Object feature processing method and apparatus, electronic device, medium, and program product

Abstract

Embodiments of the present application disclose an object feature processing method and device, electronic equipment, medium and program product. The method comprises: for a target candidate object, based on object feature information corresponding to the target candidate object, performing gating mapping processing on the degree of inter-object association between the target candidate object and each candidate object in a candidate object group to obtain gating weight information corresponding to the target candidate object; performing difference mapping processing on the inter-object feature difference between the target candidate object and each candidate object to obtain difference feature information corresponding to the target candidate object; performing weighted fusion processing on the difference feature information corresponding to the target candidate object to obtain relative ranking feature information corresponding to the target candidate object; and performing feature fusion processing based on context feature information and the relative ranking feature information to determine fusion feature information corresponding to the target candidate object. The embodiments of the present application can improve the representation accuracy of the fusion feature information corresponding to the candidate object.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to an object feature processing method, apparatus, electronic device, medium, and program product. Background Technology

[0002] In related technologies, industrial recommendation, search, or advertising ranking systems typically employ a multi-stage architecture of recall, coarse ranking, fine ranking, and re-ranking. The re-ranking stage often uses a Transformer-based multi-head attention mechanism to model the contextual relationships between candidate objects. However, multi-head attention mechanisms usually focus on global association modeling, easily incorporating all candidate objects in a group into the same comparison scope. This can lead to inaccurate feature representations and ranking relationships of candidate objects, thus affecting the accuracy of subsequent query, recommendation, filtering, or ranking results. Summary of the Invention

[0003] This application provides an object feature processing method, apparatus, electronic device, medium, and program product, which can improve the accuracy of representing the fused feature information corresponding to each candidate object. The above technical solution is as follows: In a first aspect, embodiments of this application provide an object feature processing method, including: Obtain the object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object. For the target candidate object in the above candidate object group, based on the object feature information corresponding to the target candidate object, the degree of object association between the target candidate object and each candidate object in the above candidate object group is subjected to gating mapping processing to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the above candidate object group. Based on the above object feature information, the inter-object feature differences between the above target candidate objects and the above candidate objects are processed by differential mapping to obtain the difference feature information corresponding to the above target candidate objects. Based on the gating weight information corresponding to the above target candidate objects, the differential feature information corresponding to the above target candidate objects is weighted and fused to obtain the relative ranking feature information corresponding to the above target candidate objects. Based on the contextual feature information and relative ranking feature information corresponding to the above target candidate objects, feature fusion processing is performed to determine the fused feature information corresponding to the above target candidate objects.

[0004] In one possible implementation, based on the object feature information corresponding to the target candidate object, the degree of inter-object association between the target candidate object and each candidate object in the candidate object group is subjected to gating mapping processing to obtain the gating weight information corresponding to the target candidate object, including: The object feature information corresponding to each candidate object in the above candidate object group is associated with the mapping process to obtain the associated query feature information and associated matching feature information corresponding to each candidate object. Based on the association query feature information corresponding to the above target candidate objects and the association matching feature information corresponding to each of the above candidate objects, the inter-object association information between the above target candidate objects and each of the above candidate objects is determined. The inter-object association information is used to characterize the feature matching degree between the above target candidate objects and each of the above candidate objects. Based on the above-mentioned inter-object association information, the gating weight information corresponding to the above-mentioned target candidate objects is determined.

[0005] In one possible implementation, determining the gating weight information corresponding to the target candidate object based on the inter-object association information includes: The above-mentioned inter-object association information is scaled to obtain scaled inter-object association information; By applying gating activation to the scaled inter-object association information, the gating weight information corresponding to the target candidate objects is obtained.

[0006] In one possible implementation, based on the object feature information, the inter-object feature differences between the target candidate object and each of the candidate objects are processed by differential mapping to obtain the difference feature information corresponding to the target candidate object, including: Based on the object feature information corresponding to each of the above candidate objects, a difference comparison mapping process is performed to obtain the comparison feature information corresponding to each of the above candidate objects. The comparison feature information corresponding to the above target candidate object is compared with the comparison feature information corresponding to each of the above candidate objects to obtain the comparison difference information between the above target candidate object and each of the above candidate objects. The above-mentioned comparative difference information is processed by nonlinear mapping to obtain the difference feature information corresponding to the above-mentioned target candidate objects.

[0007] In one possible implementation, the differential feature information corresponding to the target candidate object is weighted and fused based on the gating weight information corresponding to the target candidate object to obtain the relative ranking feature information corresponding to the target candidate object, including: Based on the gating weight information corresponding to the above target candidate objects, the difference feature information corresponding to the above target candidate objects is weighted and summed to obtain the weighted difference features corresponding to the above target candidate objects. Based on the gating weight information corresponding to the above target candidate objects, the above weighted difference features are normalized to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0008] In one possible implementation, the feature fusion processing based on the contextual feature information and the relative ranking feature information corresponding to the target candidate object, to determine the fused feature information corresponding to the target candidate object, includes: The relative ranking feature information corresponding to the above target candidate objects is processed by distribution transformation to obtain the distribution-aligned relative ranking feature information; The contextual feature information corresponding to the above target candidate objects and the corresponding relative ranking feature information after distribution alignment are subjected to feature concatenation processing to obtain the fused feature information corresponding to the above target candidate objects.

[0009] In one possible implementation, the above-mentioned acquisition of object feature information corresponding to each candidate object in the candidate object group to be processed, and determination of context feature information corresponding to each candidate object, includes: Obtain a group of candidate objects to be processed, which includes multiple candidate objects; The feature encoding process is performed on each candidate object in the above candidate object group to obtain the object feature information corresponding to each candidate object. Based on the object feature information corresponding to each of the above candidate objects, generate context query feature information, context matching feature information, and context value feature information corresponding to each of the above candidate objects; Based on the context query feature information corresponding to the above target candidate objects and the context matching feature information corresponding to each of the above candidate objects, the context association weight between the above target candidate objects and each of the above candidate objects is determined. Based on the aforementioned contextual association weights, the contextual value feature information corresponding to each of the above candidate objects is weighted and fused to obtain the contextual feature information corresponding to the above target candidate objects.

[0010] In one possible implementation, before obtaining the object feature information corresponding to each candidate object in the candidate object group to be processed, and determining the context feature information corresponding to each candidate object, the method further includes: Receive service request information, which can be any of the following: object query request, object recommendation request, object filtering request, or object sorting request; Based on the above service request information, multiple candidate objects are determined from the preset object resource library; Based on the above multiple candidate objects, a group of candidate objects to be processed is determined.

[0011] In one possible implementation, the above method also includes: Based on the above fusion feature information, the object processing score corresponding to each of the above candidate objects is determined. Based on the above service request information and the above object processing score, request adaptation processing is performed on each of the above candidate objects to obtain the request processing result corresponding to the above service request information. Specifically, when the service request information is an object query request, the request processing result includes the object query result; when the service request information is an object recommendation request, the request processing result includes the object recommendation result; when the service request information is an object filtering request, the request processing result includes the object filtering result; and when the service request information is an object sorting request, the request processing result includes the object sorting result.

[0012] Secondly, embodiments of this application provide an object feature processing apparatus, including: The acquisition module is used to acquire the object feature information corresponding to each candidate object in the candidate object group to be processed, and to determine the context feature information corresponding to each candidate object. The first processing module is used to perform gating mapping processing on the inter-object correlation degree between the target candidate object and each candidate object in the candidate object group based on the object feature information corresponding to the target candidate object in the above candidate object group, so as to obtain the gating weight information corresponding to the target candidate object, wherein the target candidate object is any candidate object in the above candidate object group. The second processing module is used to perform differential mapping processing on the inter-object feature differences between the target candidate object and each of the candidate objects based on the above object feature information, so as to obtain the difference feature information corresponding to the target candidate object. The third processing module is used to perform weighted fusion processing on the difference feature information corresponding to the above target candidate objects based on the gating weight information corresponding to the above target candidate objects, so as to obtain the relative ranking feature information corresponding to the above target candidate objects. The fusion module is used to perform feature fusion processing based on the context feature information and relative ranking feature information corresponding to the above target candidate objects to determine the fused feature information corresponding to the above target candidate objects.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory; wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps provided in the first aspect of embodiments of this application.

[0014] Fourthly, embodiments of this application provide a computer storage medium storing multiple instructions adapted for loading and executing the method steps provided in the first aspect of embodiments of this application by a processor.

[0015] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer or processor, cause the computer or processor to execute the method steps provided in the first aspect of embodiments of this application.

