Multimedia resource recommendation method and device, computer equipment and storage medium

Abstract

The invention discloses a multimedia resource recommendation method and device, computer equipment and a storage medium, and belongs to the technical field of computers. The method comprises the steps of obtaining task features of a target recommendation task; performing feature extraction on the task features through a scene expert network in the recommendation model to obtain scene granularity features; performing feature extraction on the task features through a task expert network in the recommendation model to obtain task granularity features; performing feature extraction on the task features through a global expert network in the recommendation model to obtain global granularity features; based on the scene granularity feature, the task granularity feature and the global granularity feature, a prediction result is determined, and the prediction result is used for representing the interaction condition of the object and the multimedia resource under the condition that the multimedia resource is recommended to the object. The problem that the recommendation effect of a recommendation model shared by multiple tasks is poor can be effectively relieved, and the accuracy of multimedia resource recommendation is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a multimedia resource recommendation method, apparatus, computer equipment, and storage medium. Background Art

[0002] With the rapid development of computer technology, artificial intelligence technology is being used more and more widely in the field of recommendation. In the process of recommending multimedia resources to users, users will provide feedback on the multimedia resources they receive. The feedback data can be used as the basis for the recommendation system to make recommendations to users.

[0003] However, due to the diversity of different recommendation scenarios and the increasing variety of different recommendation tasks in the same recommendation scenario, the differences between different recommendation tasks are becoming increasingly larger, resulting in poorer recommendation effects using the same recommendation system and affecting the accuracy of multimedia resource recommendations. Summary of the Invention

[0004] The embodiments of the present application provide a multimedia resource recommendation method, apparatus, computer device, and storage medium, which can improve the accuracy of multimedia resource recommendations. The technical solution is as follows:

[0005] In one aspect, a multimedia resource recommendation method is provided, the method comprising:

[0006] Obtaining task features of a target recommendation task, the task features including object features of an object and resource features of a multimedia resource, the target recommendation task belonging to any one of a plurality of recommendation scenarios, and each recommendation scenario corresponding to at least one recommendation task;

[0007] The task features are extracted through the scene expert network in the recommendation model to obtain scene granularity features;

[0008] Extracting the task features through the task expert network in the recommendation model to obtain task granularity features;

[0009] Extracting the task features through the global expert network in the recommendation model to obtain global granular features;

[0010] A prediction result is determined based on the scene granularity feature, the task granularity feature, and the global granularity feature, where the prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0011] In another aspect, a multimedia resource recommendation device is provided, the device comprising:

[0012] A feature acquisition module is used to acquire task features of a target recommendation task, wherein the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any of a plurality of recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0013] A feature extraction module is used to extract the task features through the scene expert network in the recommendation model to obtain scene granularity features;

[0014] The feature extraction module is further configured to extract the task features through the task expert network in the recommendation model to obtain task granularity features;

[0015] The feature extraction module is further configured to extract the task features through the global expert network in the recommendation model to obtain global granularity features;

[0016] A prediction module is used to determine a prediction result based on the scene granularity feature, the task granularity feature and the global granularity feature, wherein the prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0017] Optionally, the scene expert network includes a scene memory expert network and a scene generalization expert network; the feature extraction module is used to:

[0018] Obtaining a memory bias feature and a generalization bias feature from the task features, wherein the memory bias feature refers to a feature affected by the interaction between the object and the multimedia resource, and the generalization bias feature refers to a feature not affected by the interaction between the object and the multimedia resource;

[0019] Extracting the memory bias features through the scene memory expert network to obtain scene granularity memory features;

[0020] The generalization bias feature is extracted through the scene generalization expert network to obtain a scene granularity generalization feature, where the scene granularity feature includes the scene granularity memory feature and the scene granularity generalization feature.

[0021] Optionally, the task expert network is a task memory expert network; and the feature extraction module is used to:

[0022] Acquire a memory bias feature from the task feature, wherein the memory bias feature refers to a feature affected by the interaction between the object and the multimedia resource;

[0023] The memory bias feature is extracted through the task memory expert network to obtain a task granularity memory feature, wherein the task granularity feature includes the task granularity memory feature.

[0024] Optionally, the global expert network is a global generalized expert network; and the feature extraction module is used to:

[0025] Obtaining a generalized bias feature from the task feature, wherein the generalized bias feature refers to a feature that is not affected by the interaction between the object and the multimedia resource;

[0026] The generalization bias feature is extracted through the global generalization expert network to obtain a global granularity generalization feature, which includes the global granularity generalization feature.

