have used WeChat for more than 24 hours, etc.; then, based on a clustering metric learning framework, the reordering features of samples in the metric space are obtained to improve the information representation ability of the features; then, the metric learning framework method optimized by co-coefficient matrix is ​​used, combined with the analytic hierarchy process, to pool and weight the values ​​of the samples on different distance metric dimensions, and finally, the probability of a sample belonging to a positive example is fitted.

[0369] It should be noted that by using metric learning algorithms to process the features of the initial samples, the sample data with relatively small amounts of information can be expanded and extended. This improves the training quality of the information recognition model and thus enhances the accuracy of information recognition, given limited training resources.

[0370] For specific embodiments, please refer to the examples shown in the above information recognition device; these will not be repeated here.

[0371] The embodiments provided in this application obtain account behavior information corresponding to multiple sample accounts; based on the account behavior information corresponding to the multiple sample accounts, multiple initial sample features are obtained, wherein each initial sample feature corresponds one-to-one with each sample account; the multiple initial sample features are processed using a metric learning algorithm to obtain multiple target sample features; the multiple target sample features are input into an initial education level recognition model for training to obtain a trained education level recognition model; based on the trained education level recognition model, the account education information associated with the account to be identified is identified. By using a metric learning algorithm to process the initial sample features, the sample data with a small amount of information is expanded and extended, thereby achieving the technical objective of improving the training quality of the information recognition model on the basis of limited training resources, and thus realizing the technical effect of improving the accuracy of information recognition.

[0372] As an optional solution, the third processing unit 1206 includes:

[0373] The partitioning module is used to divide multiple initial sample features into training and test sets;

[0374] The first training module is used to train the initial sample features in the training set using a metric learning algorithm to obtain the mapping matrix W and the kernel matrix M.

[0375] The first calculation module is used to calculate the original distance of each initial sample feature in the training set in the training set based on the mapping matrix W and the kernel matrix M.

[0376] The clustering module is used to cluster the initial sample features in the training set using the original distance, and obtain K cluster centers, where K is a natural number;

[0377] The second calculation module is used to calculate the first distance from each initial sample feature in the training set to the cluster center, and sort the initial sample features in the training set based on the feature similarity corresponding to the first distance to obtain the first sorting result;

[0378] The module is used to take the initial sample features in the training set as the target sample features and add them to the target feature space according to the first sorting result.

[0379] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0380] As an optional solution, the second input unit 1208 includes:

[0381] The second training module is used to input the features of multiple target samples into the initial education level recognition model for training until the training convergence condition is met:

[0382] Obtain the current academic qualification recognition model and input the target feature space into the current academic qualification recognition model;

[0383] Using a first distance unit, clustering is performed on each target sample feature in the target feature space to obtain M first clustering results, where M is a natural number; using a second distance unit, clustering is performed on each target sample feature in the target feature space to obtain M second clustering results;

[0384] Calculate the first co-co-relation matrix by performing calculations on the M first clustering results; calculate the second co-co-relation matrix by performing calculations on the M second clustering results;

[0385] Based on the first co-co-correlation matrix and the second co-co-correlation matrix, the target clustering result of each target sample feature in the target feature space is obtained, wherein the target clustering result is used to indicate the probability that the target sample feature belongs to the feature corresponding to the target education level;

[0386] If the target clustering results meet the training convergence condition, the current academic qualification recognition model is determined as the trained academic qualification recognition model.

[0387] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0388] As an optional approach, based on the first and second co-co ...

[0389] Use the first threshold corresponding to the first co-co ...

[0390] If the elements in the first distance matrix and / or the elements in the second distance matrix need to be adjusted, and the current iteration number is less than or equal to the target threshold, the second threshold is used to determine whether the elements in the second distance matrix need to be adjusted, until the elements in the first distance matrix and / or the elements in the second distance matrix do not need to be adjusted, or the current iteration number is greater than the target threshold.