[0016] This application embodiment obtains object feature information corresponding to each candidate object in the candidate object group to be processed, and determines the context feature information corresponding to each candidate object. Then, for the target candidate object in the candidate object group, based on the object feature information corresponding to the target candidate object, a gating mapping process is performed on the inter-object association degree between the target candidate object and each candidate object in the candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the candidate object group. Based on the object feature information, a differential mapping process is performed on the inter-object feature differences between the target candidate object and each candidate object to obtain the difference feature information corresponding to the target candidate object. Based on the gating weight information corresponding to the target candidate object, a weighted fusion process is performed on the difference feature information corresponding to the target candidate object to obtain the relative ranking feature information corresponding to the target candidate object. Finally, feature fusion processing is performed based on the context feature information and the relative ranking feature information corresponding to the target candidate object to determine the fused feature information corresponding to the target candidate object. Thus, relative ranking feature information is generated by combining gating weight information and difference feature information on the basis of context feature information. Among them, gating weight information can control the contribution of different candidate objects in the difference fusion process and reduce the interference of weakly related candidate objects; difference feature information can reflect the relative feature differences between candidate objects. By fusing relative ranking feature information with context feature information, the fused feature information can simultaneously contain the global contextual relationship and local relative comparison relationship of the candidate object group, thereby improving the representation accuracy of the fused feature information corresponding to each candidate object, which is conducive to improving the accuracy and stability of subsequent object processing results such as querying, recommendation, filtering or ranking. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 An exemplary system architecture diagram of an object feature processing method provided in this application embodiment; Figure 2 A flowchart illustrating an object feature processing method provided in an embodiment of this application; Figure 3 A flowchart illustrating another object feature processing method provided in an embodiment of this application; Figure 4 A flowchart illustrating another object feature processing method provided in an embodiment of this application; Figure 5 A flowchart illustrating another object feature processing method provided in an embodiment of this application; Figure 6 A flowchart illustrating another object feature processing method provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an object feature processing device provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the features and advantages of the embodiments of this application more apparent and understandable, 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 protection scope of the embodiments of this application.

[0020] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims. Furthermore, in the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B; the word "and / or" in the text is merely a description of the association relationship between related objects, indicating that three relationships can exist, for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Additionally, in the description of the embodiments of this application, "multiple" refers to two or more.

[0021] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0022] In related technologies, industrial recommendation, search, or advertising ranking systems typically employ a multi-stage architecture of recall, coarse ranking, fine ranking, and re-ranking. The re-ranking stage often uses a multi-head attention mechanism based on Transformers to model the contextual relationships between recommended objects. However, multi-head attention mechanisms usually focus on global association modeling, which can easily lead to all candidate objects in a candidate object group being included in the same comparison scope. This may result in inaccurate feature representations and ranking relationship representations of candidate objects, thus affecting the accuracy of subsequent query, recommendation, filtering, or ranking results.

[0023] Therefore, this application provides an object feature processing method to solve the technical problem of low feature representation accuracy of candidate objects.

[0024] Please see Figure 1 This is a schematic diagram of a system architecture for an object feature processing method provided in an exemplary embodiment of this application. The server 10 communicates with the user terminal 20 via a network. A data storage system can store the data that the server 10 needs to process. The data storage system can be integrated onto the server 10, or it can be located in the cloud or on another network server.

[0025] In some possible embodiments, server 10 obtains object feature information corresponding to each candidate object in the candidate object group to be processed, and determines context feature information corresponding to each candidate object; then, for the target candidate object in the candidate object group, based on the object feature information corresponding to the target candidate object, performs gating mapping processing on the degree of object association between the target candidate object and each candidate object in the candidate object group to obtain gating weight information corresponding to the target candidate object, wherein the target candidate object is any candidate object in the candidate object group; and based on the object feature information, performs differential mapping processing on the object feature differences between the target candidate object and each candidate object to obtain difference feature information corresponding to the target candidate object; and based on the gating weight information corresponding to the target candidate object, performs weighted fusion processing on the difference feature information corresponding to the target candidate object to obtain relative ranking feature information corresponding to the target candidate object; and then performs feature fusion processing based on the context feature information and the relative ranking feature information corresponding to the target candidate object to determine the fused feature information corresponding to the target candidate object. Furthermore, the server 10 can perform downstream object processing based on the aforementioned fused feature information to obtain the corresponding object processing result, and can send the object processing result to the user terminal 20 so that the user terminal 20 can display the object processing result.

[0026] Understandably, server 10 can be implemented using a standalone server or a server cluster consisting of multiple servers. User terminal 20 can be, but is not limited to, various smartphones, tablets, personal computers, laptops, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc.

[0027] Optionally, the user terminal 20 may include the terminal device used by the user who requests the object, such as a terminal device used for searching, browsing, recommending, filtering or sorting candidate objects such as goods, content, services, advertisements, and resources; it may also include the terminal device used by the user who provides the object, such as a merchant-side terminal device, a content publishing-side terminal device or a service providing-side terminal device used for uploading, publishing, configuring or managing the aforementioned candidate objects.

[0028] It should be understood that Figure 1 The number of servers and user terminals shown is only illustrative; the number of servers and user terminals can be any number depending on the implementation requirements.

[0029] In one embodiment, such as Figure 2 As shown, an object feature processing method is provided, including the following steps: S201: Obtain the object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object.

[0030] The aforementioned candidate object group can refer to a collection of multiple candidate objects determined for a single service request. Optionally, the candidate objects in the candidate object group can be products, advertisements, content, videos, articles, services, tasks, resources, documents, users, or devices, etc. For example, in a search scenario, the candidate object group can be multiple search results retrieved based on relevant search terms; in a recommendation scenario, the candidate object group can be multiple candidate contents retrieved by the recommendation system for the user; in a resource scheduling scenario, the candidate object group can be multiple computing resources or task resources to be allocated.

[0031] Optionally, the aforementioned object feature information can be information used to characterize the candidate object's own attributes and attributes related to the current service scenario. Object feature information may include the candidate object's basic attribute features, statistical features, contextual features, user interaction features, request-related features, etc. For example, in a product search scenario, object feature information may include product category, product brand, product price, product sales volume, product review information, product inventory information, product click-through rate, product conversion rate, the matching degree between the product and search terms, user browsing history information, user purchase history information, etc.; in a content recommendation scenario, object feature information may include content category, publication time, click-through rate, user interest matching degree, etc.; in an advertising scenario, object feature information may include advertising category, bidding information, estimated click-through rate, user profile matching information, etc.; in a resource scheduling scenario, object feature information may include resource load, task priority, processing latency, availability status, etc. The specific content of the aforementioned object feature information can be set according to different service scenarios, and this application embodiment does not limit this.

[0032] It should be noted that the aforementioned contextual feature information can be obtained by performing intra-group contextual modeling on the corresponding candidate objects, and can be used to characterize the contextual relationships of candidate objects within the overall environment of the candidate object group. For example, multi-head attention processing can be used to model the global relationships between candidate objects in the candidate object group, thereby obtaining the contextual feature information corresponding to each candidate object. This contextual feature information can reflect the overall association and influence between a candidate object and all other candidate objects in the candidate object group (including itself).

[0033] In one specific embodiment, feature processing can be performed on each candidate object in the candidate object group to obtain object feature information. Further, based on the object feature information, context encoding processing can be performed on the candidate object group through a relevant context encoding module to obtain context feature information. This context feature information can be a feature representation of each candidate object obtained after fusing the overall context information of the candidate object group. For example, the candidate object group includes A, B, C, D, and E. For candidate object A, the context encoding module can calculate the degree of association between A and A, B, C, D, and E, and fuse the feature information of A, B, C, D, and E based on this degree of association to obtain the context feature information corresponding to A. Therefore, the context feature information corresponding to A not only includes information about A itself, but also includes the contextual influence of other candidate objects in the candidate object group on A.

[0034] S202: For the target candidate object in the above candidate object group, based on the object feature information corresponding to the target candidate object, perform gating mapping processing on the degree of object association between the target candidate object and each candidate object in the above candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the above candidate object group.

[0035] Gating weights are used to control the degree of contribution of candidate objects in comparisons. For any candidate object in the candidate object group, i.e., the target candidate object, the gating weights characterize the degree of comparability or correlation between the target candidate object and other candidate objects in the group. For example, in recommendation scenarios, technology-related content is more suitable for comparison with content of the same topic or similar popularity; in resource scheduling scenarios, a high-priority task is more suitable for comparison with tasks of the same priority or similar resource requirements. Gating weights can reduce the interference of weakly related objects on the comparison results.

