Object classification method, apparatus, storage medium, device, and product

Abstract

The application discloses an object classification method and device, a storage medium, equipment and products, relates to the technical field of artificial intelligence, and can be applied to the technical field of blockchains, map Internet of Vehicles and the like. The method comprises the following steps: acquiring object information of at least one object, wherein the object information comprises at least one product category; respectively classifying objects associated with each product category to obtain same-category objects of each object under each associated product category; taking each object as a target object, respectively generating enhanced feature data under each associated product category based on the object information of the same-category objects of the target object under each associated product category; performing calculation and processing on the enhanced feature data under each associated product category and the same-category object weight to obtain object feature data of the target object; and obtaining a first classification result of the target object according to the object feature data of the target object. The application can effectively improve the accuracy of object classification.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an object classification method, apparatus, storage medium, device, and product. Background Technology

[0002] Object classification is the process of categorizing objects into classes. For example, object classification can determine the default category or consumption intention category of an object. Current object classification methods employ classification models based on machine learning and deep learning algorithms.

[0003] However, current object classification methods mostly rely heavily on the extraction of static features and temporal behavioral features of the objects to be classified, resulting in low object classification accuracy. Summary of the Invention

[0004] This application provides an object classification method, apparatus, storage medium, device, and product, which can effectively improve the accuracy of object classification.

[0005] To address the aforementioned technical problems, this application provides the following technical solutions:

[0006] According to one embodiment of this application, an object classification method is provided, the method comprising: obtaining object information of at least one object, the object information including at least one product category; classifying the objects associated with each product category to obtain similar objects of each object under each associated product category, wherein the associated product category is the product category associated with each object; taking each object as a target object, generating enhanced feature data under each associated product category based on the object information of similar objects of the target object under each associated product category; calculating the enhanced feature data under each associated product category with the weights of similar objects to obtain object feature data of the target object, wherein the weights of similar objects are the weights of similar objects under the corresponding associated product category; and obtaining a first classification result of the target object based on the object feature data of the target object.

[0007] According to one embodiment of this application, an object classification apparatus includes: an acquisition module for acquiring object information of at least one object, the object information including at least one product category; a classification module for classifying objects associated with each product category to obtain similar objects of each object under each associated product category, the associated product category being the product category associated with each object; a generation module for generating enhanced feature data for each associated product category, based on the object information of similar objects of the target object under each associated product category, with each object as the target object; a calculation module for calculating the enhanced feature data under each associated product category and the weights of similar objects to obtain object feature data of the target object, the weights of similar objects being the weights of similar objects under the corresponding associated product category; and an output module for obtaining a first classification result of the target object based on the object feature data of the target object.

[0008] In some embodiments of this application, the weight of similar objects is an object attention weight; the calculation module includes: a transformation unit, used to perform linear transformation processing on the enhanced feature data under each associated product category to obtain transformed feature data under each associated product category; an object attention weight calculation unit, used to perform attention calculation processing on the transformed feature data under each associated product category and a preset object attention parameter to obtain the object attention weight of neighboring objects under each associated product category; a calculation unit, used to calculate processing on the transformed feature data under each associated product category and the object attention weight to obtain first feature data under each associated product category; and a generation unit, used to generate object feature data of the target object based on the first feature data.

[0009] In some embodiments of this application, the generation unit includes: a category weight calculation subunit, used to calculate and process the first feature data under each associated product category and the category weight corresponding to each associated product category respectively to obtain the second feature data under each associated product category; and a feature data generation subunit, used to generate object feature data of the target object based on the second feature data under each associated product category.

[0010] In some embodiments of this application, the object information includes object behavior information, and the category weight is a category attention weight. The category weight calculation subunit is used to: generate behavioral feature data based on the object behavior information under each associated product category to obtain the behavioral feature dataset of the target object; determine the non-core category in the associated product category corresponding to the target object based on the behavioral feature data under each associated product category; hide the behavioral feature data corresponding to the non-core category in the behavioral feature dataset to obtain a reset behavioral feature dataset; perform attention calculation processing based on the reset behavioral feature dataset, preset category attention parameters, and the first feature data to obtain the category attention weight of each associated product category; and calculate the first feature data and the category attention weight under each associated product category to obtain the second feature data under each associated product category.

[0011] In some embodiments of this application, the category weight calculation subunit is configured to: determine the object corresponding to each of the associated product categories from the at least one object, to obtain an object set corresponding to each of the associated product categories; for each object set corresponding to the associated product category, compare and sort the objects in the object set according to the behavioral feature data of each object, to obtain the relative ranking of each object in each object set; determine the relative ranking of the target object in each object set based on the relative ranking of each object in each object set; and determine the non-core category in the associated product categories corresponding to each object set based on the relative ranking of the target object in each object set.

[0012] In some embodiments of this application, the feature data generation subunit is used to: add the second feature data under each associated product category to obtain superimposed second feature data; and use the superimposed second feature data as the object feature data of the target object.

[0013] In some embodiments of this application, the first classification result is obtained using a first classification model, and the device further includes an auxiliary classification module for: generating target feature data based on the object information of the target object; classifying the target feature data using a second classification model to obtain a second classification result corresponding to the target object, wherein the second classification model is trained by mutual learning with the first classification model; and determining the object category of the target object based on at least one of the first classification result and the second classification result.

[0014] In some embodiments of this application, the apparatus further includes a first classification model training module, configured to: classify training object samples using a preset first classification model to obtain a first predicted classification result corresponding to the training object samples, wherein the training object samples are labeled with a predetermined classification result; classify training object samples using a preset second classification model to obtain a second predicted classification result corresponding to the training object samples; analyze a first classification loss of the preset first classification model based on the predetermined classification result and the first predicted classification result; analyze a first matching loss between the preset first classification model and the preset second classification model based on the first predicted classification result and the second predicted classification result; calculate a first total classification loss of the preset first classification model based on the first classification loss and the first matching loss; and adjust the preset first classification model based on the first total classification loss until predetermined training conditions are met to obtain the trained first classification model.

[0015] In some embodiments of this application, the apparatus further includes a second classification model training module, configured to: analyze the second classification loss of the preset second classification model based on the predetermined classification result and the second predicted classification result; analyze the second matching loss between the preset second classification model and the preset first classification model based on the second predicted classification result and the first predicted classification result; calculate the second total classification loss of the preset second classification model based on the second classification loss and the second matching loss; and adjust the preset second classification model based on the second total classification loss until predetermined training conditions are met, thereby obtaining the trained second classification model.

[0016] In some embodiments of this application, the first classification result is a first default probability, and the second classification result is a second default probability; the auxiliary classification module is used to: determine the default category of the target object based on at least one of the first default probability and the second default probability.

[0017] In some embodiments of this application, the partitioning module is configured to: determine objects corresponding to each product category from the at least one object, and obtain a set of sub-objects corresponding to each product category; classify the objects in the set of sub-objects for each product category to obtain at least one cluster of similar objects under each product category; and determine the other objects in each cluster of similar objects (excluding the object itself) as similar objects under each associated product category, wherein the associated product category is the product category corresponding to the cluster of similar objects.

[0018] In some embodiments of this application, each associated product category includes at least one of the aforementioned similar objects; the generation module is configured to: generate similar object feature data for each similar object based on the object information of each similar object; and combine the similar object feature data of at least one similar object under each associated product category to obtain enhanced feature data for each associated product category.

[0019] In some embodiments of this application, the output module is configured to: perform a fully connected operation on the object feature data of the target object to obtain the fully connected operation result; and classify the fully connected operation result to obtain a first classification result of the target object.

[0020] According to another embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a computer's processor, causes the computer to perform the methods described in the embodiments of this application.

[0021] According to another embodiment of this application, an electronic device includes: a memory storing a computer program; and a processor reading the computer program stored in the memory to execute the methods described in the embodiments of this application.

[0022] According to another embodiment of this application, a computer program product or computer program 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 methods provided in the various optional implementations described in the embodiments of this application.

[0023] In this embodiment, object information of at least one object is obtained, the object information including at least one product category; objects associated with each product category are classified to obtain similar objects of each object under each associated product category, the associated product category being the product category associated with each object; taking each object as the target object, enhanced feature data under each associated product category is generated based on the object information of similar objects of the target object under each associated product category; the enhanced feature data under each associated product category is calculated and processed with the weight of similar objects to obtain object feature data of the target object, the weight of similar objects being the weight of similar objects under the corresponding associated product category; based on the object feature data of the target object, a first classification result of the target object is obtained.

[0024] In this way, each object is identified as belonging to the same category of related products. Based on the object information of the same category, enhanced feature data is generated for each related product category. The enhanced feature data can reflect the relationship between objects in different product categories and contains rich and effective representation information. Furthermore, object feature data is calculated by combining the weights of the same category objects with the enhanced feature data, so that the object feature data contains rich and accurate representation information. Based on the object feature data, an accurate first classification result can be obtained, thereby effectively improving the accuracy of object classification. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A schematic diagram of a system to which embodiments of this application can be applied is shown.