[0391] Without needing to adjust the elements in the first distance matrix and the elements in the second distance matrix, the hierarchical clustering algorithm is used to obtain the target clustering result.

[0392] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0393] As an optional solution, the device also includes:

[0394] The fifth acquisition unit is used to acquire account behavior information corresponding to each of the multiple sample accounts;

[0395] The sixth acquisition unit is used to acquire multiple first sample features based on the account behavior information corresponding to each of the multiple first sample accounts, wherein each of the multiple first sample features corresponds one-to-one with each of the multiple first sample accounts.

[0396] The fourth processing unit is used to process multiple first sample features using a metric learning algorithm to obtain multiple second sample features;

[0397] The third input unit is used to input multiple second sample features into the initial gender recognition model for training, so as to obtain a trained gender recognition model.

[0398] The second identification unit is used to identify the educational information of the account associated with the account to be identified based on the trained gender identification model.

[0399] For specific implementation examples, please refer to the examples shown in the above information recognition method; these examples will not be repeated here.

[0400] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described information identification method is also provided, such as... Figure 13 As shown, the electronic device includes a memory 1302 and a processor 1304. The memory 1302 stores a computer program, and the processor 1304 is configured to execute the steps of any of the above method embodiments through the computer program.

[0401] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0402] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0403] S1, Obtain account behavior information corresponding to the target account to be identified;

[0404] S2, construct the basic profile features of the target account based on the historical behavior information in the account behavior information, and construct the business vertical features of the target account based on the business behavior information in the account behavior information. The historical behavior information is the behavior information that the target account has performed within a preset time range, and the business behavior information is the behavior information that the target account has performed on the target object within a preset range.

[0405] S3, aggregate the basic profile features and business vertical features in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0406] S4. After obtaining the target features based on the aggregated portrait features, the target features are input into the education recognition model. The information recognition model is a neural network model trained using the target sample features to identify the education level of the account. The target sample features are the sample features obtained after processing the initial sample features using a metric learning algorithm.

[0407] S5, obtain the output of the education level recognition model, where the output includes the education level information of the account associated with the target account. Or,

[0408] S1, obtain the account behavior information corresponding to each of the multiple sample accounts;

[0409] S2, based on the account behavior information corresponding to each of the multiple sample accounts, obtain multiple initial sample features, wherein each initial sample feature in the multiple initial sample features corresponds one-to-one with each sample account in the multiple sample accounts;

[0410] S3, using a metric learning algorithm to process multiple initial sample features to obtain multiple target sample features;

[0411] S4. Input the features of multiple target samples into the initial academic qualification recognition model for training to obtain the trained academic qualification recognition model.

[0412] S5 identifies the educational background information of the accounts associated with the accounts to be identified based on the trained educational background recognition model.

[0413] Alternatively, as those skilled in the art will understand, Figure 13 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 13 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 13 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 13 The different configurations shown.

[0414] The memory 1302 can be used to store software programs and modules, such as the program instructions / modules corresponding to the information identification method and device in this embodiment. The processor 1304 executes various functional applications and data processing by running the software programs and modules stored in the memory 1302, thereby realizing the aforementioned information identification method. The memory 1302 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1302 may further include memory remotely located relative to the processor 1304, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1302 may be used, but is not limited to, to store account behavior information, aggregated profile features, and account education information, or account behavior information, initial sample features, and account education information, etc. As an example, such as... Figure 13 As shown, the memory 1302 may include, but is not limited to, the first acquisition unit 1102, construction unit 1104, aggregation unit 1106, first input unit 1108, and second acquisition unit 1110 (or the third acquisition unit 1202, fourth acquisition unit 1204, third processing unit 1206, second input unit 1208, and first identification unit 1210 not shown) of the information recognition device. Furthermore, it may include, but is not limited to, other module units of the information recognition device, which will not be elaborated upon in this example.

[0415] Optionally, the transmission device 1306 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1306 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1306 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0416] In addition, the aforementioned electronic device also includes: a display 1308 for displaying the aforementioned account behavior information, aggregated profile features and account education information, or account behavior information, initial sample features and account education information, etc.; and a connection bus 1310 for connecting the various module components in the aforementioned electronic device.