[0036] In one embodiment, for any candidate object (target candidate object) in the candidate object group, the corresponding gating weight can be determined based on the feature matching degree between the target candidate object and each candidate object in the candidate object group. Optionally, the above-mentioned gating weight information can be used to characterize the contribution degree among each candidate object in the candidate object group during the difference comparison process.

[0037] In one embodiment, for a target candidate object in a candidate object group, a gating weight can be determined between the target candidate object and each candidate object in the candidate object group. A larger gating weight indicates a higher degree of difference fusion contribution of the corresponding candidate object to the target candidate object; a smaller gating weight indicates a lower degree of difference fusion contribution of the corresponding candidate object to the target candidate object.

[0038] For example, the candidate object group includes A, B, C, D, and E. For the target candidate object A, the gating weight information can include the gating weights corresponding to A and A, B, C, D, and E respectively, which are used to control the degree of attention paid to the differential feature information of candidate objects A, B, C, D, and E when generating relative ranking feature information for the target candidate object A.

[0039] S203: Based on the above object feature information, perform differential mapping processing on the inter-object feature differences between the above target candidate object and each of the above candidate objects to obtain the difference feature information corresponding to the above target candidate object.

[0040] The aforementioned differential feature information can be used to characterize the relative differences between candidates in the feature space within a candidate group. For example, in advertising scenarios, differential feature information can reflect differences in click probability, bid, audience matching, etc., among different ads; in content filtering scenarios, differential feature information can reflect differences in theme, quality, popularity, or user interest matching among different content.

[0041] S204: Based on the gating weight information corresponding to the above target candidate objects, the differential feature information corresponding to the above target candidate objects is weighted and fused to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0042] Among them, the aforementioned relative ranking feature information can be used to characterize the relative performance of the corresponding candidate object within the corresponding fit comparison range.

[0043] Specifically, for a target candidate in the candidate object group, the gating weights between the target candidate and each candidate in the candidate object group can be used to perform weighted summation or weighted normalization of the difference feature information between the target candidate and each candidate to obtain the relative ranking feature information corresponding to the target candidate.

[0044] It should be noted that the aforementioned relative ranking features can be soft relative positional features, and are not limited to discrete rankings. For example, relative ranking features can be used to characterize the relative advantages, disadvantages, or positions of a target candidate relative to comparable objects.

[0045] S205: Based on the contextual feature information and relative ranking feature information corresponding to the above target candidate objects, feature fusion processing is performed to determine the fused feature information corresponding to the above target candidate objects.

[0046] The fused feature information corresponding to the target candidate object can be used for subsequent processing such as querying, recommendation, filtering, sorting, prediction, or resource allocation.

[0047] Understandably, fused feature information can simultaneously include both the global contextual relationships and local relative comparison relationships of the target candidate object. Specifically, for a target candidate object, the contextual feature information and relative ranking feature information corresponding to that target candidate object can be added, concatenated, mapped and fused, or attention-fused to obtain the fused feature information corresponding to that target candidate object.

[0048] This application embodiment obtains object feature information corresponding to each candidate object in the candidate object group to be processed, and determines the context feature information corresponding to each candidate object. Then, for the target candidate object in the candidate object group, based on the object feature information corresponding to the target candidate object, a gating mapping process is performed on the inter-object association degree between the target candidate object and each candidate object in the candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the candidate object group. Based on the object feature information, a differential mapping process is performed on the inter-object feature differences between the target candidate object and each candidate object to obtain the difference feature information corresponding to the target candidate object. Based on the gating weight information corresponding to the target candidate object, a weighted fusion process is performed on the difference feature information corresponding to the target candidate object to obtain the relative ranking feature information corresponding to the target candidate object. Finally, feature fusion processing is performed based on the context feature information and the relative ranking feature information corresponding to the target candidate object to determine the fused feature information corresponding to the target candidate object. Thus, relative ranking feature information is generated by combining gating weight information and difference feature information on the basis of context feature information. Among them, gating weight information can control the contribution of different candidate objects in the difference fusion process and reduce the interference of weakly related candidate objects; difference feature information can reflect the relative feature differences between candidate objects. By fusing relative ranking feature information with context feature information, the fused feature information can simultaneously contain the global contextual relationship and local relative comparison relationship of the candidate object group, thereby improving the representation accuracy of the fused feature information corresponding to each candidate object, which is conducive to improving the accuracy and stability of subsequent object processing results such as querying, recommendation, filtering or ranking.

[0049] Please see Figure 3 , Figure 3 This is a flowchart illustrating another object feature processing method provided in an embodiment of this application. Figure 3 As shown, object feature processing methods can include at least: S301: Obtain the object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object.

[0050] Specifically, S301 is the same as S201, and will not be repeated here.

[0051] S302: Perform association mapping processing on the object feature information corresponding to each candidate object in the above candidate object group to obtain the association query feature information and association matching feature information corresponding to each candidate object.

[0052] Among them, the aforementioned association query feature information can be the feature information of the candidate object actively initiating matching in the object association calculation, specifically the query (Q) vector in the multi-head attention network; the aforementioned association matching feature information can be the feature information of the candidate object as the matched object in the object association calculation, specifically the key (K) vector in the multi-head attention network.

[0053] In one specific implementation, the object feature information corresponding to the candidate object group can be input into the association mapping layer of a multi-head attention network. This multi-head attention network can include multiple attention heads, each of which can map the object feature information of each candidate object to generate association query feature information and association matching feature information under that attention head. In other words, for the same candidate object, multiple sets of association query feature information and association matching feature information can be obtained under different attention heads, and different attention heads can be used to represent the association relationship between objects from different latent feature dimensions.

[0054] For example, in a search scenario, candidate objects can be multiple search results. Different attention heads can learn association features under different implicit dimensions such as request matching degree, object quality, object popularity, and category level. In a recommendation scenario, candidate objects can be multiple candidate content, and different attention heads can learn association features under different dimensions such as content topic, user interest matching, content popularity, and publication time. In an advertising scenario, different attention heads can learn association features under different dimensions such as ad audience matching, bidding characteristics, estimated click-through rate, and ad category.

[0055] S303: Based on the association query feature information corresponding to the target candidate object and the association matching feature information corresponding to each candidate object, determine the inter-object association information between the target candidate object and each candidate object. The inter-object association information is used to characterize the feature matching degree between the target candidate object and each candidate object.

[0056] The feature matching degree can be used to indicate whether two candidate objects have a strong correlation or comparability under the latent feature dimension corresponding to the current attention head.

[0057] Optionally, for each attention head, the association query feature information of the target candidate object under that attention head can be matched and calculated with the association matching feature information of each candidate object in the candidate object group under that attention head to obtain the inter-object association information under that attention head. Different attention heads can each obtain a set of inter-object association information, thus forming multi-head inter-object association information.

[0058] For example, in a search result processing scenario, if the current target candidate is a search result of a certain high-end type, then in the attention head representing the object level or object type, the inter-object association information between it and candidate objects of similar level or type can be high; while the inter-object association information between it and candidate objects with significant differences in level or type can be low. In a content recommendation scenario, if the current target candidate is technology content, then the inter-object association information between it and content with similar themes or similar user interest distributions can be high, while the inter-object association information between it and content with significant differences in theme can be low.

[0059] S304: The above-mentioned inter-object association information is scaled to obtain scaled inter-object association information.

[0060] Among them, scaling can be used to adjust the numerical scale of the association information between objects, reduce the influence of feature dimension, feature magnitude or matching calculation method on the numerical value of the association information between objects, and make the subsequent gating activation process more stable.

[0061] Optionally, for each attention head, a scaling factor can be determined based on the dimension of the associated query feature information or associated matching feature information under that attention head, and the inter-object association information under that attention head can be scaled based on the scaling factor. This avoids the inter-object association information values ​​from being too large, causing subsequent gating activation results to prematurely approach their maximum or minimum values.

[0062] For example, in a multi-head attention network, if the dimensionality of the associated query feature information and associated matching feature information output by a certain attention head is high, the inter-object association information calculated by matching may have a large numerical range. By scaling, the inter-object association information can be placed within a numerical range more suitable for gating activation, thereby improving the computational stability of the gating weight information.

[0063] S305: Gating activation is performed on the scaled inter-object association information to obtain the gating weight information corresponding to the target candidate objects.