[0027] Figure 2 A flowchart of an object classification method according to an embodiment of this application is shown.

[0028] Figure 3 A flowchart of an object feature data generation method according to an embodiment of this application is shown.

[0029] Figure 4 A flowchart of an object classification method according to another embodiment of this application is shown.

[0030] Figure 5 A flowchart illustrating object classification using embodiments of this application in a given scenario is shown.

[0031] Figure 6 A block diagram of an object classification apparatus according to another embodiment of this application is shown.

[0032] Figure 7 A block diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] It is understood that in the specific implementation of this application, data such as object information are involved. When the above embodiments of this application are applied to specific products or technologies, user 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.

[0035] Figure 1 A schematic diagram of a system 100 to which embodiments of this application can be applied is shown. For example... Figure 1 As shown, system 100 may include server 110 and / or terminal 120.

[0036] Server 110 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides 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.

[0037] Terminal 120 can be any device, including but not limited to mobile phones, computers, smart voice interaction devices, smart home appliances, in-vehicle terminals, VR / AR devices, smartwatches, and computers, etc. In some embodiments, terminal 120 and server 110 can be nodes in a blockchain network.

[0038] In one embodiment of this example, server 110 or terminal 120 may: obtain object information of at least one object, the object information including at least one product category; classify the objects associated with each product category to obtain similar objects of each object under each associated product category, the associated product category being the product category associated with each object; take each object as the target object, and based on the object information of similar objects of the target object under each associated product category, generate enhanced feature data for each associated product category; calculate and process the enhanced feature data under each associated product category with the weights of similar objects to obtain object feature data of the target object, the weights of similar objects being the weights of similar objects under the corresponding associated product category; and obtain a first classification result of the target object based on the object feature data of the target object.

[0039] In some implementations, at least one object may be a relevant user on a map-based vehicle networking platform. In some implementations, at least one object may be a relevant user on a shopping platform. It is understood that in other implementations, at least one object may be selected based on the actual situation.

[0040] Figure 2A flowchart illustrating an embodiment of an object classification method according to this application is shown schematically. The entity executing this object classification method can be any device, such as... Figure 1 The server 110 or terminal 120 shown.

[0041] like Figure 2 As shown, the object classification method may include steps S210 to S250.

[0042] Step S210: Obtain object information for at least one object, wherein the object information includes at least one product category;

[0043] Step S220: Classify the objects associated with each product category to obtain the same type of objects for each object under each associated product category. The associated product category is the product category associated with each object.

[0044] Step S230: Taking each object as the target object, based on the object information of the target object in the same type of object under each associated product category, generate enhanced feature data for each associated product category respectively;

[0045] Step S240: Calculate and process the enhanced feature data of each associated product category and the weight of similar objects to obtain the object feature data of the target object. The weight of similar objects is the weight of similar objects in the corresponding associated product category.

[0046] Step S250: Based on the object feature data of the target object, obtain the first classification result of the target object.

[0047] At least one object's information can be retrieved from a predetermined location. Each object has its own unique object information, which includes static attribute information and behavioral information. Static attribute information includes fields such as gender, age, marital status, and education level. Behavioral information includes fields such as transaction time, transaction amount, product category, and merchant name, as well as average transaction amount and number of transactions, average transaction amount and number of transactions under a predetermined number of product categories, average transaction amount and number of transactions on weekdays or weekends, and average transaction amount and number of transactions during the day or night.

[0048] Product categories refer to the categories of products that an object is associated with, such as the categories of products that the object has consumed. Specific product categories include internet services, leisure, and public welfare. Each object's object information can include at least one product category. It can be understood that different objects can have the same, different, or partially the same product categories.

[0049] For example, at least one object includes object A and object B. The object information of A includes product categories LM1, LM2 and LM3, and the object information of B includes product categories LM1, LM2 and LM4. In this case, A and B may have some of the same product categories LM1 and LM2.

[0050] For each product category included in the object information of at least one object, the objects associated with each product category in at least one object can be classified separately. For example, all objects (such as A and B and other objects) corresponding to product category LM1 in at least one object can be classified to obtain at least one cluster of objects of the same type under product category LM1. Similarly, at least one cluster of objects of the same type under other product categories can be obtained.

[0051] Furthermore, it is possible to determine the similar objects of each object under each associated product category. For example, if a cluster of similar objects under product category LM1 includes objects A, B, and F, then objects A under its associated product category LM1 include objects B and F.

[0052] For each object as the target object, based on the object information of the target object in each related product category, enhanced feature data is generated for each related product category. For example, if object A has objects B and F in its related product category LM1, enhanced feature data for object A in related product category LM1 can be generated based on the object information of objects B and F. In this way, enhanced feature data can be generated for each related product category of object A.

[0053] The weight of objects of the same category is the weight of objects of the same category under the corresponding related product category. For example, objects of the same category under the related product category LM1 include B and F. B and F can each correspond to a weight of objects of the same category under the related product category LM1. Objects of the same category under the related product category LM2 include D and E. D and E can each correspond to a weight of objects of the same category under the related product category LM2.

[0054] The enhanced feature data of each associated product category is calculated and processed with the weights of similar objects to obtain the object feature data of the target object. For example, the enhanced feature data of object A under associated product category LM1 is calculated and processed with the weights of similar objects B and F under associated product category LM1, and the enhanced feature data of object A under associated product category LM2 is calculated and processed with the weights of similar objects D and E under associated product category LM2 to obtain the object feature data of object A.

[0055] After obtaining the object feature data of the target objects (i.e., each object in at least one category), the object feature data can accurately characterize the target objects, and the first classification result of the target objects can be accurately obtained based on the object feature data. The first classification result is, for example, the probability of default or the probability of purchasing an item.

[0056] In this way, based on steps S210 to S250, the similar objects of each object under each associated product category are determined. Based on the object information of the similar objects, enhanced feature data under each associated product category is generated. The enhanced feature data can reflect the association relationship of objects under different product categories. The enhanced feature data contains rich and effective representation information. Furthermore, object feature data is calculated by combining the weights of similar objects under the associated product categories with the enhanced feature data, so that the object feature data contains rich and accurate representation information. Based on the object feature data, an accurate first classification result can be obtained, thereby effectively improving the accuracy of object classification.

[0057] The following description Figure 2 Other further embodiments include those that perform object classification in the examples.

[0058] In one embodiment, step S220 involves classifying the objects associated with each product category to obtain similar objects for each associated product category, including:

[0059] Identify the objects corresponding to each product category from at least one object to obtain a set of sub-objects corresponding to each product category; classify the objects in the set of sub-objects for each product category to obtain at least one cluster of objects of the same type under each product category; identify the other objects in each cluster of objects of the same type (excluding the object itself) as objects of the same type under each associated product category, and the associated product category is the product category corresponding to the cluster of objects of the same type.

[0060] For example, if product category LM1 corresponds to all objects including A, B, F, and X in at least one object (which may include A, B, etc.), then the sub-object set corresponding to product category LM1 includes objects A, B, F, and X. Similarly, the sub-object set corresponding to each product category can be determined.

[0061] For each product category, the objects in the sub-object set are classified. For example, objects A, B, F and X in the sub-object set corresponding to product category LM1 are classified to obtain at least one cluster of objects of the same type under product category LM1. Similarly, at least one cluster of objects of the same type under other product categories can be obtained.

[0062] Furthermore, the other objects in each cluster of objects of the same type, excluding the object itself, can be identified as the same type of objects under each associated product category. For example, the associated product categories of object A include LM1 and LM5. If a cluster of objects of the same type 1 under product category LM1 includes objects A, B, and F, then the same type of objects of object A under its associated product category LM1 include B and F. B and F are the other objects in the same cluster of objects of the same type 1, excluding object A itself. In other words, the associated product category LM1 is the product category corresponding to the same cluster of objects of the same type 1. If a cluster of objects of the same type 5 under product category LM5 includes objects A, O, and H, then the same type of objects of object A under its associated product category LM5 include O and H. O and H are the other objects in the same cluster of objects of the same type 5, excluding object A itself. In other words, the associated product category LM5 is the product category corresponding to the same cluster of objects of the same type 5.

[0063] In one embodiment, when classifying the objects in the sub-object set corresponding to each product category to obtain at least one cluster of objects of the same type under each product category, the k-Nearest Neighbor algorithm can be used to classify the sub-object set U under each product category. C The objects are classified separately to obtain at least one cluster of objects of the same type under each product category.

[0064] In other embodiments, other existing clustering algorithms can be used to classify objects in the sub-object set.

[0065] In one embodiment, each associated product category includes at least one object of the same type; step S230, based on the object information of the target object in each associated product category, generates enhanced feature data for each associated product category, including:

[0066] Based on the object information of each object of the same type, generate object feature data of the same type for each object of the same type; combine the object feature data of at least one object of the same type under each related product category to obtain enhanced feature data under each related product category.

