Resource recommendation method and device based on large model, intelligent agent, electronic equipment and storage medium

Through the resource recommendation method based on the big model, the similarity between historical interactive resources and initial resources and the big model to process target historical resources is solved, and the problem that resource recommendation in the existing technology is difficult to meet user needs is achieved, and higher recommendation accuracy is achieved. Match user preferences.

CN119089047APending Publication Date: 2024-12-06BEIJING BAIDU NETCOM SCI & TECH CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411288546.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In the scenarios of video resources, news resources and product resources recommendation, the existing technology is difficult to accurately meet the actual needs of users, and the accuracy of resource recommendation is low.

Method used

Using a resource recommendation method based on the big model, by obtaining the similarity between the user's historical interaction resources and the initial resources, the big model uses the target historical resources to determine the expected interaction attributes of the initial resources, so as to recommend the target to the user more in line with the preferences. resource.

Benefits of technology

It improves the accuracy of resource recommendations, can more accurately meet users' interaction preferences, and improves the accuracy and user experience of resource recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089047A_ABST
    Figure CN119089047A_ABST
Patent Text Reader

Abstract

The invention provides a resource recommendation method and device based on a large model, an intelligent agent, electronic equipment and a storage medium, relates to the technical field of artificial intelligence, in particular to the technical fields of large models, deep learning, intelligent search and the like, and can be applied to application scenes such as intelligent e-commerce and intelligent video recommendation. The resource recommendation method based on the large model comprises the steps of obtaining a plurality of initial resources to be processed and a plurality of historical interaction resources of a target object; determining a target historical resource from the plurality of historical interaction resources based on the similarity between the plurality of historical interaction resources and the at least one initial resource; processing the target historical resource by using the large model to obtain an expected interaction attribute of the initial resource, the expected interaction attribute representing an attribute of an interaction operation executed by the target object for the initial resource; and recommending a target resource in the plurality of initial resources to the target object based on the expected interaction attributes of the plurality of initial resources.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to technical fields such as large models, deep learning, and intelligent search, and can be applied to application scenarios such as intelligent e-commerce and intelligent video recommendation. Background Art

[0002] With the rapid development of Internet technology, users can easily browse news, videos and other resources through terminal devices such as smart phones. Relevant resource platforms can also recommend resources to users to meet their actual needs. Summary of the invention

[0003] The present disclosure provides a resource recommendation method, intelligent agent, device, electronic device, storage medium and program product based on a large model.

[0004] According to one aspect of the present disclosure, a resource recommendation method based on a big model is provided, comprising: obtaining multiple initial resources to be processed and multiple historical interaction resources of a target object; determining a target historical resource from the multiple historical interaction resources based on the similarity between the multiple historical interaction resources and at least one initial resource; processing the target historical resource using the big model to obtain expected interaction attributes of the initial resource, the expected interaction attributes characterizing the attributes of the interaction operation performed by the target object on the initial resource; and recommending a target resource from the multiple initial resources to the target object based on the expected interaction attributes of each of the multiple initial resources.

[0005] According to another aspect of the present disclosure, an artificial intelligence agent is provided, which is configured to execute the method provided according to the embodiment of the present disclosure.

[0006] According to another aspect of the present disclosure, a resource recommendation device based on a large model is provided, including: an acquisition module, used to acquire multiple initial resources to be processed and multiple historical interaction resources of a target object; a target historical resource acquisition module, used to determine a target historical resource from multiple historical interaction resources based on the similarity between the multiple historical interaction resources and at least one initial resource; an expected interaction attribute acquisition module, used to process the target historical resource using the large model to obtain the expected interaction attribute of the initial resource, the expected interaction attribute characterizing the attribute of the interaction operation performed by the target object on the initial resource; and a recommendation module, used to recommend a target resource from the multiple initial resources to the target object based on the expected interaction attributes of each of the multiple initial resources.

[0007] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method provided by an embodiment of the present disclosure.

[0008] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method provided by an embodiment of the present disclosure.

[0009] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, wherein the computer program implements the method provided by the embodiment of the present disclosure when executed by a processor.

[0010] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0012] Figure 1 An exemplary system architecture to which a large model-based resource recommendation method and device can be applied according to an embodiment of the present disclosure is schematically shown;

[0013] Figure 2 The flowchart of the resource recommendation method based on the big model according to the embodiment of the present disclosure is schematically shown;

[0014] Figure 3 A schematic diagram of obtaining multiple initial resources according to an embodiment of the present disclosure is shown;

[0015] Figure 4 A schematic diagram showing a principle diagram of determining a target historical resource according to an embodiment of the present disclosure is shown;

[0016] Figure 5 The schematic diagram of a large model according to an embodiment of the present disclosure is shown;

[0017] Figure 6 The schematic diagram of the resource recommendation model according to the embodiment of the present disclosure is shown;

[0018] Figure 7 The application scenario diagram of the resource recommendation method based on a large model according to an embodiment of the present disclosure is schematically shown;

[0019] Figure 8 A block diagram schematically shows a resource recommendation device based on a large model according to an embodiment of the present disclosure; and

[0020] Fig. 9 A block diagram of an electronic device suitable for implementing a resource recommendation method based on a large model according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0021] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0022] In the technical solution disclosed in the present invention, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations, necessary confidentiality measures are taken, and do not violate public order and good morals.

[0023] The inventors have discovered that in resource recommendation scenarios such as video resources, news resources, and commodity resources, the resources recommended to users are difficult to meet the actual needs of users.

[0024] The embodiments of the present disclosure provide a resource recommendation method, intelligent agent, device, electronic device, storage medium and program product based on a large model. The resource recommendation method based on a large model includes: obtaining multiple initial resources to be processed and multiple historical interaction resources of a target object; determining a target historical resource from multiple historical interaction resources based on the similarity between multiple historical interaction resources and at least one initial resource; processing the target historical resource using a large model to obtain the expected interaction attribute of the initial resource, the expected interaction attribute characterizing the attribute of the interaction operation performed by the target object on the initial resource; and recommending a target resource from multiple initial resources to the target object based on the expected interaction attribute of each of the multiple initial resources.

[0025] According to the embodiments of the present disclosure, by using the similarity between historical interaction resources and initial resources to select target historical resources, the resource interaction preferences of the target object in the historical stage can be captured for the initial resources to be processed. By using the large model to process the target historical resources to determine the expected interaction attributes of the initial resources, the interaction preferences of the target object for the initial resources can be more accurately predicted based on the relatively powerful understanding ability of the large model, so as to use the expected interaction attributes to recommend the target resource to the target object from multiple initial resources, so that the target resource can more accurately meet the interaction preferences of the target object and improve the accuracy of resource recommendation.

[0026] Figure 1 An exemplary system architecture to which the large model-based resource recommendation method and apparatus can be applied according to an embodiment of the present disclosure is schematically shown.

[0027] It should be noted that Figure 1 The examples shown are only examples of system architectures to which the embodiments of the present disclosure can be applied, in order to help those skilled in the art understand the technical content of the present disclosure, but do not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios. For example, in another embodiment, an exemplary system architecture to which the large model-based resource recommendation method and apparatus can be applied may include a terminal device, but the terminal device may implement the large model-based resource recommendation method and apparatus provided by the embodiments of the present disclosure without interacting with the server.

[0028] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0029] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients and / or social platform software, etc. (only as examples).

[0030] The terminal devices 101 , 102 , and 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0031] The server 105 may be a server that provides various services, such as a background management server (only an example) that provides support for the content browsed by the user using the terminal devices 101, 102, and 103. The background management server may analyze and process the received data such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to the user request) to the terminal device.