[0027] Optionally, the number of the scene expert networks is m, the number of the scene granularity features is m, and the network parameters of each scene expert network are different; the number of the task expert networks is n, the number of the task granularity features is n, and the network parameters of each task expert network are different; the number of the global expert networks is k, the number of the global granularity features is k, and the network parameters of each global expert network are different; wherein m, n and k are all positive integers.

[0028] Optionally, the recommendation model further includes a fusion network; the prediction module is configured to:

[0029] a fusion unit, configured to fuse the scene granularity features, the task granularity features, and the global granularity features through a fusion network in the recommendation model to obtain a target fusion feature;

[0030] A prediction unit is used to generate the prediction result based on the target fusion feature.

[0031] Optionally, the scene granularity features include scene granularity memory features and scene granularity generalization features, the task granularity features include task granularity memory features, and the global granularity features include global granularity generalization features; the fusion network includes a memory fusion network, a generalization fusion network, and a target fusion network;

[0032] The fusion unit is used to:

[0033] The scene granularity memory feature and the task granularity memory feature are fused through the memory fusion network to obtain a memory fusion feature;

[0034] The scene granularity generalized features and the global granularity generalized features are fused through the generalized fusion network to obtain a generalized fusion feature;

[0035] The memory fusion feature and the generalization fusion feature are fused through the target fusion network to obtain the target fusion feature.

[0036] Optionally, the task features include memory bias features and generalization bias features, wherein the memory bias features refer to features affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resource; the fusion unit is configured to:

[0037] Normalizing the concatenation result of the memory bias feature and the generalization bias feature through the memory fusion network to obtain a first memory weight and a second memory weight;

[0038] The memory fusion network performs weighted fusion on the scene granularity memory feature and the task granularity memory feature based on the first memory weight and the second memory weight to obtain the memory fusion feature.

[0039] Optionally, the task features include memory bias features and generalization bias features, wherein the memory bias features refer to features affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resource; the fusion unit is configured to:

[0040] Normalizing the concatenation result of the memory bias feature and the generalization bias feature through the generalization fusion network to obtain a first generalization weight and a second generalization weight;

[0041] The generalized fusion network performs weighted fusion on the scene granularity generalized features and the global granularity generalized features based on the first generalization weight and the second generalization weight to obtain the generalized fusion features.

[0042] Optionally, the fusion unit is used to:

[0043] Acquire a memory-biased resource feature from the multimedia resource features, wherein the memory-biased resource feature refers to a multimedia resource feature affected by an interaction between the object and the multimedia resource;

[0044] Normalizing based on the memory bias resource characteristics through the target fusion network to obtain a first target weight and a second target weight;

[0045] The target fusion network performs weighted fusion on the memory fusion feature and the generalization fusion feature based on the first target weight and the second target weight to obtain the target fusion feature.

[0046] Optionally, the number of the scene granularity memory features, the scene granularity generalization features, the task granularity memory features, and the global granularity generalization features is multiple; the fusion network further includes a first scene granularity fusion network, a second scene granularity fusion network, a task granularity fusion network, and a global granularity fusion network; the fusion unit is further used to:

[0047] fusing multiple scene granularity memory features through the first scene granularity fusion network to obtain a fused scene granularity memory feature;

[0048] fusing multiple scene granularity generalization features through the second scene granularity fusion network to obtain a fused scene granularity generalization feature;

[0049] By using the task granularity fusion network, multiple task granularity memory features are fused to obtain fused task granularity memory features;

[0050] By using the global granularity fusion network, multiple global granularity generalization features are fused to obtain a fused global granularity generalization feature;

[0051] Among them, the input of the memory fusion network is the fused scene granularity memory features and the fused task granularity memory features, and the input of the generalization fusion network is the fused scene granularity generalization features and the fused global granularity generalization features.

[0052] Optionally, the fusion unit is used to:

[0053] For any scene granularity memory feature, normalizing the scene granularity memory feature by the first scene granularity fusion network based on the product of the scene granularity memory feature and the transpose of the scene granularity memory feature to obtain a first scene weight of the scene granularity memory feature;

[0054] The first scene granularity fusion network performs weighted fusion on the multiple scene granularity memory features based on their first scene weights to obtain the fused scene granularity memory features.