[0064] Gated activation can be a non-linear processing procedure that converts scaled inter-object correlation information into gating weight information. This gating weight information can be used to control the comparative contribution of the target candidate object and other candidate objects in the subsequent differential feature fusion process.

[0065] Optionally, gating activation can be applied to the scaled inter-object association information under each attention head to obtain gating weight information for each attention head. Specifically, the gating activation can be implemented using the sigmoid activation function, so that each candidate object and each other obtains a gating weight. This gating weight can independently represent the contribution degree of the relationship between the corresponding objects; the gating weight can be within a preset numerical range, such as between 0 and 1.

[0066] For example, in object ranking, object recommendation, or object filtering scenarios, if the current target candidate object is closer to some candidate objects in terms of category, level, user interest matching degree, or request matching degree, the corresponding gating weight can be larger; if the current target candidate object differs significantly from some candidate objects, the corresponding gating weight can be smaller. Therefore, when subsequently weighting and fusing the difference feature information, the contribution of the difference features corresponding to strongly correlated candidate objects can be enhanced, while the interference of the difference features corresponding to weakly correlated candidate objects can be reduced, making the obtained relative ranking feature information more reflective of the relative performance of the current candidate object within the adaptation comparison range.

[0067] S306: Based on the above object feature information, perform differential mapping processing on the inter-object feature differences between the above target candidate object and each of the above candidate objects to obtain the difference feature information corresponding to the above target candidate object.

[0068] Specifically, S306 is the same as S203, and will not be repeated here.

[0069] S307: Based on the gating weight information corresponding to the above target candidate objects, the differential feature information corresponding to the above target candidate objects is weighted and fused to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0070] Specifically, S307 is the same as S204, and will not be repeated here.

[0071] S308: Based on the contextual feature information and relative ranking feature information corresponding to the above target candidate objects, feature fusion processing is performed to determine the fused feature information corresponding to the above target candidate objects.

[0072] Specifically, S308 is the same as S205, and will not be repeated here.

[0073] In this embodiment, a multi-head attention network is used to perform association mapping on each candidate object, obtaining association query feature information and association matching feature information under different attention heads, and determining the association information between objects based on the two. Further, scaling and gating activation processing are used to obtain gating weight information to control the degree of comparison contribution between candidate objects. Since this gating weight information can characterize the matching degree between the target candidate object and each candidate object under different latent feature dimensions, it can enhance the contribution of the differential features corresponding to strongly correlated candidate objects and reduce the interference of the differential features corresponding to weakly correlated candidate objects when performing weighted fusion of differential feature information. The resulting relative ranking feature information can more accurately reflect the relative performance of candidate objects within the fitting comparison range, thereby improving the ability of fused feature information to represent local comparison relationships and global contextual relationships within the candidate object group, which is beneficial to improving the accuracy and stability of subsequent object query, recommendation, filtering, or ranking processing results.

[0074] Please see Figure 4 , Figure 4 This is a flowchart illustrating another object feature processing method provided in an embodiment of this application. Figure 4 As shown, object feature processing methods can include at least: S401: Obtain the object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object.

[0075] Specifically, S401 is the same as S201, and will not be repeated here.

[0076] S402: For the target candidate object in the above candidate object group, based on the object feature information corresponding to the target candidate object, perform gating mapping processing on the degree of object association between the target candidate object and each candidate object in the above candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the above candidate object group.

[0077] Specifically, S402 is the same as S202, and will not be repeated here.

[0078] S403: Perform difference comparison mapping processing based on the object feature information corresponding to each of the above candidate objects to obtain the comparison feature information corresponding to each of the above candidate objects.

[0079] The aforementioned comparative feature information can be a feature representation used for comparing differences between objects. Optionally, the object feature information of each candidate object can be mapped through the value vector mapping branch in a multi-head attention network to obtain the comparative feature information corresponding to each candidate object under each attention head. That is, the same candidate object can correspond to multiple sets of comparative feature information, and different attention heads can be used to represent the comparative features of candidate objects from different latent feature dimensions.

[0080] For example, in a search scenario, comparative feature information can be used to characterize the overall performance of search results in dimensions such as request matching degree, object quality, object popularity, and object hierarchy; in a recommendation scenario, comparative feature information can be used to characterize the overall performance of candidate content in dimensions such as content topic, user interest matching degree, content popularity, and publication time; in an advertising scenario, comparative feature information can be used to characterize the overall performance of candidate ads in dimensions such as audience matching, bidding characteristics, estimated click-through rate, and ad category.

[0081] S404: Perform differential processing on the comparison feature information corresponding to the target candidate object and the comparison feature information corresponding to each candidate object to obtain the comparison difference information between the target candidate object and each candidate object.

[0082] Optionally, for the target candidate in the above candidate object group, the comparison feature information corresponding to the target candidate can be differentially processed with the comparison feature information corresponding to each candidate in the above candidate object group to obtain the comparison difference information between the target candidate and each candidate.

[0083] The aforementioned comparative difference information can be used to characterize the relative differences between the target candidate object and each other in the comparative feature space. This comparative difference information can be obtained in the latent feature space based on the mapped comparative feature information.

[0084] In one specific embodiment, for each attention head, the difference between the comparative feature information of the target candidate object under that attention head and the comparative feature information of each candidate object in the candidate object group under that attention head can be calculated to obtain the comparative difference information under that attention head. Different attention heads can generate comparative difference information under different latent feature dimensions.

[0085] For example, in a product search scenario, comparison difference information can reflect the relative differences between a candidate product (candidate object) and other candidate products in terms of implicit comparison dimensions such as price level, brand level, sales popularity, and configuration level; in a resource scheduling scenario, comparison difference information can reflect the relative differences between a candidate resource or task and other candidate resources or tasks in terms of dimensions such as load, priority, processing latency, and resource requirements.

[0086] S405: Perform nonlinear mapping processing on the above comparison difference information to obtain the difference feature information corresponding to the above target candidate objects.

[0087] Nonlinear mapping processing can be used to transform, compress, or smooth the comparison difference information, converting it into difference feature information more suitable for subsequent gated weighted fusion. Optionally, nonlinear mapping processing can be implemented using hyperbolic tangent function, sigmoid function, gate function, normalization function, or other nonlinear mapping methods; this application embodiment does not limit this.

[0088] In one specific embodiment, the comparative difference information under each attention head can be processed by nonlinear mapping to obtain the difference feature information under each attention head. This reduces the impact of excessively large or abnormal difference values ​​on the subsequent fusion process, making the numerical range of the difference feature information more stable.

[0089] For example, in content recommendation scenarios, different pieces of content may vary significantly in terms of topic, quality, popularity, or degree of matching with user interests. Through nonlinear mapping, these differences can be transformed into stable differential feature information, which can then be weighted and fused using gating weights.

[0090] S406: Based on the gating weight information corresponding to the above target candidate objects, perform weighted summation on the difference feature information corresponding to the above target candidate objects to obtain the weighted difference features corresponding to the above target candidate objects.

[0091] The gating weight information can be used to control the contribution of the differential feature information corresponding to different candidate objects in the weighted fusion process. If the gating weight between the target candidate object and a certain candidate object is large, the differential feature information between that candidate object and the target candidate object will contribute more; if the gating weight is small, the differential feature information between that candidate object and the target candidate object will contribute less.

[0092] In one specific embodiment, under each attention head, the difference feature information between the target candidate object and each other candidate object can be multiplied by the corresponding gating weight, and the weighted difference feature information can be summed to obtain the weighted difference feature under that attention head. Multiple attention heads can each obtain their corresponding weighted difference features. In this way, the weighted difference features can focus more on reflecting the difference information corresponding to candidate objects that are highly correlated or comparable to the current target candidate object, reducing the interference of weakly correlated candidate objects on the difference fusion result.

[0093] S407: Based on the gating weight information corresponding to the above target candidate objects, the above weighted difference features are normalized to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0094] The normalization process described above can be used to reduce the impact of the overall size of the gating weights on the weighted difference features, making the relative ranking feature information corresponding to different candidate objects have a more stable numerical scale. Optionally, the weighted difference features can be normalized based on the sum of the gating weights between the target candidate object and each candidate object in the candidate object group.

[0095] Understandably, relative ranking features can be used to characterize the relative performance of a target candidate within its fit comparison range. Relative ranking can be understood as a soft relative position, relative advantage, or relative disadvantage, rather than being limited to a discrete ranking. For example, in object ranking, object recommendation, or object selection scenarios, relative ranking features can reflect the relative performance of the current candidate relative to strongly related candidates, thus allowing the model to focus not only on the candidate's own features but also on its local comparative relationships within the candidate group.