[0032] The server 105 may also be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0033] It should be noted that the resource recommendation method based on the big model provided in the embodiment of the present disclosure can generally be executed by the server 105. Accordingly, the resource recommendation device based on the big model provided in the embodiment of the present disclosure can generally be set in the server 105. The resource recommendation method based on the big model provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the resource recommendation device based on the big model provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0034] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0035] Figure 2 The flowchart of the resource recommendation method based on a large model according to an embodiment of the present disclosure is schematically shown.

[0036] like Figure 2 As shown, the resource recommendation method based on a large model includes operations S210 to S240.

[0037] In operation S210, a plurality of initial resources to be processed and a plurality of historical interaction resources of a target object are obtained;

[0038] In operation S220 , a target historical resource is determined from the plurality of historical interaction resources based on similarities between the plurality of historical interaction resources and at least one initial resource.

[0039] In operation S230, the target historical resource is processed using the large model to obtain the expected interactive attributes of the initial resource.

[0040] In operation S240 , a target resource among the plurality of initial resources is recommended to the target object based on the expected interaction attributes of each of the plurality of initial resources.

[0041] According to an embodiment of the present disclosure, the initial resource may be a resource to be selected and recommended. The historical interaction resource may include resources on which the target object has performed interaction operations in a historical period. For example, the historical interaction resource may be a resource that the target object has clicked on in a historical period, a resource that has been browsed for a long time, or a resource on which interaction operations such as likes and favorites have been performed. The historical interaction resource may characterize the interaction preference of the target object in a historical period.

[0042] It should be noted that the resources involved in the embodiments of the present disclosure may include any type of resources such as news resources, video resources, commodity resources, etc. that can be browsed by terminal devices such as smart phones. The embodiments of the present disclosure do not limit the specific types of resources.

[0043] According to an embodiment of the present disclosure, the target historical resource may be a resource among multiple historical interaction resources whose similarity with the initial resource is greater than or equal to a preset similarity threshold. By determining the target historical resource with a higher similarity with the initial resource from the historical interaction resources, the similarity between the historical interaction preference of the target object and the initial resource to be processed can be captured based on the target historical resource.

[0044] According to an embodiment of the present disclosure, the similarity between multiple historical interaction resources and the initial resource can be obtained by processing the historical interaction resources and the initial resource based on any type of algorithm. For example, the resource information of the historical interaction resources and the resource information of the initial resource can be processed based on a neural network algorithm to obtain a similarity classification result. However, it is not limited to this, and the resource information of the historical interaction resources and the resource information of the initial resource can also be processed based on any other type of similarity algorithm such as a cosine similarity algorithm to obtain similarity. The resource information may include any type of information related to the historical interaction resource or the initial, for example, the resource information may include the title, subtitle, page content, resource image, resource category, resource theme, etc. of the historical interaction resource. The embodiment of the present disclosure does not limit the specific type of resource information.

[0045] According to an embodiment of the present disclosure, a large model may refer to a deep learning model with large-scale model parameters, and a large model generally contains hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than ten trillion model parameters. The large model may include multimodal large models such as the VideoChat model and the Video-LlaMA model. Alternatively, the large model may also be a large language model (Large Language Model) model, a GPT (Generative Pre-trained Transformer), a visual large model, and the like. The large model involved in the embodiments of the present disclosure may be a general large model, or may be an expert large model obtained after fine-tuning (Fine Tune) based on demand, and the embodiments of the present disclosure do not limit this. The large model can be used to process information of any modality such as images, text, video, and audio.

[0046] According to the embodiments of the present disclosure, by using a big model to process the target historical resources, it is possible to capture the relationship between the resource semantic attributes in the target historical resources and the interaction preference semantic attributes of the target object based on the big model's relatively powerful semantic attribute understanding ability and task execution ability, thereby achieving a more accurate prediction of the target object's expected interaction attributes for the initial resource.

[0047] According to an embodiment of the present disclosure, the expected interaction attribute represents the attribute of the interactive operation performed by the target object on the initial resource. For example, the expected interaction attribute may represent whether the target object performs a click operation on the initial resource. For another example, the expected interaction attribute may also represent the attribute of the interactive operation such as the browsing time, the probability of liking, the probability of collecting, etc. of the target object on the initial resource. The embodiment of the present disclosure does not limit the attribute type of the specific expected interactive operation represented by the expected interaction attribute.

[0048] According to an embodiment of the present disclosure, based on the expected interaction attributes of each of the multiple initial resources, recommending a target resource from multiple initial resources to a target object may include sorting the multiple initial resources based on the expected interaction attributes of each of the multiple initial resources. An initial resource with a relatively high sorting position is recommended to the target object as a target resource. However, it is not limited to this, and the initial resource that matches the interaction attribute type preferred by the target object may also be recommended to the target object as a target resource based on matching the attribute type of the expected interaction attribute with the interaction attribute type preferred by the target object. The embodiment of the present disclosure does not limit the specific method for determining the target resource, and those skilled in the art may make a selection based on actual needs.

[0049] According to an embodiment of the present disclosure, recommending a target resource to a target object may include pushing text messages and notification messages to terminal devices such as smart phones and tablet computers related to the target object to facilitate the target object to browse the target resource. Alternatively, it may also include pushing the target resource to the target object under the condition that the target object authorizes and the target object uses a resource service application (such as a short video application). The embodiments of the present disclosure do not limit the specific method of recommending target resources.

[0050] According to an embodiment of the present disclosure, obtaining multiple initial resources to be processed may include: semantically fusing the obtained multiple historical interaction resources to obtain historical resource features; and performing resource retrieval based on the historical resource features to obtain multiple initial resources.

[0051] According to the embodiments of the present disclosure, semantic fusion can be performed by semantically fusion encoding multiple historical interaction resources. For example, multiple historical interaction resources can be processed based on the attention network algorithm to obtain historical resource features. However, it is not limited to this, and multiple historical interaction resources can also be semantically fused based on other methods. The embodiments of the present disclosure are not limited to this. By semantically fusing multiple historical interaction resources, the historical resource features can be fused with the resource semantic attributes of multiple historical interaction resources, so that the multiple initial resources determined based on the historical resource features can be similar to the resource semantic attributes of the multiple historical interaction resources. In this way, the target object's preference for resource semantic attributes in the historical period can be captured by determining multiple initial resources.

[0052] In one example, multiple historical interaction resources may be processed based on RQ-VAE (Riemannian Quantization Variational Autoencoder) to obtain historical resource features. The historical resource features may be 4-bit codes based on any value from 0 to 1023 as coding characters.

[0053] According to an embodiment of the present disclosure, resource retrieval based on historical resource features may include processing historical resource features based on a neural network algorithm to obtain an initial resource code representing a semantic attribute of the resource, and searching for an initial resource corresponding to the semantic attribute of the resource through the initial resource code. However, this is not limited to the above, and historical resource features may also be processed based on a preset search engine to retrieve multiple initial resources with a high degree of similarity to the semantic attributes of the resource represented by the historical resource features. The embodiment of the present disclosure does not limit the specific method of performing resource retrieval, as long as the initial resources related to the semantic attributes of the resource represented by the historical resource features can be retrieved.

[0054] Figure 3 The schematic diagram shows a principle diagram of acquiring multiple initial resources according to an embodiment of the present disclosure.

[0055] like Figure 3 As shown, multiple historical interaction resources T301 can be input into the encoder 310, and the historical resource features F310 can be output. The encoder 310 can be constructed based on the encoder of the bidirectional Transformer network model. The historical resource features F310 can be based on four coding identifiers F311, F312, F313 and F314. The four coding identifiers F311, F312, F313 and F314 can be represented as 1023, 12, 2 and 513 respectively. The four coding identifiers F311, F312, F313 and F314 are determined from the preset coding table respectively. For example, the coding identifier F311 is determined from the preset coding identifiers 0 to 1023 in the preset coding table K301. The historical resource features F310 are input into the decoder 320, and multiple initial resources T302 can be output. The decoder 320 can be constructed based on a multi-layer attention head, so that the historical resource features can be subjected to multiple attention fusions to predict the initial resources related to the resource semantic attributes represented by the historical resource features.