[0067] Generate characteristic data for each object of the same type based on the object information of each object of the same type. For example, target object u i All corresponding related product categories (i.e., the set of related product categories) It includes at least one related product category c, and the target object u. i Similar objects under a related product category c are represented as follows: Based on each object of the same type The object information can be used to generate object u. iFeature data of each similar object under the related product category C in Represents the j-th object of the same type. The characteristic data of objects of the same type, where p is the number of objects of the same type.

[0068] Object u i At least one type of object characteristic data of the same type of object Object u can be generated by combining elements in ways such as addition or concatenation. i Enhanced feature data under the associated product category c. In one embodiment, according to the formula... Collect the characteristic data of at least one similar object under the related product category c. By concatenating and combining the data, we can obtain the enhanced feature data under the related product category c. in, It can represent a series connection.

[0069] In one embodiment, see Figure 3 The weights of similar objects are the object attention weights; step S240, the enhanced feature data under each associated product category and the weights of similar objects are calculated and processed to obtain the object feature data of the target object, including:

[0070] Step S241: Perform linear transformation on the enhanced feature data under each associated product category to obtain the transformed feature data under each associated product category; Step S242: Perform attention operation on the transformed feature data under each associated product category and the preset object attention parameters to obtain the object attention weight of the neighboring objects under each associated product category; Step S243: Calculate the transformed feature data and the object attention weight under each associated product category to obtain the first feature data under each associated product category; Step S244: Generate the object feature data of the target object based on the first feature data.

[0071] By using a linear transformation function (such as the ReLU function), the enhanced feature data under each associated product category can be linearly transformed to obtain the transformed feature data under each associated product category.

[0072] In one implementation, enhanced feature data under the associated product category c can be... Through formula Perform linear transformation to obtain enhanced feature data under the associated product category c. in, and Let these represent the weight matrix and bias vector, respectively, and D be the augmented feature data. The feature dimension is d, where d is the dimension of the latent representation.

[0073] The attention mechanism allows for the processing of transformed feature data and preset object attention parameters under each related product category through attention operations.

[0074] In one implementation, the formula for the attention mechanism implemented using the softmax function is... and Transformed feature data for each associated product category c With preset object attention parameters ( and Attention operations are performed to obtain the object attention weights of objects j of the same type under each associated product category c. in, and These are the weight matrix and bias vector, respectively; tanh is the hyperbolic tangent function; and P is the value of the object u. i The number of similar objects under the associated product category c For P objects of the same type, the j-th neighbor object For P objects of the same type, the l-th neighbor object

[0075] By performing calculations such as dot product or summation on the transformed feature data and object attention weights under each related product category, the first feature data under each related product category can be obtained.

[0076] In one implementation, according to the formula Transformed feature data for each associated product category c With object attention weights Perform dot product calculations separately to obtain the first feature data under each associated product category c. Where, ⊙ represents and Multiply the elements one by one. The j-th element can be The j-th element can be middle Obtained after linear transformation.

[0077] When generating object feature data of a target object based on the first feature data, the first feature data under each associated product category can be directly combined (e.g., concatenated) as object feature data. Alternatively, the object feature data of the target object can be generated after calculating the second feature data under each associated product category based on subsequent embodiments.

[0078] In this way, the attention weight of neighboring objects under each associated product category is calculated through attention operations. The object attention weight can accurately represent the contribution of neighboring objects to the object classification under the corresponding associated product category. Based on the attention weight, the first feature data under each associated product category is calculated, and then the object feature data is generated, which can further improve the accuracy of the object feature data in representing the target object.

[0079] In one embodiment, the weight of objects of the same type is a preset object weight; step S240, calculates and processes the enhanced feature data of each associated product category with the weight of objects of the same type to obtain the object feature data of the target object, including: multiplying the enhanced feature data of each associated product category with the preset object weight to obtain the object feature data of the target object.

[0080] In one embodiment, step S244, generating object feature data of the target object based on the first feature data, includes: calculating and processing the first feature data under each associated product category and the category weight corresponding to each associated product category to obtain the second feature data under each associated product category; and generating object feature data of the target object based on the second feature data under each associated product category.

[0081] Each associated product category can correspond to a category weight, which reflects the degree to which the associated product category contributes to object classification. By calculating and processing the first feature data and the corresponding category weight for each associated product category, the first feature data under each associated product category can be further adjusted at the product category level to obtain the second feature data for each associated product category. Based on the second feature data, object feature data can be generated more accurately, improving the accuracy of object classification.

[0082] In one embodiment, the object information includes object behavior information, and the category weight is the category attention weight; the first feature data under each associated product category and the category weight corresponding to each associated product category are calculated and processed respectively to obtain the second feature data under each associated product category, including:

[0083] Behavioral feature data is generated based on the object behavior information under each associated product category to obtain the target object's behavioral feature dataset. Based on the behavioral feature data under each associated product category, non-core categories within the associated product categories corresponding to the target object are determined. The behavioral feature data corresponding to the non-core categories in the behavioral feature dataset is hidden to obtain a reset behavioral feature dataset. Attention calculation is performed based on the reset behavioral feature dataset, preset category attention parameters, and the first feature data to obtain the category attention weight for each associated product category. The first feature data and category attention weight under each associated product category are calculated separately to obtain the second feature data for each associated product category.

[0084] Based on target object u i Behavioral feature data can be generated from the object behavior information under each associated product category c. Get the target object u i Behavioral feature dataset and then,

[0085] Based on the behavioral characteristic data of each associated product category c By conducting analysis, the target object u can be determined. i All corresponding related product categories (i.e., the set of related product categories) Non-core categories that fall below the predetermined importance criteria; non-core categories include all related product categories. Related product category c that is below the predetermined importance condition.

[0086] behavioral feature dataset Behavioral characteristic data corresponding to core categories in China and Africa Hiding the data and obtaining the reset behavioral feature dataset. Among them, behavioral characteristic data corresponding to non-core categories The method of hiding can be to hide the behavioral feature data corresponding to non-core categories. Reset to 0 or a specific value.

[0087] Based on the attention mechanism, and using the reset behavioral feature dataset Preset category attention parameters and first feature data By performing attention calculations, the target object u can be obtained. i The category attention weight β for each associated product category c i,c .

[0088] In one implementation, the formula for the attention mechanism implemented using the softmax function is... and Reset behavioral feature dataset Preset category attention parameter W for each associated product category c c and first feature data By performing attention calculations separately, the target object u can be obtained. i The category attention weight β for each associated product category c i,c ,in, For the reset behavioral feature dataset target object u i In the behavioral feature data under the associated product category c, tanh is the hyperbolic tangent function, and g(·) is the ReLU function.

[0089] Furthermore, the first feature data under each associated product category c... Category attention weight β i,c The second feature data for each associated product category c can be obtained by multiplying them separately.

[0090] In this way, the category attention weight of each associated product category corresponding to the target object is calculated by using an attention mechanism and hiding the behavioral feature data corresponding to non-core categories. The category attention weight of each associated product category can further accurately reflect the contribution of associated product categories to object classification, thereby further improving the accuracy of object classification. Furthermore, by selectively ignoring unimportant representation information through hiding, the overall classification efficiency is improved.

[0091] In one embodiment, based on behavioral characteristic data under each associated product category, the non-core categories in the associated product categories corresponding to the target object are determined, including:

[0092] Identify the objects corresponding to each associated product category from at least one object, thus obtaining an object set corresponding to each associated product category; for each object set corresponding to an associated product category, compare and sort the objects in the object set according to the behavioral characteristic data of each object, thus obtaining the relative ranking of each object in each object set; based on the relative ranking of each object in each object set, determine the relative ranking of the target object in each object set; based on the relative ranking of the target object in each object set, determine the non-core categories in the associated product categories corresponding to each object set.

[0093] Let C be the set of all product categories included in the training object samples of at least one object, and let u be the target object. i The set of all related product categories c is represented as follows: From at least one object, the object corresponding to each associated product category c can be determined, resulting in the object set U corresponding to each associated product category c.C ,

[0094] For each associated product category c, the corresponding object set U C , to set the object U C The objects are compared and sorted according to their behavioral characteristics under the associated product category c, resulting in object sets U. C The relative ranking of each object in the set U. C The behavioral feature data of each object in the dataset are processed by weighted summation or normalization, and then ranked by comparison to obtain the object set U. C Each object in the object set U C The relative ranking within.

[0095] Therefore, the relative ranking of each object within all its associated product categories can be calculated. Then, based on the relative ranking of each object in each object set, the target object u can be determined. i Relative ranking in each object set

[0096] According to the target object u i Relative ranking in each object set It can generate a ranking vector r i According to the ranking vector r i It can calculate the target object u i exist The ranking of all related product categories c within the target object u. i Relative ranking in each object set Sort the objects in descending order to obtain the ranking of each object and its corresponding associated product category c. The top N ranked associated product categories can be selected as the user's primary product categories, denoted as N. cate If related product categories It does not belong to the target object u i Main product category N cate If so, then the associated product category c is a non-core category.