[0096] Optionally, if the multi-head attention network includes multiple attention heads, the relative ranking feature information corresponding to each target candidate object can include the relative ranking features under multiple attention heads. Different attention heads can characterize the relative performance of candidate objects from different latent feature dimensions.

[0097] S408: Based on the contextual feature information and relative ranking feature information corresponding to the above target candidate objects, feature fusion processing is performed to determine the fused feature information corresponding to the above target candidate objects.

[0098] Specifically, S408 is the same as S205, and will not be repeated here.

[0099] In this embodiment, by performing difference comparison mapping on each candidate object, comparative feature information can be obtained to characterize the comparative performance of candidate objects under different implicit feature dimensions. By performing difference processing on the comparative feature information between any candidate object and other candidate objects, the relative differences between candidate objects can be obtained. Furthermore, by performing nonlinear mapping processing, the numerical range of the difference feature information can be made more stable, facilitating subsequent fusion processing. Further, by performing weighted summation and normalization processing on the difference feature information using gating weight information, the relative ranking feature information can better reflect the difference information corresponding to candidate objects with a high degree of correlation and comparability with the current candidate object, reducing the interference of weakly correlated candidate objects on the relative ranking feature information. Therefore, the obtained relative ranking feature information can more accurately characterize the relative performance of candidate objects within the suitable comparison range. Moreover, by fusing the relative ranking feature information with contextual feature information, the fused feature information can simultaneously include the global contextual relationship and local relative comparison relationship of the candidate object group, thereby improving the representation accuracy of the fused feature information corresponding to each candidate object, which is beneficial to improving the accuracy and stability of subsequent object query, recommendation, filtering, or sorting processing results.

[0100] Please see Figure 5 , Figure 5 This is a flowchart illustrating another object feature processing method provided in an embodiment of this application. Figure 5 As shown, object feature processing methods can include at least: S501: Obtain the candidate object group to be processed, which includes multiple candidate objects.

[0101] The aforementioned candidate object group can be multiple objects to be processed identified for a single service request. For example, in an object query scenario, the candidate object group can be multiple search results retrieved from the object resource library based on the query request; in an object recommendation scenario, the candidate object group can be multiple candidate contents identified by the recommendation system for the target user; in an advertising delivery scenario, the candidate object group can be multiple candidate advertisements that meet the delivery conditions; and in a resource scheduling scenario, the candidate object group can be multiple resources or tasks to be allocated.

[0102] S502: Perform feature encoding processing on each candidate object in the above candidate object group to obtain the object feature information corresponding to each candidate object.

[0103] The aforementioned feature encoding process can be used to convert the candidate object's original attributes, statistical data, user interaction data, request-related data, etc., into a feature representation that the model can process. This object feature information can serve as a unified feature representation for the candidate object before it enters the subsequent context encoding module.

[0104] For example, in a search scenario, object feature information may include the matching features between search results and query terms, object quality features, popularity features, etc.; in a recommendation scenario, object feature information may include content category, user interest matching degree, click-through rate, publication time, etc.

[0105] S503: Based on the object feature information corresponding to each of the above candidate objects, generate context query feature information, context matching feature information, and context value feature information corresponding to each of the above candidate objects.

[0106] Among them, the aforementioned context query feature information can correspond to the query vector in multi-head attention, which is used to represent the features that the current candidate object wants to focus on when modeling the context; the aforementioned context matching feature information can correspond to the key vector, which is used to represent the features that the candidate object can be matched by other candidate objects; the aforementioned context value feature information can correspond to the value vector, which is used to represent the information that the candidate object provides to other candidate objects during the context fusion process.

[0107] Specifically, the object feature information of each candidate object can be mapped to context query feature information, context matching feature information, and context value feature information through fully connected layers, linear mapping layers, or mapping layers in multi-head attention networks.

[0108] S504: Based on the context query feature information corresponding to the target candidate object and the context matching feature information corresponding to each candidate object, determine the context association weight between the target candidate object and each candidate object.

[0109] The aforementioned contextual association weights can be used to characterize the degree of attention the target candidate object pays to each candidate object during the contextual modeling process. The larger the contextual association weight, the greater the influence of the corresponding candidate object on the contextual features of the target candidate object.

[0110] Specifically, the context query feature information corresponding to the target candidate object can be compared with the context matching feature information corresponding to each candidate object to calculate the similarity and obtain the context association score; then the context association score is normalized to obtain the context association weight.

[0111] For example, in recommendation scenarios, target candidate content can have a higher contextual relevance weight to candidate content with the same topic, similar popularity, or similar user interest distribution; in search scenarios, target search results can have a higher contextual relevance weight to search results with the same category, similar quality, or similar matching degree.

[0112] S505: Based on the above context association weights, the context value feature information corresponding to each of the above candidate objects is weighted and fused to obtain the context feature information corresponding to the above target candidate objects.

[0113] Among them, the context value feature information can represent the content information provided by each candidate object during the context fusion process; the aforementioned context feature information can represent the feature representation of the target candidate object after fusing the overall context relationship of the candidate object group.

[0114] Specifically, the contextual association weights between the target candidate object and each other can be applied to the contextual value feature information corresponding to each candidate object, and the weighted contextual value feature information can be summed or fused to obtain the contextual feature information corresponding to the target candidate object. In this way, the contextual feature information corresponding to each candidate object no longer only contains the features of the candidate object itself, but also further incorporates relevant information from other candidate objects in the candidate object group.

[0115] S506: For the target candidate object in the above candidate object group, based on the object feature information corresponding to the target candidate object, perform gating mapping processing on the degree of object association between the target candidate object and each candidate object in the above candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the above candidate object group.

[0116] Specifically, S506 is the same as S202, and will not be repeated here.

[0117] S507: Based on the above object feature information, perform differential mapping processing on the inter-object feature differences between the above target candidate object and each of the above candidate objects to obtain the difference feature information corresponding to the above target candidate object.

[0118] Specifically, S507 is the same as S203, and will not be repeated here.

[0119] S508: Based on the gating weight information corresponding to the above target candidate objects, the difference feature information corresponding to the above target candidate objects is weighted and fused to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0120] Specifically, S508 is the same as S204, and will not be repeated here.

[0121] S509: Perform distribution transformation processing on the relative ranking feature information corresponding to the above target candidate objects to obtain the distribution-aligned relative ranking feature information.

[0122] The aforementioned relative ranking feature information can be relative comparison features obtained through gating mapping, differential mapping, and weighted fusion processing. The distribution transformation processing described above can be used to adjust the numerical distribution of the relative ranking feature information, making the processed relative ranking feature information more consistent with the numerical distribution of the context feature information. This can reduce the feature shift problem caused by distribution differences during subsequent feature fusion.

[0123] Optionally, the relative ranking feature information can first be processed by interval mapping to ensure that the relative ranking feature information lies within a preset numerical range; further, the relative ranking feature information after interval mapping can be processed by distribution restoration to ensure that the processed features conform to a preset distribution. For example, the relative ranking feature information can be converted into a feature representation that approximates a standard normal distribution by using the inverse function of the error function.

[0124] S510: Perform feature concatenation processing on the contextual feature information corresponding to the above target candidate objects and the corresponding relative ranking feature information after the above distribution alignment to obtain the fused feature information corresponding to the above target candidate objects.

[0125] Optionally, the contextual feature information and the distribution-aligned relative ranking feature information can be concatenated in the feature dimension, head dimension, or channel dimension. Furthermore, the concatenated features can be mapped, compressed, or normalized to obtain the final fused feature information.

[0126] Understandably, the fused feature information can be further input into downstream object processing modules to determine object processing scores, query results, recommendation results, filtering results, or ranking results. Fused feature information can reflect both the overall contextual relationships within a candidate object group and the relative performance of each candidate object within the adaptation comparison range, thereby improving the accuracy of candidate object feature representation and the stability of subsequent object processing results.

[0127] In this embodiment, by performing context encoding on candidate objects in a candidate object group, each candidate object can obtain contextual feature information that integrates the overall relationship of the candidate object group. Furthermore, through gating mapping, differential mapping, and weighted fusion processing, relative ranking feature information reflecting the relative performance of the target candidate object within the adaptation comparison range can be obtained. By performing distribution transformation processing on the relative ranking feature information, the relative ranking feature information and contextual feature information can have a more consistent feature distribution, thereby reducing the distribution offset problem in the feature fusion process. Furthermore, by performing feature concatenation processing on the contextual feature information and the distribution-aligned relative ranking feature information, fused feature information that simultaneously includes global contextual relationships and local relative comparison relationships can be obtained, thereby improving the accuracy of candidate object feature representation and the stability of subsequent object ranking, object recommendation, object filtering, or object query results.