[0056] Figure 4 The schematic diagram shows a principle diagram of determining a target historical resource according to an embodiment of the present disclosure.

[0057] like Figure 4 As shown, for the initial resource T401, the initial resource T401 can be input into the semantic coding network layer 410, and the initial resource feature F401 is output. Multiple historical interaction resources S1, S2, ... to Sn in the multiple historical interaction resources T4100 can be respectively input into the semantic coding network layer 410, and multiple historical interaction resource features F4101, F4102, ... to F410n are output. The initial resource feature F401 and the multiple historical interaction resource features F4101, F4102, ... to F410n are input into the resource screening layer 420, and the resource screening layer 420 can calculate the similarity between the initial resource feature F401 and each of the multiple historical interaction resource features F4101, F4102, ... to F410n, and obtain the similarity of each of the multiple historical interaction resources S1, S2, ... to Sn. Based on a preset similarity distance threshold, the historical interaction resources S1, S2, ... to Sn with a similarity higher than the preset similarity threshold can be determined as target historical resources, so that the resource screening layer outputs multiple target historical resources T4110 corresponding to the initial resource T401. The multiple target historical resources T4110 may include historical interaction resources S1, S3 ... to Sn.

[0058] It should be noted that, among the multiple initial resources, other initial resources except the initial resource T401 can also be based on Figure 4The execution operation shown is to determine a plurality of target historical resources corresponding to the other initial resources, which will not be described in detail in the embodiment of the present disclosure.

[0059] According to the embodiments of the present disclosure, by determining the target historical resources with a high similarity to the initial resources from multiple historical interaction resources, it is possible to preliminarily mine the correlation between the historical interaction preferences of the target object and the initial resources through the target historical resources, so as to facilitate the subsequent determination of the target resources that match the interaction preferences of the target object from multiple initial resources. At the same time, the historical resources that the target object has interacted with in the historical stage can be preliminarily screened, for example, the historical interaction resources of tens of thousands can be compressed to the target historical resources of hundreds. This can save the amount of data processing for the subsequent large model and improve the computing efficiency.

[0060] According to an embodiment of the present disclosure, a large model may include a feature fusion network and multiple expert network layers. The feature fusion network may be constructed based on an attention network algorithm. Multiple expert network layers may be constructed based on any type of deep learning algorithm, for example, an expert network layer may be constructed based on a deep learning algorithm such as a convolutional neural network algorithm, a long short-term memory network algorithm, etc. The model parameters between multiple expert network layers may be different.

[0061] According to an embodiment of the present disclosure, using a large model to process the target historical resources to obtain the expected interaction attributes of the initial resources may include: using a feature fusion network to process the object attribute information of the target historical resources and the target object to obtain an initial fusion feature; based on a target expert network layer corresponding to the recommendation scenario of the expected interaction attribute among multiple expert network layers, processing the initial fusion feature to obtain at least one expected interaction attribute.

[0062] According to an embodiment of the present disclosure, the object attribute information represents the interactive operation preference or resource semantic attribute preference of the target object. For example, the object attribute information may be attribute information such as the browsing time and completion rate of the target object. The embodiment of the present disclosure does not limit the specific type of the object attribute information, as long as it can represent the interactive operation preference or resource semantic attribute preference of the target object.

[0063] According to an embodiment of the present disclosure, the target historical resources and object attribute information can be integrated based on the attention network algorithm, so that the initial fusion features can fully learn the interactive operation preferences of the target object and the resource semantic attributes of the target historical resources.

[0064] According to the embodiments of the present disclosure, the recommendation scenario may represent the scenario in which the target object browses resources, such as a video resource browsing scenario, a commodity resource browsing scenario, etc. However, this is not limited thereto, and specific recommendation scenarios may also be divided based on other methods, such as a main channel browsing scenario, a follow channel browsing scenario, etc. The embodiments of the present disclosure do not limit the method for determining the recommendation scenario.

[0065] According to an embodiment of the present disclosure, by determining a target expert network layer corresponding to a recommended scenario to process initial fusion features, a target expert network layer that is fully learned based on the recommended scenario can be used to detect expected interaction attributes in the recommended scenario, thereby improving the detection accuracy of the interaction operation attributes of the target object in the recommended scenario.

[0066] In one example, the number of target expert network layers is less than the number of expert network layers. In this way, some expert network layers in the large model can be activated based on the recommendation scenario to perform expected interaction attribute detection, which can reduce the number of model parameters of the large model that simultaneously participate in the expected interaction attribute detection, improve the computational efficiency of the large model, and reduce computational overhead.

[0067] It should be noted that the expert network layer can be constructed based on any type of neural network algorithm, and the implementation of the present disclosure is not limited to this.

[0068] According to an embodiment of the present disclosure, using a feature fusion network to process the target historical resources and the object attribute information of the target object to obtain an initial fusion feature includes: fusing the target historical resources and the object attribute information based on an attention mechanism to obtain an initial fusion feature.

[0069] According to an embodiment of the present disclosure, the target historical resource feature representing the target historical resource and the object attribute feature representing the object attribute information may be fused based on an attention network algorithm to obtain an initial fused feature.

[0070] In one example, the initial fusion features may also be obtained by associating the initial resources, object attribute information, and target historical resources based on an attention network algorithm.

[0071] According to an embodiment of the present disclosure, using a feature fusion network to process the object attribute information of the target historical resource and the target object to obtain an initial fusion feature may also include: performing feature fusion on the target historical resource feature and at least one initial resource feature based on an attention mechanism to obtain a first resource fusion feature; fusing the scene attribute information related to the recommended scene in the object attribute information with the first resource fusion feature to obtain an initial fusion feature.

[0072] According to an embodiment of the present disclosure, the initial resource feature may be determined based on the initial resource corresponding to the target historical resource. For example, the feature of the initial resource corresponding to the target historical resource may be extracted based on a neural network algorithm to obtain the initial resource feature.

[0073] According to an embodiment of the present disclosure, the scene attribute information may characterize the interaction preference of the target object in a specified recommendation scene. For example, the scene attribute information in the discovery scene may characterize the click operation preference, play duration preference, completion rate preference, etc. of the target object in the discovery scene in the historical stage.

[0074] According to an embodiment of the present disclosure, fusing the first resource fusion feature with the scene attribute information may include processing the first resource fusion feature and the scene attribute information based on a fusion function. Alternatively, it may also include processing the first resource fusion feature and the scene attribute information based on a deep learning algorithm. The embodiment of the present disclosure does not limit the specific algorithm type for fusing the first resource fusion feature with the scene attribute information.

[0075] According to an embodiment of the present disclosure, by fusing the initial resource feature with the target historical resource feature, the obtained first resource feature can fully characterize the resource semantic attributes between the associated initial resource and the target historical resource. By fusing the first resource feature and the scene attribute information, the obtained initial fusion feature can semantically fuse the resource semantic attributes between the initial resource and the target historical resource, as well as the target object's preference for the recommendation scene represented by the scene attribute information. In this way, by determining the expected interaction attributes based on the initial fusion features, it is possible to more accurately characterize the target object's interactive operation attributes for the initial resource in the recommendation scene, so that the target resource can be more accurately recommended to the target object based on the expected interaction attributes, thereby improving the accuracy of resource recommendations.