[0417] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer (P2P) network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0418] According to one aspect of this application, a computer program product is provided, comprising a computer program / instructions containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit, it performs various functions provided in embodiments of this application.

[0419] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0420] It should be noted that the computer system of the electronic device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0421] A computer system includes a Central Processing Unit (CPU), which performs various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) or loaded from RAM. ROM also stores various programs and data required for system operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output interfaces (I / O interfaces) are also connected to the bus.

[0422] The following components are connected to the input / output interface: input sections including keyboards, mice, etc.; output sections including cathode ray tubes (CRTs), liquid crystal displays (LCDs), and speakers; storage sections including hard drives; and communication sections including network interface cards such as LAN cards and modems. The communication section performs communication processing via a network such as the Internet. Drives are also connected to the input / output interface as needed. Removable media, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage section as required.

[0423] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit, it performs various functions defined in the system of this application.

[0424] According to one aspect of this application, a computer-readable storage medium is provided, wherein a processor of a computer device reads computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.

[0425] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store a computer program for performing the following steps:

[0426] S1, Obtain account behavior information corresponding to the target account to be identified;

[0427] S2, construct the basic profile features of the target account based on the historical behavior information in the account behavior information, and construct the business vertical features of the target account based on the business behavior information in the account behavior information. The historical behavior information is the behavior information that the target account has performed within a preset time range, and the business behavior information is the behavior information that the target account has performed on the target object within a preset range.

[0428] S3, aggregate the basic profile features and business vertical features in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions.

[0429] S4. After obtaining the target features based on the aggregated portrait features, the target features are input into the education recognition model. The information recognition model is a neural network model trained using the target sample features to identify the education level of the account. The target sample features are the sample features obtained after processing the initial sample features using a metric learning algorithm.

[0430] S5, obtain the output of the education level recognition model, where the output includes the education level information of the account associated with the target account. Or,

[0431] S1, obtain the account behavior information corresponding to each of the multiple sample accounts;

[0432] S2, based on the account behavior information corresponding to each of the multiple sample accounts, obtain multiple initial sample features, wherein each initial sample feature in the multiple initial sample features corresponds one-to-one with each sample account in the multiple sample accounts;

[0433] S3, using a metric learning algorithm to process multiple initial sample features to obtain multiple target sample features;

[0434] S4. Input the features of multiple target samples into the initial academic qualification recognition model for training to obtain the trained academic qualification recognition model.

[0435] S5 identifies the educational background information of the accounts associated with the accounts to be identified based on the trained educational background recognition model.

[0436] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0437] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0438] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0439] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0440] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.

[0441] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0442] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0443] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An information identification method, characterized in that, include: Obtain account behavior information corresponding to the target account to be identified; The basic profile features of the target account are constructed based on the historical behavior information in the account behavior information, and the business vertical features of the target account are constructed based on the business behavior information in the account behavior information. The historical behavior information is the behavior information that the target account has performed within a preset time range, and the business behavior information is the behavior information that the target account has performed on target objects within a preset range. The basic profile features and the business vertical features are aggregated in at least two time dimensions to obtain at least two aggregated profile features in different time dimensions. When the target features are obtained based on the aggregated profile features, the target features are input into the education level recognition model, wherein the education level recognition model is a neural network model trained using multiple target sample features to identify the education level of an account, and the target sample features are sample features obtained by processing the initial sample features using a metric learning algorithm; Obtain the output result of the educational qualification recognition model, wherein the output result includes the educational qualification information of the account associated with the target account; Before inputting the target features into the educational qualification recognition model, the method further includes: inputting the multiple target sample features into an initial educational qualification recognition model for training until the training convergence condition is met. Obtain the current academic qualification recognition model, and input the target feature space with the added features of the multiple target samples into the current academic qualification recognition model; Using a first distance unit, clustering is performed on each target sample feature in the target feature space to obtain M first clustering results, where M is a natural number and the first distance unit is the distance unit of the original distance; using a second distance unit, clustering is performed on each target sample feature in the target feature space to obtain M second clustering results, where the second distance unit is the distance unit of Mahalanobis distance; The first co-clustering matrix is ​​obtained by calculating the M first clustering results; the second co-clustering matrix is ​​obtained by calculating the M second clustering results. Based on the first co-co-relation matrix and the second co-co-relation matrix, the target clustering result of each target sample feature in the target feature space is obtained, wherein the target clustering result is used to indicate the probability that the target sample feature belongs to the feature corresponding to the target education level; If the target clustering result meets the training convergence condition, the current academic qualification recognition model is determined as the trained academic qualification recognition model.