[0097] In one embodiment, generating object feature data of a target object based on the second feature data under each associated product category includes: adding the second feature data under each associated product category to obtain superimposed second feature data; and using the superimposed second feature data as the object feature data of the target object.

[0098] For example, you can follow the formula target object u i All related product categories Second feature data under each associated product category c Add them together to obtain the superimposed second feature data v. i The superimposed second feature data v i As object feature data of the target object.

[0099] In other embodiments, generating object feature data of the target object based on the second feature data under each associated product category may include: concatenating the second feature data under each associated product category and using it as the object feature data of the target object.

[0100] In one embodiment, obtaining a first classification result of the target object based on the object feature data of the target object includes: performing a fully connected operation on the object feature data of the target object to obtain a fully connected operation result; and performing classification processing on the fully connected operation result to obtain a first classification result of the target object.

[0101] Inputting the object feature data of the target object into a fully connected neural network allows for fully connected computation to be performed, yielding the fully connected computation result. The fully connected computation result is then classified based on a classification function (such as the Softmax function) to obtain the first classification result of the target object, such as the probability of default or the probability of purchase.

[0102] For further details, please refer to [link / reference]. Figure 4 In one embodiment, the first classification result is obtained by using a first classification model. The object classification method further includes: step S310, generating target feature data based on the object information of the target object; step S320, classifying the target feature data using a second classification model to obtain a second classification result corresponding to the target object, wherein the second classification model is trained by mutual learning with the first classification model; and step S330, determining the object category of the target object based on at least one of the first classification result and the second classification result.

[0103] The first classification model can execute steps S220 to S250 to obtain the first classification result. The second classification model is trained in advance by mutual learning with the first classification model. The second classification model is used to classify the target feature data to obtain the second classification result corresponding to the target object. The second classification result can be, for example, the default probability or purchase probability of the target object.

[0104] In this way, a mutual learning strategy is used to improve the classification performance of the two classification models by enabling them to learn from each other. The two models can accurately classify objects by considering completely different information dimensions. Furthermore, the object category of the target object can be more accurately determined based on at least one of the first and second classification results.

[0105] In one embodiment, the second classification model is a neural network model incorporating an attention mechanism (e.g., an LSTM model). When generating target feature data based on the object information of the target object, the feature vectors of each piece of information in the object information can be concatenated according to the time nodes of the information to obtain the target feature data. The target feature data can be represented as {x} i1 ,x i2 ,x i3 ,……,x iK}, where K is the length of the sequence, x ik Represents the target object u i Based on the feature vector of information at time node k, the second classification model can extract the dynamic behavior features of the target object based on the serialized target feature data, and classify it based on the dynamic behavior features to output the second classification result.

[0106] In one embodiment, the method further includes training a first classification model through mutual learning according to the following steps: classifying training object samples using a preset first classification model to obtain a first predicted classification result corresponding to the training object samples, wherein the training object samples are labeled with a predetermined classification result; classifying training object samples using a preset second classification model to obtain a second predicted classification result corresponding to the training object samples; analyzing a first classification loss of the preset first classification model based on the predetermined classification result and the first predicted classification result; analyzing a first matching loss between the preset first classification model and the preset second classification model based on the first predicted classification result and the second predicted classification result; calculating a first total classification loss of the preset first classification model based on the first classification loss and the first matching loss; and adjusting the preset first classification model based on the first total classification loss until predetermined training conditions are met to obtain a trained first classification model.

[0107] The preset first classification model is the first classification model to be trained, and the preset second classification model is the second classification model to be trained. The training object samples are the object information collected for training the models. Each training object sample is labeled with a predetermined classification result, such as a predetermined default probability.

[0108] The preset first classification model can classify the training object samples according to the classification logic of steps S220 to S250 to obtain the first predicted classification result corresponding to the training object sample, such as the first predicted default probability.

[0109] The training object samples can be classified according to the classification logic of the preset second classification model to obtain the second predicted classification result corresponding to the training object samples. The second predicted classification result is, for example, the second predicted default probability.

[0110] Based on the predetermined classification result and the first predicted classification result, the first classification loss of the preset first classification model is analyzed. The first classification loss is the traditional supervised model loss. The predetermined classification result and the first predicted classification result can be calculated and processed based on the supervised model loss function to obtain the first classification loss of the preset first classification model.

[0111] In one approach, the supervised model loss function uses the cross-entropy loss function, which can be based on the cross-entropy loss function. For the predetermined classification result y i and the first predicted classification results The calculation process is performed to obtain the first classification loss of the preset first classification model. Where M is the number of first predicted classification results.

[0112] Based on the difference between the first and second predicted classification results, the first matching loss between the preset first classification model and the preset second classification model can be analyzed.

[0113] In one approach, the Kullback-Leibler divergence can be used to calculate the first matching loss between a pre-defined first classification model and a pre-defined second classification model. The KL divergence can measure the first predicted classification result. and the second predicted classification results The degree of matching between two probability distributions; the greater the difference between the two distributions, the greater the value of the KL divergence. Specifically, this can be based on the formula... Measure the first predicted classification result and the second predicted classification results KL divergence D KL (p2||p1), KL divergence D KL (p2||p1) is the first matching loss. In this formula, N is the total number of training samples, and M is the number of first and second predicted classification results.

[0114] By summing or adding the first classification loss and the first matching loss, the first total classification loss of the preset first classification model can be accurately calculated.

[0115] In one approach, it can be done according to the formula. Classification of loss and the first matching loss The summation calculation is performed to obtain the first total classification loss. Where λ is the adjustment parameter.

[0116] The model parameters of the preset first classification model are adjusted based on the first total classification loss until the predetermined training conditions (such as the learning rate or classification accuracy meeting the requirements) are met, and the trained first classification model is obtained. In this way, a first classification model with superior classification performance can be obtained.

[0117] In one embodiment, the method further includes training a second classification model through mutual learning according to the following steps:

[0118] Based on the predetermined classification result and the second predicted classification result, analyze the second classification loss of the preset second classification model; based on the second predicted classification result and the first predicted classification result, analyze the second matching loss between the preset second classification model and the preset first classification model; calculate the second total classification loss of the preset second classification model based on the second classification loss and the second matching loss; adjust the preset second classification model based on the second total classification loss until the predetermined training conditions are met, and obtain the trained second classification model.

[0119] Based on the predetermined classification result and the second predicted classification result, the second classification loss of the preset second classification model is analyzed. The second classification loss is the traditional supervised model loss. It can be calculated and processed based on the predetermined classification result and the second predicted classification result to obtain the second classification loss of the preset second classification model.

[0120] In one approach, the supervised model loss function uses the cross-entropy loss function, which can be based on the cross-entropy loss function. For the predetermined classification result y i and the second predicted classification results The calculation process is performed to obtain the second classification loss of the preset second classification model. Where M is the number of second predicted classification results.

[0121] By performing subtraction and other analyses based on the second and first predicted classification results, the second matching loss between the preset second classification model and the preset first classification model can be analyzed.

[0122] In one approach, the Kullback-Leibler divergence (KL divergence) can be used to calculate the second matching loss between the pre-defined second classification model and the pre-defined first classification model. The KL divergence can measure the second predicted classification result. and the first predicted classification results The degree of matching between two probability distributions; the greater the difference between the two distributions, the greater the value of the KL divergence. Specifically, it can be based on the formula... Measuring the second predicted classification result and the first predicted classification results KL divergence D KL(p1||p2), KL divergence D KL (p1||p2) is the second matching loss. In this formula, N is the total number of training samples, and M is the number of first and second predicted classification results.

[0123] By summing or adding the second classification loss and the second matching loss, the second total classification loss of the preset second classification model can be accurately calculated.

[0124] In one approach, it can be done according to the formula. Second category loss and the second matching loss The summation calculation is performed to obtain the second total category loss. Where λ is the adjustment parameter.

[0125] The model parameters of the preset second classification model are adjusted based on the second total classification loss until the predetermined training conditions (such as the learning rate or classification accuracy meeting the requirements) are met, and the trained second classification model is obtained. In this way, a second classification model with superior classification performance can be obtained.

[0126] In one embodiment, the first classification result is a first default probability, and the second classification result is a second default probability; determining the object category of the target object based on at least one of the first classification result and the second classification result includes: determining the default category of the target object based on at least one of the first default probability and the second default probability.

[0127] In one implementation, the default category of the target object can be determined as a default type when at least one of the first default probability and the second default probability is higher than a predetermined threshold.

[0128] In one implementation, determining the object category of a target object based on at least one of a first classification result and a second classification result includes: inputting the first classification result and the second classification result into a logistic regression model jointly trained with the first classification model and the second classification model for analysis and processing to obtain the object category of the target object. In this way, it is possible to avoid assigning weights to the first and second classification results, thereby improving the accuracy of object classification.

[0129] The above embodiments are further described below with reference to an application scenario for object classification. In this scenario, object classification is performed by applying the above embodiments of this application to predict the default type of the objects. See also Figure 5 The flowchart for object classification in this scenario is shown.