[0076] It should be noted that the resource semantic attributes in the embodiments of the present disclosure can represent the semantics of any resource information such as text semantics, image semantics, theme semantics, etc. of the resource information of resources (such as initial resources, historical interaction resources, etc.), and the embodiments of the present disclosure will not go into details about this.

[0077] According to an embodiment of the present disclosure, processing initial fusion features based on a target expert network layer corresponding to a recommendation scenario of expected interaction attributes in multiple expert network layers may include: using multiple target expert network layers to process the initial fusion features respectively to obtain multiple intermediate fusion features; fusing the multiple intermediate fusion features based on the gated weights corresponding to the recommendation scenario to obtain a first target fusion feature; and determining at least one expected interaction attribute based on the first target fusion feature.

[0078] According to the embodiments of the present disclosure, the gating weights can characterize the ability of multiple target expert network layers to extract the semantic attributes of resource semantic attributes and object attribute information from the initial fusion features. By fusing multiple intermediate fusion features through gating weights, the feature extraction capabilities of multiple target expert network layers can be balanced, thereby making the initial fusion features more accurately characterize the resource semantic attributes of the target historical resources and the information semantic attributes of the object attribute information.

[0079] In one example, the gating weight corresponding to the recommended scenario may be determined by fine-tuning the large model.

[0080] In one example, the gating weight may also be determined by a gating selection network layer of a large model processing multiple intermediate fusion features.

[0081] According to an embodiment of the present disclosure, determining at least one expected interaction attribute based on the first target fusion feature may include inputting the first target fusion feature into at least one expected attribute detection head to obtain the at least one expected interaction attribute.

[0082] For example, the first target fusion feature can be input into the click rate detection head to output the expected click rate.

[0083] In one example, the gating weight may be set in an expected attribute detection head corresponding to the expected interaction attribute, and the expected attribute detection head is used to fuse multiple intermediate fusion features according to the gating weight to obtain the expected interaction attribute.

[0084] According to an embodiment of the present disclosure, the expected interaction attribute may include at least one of the following: click-through rate, video completion rate, and browsing time.

[0085] According to an embodiment of the present disclosure, the click rate may represent the expected probability of a target object clicking on a target resource when the initial resource is used as the target resource.

[0086] According to an embodiment of the present disclosure, when the target resource includes a video, the video completion rate may represent the expected probability that the video is completed during the process of the target object traffic targeting the target resource.

[0087] According to an embodiment of the present disclosure, the browsing duration may represent the expected browsing duration of the target object for the target resource when the initial resource is used as the target resource. Alternatively, the browsing duration may represent the expected browsing duration of the target object for the partial resource information such as the video and image of the target resource when the initial resource is used as the target resource.

[0088] According to an embodiment of the present disclosure, the large model may further include multiple attribute detection heads.

[0089] According to an embodiment of the present disclosure, determining at least one expected interaction attribute based on the first target fusion feature may include: processing the first target fusion feature based on a target attribute detection head corresponding to the recommended scene among multiple attribute detection heads to obtain at least one expected interaction attribute.

[0090] According to an embodiment of the present disclosure, a recommendation scenario may include one or more target attribute detection heads, and the target attribute detection head of the recommendation scenario is used to detect expected interaction attributes of a target object in the recommendation scenario.

[0091] According to an embodiment of the present disclosure, different recommendation scenarios can each correspond to a specified target attribute detection head. For example, a product recommendation scenario corresponds to a first click rate detection head and a first browsing duration detection head. A video playback scenario can correspond to a second click rate detection head and a second browsing duration detection head. The first click rate detection head and the second click rate detection head have different model parameters, and the first browsing duration detection head and the second browsing duration detection head have different model parameters.

[0092] According to an embodiment of the present disclosure, the recommendation scenario includes at least one of the following: a primary feedback scenario, a resource discovery scenario, and a resource two-hop scenario.

[0093] According to an embodiment of the present disclosure, the main feedback scene (or main feed scene) can represent the information flow that the target application continuously updates and displays to the user. For example, the main feedback scene can be a follow page that continuously updates messages in a social application. For another example, the main feedback scene can also be a short video playback interface that continuously provides a short video playback information flow in a short video application. The main feedback scene can, for example, be represented based on a single-column information flow page.

[0094] According to an embodiment of the present disclosure, the resource discovery scenario may include a scenario in which the target object browses resources in a discovery channel of a target application. The discovery channel may include, for example, a video recommendation column (or video recommendation column) of a short video application, or a news recommendation column of a news application.

[0095] According to an embodiment of the present disclosure, the resource two-hop scenario can represent a scenario in which a target object browses a jump page after clicking on a resource. The target attribute detection head corresponding to the resource two-hop scenario can output the target object's expected click rate, expected browsing time, and other expected interactive attributes for the initial resource in the resource two-hop scenario.

[0096] Figure 5 The schematic diagram shows a principle diagram of a large model according to an embodiment of the present disclosure.

[0097] like Figure 5 As shown, the large model 500 may include a feature fusion network 510, multiple expert network layers, and multiple target attribute detection heads. The multiple expert network layers include the first expert network layer 521, the second expert network layer 522, ... to the nth expert network layer 52n. The multiple target attribute detection heads may include a click rate detection head 531, a browsing time detection head 532, ... to an interactive operation detection head 53n.

[0098] When the current recommendation scenario is a two-hop resource scenario, multiple target historical resources T5100 and object attribute information A510 corresponding to the specified initial resource are input into the feature fusion network 510 to obtain the initial fusion feature. The initial fusion feature is input into two target expert network layers corresponding to the two-hop resource scenario, the two target expert network layers are the first expert network layer 521 and the second expert network layer 522, respectively, to obtain the first intermediate fusion feature and the second intermediate fusion feature.

[0099] like Figure 5 As shown, the first intermediate fusion feature and the second intermediate fusion feature are input into the click rate detection head 531. The click rate detection head 531 fuses the first intermediate fusion feature and the second intermediate fusion feature based on the click rate gating weight to obtain a first target fusion feature corresponding to the expected click rate. A linear regression is performed based on the first target fusion feature to obtain the expected click rate.

[0100] like Figure 5 As shown, the first intermediate fusion feature and the second intermediate fusion feature are input into the browsing duration detection head 532. The browsing duration detection head 532 fuses the first intermediate fusion feature and the second intermediate fusion feature based on the browsing duration gating weight to obtain a first target fusion feature corresponding to the expected browsing duration. A linear regression is performed based on the first target fusion feature corresponding to the expected browsing duration to obtain the expected browsing duration.

[0101] like Figure 5 As shown, the first intermediate fusion feature and the second intermediate fusion feature are input into the interactive operation detection head 533. The interactive operation detection head 533 fuses the first intermediate fusion feature and the second intermediate fusion feature based on the interactive operation gating weight to obtain the first target fusion feature corresponding to the expected interactive operation. A linear regression is performed based on the first target fusion feature corresponding to the expected interactive operation to obtain the expected interactive operation. Thus, the expected click rate, expected browsing time and expected interactive operation corresponding to the initial resource can be obtained.

[0102] It should be understood that Figure 5 The large model shown can be used to process the target historical resources and object attribute information corresponding to each of the multiple initial resources to obtain the expected click-through rate, expected browsing time and expected interactive operation corresponding to each of the multiple initial resources. Based on the expected click-through rate, expected browsing time and expected interactive operation corresponding to each of the multiple initial resources, resource screening can be performed to determine at least one target resource from the multiple initial resources and recommend the target resource to the target object.