2. The method according to claim 1, characterized in that, The aggregation process of the basic profile features and the business vertical features in at least two time dimensions to obtain aggregated profile features in at least two different time dimensions includes: The system acquires basic profile features and business vertical features within a first time period. The preset time range includes the first time period. The basic profile features within the first time period are the behavioral information performed by the target account within the first time period. The business behavioral information is the behavioral information performed by the target account on target objects within the preset range within the first time period. The system aggregates the basic profile features and business vertical features within the first time period to obtain aggregated profile features in the first time dimension. The aggregated profile features in at least two different time dimensions include the aggregated profile features in the first time dimension. The system acquires basic profile features and business vertical features within a second time period. The preset time range includes the second time period. The basic profile features within the second time period are the behavioral information performed by the target account within the second time period, and the business behavioral information is the behavioral information performed by the target account on target objects within the preset range within the second time period. The system aggregates the basic profile features and the business vertical features within the second time period to obtain aggregated profile features in the second time dimension. The aggregated profile features in at least two different time dimensions include the aggregated profile features in the second time dimension.

3. The method according to claim 2, characterized in that, The aggregation of the basic profile features and the business vertical features within the first time period to obtain the aggregated profile features in the first time dimension includes: The basic profile features and business vertical features within the first time period are aggregated using an aggregation function. The aggregation function can be summarized in at least one of the following ways: summation, median, or standard deviation.

4. The method according to claim 1, characterized in that, Before inputting the target features into the education level recognition model, the method includes at least one of the following: The aggregated image features are subjected to normalized numerical feature processing; The aggregated image features are discretized into non-numerical features.

5. The method according to claim 4, characterized in that, The discretization of the aggregated image features into a non-numerical feature processing method includes at least one of the following: The features belonging to the classification value in the aggregated portrait features are processed by feature digitization; Replace the category to which the classification feature in the aggregated image features belongs with the frequency of the occurrence of the classification feature; The high-dimensional sparse classification variables in the aggregated image features are transformed into low-dimensional dense continuous variables; For missing values ​​of continuous features in the aggregated portrait features, the average value of the feature values ​​of all continuous features in the aggregated portrait features is selected to fill the missing values ​​of the continuous features; or, the median value of the feature values ​​of all continuous features in the aggregated portrait features is selected to fill the missing values ​​of the continuous features. For the missing values ​​of discrete features in the aggregated portrait features, select the feature value that appears most frequently among all discrete features in the aggregated portrait features and fill the missing values ​​of the continuous features. Multiple values ​​of variables belonging to the same category in the aggregated portrait features are grouped into the same information.

6. The method according to claim 5, characterized in that, The process of transforming the high-dimensional sparse classification variables in the aggregated image features into low-dimensional dense continuous variables includes: Based on the DNN model, feature embedding is performed on the categorical variables used to represent the account behavior trajectory of the target account to obtain continuous variables used to represent the behavior pattern of the account behavior trajectory.

7. The method according to any one of claims 1 to 6, characterized in that, The process of constructing the business vertical characteristics of the target account based on the business behavior information in the account behavior information includes: Obtain media information of a specific type associated with the account's educational background information; The target account's behavior information on media information of a specific type is obtained from the account behavior information, and the business vertical feature is constructed based on the target account's behavior information on media information of the specific type.