[0130] See Figure 5In this scenario, the object classification process may include steps S410 to S440.

[0131] Step S410: Obtain training object samples of at least one object. Each training object sample includes at least one product category. The training object sample is the object information of the collected objects. Each training object sample is labeled with a predetermined classification result, which is a predetermined default probability.

[0132] Object information can include static attribute information and object behavior information. Static attribute information can include gender, age, marital status, education level, etc. Object behavior information can include fields such as transaction time, transaction amount, product category, merchant name, as well as average transaction amount and number of transactions, average transaction amount and number of transactions under a predetermined number of product categories, average transaction amount and number of transactions on weekdays or weekends, and average transaction amount and number of transactions during the day or night. All object information is collected with user authorization, and all users' personal privacy information has been anonymized.

[0133] Step S420: The training object samples are classified using a preset first classification model to obtain the first predicted classification result corresponding to the training object samples.

[0134] Specifically, the first classification model is preset as follows: Figure 5 The Sub-modelθ1 shown can execute steps S421 to S425, classifying the training object samples according to the classification logic of steps S220 to S250.

[0135] In step S421, a Consumption-based Relation Matrix is ​​constructed.

[0136] Specifically, (1) the objects associated with each product category are classified and processed separately to obtain the same type of object under each associated product category. The associated product category is the product category associated with each object.

[0137] The objects associated with each product category are classified separately to obtain the same type of objects for each object under each associated product category. This further includes: determining the objects corresponding to each product category from at least one object to obtain a set of sub-objects corresponding to each product category; classifying the objects in the sub-object sets corresponding to each product category to obtain at least one cluster of objects of the same type under each product category; and determining the other objects in each cluster of objects of the same type (excluding the object itself) as the same type of objects (i.e., neighboring objects) of each object under each associated product category, where the associated product category is the product category corresponding to the cluster of objects of the same type.

[0138] For each product category's corresponding sub-object set, the objects in the sub-object set are classified to obtain at least one cluster of objects of the same type under each product category. This further includes: using the k-Nearest Neighbor algorithm to classify the sub-object set U under each product category... C The objects are classified separately to obtain at least one cluster of objects of the same type under each product category.

[0139] Furthermore, for each product category in the training object sample of at least one object, the objects associated with each product category in at least one object can be classified separately to obtain at least one cluster of objects of the same type under each product category.

[0140] For example, classifying the objects associated with product category LM1 in at least one object (such as A and B, and other objects) yields at least one cluster of similar objects under product category LM1. Similarly, at least one cluster of similar objects under other product categories can be obtained. If a cluster of similar objects under product category LM1 includes objects A, B, and F, then object A's associated similar objects (i.e., neighboring objects) under its associated product category LM1 include B and F.

[0141] (2) Taking each object as the target object, based on the training object samples of the same type of object under each related product category, generate enhanced feature data for each related product category.

[0142] For each object as the target object, enhanced feature data of the target object under each associated product category is generated. That is, for each object in at least one object, corresponding enhanced feature data can be generated under each associated product category associated with each object.

[0143] The target object includes at least one similar object (i.e., a neighboring object) under each associated product category; based on the object information of the similar objects (i.e., neighboring objects) of the target object under each associated product category, enhanced feature data for each associated product category is generated, further including: generating similar object feature data for each similar object (i.e., neighboring object) based on the training object samples of each similar object (i.e., neighboring object); combining the similar object feature data of at least one similar object (i.e., neighboring object) under each associated product category to obtain enhanced feature data for each associated product category.

[0144] For example, target object u i All corresponding related product categories (i.e., the set of related product categories) It includes at least one related product category c, and the target object u. i Similar objects under a related product category c are represented as follows: Based on each object of the same type The object information can be used to generate object u. i Feature data of each similar object under the related product category C in Represents the j-th object of the same type. The characteristic data of similar objects, where p is the number of similar objects. According to the formula... Collect the characteristic data of at least one similar object under the related product category c. By performing a series combination, the target object u is obtained. i Enhanced feature data under related product category C in, It can represent a series connection.

[0145] Similarly, enhanced feature data for each object can be generated for each associated product category in all associated product categories (i.e., the associated product category set).

[0146] (3) Further, by taking the enhanced feature data of each object under each associated product category as the matrix element in the matrix, a Consumption-based Relation Matrix can be constructed.

[0147] In step S422, in this scenario, the Consumption-based RelationMatrix may optionally undergo further feature extraction processing.

[0148] See Figure 5The consumption-based relation matrix is ​​then input into a multilayer perceptron (MLP) for further feature extraction to obtain the relation matrix to be analyzed for each object.

[0149] In the relation matrix to be analyzed, the vertical axis represents product categories, and the horizontal axis represents objects of the same category (Neighbor objects). See [link / reference]. Figure 5 This shows a relation matrix to be analyzed for a target object. In this relation matrix, each element can be the "initial feature data of the target object's neighboring objects" (Neighbor objects). "Feature data of similar objects to be analyzed obtained through feature extraction" Each line represents the initial enhanced feature data of the target object under the associated product category c. "Enhanced feature data to be analyzed" obtained through feature extraction ".

[0150] In step S423, the Neighbor-Level PsratialAttention mechanism is executed, where the neighbor dimension is the dimension of similar objects (neighbor objects).

[0151] Specifically, the enhanced feature data under each associated product category and the weight of similar objects are calculated and processed to obtain the object feature data of the target object. The weight of similar objects is the weight of similar objects under the corresponding associated product category, and the weight of similar objects is the object attention weight.

[0152] The enhanced feature data under each associated product category is calculated and processed with the weights of similar objects to obtain the object feature data of the target object. This further includes: performing linear transformation on the enhanced feature data under each associated product category to obtain transformed feature data under each associated product category; performing attention operations on the transformed feature data under each associated product category and preset object attention parameters to obtain the object attention weights of neighboring objects under each associated product category; calculating and processing the transformed feature data under each associated product category and the object attention weights to obtain the first feature data under each associated product category; and generating the object feature data of the target object based on the first feature data.

[0153] Specifically, in this scenario, the enhanced feature data under each associated product category c will be... Through formula (i.e., the ReLU function) is used for linear transformation to obtain enhanced feature data for each associated product category c. in, and Let these represent the weight matrix and bias vector, respectively, and D be the augmented feature data. The feature dimension, d is the dimension of the latent representation.

[0154] And, the formula for the attention mechanism implemented using the softmax function. and Transformed feature data for each associated product category c With preset object attention parameters ( and Attention operations are performed to obtain the object attention weights of objects j of the same type under each associated product category c. in, and These are the weight matrix and bias vector, respectively; tanh is the hyperbolic tangent function; and P is the value of the object u. i The number of similar objects under the associated product category c For P objects of the same type, the j-th neighbor object For P objects of the same type, the l-th neighbor object

[0155] And, according to the formula Transformed feature data for each associated product category c With object attention weights Perform dot product calculations separately to obtain the first feature data under each associated product category c. Where, ⊙ represents and Multiply the elements one by one. The j-th element can be The j-th element can be middle Obtained after linear transformation.

[0156] In step S424, an attention mechanism (Category-Level PsratialAttention) is executed on the product category dimension, which is the dimension of the product category.

[0157] Specifically, generating object feature data of the target object based on the first feature data further includes: calculating and processing the first feature data under each associated product category and the category weight corresponding to each associated product category to obtain the second feature data under each associated product category; and generating object feature data of the target object based on the second feature data under each associated product category.

[0158] Category weight is a category attention weight. It involves calculating and processing the first feature data under each associated product category and the corresponding category weight for each associated product category to obtain the second feature data under each associated product category, which further includes:

[0159] Behavioral feature data is generated based on the object behavior information under each associated product category to obtain the target object's behavioral feature dataset. Based on the object behavior information under each associated product category, the non-core categories within the associated product categories corresponding to the target object are determined. The behavioral feature data corresponding to the non-core categories in the behavioral feature dataset is hidden to obtain a reset behavioral feature dataset. Attention calculation is performed based on the reset behavioral feature dataset, preset category attention parameters, and the first feature data to obtain the category attention weight for each associated product category. The first feature data and category attention weight under each associated product category are calculated separately to obtain the second feature data for each associated product category.

[0160] Specifically, based on the target object u i Behavioral feature data can be generated from the object behavior information under each associated product category c. Get the target object u i Behavioral feature dataset and then, Object behavior information refers to the object behavior information in the training object samples.

[0161] By analyzing the object behavior information under each associated product category c, the target object u can be determined. i All corresponding related product categories (i.e., the set of related product categories) Non-core categories that fall below the predetermined importance criteria; non-core categories include all related product categories. Related product category c that is below the predetermined importance condition.

[0162] behavioral feature dataset Behavioral characteristic data corresponding to core categories in China and Africa Hidden processing yields the reset behavioral feature dataset. Among them, behavioral characteristic data corresponding to non-core categories The method of hiding can be to hide the behavioral feature data corresponding to non-core categories. Reset to 0 or a specific value.