[0103] According to an embodiment of the present disclosure, recommending a target resource from a plurality of initial resources to a target object based on the expected interaction attributes of each of the plurality of initial resources may include: determining a plurality of candidate resources from the plurality of initial resources based on the plurality of expected interaction attributes; processing the plurality of candidate resources using a resource recommendation large model to obtain a target resource; and recommending at least one target resource to the target object.

[0104] According to an embodiment of the present disclosure, multiple initial resources may be sorted based on their respective expected interaction attributes, thereby screening out candidate resources with higher expected interaction operations of the target object.

[0105] According to an embodiment of the present disclosure, the resource recommendation model can be a deep learning model with large-scale parameters based on any type of large language model, multimodal large model, etc. The resource recommendation model can be determined after fine-tuning based on the expected interaction properties of sample resources and tags.

[0106] According to an embodiment of the present disclosure, candidate resources are determined from initial resources, and a plurality of candidate resources are processed using a resource recommendation large model to obtain a target resource, which can further improve the matching degree between the target resource and the historical interaction preference of the target object, thereby improving the recommendation accuracy of the target resource.

[0107] According to an embodiment of the present disclosure, using a resource recommendation big model to process multiple candidate resources to obtain a target resource may include: for the same candidate resource, using a resource feature fusion network of the resource recommendation big model to perform feature fusion on multiple candidate resources to obtain a candidate resource fusion feature corresponding to the same candidate resource; determining candidate expected interaction attributes based on the candidate resource fusion feature; and determining the target resource based on the candidate expected interaction attribute of each of at least one candidate resource.

[0108] According to an embodiment of the present disclosure, a resource feature fusion network can be constructed based on an attention network algorithm. For a target candidate resource among multiple candidate resources, the resource feature fusion network can set a specified attention weight for the target candidate resource, so that the resource semantic attributes of the target candidate resource can be fully integrated with the resource semantic attributes of other candidate resources, and fully learn the resource semantic attribute relationship between the target candidate resource and other candidate resources, so as to improve the representation accuracy of the candidate resource fusion feature for the semantic attributes of the target candidate resource and the semantic attribute relationship between all multiple candidate resources.

[0109] According to an embodiment of the present disclosure, determining candidate expected interactive attributes based on candidate resource fusion features may include processing candidate resource fusion features based on an expected attribute detection head of a resource recommendation large model to obtain candidate expected interactive attributes of the candidate resource. The candidate expected interactive attributes may include expected browsing time, expected click-through rate, expected completion rate, expected interactive operation, etc.

[0110] According to an embodiment of the present disclosure, multiple candidate resources may be ranked based on their respective candidate expected interaction attributes to select a candidate resource with a higher ranking as a target resource, thereby more accurately detecting a resource having an interactive operation preference for the target object.

[0111] According to an embodiment of the present disclosure, a resource feature fusion network of a resource recommendation large model is used to perform feature fusion on multiple candidate resources, and the candidate resource fusion features corresponding to the same candidate resource include: using the resource feature fusion network to process multiple candidate resources and set semantic attributes to obtain candidate resource fusion features.

[0112] According to an embodiment of the present disclosure, a resource feature fusion network can be constructed based on an attention network algorithm, for example, a resource feature fusion network can be constructed based on a Transformer algorithm.

[0113] According to an embodiment of the present disclosure, a collection of semantic attributes is determined by semantically encoding multiple candidate resources. The collection of semantic attributes can represent the global semantic attributes of multiple candidate resources, and the resource semantic attribute relationship between multiple candidate resources in batches can be represented through the collection of semantic attributes. By using a resource feature fusion network to process multiple candidate resources and collection semantic attributes, the semantic attributes and global semantic attributes of the target candidate resources can be fused based on the attention mechanism, as well as the semantic attributes of the target candidate resources and the local semantic attributes of other candidate resources. In this way, the candidate resource fusion features can more accurately represent the semantic attribute relationship between the target candidate resource and other candidate resources, thereby fine-grainedly representing the target object's interactive operation preference for the target candidate resource, so as to improve the detection accuracy of the candidate's expected interactive attributes.

[0114] According to an embodiment of the present disclosure, the resource feature fusion network includes an attention network layer, a recommendation gating layer, and multiple recommendation expert networks.

[0115] According to an embodiment of the present disclosure, using a resource feature fusion network to process multiple candidate resources and resource set semantic attributes to obtain candidate resource fusion features may include: using an attention network layer to fuse multiple candidate resources and resource set semantic attributes to obtain intermediate resource fusion features; using a recommendation gating layer to process intermediate candidate resource fusion features to obtain recommendation gating features; and using a target recommendation expert network corresponding to the recommendation gating features in multiple recommendation expert networks to process intermediate candidate resource fusion features to obtain candidate resource fusion features.

[0116] According to an embodiment of the present disclosure, the recommendation gating layer may include one or more. The recommendation gating layer may be constructed based on any type of neural network algorithm such as a long short-term memory network, a convolutional neural network, etc. The recommendation gating feature may indicate the recommendation gating weights of multiple recommendation expert networks, and a target recommendation expert network may be determined from multiple recommendation expert networks based on the recommendation gating weights.

[0117] According to an embodiment of the present disclosure, when a target recommendation expert network includes multiple target recommendation expert networks, the multiple target recommendation expert networks can each process an intermediate resource fusion feature to obtain multiple intermediate candidate resource fusion features. The multiple intermediate candidate resource fusion features are fused based on multiple recommendation gating weights (for example, multiple intermediate candidate resource fusion features are weighted averaged based on multiple recommendation gating weights) to obtain a candidate resource fusion feature.

[0118] According to the embodiments of the present disclosure, the target recommendation expert network is selected to process the intermediate resource fusion features according to the recommendation gating layer, which can reduce the number of model parameters activated simultaneously in the resource recommendation large model and reduce the computational overhead. The candidate expected interaction attributes can be further accurately predicted based on the learning ability of the target recommendation expert network.

[0119] According to an embodiment of the present disclosure, determining a target resource based on a candidate expected interaction attribute of each of at least one candidate resource may include: determining the target resource based on at least one candidate expected interaction attribute and a set expected interaction attribute.

[0120] According to an embodiment of the present disclosure, the expected interaction attribute of a set is determined by processing a plurality of candidate resources and the semantic attribute of the set using a resource feature fusion network with respect to the semantic attribute of the set.

[0121] In one example, the collection semantic attribute can be used as the target candidate resource, and the resource feature fusion network can be used to process multiple candidate resources and the collection semantic attribute to obtain the collection fusion feature corresponding to the collection semantic attribute. The candidate expected interaction attribute detection head of the resource recommendation model is used to process the collection fusion feature to obtain the collection expected interaction attribute.

[0122] According to an embodiment of the present disclosure, determining a target resource based on at least one candidate expected interaction attribute and a set expected interaction attribute may include fusing the candidate expected interaction attribute and the set expected interaction attribute to obtain target candidate expected interaction attributes of multiple target candidate resources. In the case where multiple candidate resources are used as a candidate resource set, a comprehensive sorting can be performed based on the target candidate expected interaction attributes of each candidate resource in the multiple candidate resource sets, thereby determining the target resource that best meets the interaction attribute preference of the target object, thereby further improving the target resource recommendation accuracy.

[0123] In one example, a sample candidate resource set including multiple sample candidate resources and a sample set semantic attribute can be used as training data, and a tag interaction attribute in a historical period of a target object can be used as a tag. The training data and tags are used to fine-tune the resource recommendation model to obtain a trained resource recommendation model. The tag interaction attribute may include a click tag that represents whether it is clicked, a viewing time tag, a follow operation tag, a sharing operation tag, a like operation tag, a favorite operation tag, and a comment operation tag. For the sample set semantic attribute, a click tag that represents whether it is clicked and a tag that represents whether an interactive operation is performed can be determined. The training data and tags are processed based on a loss function to obtain a loss value. The model parameters of the resource recommendation model are adjusted according to the loss value to obtain a trained resource recommendation model. The trained resource recommendation model can be applied to the method provided in the embodiment of the present disclosure.