[0128] Please see Figure 6 , Figure 6 This is a flowchart illustrating another object feature processing method provided in an embodiment of this application. Figure 6 As shown, object feature processing methods can include at least: S601: Receive service request information, which can be any of the following: object query request, object recommendation request, object filtering request, or object sorting request.

[0129] For example, in an object query scenario, the service request information can be an object query request, which may include the user's input query terms, filter conditions, sorting preferences, etc.; in an object recommendation scenario, the service request information can be an object recommendation request, which may include user identifiers, user profiles, currently visited pages, historical interaction behaviors, etc.; in an object filtering scenario, the service request information can be an object filtering request, which may include filtering rules, filtering conditions, or target object types, etc.; in an object sorting scenario, the service request information can include the range of objects to be sorted, sorting targets, or service constraint information, etc.

[0130] It is understandable that by receiving service request information, the scope and target of the objects that need to be processed can be determined, providing a basis for subsequently identifying candidate objects from the object resource library.

[0131] S602: Based on the above service request information, determine multiple candidate objects from the preset object resource library.

[0132] The aforementioned object resource library can be used to store object data that can be queried, recommended, filtered, or sorted. Objects in the object resource library may include goods, content, advertisements, services, resources, tasks, documents, videos, users, or devices, etc., but this embodiment does not limit this.

[0133] Specifically, multiple candidate objects can be recalled or filtered from the object resource library based on the request content, request type, user characteristics, scenario characteristics, or service rules in the service request information. For example, in an object query scenario, multiple candidate objects matching the user's query terms can be recalled from the object resource library; in an object recommendation scenario, multiple candidate recommended objects can be determined from the object resource library based on user interest characteristics; in an object filtering scenario, multiple candidate objects meeting the filtering conditions can be determined from the object resource library; and in an object ranking scenario, multiple candidate objects that have undergone recall, coarse ranking, or fine ranking can be obtained.

[0134] S603: Based on the above multiple candidate objects, determine the candidate object group to be processed.

[0135] Optionally, the multiple candidate objects mentioned above can be directly identified as a group of candidate objects to be processed; alternatively, the multiple candidate objects can be deduplicated, filtered, truncated, grouped, preliminarily sorted, or rule-validated first, and then the processed multiple candidate objects can be identified as a group of candidate objects to be processed.

[0136] For example, in recommendation or ranking scenarios, multiple candidate objects that meet quantity or quality requirements can be selected from the recall results to form a candidate object group to be processed; in advertising scenarios, multiple candidate ads can be selected from candidate ads that meet the delivery conditions to form a candidate object group to be processed; in resource scheduling scenarios, a candidate object group to be processed can be determined from multiple candidate resources or candidate tasks based on resource status or task requirements. Thus, the server can organize multiple candidate objects related to the current service request into a candidate object group, so that subsequent object feature extraction, context modeling, gating mapping, differential mapping, and feature fusion processing can be performed on this candidate object group.

[0137] S604: Obtain the object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object.

[0138] Specifically, S604 is the same as S201, and will not be repeated here.

[0139] S605: For the target candidate object in the above candidate object group, based on the object feature information corresponding to the target candidate object, perform gating mapping processing on the degree of object association between the target candidate object and each candidate object in the above candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the above candidate object group.

[0140] Specifically, S605 is the same as S202, and will not be repeated here.

[0141] S606: Based on the above object feature information, perform differential mapping processing on the inter-object feature differences between the above target candidate object and each of the above candidate objects to obtain the difference feature information corresponding to the above target candidate object.

[0142] Specifically, S606 is the same as S203, and will not be repeated here.

[0143] S607: Based on the gating weight information corresponding to the above target candidate objects, the difference feature information corresponding to the above target candidate objects is weighted and fused to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0144] Specifically, S607 is the same as S204, and will not be repeated here.

[0145] S608: Based on the contextual feature information and relative ranking feature information corresponding to the above target candidate objects, feature fusion processing is performed to determine the fused feature information corresponding to the above target candidate objects.

[0146] Specifically, S608 is the same as S205, and will not be repeated here.

[0147] S609: Based on the above fusion feature information, determine the object processing score corresponding to each of the above candidate objects.

[0148] The aforementioned fusion feature information may include fusion feature information corresponding to each candidate object. This fusion feature information includes both contextual feature information and relative ranking feature information of the candidate object. Therefore, based on the fusion feature information corresponding to each candidate object, the object processing score of each candidate object under the current service request can be determined.

[0149] Optionally, the object processing score can be used to characterize the degree of matching between the candidate object and the current service request, recommendation value, likelihood of passing the screening, ranking priority, or other service processing value. A higher object processing score indicates that the corresponding candidate object is more suitable to be the processing result of the current service request; a lower object processing score indicates that the corresponding candidate object has a lower degree of matching or processing priority.

[0150] In one specific embodiment, the fused feature information corresponding to each candidate object can be input into a prediction network, a scoring network, a classification network, or a ranking network to obtain an object processing score for each candidate object. For example, in a query scenario, the object processing score can represent the relevance between the candidate object and the query request; in a recommendation scenario, the object processing score can represent the suitability of the candidate object for recommendation to the target user; in a filtering scenario, the object processing score can represent the degree to which the candidate object meets the filtering criteria; and in a ranking scenario, the object processing score can represent the ranking priority of the candidate object.

[0151] S610: Based on the above service request information and the above object processing score, perform request adaptation processing on each of the above candidate objects to obtain the request processing result corresponding to the above service request information; wherein, when the above service request information is the above object query request, the above request processing result includes the object query result; when the above service request information is the above object recommendation request, the above request processing result includes the object recommendation result; when the above service request information is the above object filtering request, the above request processing result includes the object filtering result; when the above service request information is the above object sorting request, the above request processing result includes the object sorting result.

[0152] The aforementioned request adaptation processing can refer to processing each candidate object in the candidate object group according to the request type indicated by the service request information. For example, candidate objects can be sorted, filtered, selected, rearranged, filtered, or combined based on their object processing scores to generate a request processing result that matches the service request information.

[0153] For example, when the service request information is an object query request, the server can determine candidate objects matching the query request based on the object processing score and generate object query results. These object query results may include query objects sorted by relevance or processing score. When the service request information is an object recommendation request, the server can determine suitable candidate objects for recommendation from the candidate object group based on the object processing score and generate object recommendation results. These object recommendation results may include a set of recommended objects or a sequence of recommended objects. When the service request information is an object filtering request, the server can retain or filter candidate objects based on the object processing score and filtering conditions to obtain object filtering results. These object filtering results may include candidate objects that meet the filtering requirements. When the service request information is an object ranking request, the server can rank or rerank each candidate object in the candidate object group based on the object processing score to obtain object ranking results. Thus, the aforementioned fused feature information can be further transformed into actual request processing results, enabling the feature processing results to serve downstream object processing tasks such as querying, recommendation, filtering, or ranking.

[0154] In this embodiment, by receiving service request information and determining candidate object groups from the object resource library based on the service request information, the subsequent feature processing can be matched with specific query, recommendation, filtering, or ranking requirements. Furthermore, by performing context modeling, gating mapping, differential mapping, and feature fusion processing on the candidate object groups, fused feature information that simultaneously includes global contextual relationships and local relative comparison relationships can be obtained, thereby improving the accuracy of the feature representations corresponding to each candidate object. Based on this, the object processing score corresponding to each candidate object is determined based on the fused feature information, and request adaptation processing is performed in conjunction with the service request information, making the obtained request processing results more consistent with the processing goals of the current service request. Therefore, this is beneficial for improving the accuracy and stability of object query results, object recommendation results, object filtering results, or object ranking results.

[0155] The present application will be further described below with reference to specific embodiments.

[0156] In one specific embodiment, the relative ranking feature information corresponding to the candidate object group can be generated by the gating ranking encoding module. Assume the candidate object group to be processed is S, which includes multiple candidate objects, and the initial feature information dimension of the candidate objects is D. For a target candidate object i in the candidate object group S, gating weight information used to control the comparison range can be determined based on the association relationship between the target candidate object and each candidate object in the candidate object group.