[0163] Formula for attention mechanism implemented using softmax function and Reset behavioral feature dataset Preset category attention parameter W for each associated product category c c and first feature data By performing attention calculations separately, the target object u can be obtained. i The category attention weight β for each associated product category c i,c ,in, For the reset behavioral feature dataset target object u i In the behavioral feature data under the associated product category c, tanh is the hyperbolic tangent function, and g(·) is the ReLU function.

[0164] The first feature data under each associated product category c Category attention weight β i, The second feature data for each associated product category c can be obtained by multiplying them separately.

[0165] Specifically, based on the object behavior information under each associated product category, the non-core categories in the associated product categories corresponding to the target object are determined, including:

[0166] Identify the objects corresponding to each associated product category from at least one object, thus obtaining the object set corresponding to each associated product category; for each associated product category object set, compare and sort the objects in the object set according to the object behavior information of each object, thus obtaining the relative ranking of each object in each object set; based on the relative ranking of each object in each object set, determine the relative ranking of the target object in each object set; based on the relative ranking of the target object in each object set, determine the non-core categories in the associated product categories corresponding to each object set.

[0167] Let C be the set of all product categories included in the training object samples of at least one object, and let u be the target object. i The set of all related product categories c is represented as follows: From at least one object, the object corresponding to each associated product category c can be determined, resulting in the object set U corresponding to each associated product category c. C ,

[0168] For each associated product category c, the corresponding object set U C, to set the object U C The objects are compared and sorted according to their behavioral characteristics under the associated product category c, resulting in object sets U. C The relative ranking of each object in the set U. C The behavioral feature data of each object in the dataset are processed by weighted summation or normalization, and then ranked by comparison to obtain the object set U. C Each object in the object set U C The relative ranking within.

[0169] Therefore, the relative ranking of each object within all its associated product categories can be calculated. Then, based on the relative ranking of each object in each object set, the target object u can be determined. i Relative ranking in each object set

[0170] According to the target object u i Relative ranking in each object set It can generate a ranking vector r i According to the ranking vector r i It can calculate the target object u i exist The ranking of all related product categories c within the target object u. i Relative ranking in each object set Sort the objects in descending order to obtain the ranking of each object and its corresponding associated product category c. The top N ranked associated product categories can be selected as the user's primary product categories, denoted as N. cate If related product categories It does not belong to the target object u i Main product category N cate If so, then the associated product category c is a non-core category.

[0171] Specifically, generating object feature data for the target object based on the second feature data under each associated product category includes: adding the second feature data under each associated product category to obtain superimposed second feature data; and using the superimposed second feature data as the object feature data of the target object.

[0172] Specifically, you can follow the formula. target object u i All related product categories Second feature data under each associated product category c Add them together to obtain the superimposed second feature data v. i The superimposed second feature data v i As object feature data of the target object.

[0173] In step S425, prediction processing is performed.

[0174] Specifically, based on the object feature data of the target object, the first predicted classification result of the target object is obtained, including: performing a fully connected operation on the object feature data of the target object to obtain the fully connected operation result; and performing classification processing on the fully connected operation result to obtain the first predicted classification result of the target object.

[0175] The first predicted classification result corresponds to the training object sample of the target object, and then the first predicted classification result corresponding to each training object sample is obtained. The first predicted classification result is the first predicted default probability.

[0176] See Figure 5 The object feature data of the target object is input into a multilayer perceptron (MLP). The fully connected neural network in the MLP can perform fully connected operations to obtain the fully connected operation result. The fully connected operation result is then classified based on a classification function (e.g., the Softmax function) to obtain the first predicted classification result of the target object. The first predicted classification result is the first default probability of the target object.

[0177] Step S430: The training object samples are classified using a preset second classification model to obtain the second predicted classification result corresponding to the training object samples. The preset second classification model is as follows: Figure 5 The Sub-model θ2 is shown.

[0178] The training object samples are classified using a preset second classification model to obtain the second predicted classification result corresponding to the training object samples, including: step S431, generating target feature data based on the training object samples of the target object; step S432, classifying the target feature data using a preset second classification model to obtain the second predicted classification result corresponding to the target object.

[0179] The second prediction classification result corresponds to the training object sample of the target object, and then the second prediction classification result corresponding to each training object sample is obtained. The second prediction classification result is the second prediction default probability.

[0180] The second classification model is assumed to be a neural network model incorporating an attention mechanism (LSTM model). When generating target feature data based on training samples of the target object, the feature vectors of each piece of information in the training samples can be concatenated into a serialized feature stream according to the time nodes of the information, which can be used as the target feature data. The target feature data can be represented as {x}. i1 ,x i2 ,xi3 ,……,x ik}, where K is the length of the sequence, x ik Represents the target object u i At time node k, the feature vector of the information is used to pre-determine the second classification model to extract the dynamic behavior features of the target object based on the serialized target feature data, and to classify the target object based on the dynamic behavior features, thus outputting the second predicted classification result.

[0181] Step S440, Adjusted Mutual Learning.

[0182] Specifically, based on the predetermined classification result and the first predicted classification result, the first classification loss of the preset first classification model is analyzed; based on the first predicted classification result (prediction1) and the second predicted classification result (prediction2), the first matching loss between the preset first classification model and the preset second classification model is analyzed; based on the first classification loss and the first matching loss, the first total classification loss of the preset first classification model is calculated; based on the first total classification loss, the preset first classification model is adjusted until the predetermined training conditions are met, and the trained first classification model is obtained.

[0183] Furthermore, based on the predetermined classification result and the second predicted classification result, the second classification loss of the preset second classification model is analyzed; based on the second predicted classification result and the first predicted classification result, the second matching loss between the preset second classification model and the preset first classification model is analyzed; based on the second classification loss and the second matching loss, the second total classification loss of the preset second classification model is calculated; based on the second total classification loss, the preset second classification model is adjusted until the predetermined training conditions are met, and the trained second classification model is obtained.

[0184] Based on the predetermined classification result and the first predicted classification result, the first classification loss of the predetermined first classification model is analyzed. This first classification loss is the traditional supervised model loss, which can be calculated by processing the predetermined classification result and the first predicted classification result using the supervised model loss function. The supervised model loss function uses the cross-entropy loss function. For the predetermined classification result y i and the first predicted classification results The calculation process is performed to obtain the first classification loss of the preset first classification model. Where M is the number of first predicted classification results.

[0185] The Kullback-Leibler divergence is used to calculate the first matching loss between the preset first classification model and the preset second classification model. The Kullback-Leibler divergence can measure the first predicted classification result. and the second predicted classification results The degree of matching between two probability distributions; the greater the difference between the two distributions, the greater the value of the KL divergence. Specifically, it can be based on the formula... Measure the first predicted classification result and the second predicted classification results KL divergence D KL (p2||p1), KL divergence D KL (p2||p1) is the first matching loss. In this formula, N is the total number of training samples, and M is the number of first and second predicted classification results.

[0186] According to the formula Classification of loss and the first matching loss The summation calculation is performed to obtain the first total classification loss. Where λ is the adjustment parameter.

[0187] Based on cross-entropy loss function For the predetermined classification result y i and the second predicted classification results The calculation process is performed to obtain the second classification loss of the preset second classification model. Where M is the number of second predicted classification results.

[0188] The Kullback-Leibler divergence is used to calculate the second matching loss between the pre-defined second classification model and the pre-defined first classification model. The Kullback-Leibler divergence can measure the second predicted classification result. and the first predicted classification results The degree of matching between two probability distributions; the greater the difference between the two distributions, the greater the value of the KL divergence. Specifically, it can be based on the formula... Measuring the second predicted classification result and the first predicted classification results KL divergence D KL (p1||p2), KL divergence D KL (p1||p2) is the second matching loss. In this formula, N is the total number of training samples, and M is the number of first and second predicted classification results.

[0189] According to the formula Second category loss and the second matching loss The summation calculation is performed to obtain the second total category loss. Where λ is the adjustment parameter.

[0190] The model parameters of the preset first classification model are adjusted based on the first total classification loss until the predetermined training conditions (such as the learning rate or classification accuracy meeting the requirements) are met, and the trained first classification model is obtained. In this way, a first classification model with superior classification performance can be obtained.

[0191] The model parameters of the preset second classification model are adjusted based on the second total classification loss until the predetermined training conditions (such as the learning rate or classification accuracy meeting the requirements) are met, and the trained second classification model is obtained. In this way, a second classification model with superior classification performance can be obtained.

[0192] After training to obtain the first classification model and the second classification model, the first classification model can execute steps S220 to S250 to obtain the first classification result, which is the first default probability; generate target feature data based on the object information of the target object; use the second classification model to classify the target feature data to obtain the second classification result corresponding to the target object, which is the first default probability; and determine the default category of the target object based on at least one of the first default probability and the second default probability.