[0124] It should be noted that the interactive operation may include any type of interactive operation on the resource, such as a like operation, a favorite operation, a comment operation, a share operation, and a follow operation performed by the target object.

[0125] It should be understood that the candidate expected interaction attribute may have the same or similar specific attribute type as the expected interaction attribute. The candidate expected interaction attribute may include, for example, click-through rate, completion rate, expected browsing time, and the like.

[0126] Figure 6 The schematic diagram shows a principle diagram of a resource recommendation model according to an embodiment of the present disclosure.

[0127] like Figure 6 As shown, the resource recommendation model 600 may include an attention network layer 610, a recommendation gating layer 621, multiple recommendation expert networks, and a candidate expected attribute detection layer 640. The multiple recommendation expert networks may include a first recommendation expert network 631, a second recommendation expert network 632, ... to an nth recommendation expert network 63n.

[0128] like Figure 6As shown, the first candidate resource T601, the second candidate resource T602, ... to the nth candidate resource T60n are input into the resource semantic encoding layer, and the resource set semantic attribute T600' is output. For the first candidate resource T601, the first candidate resource T601, the second candidate resource T602, ... to the nth candidate resource T60n, and the resource set semantic attribute T600' can be input into the attention network layer 610 to obtain the intermediate resource fusion feature. The intermediate resource fusion feature is input into the recommendation gating layer 621 to obtain the recommendation gating feature. The recommendation gating feature can indicate the first recommendation expert network 631 and the second recommendation expert network 632 as the target recommendation expert network. The intermediate resource fusion feature is input into the first recommendation expert network 631 and the second recommendation expert network 632 respectively to obtain the candidate resource fusion features output by the first recommendation expert network 631 and the second recommendation expert network 632 respectively. The fusion features of the two candidate resources are input into the candidate expected attribute detection layer 640, and the first candidate expected interaction attribute 601 of the first candidate resource T601 can be output.

[0129] like Figure 6 As shown, for the second candidate resource T602 ... to the nth candidate resource T60n, the second candidate expected interaction attribute 602, ... to the nth candidate expected interaction attribute 60n of the second candidate resource T602 ... to the nth candidate resource T60n can be determined based on the same or similar operation process as that for the first candidate resource T601. The applicant will not elaborate on this.

[0130] Figure 7 The application scenario diagram of the large model-based resource recommendation method according to an embodiment of the present disclosure is schematically shown.

[0131] like Figure 7 As shown, the resource recommendation method based on the large model can be executed based on the target historical resource screening module 710, the large model module 720, the resource recommendation large model module 730, the target resource recommendation module 740 and the interaction feedback module 750. The historical interaction resources and the initial resources to be processed are determined from the resource pool. The historical interaction resources and the initial resources are transmitted to the target historical resource screening module 710, and the target historical resource is determined from multiple historical interaction resources to realize the recall of the historical interaction resources.

[0132] The initial resource and the target historical resource are transmitted to the large model module 720 to process the target historical resource and obtain the expected interaction attributes of each of the initial resources. Candidate resources are determined from the initial resource total according to the expected interaction attributes of each of the initial resources. The candidate resources are processed by the resource recommendation large model module 730 to obtain the candidate expected interaction attributes. The target resource recommendation module reorders the candidate resources based on the candidate expected interaction attributes of each of the candidate resources. The target resource is recommended to the target object based on the sorting result.

[0133] The interactive feedback module 740 can obtain the interactive operation results of the target object on the target resource with the authorization of the target object, and update the big model of the big model module 720 and the resource recommendation big model of the resource recommendation big model module 730 based on the interactive operation results, so as to timely capture the interests and preferences of the target object and improve the accuracy of subsequent resource recommendations.

[0134] According to the embodiments of the present disclosure, according to the resource recommendation method based on the big model provided by the embodiments of the present disclosure, the historical interaction resources and initial resources can be quantitatively encoded in combination with the interaction behavior of the target object in the historical time period, and the full feedback process recommendation system can be comprehensively reconstructed through autoregressive modeling to achieve multiple resource sorting and screening based on the big model, and realize an end-to-end resource recommendation system.

[0135] The embodiments of the present disclosure also provide an artificial intelligence agent, which is configured to execute the large model-based resource recommendation method provided by the embodiments of the present disclosure.

[0136] The agent mechanism of the AI ​​(Artificial Intelligence) agent according to the embodiment of the present disclosure comprehensively covers the industry's recognized capability definition of the AI ​​agent, and the functions within the modules are highly unified, the responsibilities between modules are clearly divided, and the communication mechanism is simple and effective.

[0137] In addition, the intelligent agent mechanism of the AI ​​intelligent agent according to the embodiment of the present disclosure is not designed to rely on specific application scenarios, but is a general intelligent agent mechanism suitable for various environments. Therefore, it has a wide range of application scenarios, high flexibility, and strong customizability.

[0138] At the same time, the agent mechanism of the AI ​​agent according to the embodiment of the present disclosure incorporates the evaluation ability and the reflection ability into the control ability, thereby improving the upper limit of the ability of the AI ​​agent. In addition, the concept of cognitive memory is also proposed at the same time, giving full play to the inductive reasoning ability of the AI ​​agent.

[0139] Figure 8 The block diagram of a resource recommendation device based on a large model according to an embodiment of the present disclosure is schematically shown.

[0140] like Figure 8 As shown, the resource recommendation device 800 based on the large model includes: an acquisition module 810, a target historical resource determination module 820, an expected interaction attribute acquisition module 830 and a recommendation module 840.

[0141] The acquisition module 810 is used to acquire multiple initial resources to be processed and multiple historical interaction resources of the target object.

[0142] The target historical resource determination module 820 is configured to determine a target historical resource from a plurality of historical interaction resources based on similarities between the plurality of historical interaction resources and at least one initial resource.

[0143] The expected interaction attribute acquisition module 830 is used to process the target historical resource using the large model to obtain the expected interaction attribute of the initial resource, where the expected interaction attribute represents the attribute of the interaction operation performed by the target object on the initial resource.

[0144] The recommendation module 840 is used to recommend a target resource from among the multiple initial resources to the target object based on the expected interaction attributes of each of the multiple initial resources.

[0145] According to an embodiment of the present disclosure, the large model includes a feature fusion network and multiple expert network layers.

[0146] The expected interaction attribute acquisition module includes: an initial fusion feature acquisition submodule and an expected interaction attribute acquisition submodule.

[0147] The initial fusion feature acquisition submodule is used to process the object attribute information of the target historical resources and the target object using the feature fusion network to obtain the initial fusion feature.

[0148] The expected interaction attribute acquisition submodule is used to process the initial fusion features based on the target expert network layer corresponding to the recommendation scenario of the expected interaction attribute in multiple expert network layers to obtain at least one expected interaction attribute.

[0149] According to an embodiment of the present disclosure, the expected interaction attribute obtaining submodule includes: an intermediate fusion feature obtaining unit, a first target fusion feature obtaining unit and an expected interaction attribute obtaining unit.

[0150] The intermediate fusion feature obtaining unit is used to use multiple target expert network layers to process the initial fusion features respectively to obtain multiple intermediate fusion features.

[0151] The first target fusion feature obtaining unit is used to fuse multiple intermediate fusion features based on the gated weights corresponding to the recommendation scene to obtain the first target fusion feature.

[0152] The expected interaction attribute obtaining unit is used to determine at least one expected interaction attribute based on the first target fusion feature.