[0157] In one implementation, for each initial feature information c of the candidate object, the gating attention corresponding to that initial feature information c can be calculated. Where g... ψ (S,c) can be used to characterize the comparability among candidate objects in the candidate object group S given the initial feature information c. Furthermore, it can be based on g... ψ (S,c) is used to perform weighted statistics on the relative merits of each candidate in the candidate group under the initial feature information c, and obtain the ranking code of the candidate under the initial feature information c. This method can directly express the local relative position of the candidate under a specific feature dimension, but when the initial feature information dimension D is large, it is necessary to perform gating attention calculations for a large number of initial feature information, resulting in high computational and storage costs.

[0158] In another implementation, feature-specific gating modeling for the initial feature information c can be transformed into gating modeling for the hidden representation. Specifically, D learnable comparison heads can be generated based on the hidden representation H corresponding to the candidate object group S, where D' is less than the dimension D of the initial feature information. These D' learnable comparison heads can be used to replace modeling all the initial feature information one by one, thereby reducing the computational cost. Furthermore, value vector mapping can be performed based on the hidden representation H to obtain the value projection V(H), and the comparison difference information between candidate objects can be determined based on the difference between the value projections corresponding to different candidate objects.

[0159] Specifically, for the target candidate i and each candidate j in the candidate object group, the difference information between them can be determined by projecting V_i(H) and V_j(H) based on the corresponding values ​​of i and j. For example, it can be obtained based on V_i(H) - V_j(H). Furthermore, this difference information can be processed by nonlinear mapping to obtain a differentiable relative superiority / inferiority signal. This nonlinear mapping can be implemented using the Sigmoid function, hyperbolic tangent function, or other nonlinear functions. This process avoids directly using non-differentiable indicator functions for comparative statistics, allowing the relative ranking feature information to participate in gradient updates during model training.

[0160] In one specific embodiment, the comparison difference information can first be mapped using the Sigmoid function, ensuring the mapping result lies within a preset interval, such as the (0,1) interval. Then, the mapping result is weighted and fused based on gating weight information to obtain the ranking code corresponding to each candidate object. To ensure the output distribution of the ranking code matches the feature distribution in the subsequent Transformer network, distribution alignment processing can be performed on the ranking code. Specifically, distribution restoration or transformation processing can be performed on the ranking code based on probability integral transformation. For example, the ranking code can first be mapped to the (0,1) interval using the Sigmoid function, and then converted into a feature representation approximating a standard normal distribution using the inverse function of the error function. This makes the relative ranking feature information output by the gating ranking code module more consistent with the hidden feature distribution in the Transformer network, reducing the additional learning cost of distribution shift for subsequent network layers.

[0161] Understandably, the gated ranking encoding module can generate relative ranking feature information for each candidate object in the candidate object group without significantly increasing the computational burden. This relative ranking feature information reflects the relative performance of the candidate object within its fitting comparison range. Furthermore, this relative ranking feature information can be fused with the context feature information obtained by the context encoding module to obtain the fused feature information corresponding to each candidate object, thus enabling the fused feature information to simultaneously include the global contextual relationship and the local relative comparison relationship of the candidate object group.

[0162] In this embodiment, the relative ranking modeling of candidate object groups using a gated ranking encoding module not only retains the global context modeling capability of the context encoding module but also introduces local comparison relationships between candidate objects. Specifically, by determining the fit comparison range between candidate objects through gated attention and generating differentiable relative superiority / inferiority signals based on the differences between value projections, the model can more accurately represent the relative position or relative advantage of candidate objects within their fit comparison range, reducing the interference of weakly correlated candidate objects on the ranking encoding. Furthermore, by converting feature-by-feature modeling under the original feature dimension into learnable comparison head modeling under the hidden representation, it is possible to avoid calculating gated attention separately for a large number of original features, thereby reducing computational and storage overhead and enabling the module to be flexibly inserted into different network layers. By performing distribution alignment processing on the ranking encoding, the relative ranking feature information can be better matched with the distribution of hidden features in the Transformer network, reducing the additional learning cost of distribution shift for subsequent network layers. Therefore, the embodiments of this application can enhance the ability of candidate object features to express local relative comparison relationships with a lower computational burden, so that the fused feature information simultaneously includes global contextual relationships and local relative ranking relationships, thereby improving the accuracy and stability of subsequent object processing results such as querying, recommendation, filtering or sorting.

[0163] Furthermore, it should be noted that in this embodiment, the gated ranking encoding module can operate as a plug-in module running in parallel with the context encoding module. That is, while the context encoding module performs context modeling on the candidate object group, the gated ranking encoding module can generate relative ranking feature information based on the object feature information of the same candidate object group, and fuse this relative ranking feature information with the context feature information output by the context encoding module. Thus, without significantly adjusting the overall structure of the original context encoding backbone network, it can supplement the original context modeling capability with local relative comparison modeling capability. Further, the gated ranking encoding module can also adopt a multi-head structure. Under this multi-head structure, different encoding heads can generate gate weight information, difference feature information, and relative ranking feature information from different latent feature dimensions, thereby enhancing the ability to express local comparison relationships within the candidate object group. Since the gated ranking encoding module can be used in parallel with the context encoding module, it can reduce the cost of modifying the existing model structure, training process, and online inference link, and has good module reusability and ease of engineering deployment. This pluggable parallel architecture is beneficial for integration into large-scale object processing systems such as search, recommendation, advertising, filtering, or ranking, while improving the ability to represent candidate object features and reducing engineering implementation costs and system modification complexity.

[0164] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0165] Based on the inventive concept of the above-mentioned object feature processing method, such as Figure 7 As shown in the illustration, this application also provides an object feature processing apparatus 700 for implementing the object feature processing method described above. The object feature processing apparatus 700 includes: The acquisition module 701 is used to acquire the object feature information corresponding to each candidate object in the candidate object group to be processed, and to determine the context feature information corresponding to each candidate object. The first processing module 702 is used to perform gating mapping processing on the inter-object association degree between the target candidate object and each candidate object in the candidate object group based on the object feature information corresponding to the target candidate object in the above candidate object group, so as to obtain the gating weight information corresponding to the target candidate object, wherein the target candidate object is any candidate object in the above candidate object group. The second processing module 703 is used to perform differential mapping processing on the inter-object feature differences between the target candidate object and each of the candidate objects based on the above object feature information, so as to obtain the difference feature information corresponding to the target candidate object. The third processing module 704 is used to perform weighted fusion processing on the difference feature information corresponding to the above target candidate objects based on the gating weight information corresponding to the above target candidate objects, so as to obtain the relative ranking feature information corresponding to the above target candidate objects. The fusion module 705 is used to perform feature fusion processing based on the context feature information and relative ranking feature information corresponding to the target candidate object to determine the fused feature information corresponding to the target candidate object.

[0166] In one possible implementation, the first processing module 702 is specifically used for: The object feature information corresponding to each candidate object in the above candidate object group is associated with the mapping process to obtain the associated query feature information and associated matching feature information corresponding to each candidate object. Based on the association query feature information corresponding to the above target candidate objects and the association matching feature information corresponding to each of the above candidate objects, the inter-object association information between the above target candidate objects and each of the above candidate objects is determined. The inter-object association information is used to characterize the feature matching degree between the above target candidate objects and each of the above candidate objects. Based on the above-mentioned inter-object association information, the gating weight information corresponding to the above-mentioned target candidate objects is determined.

[0167] In one possible implementation, the first processing module 702 is specifically used for: The above-mentioned inter-object association information is scaled to obtain scaled inter-object association information; By applying gating activation to the scaled inter-object association information, the gating weight information corresponding to the target candidate objects is obtained.

[0168] In one possible implementation, the second processing module 703 is specifically used for: Based on the object feature information corresponding to each of the above candidate objects, a difference comparison mapping process is performed to obtain the comparison feature information corresponding to each of the above candidate objects. The comparison feature information corresponding to the above target candidate object is compared with the comparison feature information corresponding to each of the above candidate objects to obtain the comparison difference information between the above target candidate object and each of the above candidate objects. The above-mentioned comparative difference information is processed by nonlinear mapping to obtain the difference feature information corresponding to the above-mentioned target candidate objects.

[0169] In one possible implementation, the third processing module 704 is specifically used for: Based on the gating weight information corresponding to the above target candidate objects, the difference feature information corresponding to the above target candidate objects is weighted and summed to obtain the weighted difference features corresponding to the above target candidate objects. Based on the gating weight information corresponding to the above target candidate objects, the above weighted difference features are normalized to obtain the relative ranking feature information corresponding to the above target candidate objects.