[0193] In this way, in this scenario, a first classification model of Hierarchical Partial Attention Network (HPAN) can be proposed, referred to as the HPAN model; at the same time, a deep mutual learning model framework combining temporal sequence and product category (DCTM) composed of the first classification model and the second classification model can be proposed, referred to as the DCTM framework.

[0194] The DCTM framework consists of two sub-models, one exploring product category-related information and the other exploring time-series-related information in object behavior patterns. Furthermore, a mutual learning strategy allows the sub-models to share information related to model loss and learn from each other.

[0195] The HPAN model combines feature information from neighboring objects with similar behavior patterns across different product categories, and aggregates information at the product category level to fully extract hierarchical network information based on product categories. The HPAN model effectively captures the product categories most relevant to default probability and helps users understand key information about the default prediction model's performance and assess the reliability of its output. In the second classification model, an attention-based neural network model (LSTM model) is used to capture dynamic behavioral characteristics considering time sequence from object information.

[0196] In this scenario, object classification based on the DCTM framework is used to predict the default category of an object. This approach fully leverages the relationships between object information within product categories and considers the dynamic behavior characteristics over time, enabling accurate prediction of the default category. Furthermore, a theory-based default prediction algorithm can be implemented, achieving accurate prediction while maintaining a degree of model interpretability. This addresses the limitation of modern models like deep learning models in predicting default types due to insufficient interpretability. The algorithm also effectively explores the effectiveness of mutual learning models when two sub-models (a first classification model and a second classification model) have the same task and domain but focus on different information dimensions.

[0197] To facilitate better implementation of the object classification method provided in the embodiments of this application, the embodiments of this application also provide an object classification apparatus based on the above object classification method. The meanings of the terms used are the same as in the object classification method described above, and specific implementation details can be found in the descriptions in the method embodiments. Figure 6 A block diagram of an object classification apparatus according to an embodiment of this application is shown.

[0198] like Figure 6 As shown, the object classification device 500 may include an acquisition module 510, a division module 520, a generation module 530, a calculation module 540, and an output module 550.

[0199] The acquisition module 510 can be used to acquire object information of at least one object, the object information including at least one product category; the classification module 520 can be used to classify the objects associated with each product category separately to obtain the same type of objects of each object under each associated product category, the associated product category being the product category associated with each object; the generation module 530 can be used to take each object as the target object and, based on the object information of the same type of objects of the target object under each associated product category, generate enhanced feature data under each associated product category; the calculation module 540 can be used to calculate the enhanced feature data under each associated product category and the weight of the same type of objects to obtain the object feature data of the target object, the weight of the same type of objects being the weight of the same type of objects under the corresponding associated product category; the output module 550 can be used to obtain the first classification result of the target object based on the object feature data of the target object.

[0200] In some embodiments of this application, the weight of similar objects is an object attention weight; the calculation module includes: a transformation unit, used to perform linear transformation processing on the enhanced feature data under each associated product category to obtain transformed feature data under each associated product category; an object attention weight calculation unit, used to perform attention calculation processing on the transformed feature data under each associated product category and a preset object attention parameter to obtain the object attention weight of neighboring objects under each associated product category; a calculation unit, used to calculate processing on the transformed feature data under each associated product category and the object attention weight to obtain first feature data under each associated product category; and a generation unit, used to generate object feature data of the target object based on the first feature data.

[0201] In some embodiments of this application, the generation unit includes: a category weight calculation subunit, used to calculate and process the first feature data under each associated product category and the category weight corresponding to each associated product category respectively to obtain the second feature data under each associated product category; and a feature data generation subunit, used to generate object feature data of the target object based on the second feature data under each associated product category.

[0202] In some embodiments of this application, the object information includes object behavior information, and the category weight is a category attention weight. The category weight calculation subunit is used to: generate behavioral feature data based on the object behavior information under each associated product category to obtain the behavioral feature dataset of the target object; determine the non-core category in the associated product category corresponding to the target object based on the behavioral feature data under each associated product category; hide the behavioral feature data corresponding to the non-core category in the behavioral feature dataset to obtain a reset behavioral feature dataset; perform attention calculation processing based on the reset behavioral feature dataset, preset category attention parameters, and the first feature data to obtain the category attention weight of each associated product category; and calculate the first feature data and the category attention weight under each associated product category to obtain the second feature data under each associated product category.

[0203] In some embodiments of this application, the category weight calculation subunit is configured to: determine the object corresponding to each of the associated product categories from the at least one object, to obtain an object set corresponding to each of the associated product categories; for each object set corresponding to the associated product category, compare and sort the objects in the object set according to the behavioral feature data of each object, to obtain the relative ranking of each object in each object set; determine the relative ranking of the target object in each object set based on the relative ranking of each object in each object set; and determine the non-core category in the associated product categories corresponding to each object set based on the relative ranking of the target object in each object set.

[0204] In some embodiments of this application, the feature data generation subunit is used to: add the second feature data under each associated product category to obtain superimposed second feature data; and use the superimposed second feature data as the object feature data of the target object.

[0205] In some embodiments of this application, the first classification result is obtained using a first classification model, and the device further includes an auxiliary classification module for: generating target feature data based on the object information of the target object; classifying the target feature data using a second classification model to obtain a second classification result corresponding to the target object, wherein the second classification model is trained by mutual learning with the first classification model; and determining the object category of the target object based on at least one of the first classification result and the second classification result.

[0206] In some embodiments of this application, the apparatus further includes a first classification model training module, configured to: classify training object samples using a preset first classification model to obtain a first predicted classification result corresponding to the training object samples, wherein the training object samples are labeled with a predetermined classification result; classify training object samples using a preset second classification model to obtain a second predicted classification result corresponding to the training object samples; analyze a first classification loss of the preset first classification model based on the predetermined classification result and the first predicted classification result; analyze a first matching loss between the preset first classification model and the preset second classification model based on the first predicted classification result and the second predicted classification result; calculate a first total classification loss of the preset first classification model based on the first classification loss and the first matching loss; and adjust the preset first classification model based on the first total classification loss until predetermined training conditions are met to obtain the trained first classification model.

[0207] In some embodiments of this application, the apparatus further includes a second classification model training module, configured to: analyze the second classification loss of the preset second classification model based on the predetermined classification result and the second predicted classification result; analyze the second matching loss between the preset second classification model and the preset first classification model based on the second predicted classification result and the first predicted classification result; calculate the second total classification loss of the preset second classification model based on the second classification loss and the second matching loss; and adjust the preset second classification model based on the second total classification loss until predetermined training conditions are met, thereby obtaining the trained second classification model.

[0208] In some embodiments of this application, the first classification result is a first default probability, and the second classification result is a second default probability; the auxiliary classification module is used to: determine the default category of the target object based on at least one of the first default probability and the second default probability.

[0209] In some embodiments of this application, the partitioning module is configured to: determine objects corresponding to each product category from the at least one object, and obtain a set of sub-objects corresponding to each product category; classify the objects in the set of sub-objects for each product category to obtain at least one cluster of similar objects under each product category; and determine the other objects in each cluster of similar objects (excluding the object itself) as similar objects under each associated product category, wherein the associated product category is the product category corresponding to the cluster of similar objects.

[0210] In some embodiments of this application, each associated product category includes at least one of the aforementioned similar objects; the generation module is configured to: generate similar object feature data for each similar object based on the object information of each similar object; and combine the similar object feature data of at least one similar object under each associated product category to obtain enhanced feature data for each associated product category.

[0211] In some embodiments of this application, the output module is configured to: perform a fully connected operation on the object feature data of the target object to obtain the fully connected operation result; and classify the fully connected operation result to obtain a first classification result of the target object.

[0212] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0213] Furthermore, embodiments of this application also provide an electronic device, which can be a terminal or a server, such as... Figure 7 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0214] The electronic device may include components such as a processor 601 with one or more processing cores, a memory 602 with one or more computer-readable storage media, a power supply 603, and an input unit 604. Those skilled in the art will understand that... Figure 7 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0215] The processor 601 is the control center of the electronic device. It connects to various parts of the computer device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 602, and by calling data stored in the memory 602, it performs various functions of the computer device and processes data, thereby detecting the electronic device. Optionally, the processor 601 may include one or more processing cores; preferably, the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user page, and application programs, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 601.

[0216] The memory 602 can be used to store software programs and modules. The processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 602 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 602 may also include a memory controller to provide the processor 601 with access to the memory 602.

[0217] The electronic device also includes a power supply 603 that supplies power to the various components. Preferably, the power supply 603 can be logically connected to the processor 601 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 603 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0218] The electronic device may also include an input unit 604, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0219] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 601 in the electronic device loads the executable files corresponding to the processes of one or more computer programs into the memory 602 according to the following instructions, and the processor 601 runs the computer programs stored in the memory 602, thereby realizing the various functions of the foregoing embodiments of this application.