[0153] According to an embodiment of the present disclosure, the large model further includes a plurality of attribute detection heads.

[0154] According to an embodiment of the present disclosure, the expected interaction attribute obtaining unit includes an expected interaction attribute obtaining sub-unit.

[0155] The expected interaction attribute obtaining subunit is used to process the first target fusion feature based on the target attribute detection head corresponding to the recommended scene among the multiple attribute detection heads to obtain at least one expected interaction attribute.

[0156] According to an embodiment of the present disclosure, the initial fusion feature obtaining submodule includes an initial fusion feature obtaining unit.

[0157] The initial fusion feature acquisition unit is used to fuse the target historical resources and object attribute information based on the attention mechanism to obtain the initial fusion feature.

[0158] According to an embodiment of the present disclosure, the recommendation scenario includes at least one of the following: a primary feedback scenario, a resource discovery scenario, and a resource two-hop scenario.

[0159] According to an embodiment of the present disclosure, the recommendation module includes: a candidate resource acquisition submodule, a target resource acquisition submodule and a recommendation submodule.

[0160] The candidate resource obtaining submodule is used to determine a plurality of candidate resources from a plurality of initial resources based on a plurality of expected interaction attributes.

[0161] The target resource acquisition submodule is used to use the resource recommendation model to process multiple candidate resources and obtain the target resource.

[0162] The recommendation submodule is used to recommend at least one target resource to the target object.

[0163] According to an embodiment of the present disclosure, the target resource acquisition submodule includes: a candidate resource fusion feature acquisition unit, a candidate expected interaction attribute determination unit and a target resource determination unit.

[0164] The candidate resource fusion feature acquisition unit is used to fuse the features of multiple candidate resources for the same candidate resource using the resource feature fusion network of the resource recommendation large model to obtain the candidate resource fusion features corresponding to the same candidate resource.

[0165] The candidate expected interaction attribute determination unit is used to determine the candidate expected interaction attribute based on the candidate resource fusion feature.

[0166] The target resource determination unit is used to determine the target resource based on the candidate expected interaction attribute of each of at least one candidate resource.

[0167] According to an embodiment of the present disclosure, the candidate resource fusion feature obtaining unit includes a candidate resource fusion feature obtaining sub-unit.

[0168] The candidate resource fusion feature acquisition subunit is used to process multiple candidate resources and set semantic attributes using a resource feature fusion network to obtain candidate resource fusion features, wherein the set semantic attributes are determined by semantically encoding multiple candidate resources.

[0169] The candidate resource fusion feature acquisition subunit, the resource feature fusion network includes an attention network layer, a recommendation gating layer and multiple recommendation expert networks.

[0170] According to an embodiment of the present disclosure, the candidate resource fusion feature acquisition subunit is configured as follows: using the attention network layer to fuse multiple candidate resources and resource set semantic attributes to obtain intermediate resource fusion features; using the recommendation gating layer to process the intermediate candidate resource fusion features to obtain recommendation gating features; and using the target recommendation expert network corresponding to the recommendation gating features in multiple recommendation expert networks to process the intermediate candidate resource fusion features to obtain candidate resource fusion features.

[0171] According to an embodiment of the present disclosure, the target resource determining unit includes a target resource determining subunit.

[0172] The target resource determination subunit is used to determine the target resource based on at least one candidate expected interaction attribute and a set expected interaction attribute, wherein the set expected interaction attribute is determined by processing multiple candidate resources and set semantic attributes using a resource feature fusion network for set semantic attributes.

[0173] According to an embodiment of the present disclosure, the acquisition module includes: a historical resource feature acquisition submodule and an initial resource acquisition submodule.

[0174] The historical resource feature acquisition submodule is used to semantically fuse multiple acquired historical interaction resources to obtain historical resource features.

[0175] The initial resource acquisition submodule is used to perform resource retrieval based on historical resource characteristics to obtain multiple initial resources.

[0176] According to an embodiment of the present disclosure, the expected interaction attribute includes at least one of the following: click-through rate, video completion rate, and browsing time.

[0177] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0178] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method as described above.

[0179] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the method as described above.

[0180] According to an embodiment of the present disclosure, a computer program product includes a computer program, and when the computer program is executed by a processor, the computer program implements the method as described above.

[0181] Fig. 9 A block diagram of an electronic device suitable for implementing a resource recommendation method based on a large model according to an embodiment of the present disclosure is schematically shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0182] like Fig. 9 As shown, the device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the device 900 can also be stored. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0183] A number of components in the device 900 are connected to the I / O interface 905, including: an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0184] The computing unit 901 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 901 performs the various methods and processes described above, such as a resource recommendation method based on a large model. For example, in some embodiments, the resource recommendation method based on a large model may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 908. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the resource recommendation method based on the large model described above may be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to execute the large model-based resource recommendation method in any other appropriate manner (for example, by means of firmware).

[0185] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0186] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0187] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0188] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0189] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0190] A computer system may include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises through computer programs running on respective computers and having a client-server relationship to each other. The server may be a cloud server, a server in a distributed system, or a server combined with a blockchain.

[0191] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0192] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A resource recommendation method based on a large model, comprising: Acquire multiple initial resources to be processed and multiple historical interaction resources of the target object; Determining a target historical resource from the plurality of historical interaction resources based on similarities between the plurality of historical interaction resources and at least one of the initial resources; Processing the target historical resource using the large model to obtain expected interaction attributes of the initial resource, wherein the expected interaction attributes represent attributes of the interaction operation performed by the target object on the initial resource; as well as Based on the expected interaction attributes of each of the multiple initial resources, a target resource from the multiple initial resources is recommended to the target object.

2. The method according to claim 1, wherein: The large model includes a feature fusion network and multiple expert network layers; Wherein, the use of the large model to process the target historical resource to obtain the expected interactive attributes of the initial resource includes: Using the feature fusion network to process the target historical resource and the object attribute information of the target object to obtain an initial fusion feature; Based on the target expert network layer corresponding to the recommendation scenario of the expected interaction attribute among the multiple expert network layers, the initial fusion features are processed to obtain at least one of the expected interaction attributes.

3. The method according to claim 2, wherein: The processing of the initial fusion features by a target expert network layer corresponding to the recommendation scenario of the expected interaction attribute among the plurality of expert network layers includes: Using a plurality of target expert network layers to process the initial fusion features respectively to obtain a plurality of intermediate fusion features; fusing the plurality of intermediate fused features based on the gated weights corresponding to the recommendation scenario to obtain a first target fused feature; and At least one of the expected interaction attributes is determined based on the first target fusion feature.

4. The method according to claim 3, wherein: The large model also includes a plurality of attribute detection heads; Wherein, determining at least one of the expected interaction attributes based on the first target fusion feature includes: The first target fusion feature is processed based on a target attribute detection head corresponding to the recommendation scene among the multiple attribute detection heads to obtain at least one of the expected interaction attributes.

5. The method according to claim 2, wherein: The process of processing the target historical resource and the object attribute information of the target object by using the feature fusion network to obtain the initial fusion feature includes: The target historical resource and the object attribute information are fused based on the attention mechanism to obtain the initial fusion feature.

6. The method according to claim 2, wherein: The recommended scenario includes at least one of the following: Primary feedback scenario, resource discovery scenario, and resource two-hop scenario.

7. The method according to claim 1, wherein: The recommending a target resource from among the plurality of initial resources to the target object based on the respective expected interaction attributes of the plurality of initial resources comprises: Based on the plurality of expected interaction attributes, determining a plurality of candidate resources from the plurality of initial resources; Processing the plurality of candidate resources using a resource recommendation model to obtain a target resource; and At least one of the target resources is recommended to the target object.