8. An information recognition method, characterized in that, include: Obtain account behavior information for each of the multiple sample accounts; Multiple initial sample features are obtained based on the account behavior information corresponding to each of the multiple sample accounts, wherein each initial sample feature corresponds one-to-one with each sample account among the multiple sample accounts; The features of the initial samples are processed using a metric learning algorithm to obtain the features of the target samples. The features of the multiple target samples are input into the initial educational qualification recognition model for training, resulting in a trained educational qualification recognition model, including: The features of the multiple target samples are input into the initial education level recognition model for training until the training convergence condition is met: Obtain the current academic qualification recognition model, and input the target feature space with the added features of the multiple target samples into the current academic qualification recognition model; Using a first distance unit, clustering is performed on each target sample feature in the target feature space to obtain M first clustering results, where M is a natural number and the first distance unit is the distance unit of the original distance; using a second distance unit, clustering is performed on each target sample feature in the target feature space to obtain M second clustering results, where the second distance unit is the distance unit of Mahalanobis distance; The first co-clustering matrix is ​​obtained by calculating the M first clustering results; the second co-clustering matrix is ​​obtained by calculating the M second clustering results. Based on the first co-co-relation matrix and the second co-co-relation matrix, the target clustering result of each target sample feature in the target feature space is obtained, wherein the target clustering result is used to indicate the probability that the target sample feature belongs to the feature corresponding to the target education level; If the target clustering result meets the training convergence condition, the current academic qualification recognition model is determined as the trained academic qualification recognition model. The trained educational qualification recognition model is used to identify the educational qualification information of the accounts associated with the accounts to be identified.

9. The method according to claim 8, characterized in that, The process of using a metric learning algorithm to process the multiple initial sample features yields multiple target sample features, including: The multiple initial sample features are divided into a training set and a test set; The metric learning algorithm is used to train the initial sample features in the training set to obtain the mapping matrix W and the kernel matrix M; Based on the mapping matrix W and the kernel matrix M, calculate the original distance of each initial sample feature in the training set in the training set; The initial sample features in the training set are clustered using the original distance to obtain K cluster centers, where K is a natural number; Calculate the first distance from each initial sample feature in the training set to the cluster center, and sort the initial sample features in the training set based on the feature similarity corresponding to the first distance to obtain a first sorting result; The initial sample features in the training set are used as the target sample features and added to the target feature space according to the first sorting result.

10. The method according to claim 8, characterized in that, The step of obtaining the target clustering result of each target sample feature in the target feature space based on the first co-correlation matrix and the second co-correlation matrix includes: Use a first threshold corresponding to the first co-co ... If an element in the first distance matrix and / or an element in the second distance matrix needs to be adjusted, and the current iteration number is less than or equal to the target threshold, the second threshold is used to determine whether an element in the second distance matrix needs to be adjusted, until an element in the first distance matrix and / or an element in the second distance matrix does not need to be adjusted, or the current iteration number is greater than the target threshold. Without needing to adjust the elements in the first distance matrix and the elements in the second distance matrix, the target clustering result is obtained using a hierarchical clustering algorithm.

11. The method according to any one of claims 8 to 10, characterized in that, The method further includes: Obtain account behavior information for each of the multiple sample accounts; Multiple first sample features are obtained based on the account behavior information corresponding to each of the multiple sample accounts, wherein each of the multiple first sample features corresponds one-to-one with each of the multiple sample accounts. The metric learning algorithm is used to process the multiple first sample features to obtain multiple second sample features; The multiple second sample features are input into the initial gender recognition model for training to obtain a trained gender recognition model; The trained gender recognition model is used to identify the gender information of the accounts associated with the accounts to be identified.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program can be executed by a terminal device or computer to perform the method described in any one of claims 1 to 7 or 8 to 11.

13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 7 or 8 to 11.

14. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 7 or 8 to 11 via the computer program.

Images