[0220] For example, processor 601 can perform the following operations: acquiring object information of at least one object, the object information including at least one product category; classifying the objects associated with each product category to obtain similar objects of each object under each associated product category, the associated product category being the product category associated with each object; taking each object as the target object, generating enhanced feature data under each associated product category based on the object information of similar objects of the target object under each associated product category; calculating and processing the enhanced feature data under each associated product category with the weights of similar objects to obtain object feature data of the target object, the weights of similar objects being the weights of similar objects under the corresponding associated product category; and obtaining a first classification result of the target object based on the object feature data of the target object.

[0221] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by a computer program, or by a computer program controlling related hardware. The computer program can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0222] Therefore, embodiments of this application also provide a computer-readable storage medium storing a computer program that can be loaded by a processor to perform the steps in any of the methods provided in embodiments of this application.

[0223] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0224] Since the computer program stored in the computer-readable storage medium can execute the steps in any of the methods provided in the embodiments of this application, the beneficial effects that the methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0225] According to one aspect of this application, a computer program product or computer program is provided, comprising 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 methods provided in the various optional implementations of the above embodiments of this application.

[0226] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0227] It should be understood that this application is not limited to the embodiments described above and shown in the accompanying drawings, but various modifications and changes can be made without departing from its scope.

Claims

1. An object classification method, characterized in that, include: Obtain object information for at least one object, wherein the object information includes at least one product category; the object information includes static attribute information and object behavior information of the object. The at least one object is a relevant user in the map-based vehicle networking platform or a relevant user in the shopping platform; The objects associated with each product category are classified separately to obtain the same type of objects for each object under each associated product category. The associated product category is the product category associated with each object. Taking each of the objects as the target object, based on the object information of the target object in the same type of object under each associated product category, enhanced feature data is generated for each associated product category; The enhanced feature data under each associated product category is calculated and processed with the weights of similar objects to obtain the object feature data of the target object. This includes: the weights of similar objects being object attention weights; performing linear transformation on the enhanced feature data under each associated product category to obtain transformed feature data under each associated product category; performing attention operations on the transformed feature data under each associated product category and preset object attention parameters to obtain the object attention weights of neighboring objects under each associated product category; calculating and processing the transformed feature data under each associated product category and the object attention weights to obtain first feature data under each associated product category; and generating the object feature data of the target object based on the first feature data; the weights of similar objects are the weights of the similar objects under the corresponding associated product category. Based on the object feature data of the target object, a first classification result of the target object is obtained; the first classification result is the default probability or the item purchase probability of the target object.

2. The method according to claim 1, characterized in that, The step of generating object feature data of the target object based on the first feature data includes: The first feature data under each associated product category and the category weight corresponding to each associated product category are calculated and processed separately to obtain the second feature data under each associated product category; Based on the second feature data under each associated product category, the object feature data of the target object is generated.

3. The method according to claim 2, characterized in that, The object information includes object behavior information, and the category weight is a category attention weight; the step of calculating and processing the first feature data under each associated product category and the category weight corresponding to each associated product category to obtain the second feature data under each associated product category includes: Behavioral feature data is generated based on the object behavioral information under each associated product category to obtain the behavioral feature dataset of the target object; Based on the behavioral characteristic data under each associated product category, determine the non-core categories in the associated product categories corresponding to the target object; The behavioral feature data corresponding to the non-core categories in the behavioral feature dataset is hidden to obtain the reset behavioral feature dataset. Based on the reset behavioral feature dataset, the preset category attention parameters and the first feature data, attention calculation processing is performed to obtain the category attention weight of each associated product category; The first feature data and category attention weight under each related product category are calculated and processed separately to obtain the second feature data under each related product category.

4. The method according to claim 3, characterized in that, The step of determining the non-core categories in the associated product categories corresponding to the target object based on the behavioral characteristic data under each associated product category includes: From the at least one object, determine the object corresponding to each of the associated product categories to obtain a set of objects corresponding to each of the associated product categories; For each of the associated product categories, the objects in the object set are compared and sorted according to the behavioral characteristic data of each object to obtain the relative ranking of each object in each object set; The relative ranking of the target object in each of the object sets is determined based on the relative ranking of each object in each of the object sets; Based on the relative ranking of the target object in each of the object sets, the non-core categories in the associated product categories corresponding to each of the object sets are determined.

5. The method according to claim 2, characterized in that, The process of generating object feature data for the target object based on the second feature data under each associated product category includes: The second feature data under each related product category is added together to obtain the superimposed second feature data; The superimposed second feature data is used as the object feature data of the target object.

6. The method according to any one of claims 1 to 5, characterized in that, The first classification result is obtained using a first classification model, and the method further includes: Target feature data is generated based on the object information of the target object; The target feature data is classified using a second classification model to obtain a second classification result corresponding to the target object. The second classification model is trained by mutual learning with the first classification model. The object category of the target object is determined based on at least one of the first classification result and the second classification result.

7. The method according to claim 6, characterized in that, The method further includes training the first classification model through mutual learning according to the following steps: The training object samples are classified using a preset first classification model to obtain a first predicted classification result corresponding to the training object samples, and the training object samples are labeled with a predetermined classification result; A preset second classification model is used to classify the training object samples to obtain the second predicted classification result corresponding to the training object samples; Based on the predetermined classification result and the first predicted classification result, analyze the first classification loss of the preset first classification model; Based on the first predicted classification result and the second predicted classification result, analyze the first matching loss between the preset first classification model and the preset second classification model; Calculate the first total classification loss of the preset first classification model based on the first classification loss and the first matching loss; The preset first classification model is adjusted based on the first total classification loss until the predetermined training conditions are met, thus obtaining the trained first classification model.

8. The method according to claim 7, characterized in that, The method further includes training the second classification model through mutual learning according to the following steps: Based on the predetermined classification result and the second predicted classification result, analyze the second classification loss of the preset second classification model; Based on the second predicted classification result and the first predicted classification result, analyze the second matching loss between the preset second classification model and the preset first classification model; Calculate the second total classification loss of the preset second classification model based on the second classification loss and the second matching loss; The preset second classification model is adjusted based on the second total classification loss until the predetermined training conditions are met, resulting in the trained second classification model.

9. The method according to claim 6, characterized in that, The first classification result is the first default probability, and the second classification result is the second default probability; Determining the object category of the target object based on at least one of the first classification result and the second classification result includes: The default category of the target object is determined based on at least one of the first default probability and the second default probability.

10. The method according to any one of claims 1 to 5, characterized in that, The step of classifying the objects associated with each product category to obtain the same type of objects for each associated product category includes: From the at least one object, determine the object corresponding to each product category to obtain a sub-object set corresponding to each product category; For each product category, the objects in the sub-object set are classified to obtain at least one cluster of objects of the same type under each product category. In each cluster of objects of the same kind, other objects besides the object itself are identified as objects of the same kind under each associated product category, where the associated product category is the product category corresponding to the cluster of objects of the same kind.

11. The method according to any one of claims 1 to 5, characterized in that, Each associated product category includes at least one of the aforementioned similar objects; The step of generating enhanced feature data for each associated product category based on the object information of the target object in the same category includes: Based on the object information of each object of the same type, generate characteristic data of each object of the same type. The feature data of at least one similar object under each related product category are combined and processed to obtain the enhanced feature data under each related product category.

12. The method according to any one of claims 1 to 5, characterized in that, The step of obtaining the first classification result of the target object based on the object feature data of the target object includes: Perform a fully connected operation on the object feature data of the target object to obtain the fully connected operation result; The results of the fully connected operation are classified to obtain the first classification result of the target object.

13. An object classification device, characterized in that, include: The acquisition module is used to acquire object information of at least one object, wherein the object information includes at least one product category; the object information includes static attribute information and object behavior information of the object. The at least one object is a relevant user in the map-based vehicle networking platform or a relevant user in the shopping platform; The segmentation module is used to classify the objects associated with each product category to obtain the same type of objects for each object under each associated product category, wherein the associated product category is the product category associated with each object. The generation module is used to take each of the objects as the target object and generate enhanced feature data for each associated product category based on the object information of the target object in the same type of object under each associated product category. The calculation module is used to calculate and process the enhanced feature data under each associated product category and the weights of similar objects to obtain the object feature data of the target object, including: the weight of similar objects is the object attention weight; performing linear transformation processing on the enhanced feature data under each associated product category to obtain transformed feature data under each associated product category; performing attention operation processing on the transformed feature data under each associated product category and preset object attention parameters to obtain the object attention weight of neighboring objects under each associated product category; calculating and processing the transformed feature data under each associated product category and the object attention weight to obtain the first feature data under each associated product category; generating the object feature data of the target object based on the first feature data; the weight of similar objects is the weight of the similar objects under the corresponding associated product category; The output module is used to obtain a first classification result of the target object based on the object feature data of the target object; the first classification result is the default probability or the item purchase probability of the target object.

14. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the method described in any one of claims 1 to 12.

15. An electronic device, characterized in that, include: Memory, which stores computer programs; A processor reads a computer program stored in memory to execute the method according to any one of claims 1 to 12.

16. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 12.

Images