8. The method according to claim 7, wherein: The resource recommendation model is used to process the plurality of candidate resources to obtain target resources including: For the same candidate resource, the resource feature fusion network of the resource recommendation model is used to fuse the features of multiple candidate resources to obtain candidate resource fusion features corresponding to the same candidate resource; Determining candidate expected interaction attributes based on the candidate resource fusion features; and The target resource is determined based on at least one candidate expected interaction attribute of each of the candidate resources.

9. The method according to claim 8, wherein: The resource feature fusion network of the resource recommendation large model is used to fuse features of the plurality of candidate resources to obtain candidate resource fusion features corresponding to the same candidate resource, including: The resource feature fusion network is used to process the plurality of candidate resources and the set semantic attributes to obtain the candidate resource fusion features, wherein the set semantic attributes are determined by semantically encoding the plurality of candidate resources.

10. The method according to claim 9, wherein: The resource feature fusion network includes an attention network layer, a recommendation gating layer and multiple recommendation expert networks; Wherein, the step of using the resource feature fusion network to process the plurality of candidate resources and the set semantic attributes to obtain the candidate resource fusion features includes: Using the attention network layer to fuse the plurality of candidate resources and the semantic attributes of the resource set, to obtain an intermediate resource fusion feature; Using the recommendation gating layer to process the intermediate candidate resource fusion features to obtain recommendation gating features; and The target recommendation expert network corresponding to the recommendation gating feature among the plurality of recommendation expert networks is used to process the intermediate candidate resource fusion feature to obtain the candidate resource fusion feature.

11. The method according to claim 9, wherein: The determining the target resource based on the candidate expected interaction attribute of at least one of the candidate resources comprises: The target resource is determined based on at least one of the candidate expected interaction attributes and the set expected interaction attribute, wherein the set expected interaction attribute is determined by processing a plurality of the candidate resources and the set semantic attributes using the resource feature fusion network for the set semantic attributes.

12. The method according to claim 1, wherein: The acquiring of a plurality of initial resources to be processed comprises: Performing semantic fusion on the acquired multiple historical interaction resources to obtain historical resource features; and Resource retrieval is performed based on the historical resource characteristics to obtain a plurality of the initial resources.

13. The method according to claim 1, wherein: The expected interaction attribute includes at least one of the following: Click-through rate, video completion rate, and browsing time.

14. An artificial intelligence agent configured to execute the method according to any one of claims 1 to 13.

15. A resource recommendation device based on a large model, comprising: An acquisition module, used to acquire multiple initial resources to be processed and multiple historical interaction resources of a target object; A target historical resource determination module, configured to determine a target historical resource from a plurality of the historical interaction resources based on similarities between the plurality of the historical interaction resources and at least one of the initial resources; An expected interaction attribute acquisition module, used to process the target historical resource using a large model to obtain the expected interaction attribute of the initial resource, wherein the expected interaction attribute represents the attribute of the interaction operation performed by the target object on the initial resource; as well as The recommendation module is used to recommend a target resource from among the multiple initial resources to the target object based on the expected interaction attributes of each of the multiple initial resources.

16. The device according to claim 15, wherein: The large model includes a feature fusion network and multiple expert network layers; Wherein, the expected interaction attribute acquisition module includes: An initial fusion feature acquisition submodule, used to process the target historical resource and the object attribute information of the target object using the feature fusion network to obtain an initial fusion feature; The expected interaction attribute acquisition submodule is used to process the initial fusion features based on a target expert network layer corresponding to the recommendation scenario of the expected interaction attribute among the multiple expert network layers to obtain at least one of the expected interaction attributes.

17. The device according to claim 16, wherein: The expected interaction attribute acquisition submodule includes: An intermediate fusion feature obtaining unit, used to use a plurality of target expert network layers to process the initial fusion features respectively to obtain a plurality of intermediate fusion features; A first target fusion feature obtaining unit, configured to fuse the plurality of intermediate fusion features based on a gated weight corresponding to the recommendation scenario to obtain a first target fusion feature; and The expected interaction attribute obtaining unit is used to determine at least one of the expected interaction attributes based on the first target fusion feature.

18. The device according to claim 17, wherein: The large model also includes a plurality of attribute detection heads; Wherein, the expected interaction attribute obtaining unit includes: The expected interaction attribute obtaining subunit is used to process the first target fusion feature based on the target attribute detection head corresponding to the recommended scene among the multiple attribute detection heads to obtain at least one of the expected interaction attributes.

19. The device according to claim 16, wherein: The initial fusion feature acquisition submodule includes: The initial fusion feature obtaining unit is used to fuse the target historical resource and the object attribute information based on the attention mechanism to obtain the initial fusion feature.

20. The device according to claim 16, wherein: The recommended scenario includes at least one of the following: Primary feedback scenario, resource discovery scenario, and resource two-hop scenario.

21. The device according to claim 15, wherein: The recommendation module includes: A candidate resource obtaining submodule, configured to determine a plurality of candidate resources from a plurality of the initial resources based on a plurality of the expected interaction attributes; A target resource acquisition submodule is used to process the plurality of candidate resources using a resource recommendation macromodel to obtain a target resource; and The recommendation submodule is used to recommend at least one target resource to the target object.

22. The device according to claim 21, wherein The target resource acquisition submodule includes: A candidate resource fusion feature obtaining unit is used to fuse features of a plurality of candidate resources using a resource feature fusion network of the resource recommendation macro model for the same candidate resource, so as to obtain a candidate resource fusion feature corresponding to the same candidate resource; a candidate expected interaction attribute determination unit, configured to determine a candidate expected interaction attribute based on the candidate resource fusion feature; and The target resource determination unit is used to determine the target resource based on the candidate expected interaction attribute of at least one of the candidate resources.

23. The device according to claim 22, wherein: The candidate resource fusion feature obtaining unit includes: The candidate resource fusion feature obtaining subunit is used to use the resource feature fusion network to process the plurality of candidate resources and the set semantic attributes to obtain the candidate resource fusion features, wherein the set semantic attributes are determined by semantically encoding the plurality of candidate resources.

24. The device according to claim 23, wherein: The resource feature fusion network includes an attention network layer, a recommendation gating layer and multiple recommendation expert networks; Wherein, the candidate resource fusion feature obtaining subunit is configured as follows: Using the attention network layer to fuse the plurality of candidate resources and the semantic attributes of the resource set, to obtain an intermediate resource fusion feature; Using the recommendation gating layer to process the intermediate candidate resource fusion features to obtain recommendation gating features; and The target recommendation expert network corresponding to the recommendation gating feature in the plurality of recommendation expert networks is used to process the intermediate candidate resource fusion feature to obtain the candidate resource fusion feature.

25. The device according to claim 23, wherein: The target resource determination unit comprises: The target resource determination subunit is used to determine the target resource based on at least one of the candidate expected interaction attributes and the set expected interaction attribute, wherein the set expected interaction attribute is determined by processing multiple candidate resources and set semantic attributes using the resource feature fusion network for the set semantic attribute.

26. The device according to claim 15, wherein: The acquisition module comprises: A historical resource feature acquisition submodule is used to semantically fuse the acquired multiple historical interaction resources to obtain historical resource features; and The initial resource acquisition submodule is used to perform resource retrieval based on the historical resource characteristics to obtain a plurality of the initial resources.

27. The device according to claim 15, wherein: The expected interaction attribute includes at least one of the following: Click-through rate, video completion rate, and browsing time.

28. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 14.

29. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 14.

30. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 14.

Citation Information

Cited By

  • Resource recommendation method, method and device for training deep learning model, and intelligent agent

    CN120653846A

  • Resource content understanding and resource recommendation method and device, and electronic equipment

    CN121210759A