[0170] In one possible implementation, the fusion module 705 is specifically used for: The relative ranking feature information corresponding to the above target candidate objects is processed by distribution transformation to obtain the distribution-aligned relative ranking feature information; The contextual feature information corresponding to the above target candidate objects and the corresponding relative ranking feature information after distribution alignment are subjected to feature concatenation processing to obtain the fused feature information corresponding to the above target candidate objects.

[0171] In one possible implementation, module 701 is specifically used for: Obtain a group of candidate objects to be processed, which includes multiple candidate objects; The feature encoding process is performed on each candidate object in the above candidate object group to obtain the object feature information corresponding to each candidate object. Based on the object feature information corresponding to each of the above candidate objects, generate context query feature information, context matching feature information, and context value feature information corresponding to each of the above candidate objects; Based on the context query feature information corresponding to the above target candidate objects and the context matching feature information corresponding to each of the above candidate objects, the context association weight between the above target candidate objects and each of the above candidate objects is determined. Based on the aforementioned contextual association weights, the contextual value feature information corresponding to each of the above candidate objects is weighted and fused to obtain the contextual feature information corresponding to the above target candidate objects.

[0172] In one possible implementation, the device 700 further includes: The receiving module is used to receive service request information before obtaining the object feature information corresponding to the candidate object group to be processed. The service request information can be any one of the following: object query request, object recommendation request, object filtering request, or object sorting request. The first determining module is used to determine multiple candidate objects from a preset object resource library based on the aforementioned service request information. The second determining module is used to determine the candidate object group to be processed based on the above multiple candidate objects.

[0173] In one possible implementation, the device 700 further includes: The third determining module is used to determine the object processing score corresponding to each candidate object in the above candidate object group based on the above fused feature information. The fourth processing module is used to perform request adaptation processing on each of the above candidate objects based on the above service request information and the above object processing score, so as to obtain the request processing result corresponding to the above service request information. Specifically, when the service request information is an object query request, the request processing result includes the object query result; when the service request information is an object recommendation request, the request processing result includes the object recommendation result; when the service request information is an object filtering request, the request processing result includes the object filtering result; and when the service request information is an object sorting request, the request processing result includes the object sorting result.

[0174] The division of modules in the above-described object feature processing device is for illustrative purposes only. In other embodiments, the object feature processing device can be divided into different modules as needed to complete all or part of the functions of the above-described object feature processing device. The implementation of each module in the object feature processing device provided in this application embodiment can be in the form of a computer program. This computer program can run on a terminal or server. The program modules constituted by this computer program can be stored in the memory of the terminal or server. When the computer program is executed by a processor, it implements all or part of the steps of the object feature processing method described in the embodiments of this application.

[0175] This application also provides an electronic device, which can be a server, and its internal structure diagram can be as follows: Figure 8 As shown, this electronic device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. The processor executes computer programs to implement an object feature processing method.

[0176] Those skilled in the art will understand that Figure 8 The structures shown are merely block diagrams of some structures related to the embodiments of this application and do not constitute a limitation on the electronic devices to which the embodiments of this application are applied. Specific electronic devices may include more or fewer components than those shown in the figures, or combine certain components, or have different component arrangements.

[0177] In one possible implementation, a computer storage medium is provided that stores instructions, which, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the above embodiments. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium.

[0178] In one possible implementation, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0179] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer storage medium or transmitted through the computer storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0180] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0181] The above-described embodiments are merely preferred embodiments of the present application and are not intended to limit the scope of the present application. Various modifications and improvements made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application should fall within the protection scope defined by the claims.

[0182] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0183] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the features, data, and information involved in this specification were all obtained under full authorization.

Claims

1. A method for processing object features, characterized in that, include: Obtain object feature information corresponding to each candidate object in the candidate object group to be processed, and determine the context feature information corresponding to each candidate object; For a target candidate object in the candidate object group, based on the object feature information corresponding to the target candidate object, a gating mapping process is performed on the inter-object association degree between the target candidate object and each candidate object in the candidate object group to obtain the gating weight information corresponding to the target candidate object. The target candidate object is any candidate object in the candidate object group. Based on the object feature information, the inter-object feature differences between the target candidate object and each candidate object are processed by differential mapping to obtain the difference feature information corresponding to the target candidate object; Based on the gating weight information corresponding to the target candidate object, the difference feature information corresponding to the target candidate object is weighted and fused to obtain the relative ranking feature information corresponding to the target candidate object. Based on the contextual feature information and relative ranking feature information corresponding to the target candidate object, feature fusion processing is performed to determine the fused feature information corresponding to the target candidate object.

2. The method as described in claim 1, characterized in that, The step of performing gating mapping processing on the inter-object association degree between the target candidate object and each candidate object in the candidate object group based on the object feature information corresponding to the target candidate object, to obtain the gating weight information corresponding to the target candidate object, includes: The object feature information corresponding to each candidate object in the candidate object group is associated with the object feature information to obtain the associated query feature information and associated matching feature information corresponding to each candidate object. Based on the association query feature information corresponding to the target candidate object and the association matching feature information corresponding to each candidate object, the inter-object association information between the target candidate object and each candidate object is determined. The inter-object association information is used to characterize the feature matching degree between the target candidate object and each candidate object. Based on the inter-object association information, the gating weight information corresponding to the target candidate object is determined.

3. The method as described in claim 2, characterized in that, The step of determining the gating weight information corresponding to the target candidate object based on the inter-object association information includes: The object association information is scaled to obtain scaled object association information. The scaled inter-object association information is gating activated to obtain the gating weight information corresponding to the target candidate object.

4. The method as described in claim 1, characterized in that, The step of performing differential mapping processing on the inter-object feature differences between the target candidate object and each of the candidate objects based on the object feature information to obtain the difference feature information corresponding to the target candidate object includes: Based on the object feature information corresponding to each candidate object, a difference comparison mapping process is performed to obtain the comparison feature information corresponding to each candidate object; The comparison feature information corresponding to the target candidate object is differentially processed with the comparison feature information corresponding to each candidate object to obtain the comparison difference information between the target candidate object and each candidate object. The comparison difference information is processed by nonlinear mapping to obtain the difference feature information corresponding to the target candidate object.

5. The method as described in claim 1, characterized in that, The step of performing weighted fusion processing on the differential feature information corresponding to the target candidate object based on the gating weight information corresponding to the target candidate object to obtain the relative ranking feature information corresponding to the target candidate object includes: Based on the gating weight information corresponding to the target candidate object, the difference feature information corresponding to the target candidate object is weighted and summed to obtain the weighted difference feature corresponding to the target candidate object. Based on the gating weight information corresponding to the target candidate object, the weighted difference features are normalized to obtain the relative ranking feature information corresponding to the target candidate object.

6. The method as described in claim 1, characterized in that, The step of performing feature fusion processing based on the contextual feature information and relative ranking feature information corresponding to the target candidate object to determine the fused feature information corresponding to the target candidate object includes: The relative ranking feature information corresponding to the target candidate object is subjected to distribution transformation processing to obtain the distribution-aligned relative ranking feature information; The contextual feature information corresponding to the target candidate object and the corresponding distribution-aligned relative ranking feature information are subjected to feature concatenation processing to obtain the fused feature information corresponding to the target candidate object.

7. An object feature processing device, characterized in that, include: The acquisition module is used to acquire object feature information corresponding to each candidate object in the candidate object group to be processed, and to determine the context feature information corresponding to each candidate object; The first processing module is used to perform gating mapping processing on the inter-object association degree between the target candidate object and each candidate object in the candidate object group based on the object feature information corresponding to the target candidate object, to obtain gating weight information corresponding to the target candidate object, wherein the target candidate object is any candidate object in the candidate object group; The second processing module is used to perform differential mapping processing on the inter-object feature differences between the target candidate object and each candidate object based on the object feature information, so as to obtain the difference feature information corresponding to the target candidate object; The third processing module is used to perform weighted fusion processing on the difference feature information corresponding to the target candidate object based on the gating weight information corresponding to the target candidate object, so as to obtain the relative ranking feature information corresponding to the target candidate object. The fusion module is used to perform feature fusion processing based on the context feature information and the relative ranking feature information corresponding to the target candidate object, and to determine the fused feature information corresponding to the target candidate object.

8. An electronic device, comprising: Processor and memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code stored in the memory to perform the method as described in any one of claims 1-6.

9. A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps of any one of claims 1-6.

10. A computer program product comprising instructions that, when run on a computer or processor, causes the computer or processor to perform the method steps of any one of claims 1-6.

Images