[0055] Optionally, the prediction result is a prediction probability, which is a probability that the object will interact with the multimedia resource when the multimedia resource is recommended to the object; after determining the prediction result based on the scene granularity feature, the task granularity feature, and the global granularity feature, the device further includes a recommendation module for:

[0056] When the predicted probability is greater than a preset probability threshold, the multimedia resource is recommended to the object.

[0057] Optionally, the device further includes a model training module, configured to:

[0058] Obtaining sample task features and sample labels for a sample recommendation task, wherein the sample task features include sample object features of a sample object and sample resource features of a sample multimedia resource. The sample recommendation task belongs to any recommendation task in any recommendation scenario. The sample label is used to characterize the actual interaction between the sample object and the sample multimedia resource after the sample multimedia resource is recommended to the sample object.

[0059] Extracting the sample task features through the scene expert network in the recommendation model to obtain sample scene granularity features;

[0060] Extracting the sample task features through the task expert network in the recommendation model to obtain sample task granularity features;

[0061] Extracting the sample task features through the global expert network in the recommendation model to obtain the sample global granularity features;

[0062] Determining a sample prediction result based on the sample scene granularity feature, the sample task granularity feature, and the sample global granularity feature, wherein the sample prediction result is used to characterize an interaction between the sample object and the sample multimedia resource when the sample multimedia resource is recommended to the sample object;

[0063] The recommendation model is trained based on the difference between the sample prediction result and the sample label.

[0064] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed by the multimedia resource recommendation method as described in the above aspects.

[0065] On the other hand, a computer-readable storage medium is provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the operations performed by the multimedia resource recommendation method as described in the above aspects.

[0066] On the other hand, a computer program product is provided, including a computer program, wherein the computer program is loaded and executed by a processor to implement the operations performed by the multimedia resource recommendation method as described in the above aspects.

[0067] The solution provided by the embodiment of the present application is to first obtain the task features of the recommendation task for any recommendation task in any recommendation scenario, and then use the recommendation model to make predictions based on the task features. In the recommendation model, three expert networks of different granularities are set for the three levels of scene granularity, task granularity and global granularity respectively. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities. Then, based on the features of the three different granularities, the interaction between the object and the multimedia resource is predicted when the multimedia resource is recommended to the object, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0069] Figure 1 This is a schematic diagram of an implementation environment provided by an embodiment of the present application;

[0070] Figure 2 This is a flowchart of a multimedia resource recommendation method provided by an embodiment of the present application;

[0071] Figure 3 This is a schematic diagram of the structure of a recommendation model provided in an embodiment of the present application;

[0072] Figure 4 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application;

[0073] Figure 5 This is a schematic diagram of the structure of a memory expert network and a generalized expert network provided in an embodiment of the present application;

[0074] Figure 6 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application;

[0075] Figure 7 This is a flow chart of a fusion method provided in an embodiment of the present application;

[0076] Figure 8 This is a schematic diagram of the structure of a memory fusion network provided in an embodiment of the present application;

[0077] Figure 9 This is a schematic diagram of the structure of a target fusion network provided in an embodiment of the present application;

[0078] Figure 10 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application;

[0079] Figure 11 This is a schematic diagram of the structure of a first scenario granularity fusion network provided in an embodiment of the present application;

[0080] Figure 12 This is a flowchart of a training method for a recommendation model provided in an embodiment of the present application;

[0081] Figure 13 This is a structural diagram of a multimedia resource recommendation device provided in an embodiment of the present application;

[0082] Figure 14 This is a structural diagram of another multimedia resource recommendation device provided in an embodiment of the present application;

[0083] Figure 15 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present application;

[0084] Figure 16 This is a structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0085] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0086] It is understood that the terms "first," "second," and the like used herein may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are merely used to distinguish one concept from another. For example, a first target weight may be referred to as a second target weight, and similarly, a second target weight may be referred to as a first target weight, without departing from the scope of this application.

[0087] Here, "at least one" refers to one or more than one. For example, the at least one expert network can be one expert network, two expert networks, three expert networks, or any other integer greater than or equal to one. "Multiple" refers to two or more than two. For example, the multiple expert networks can be two expert networks, three expert networks, or any other integer greater than or equal to two. "Each" refers to each of the at least one. For example, "each expert network" refers to each expert network in the multiple expert networks. If the multiple expert networks are three expert networks, "each expert network" refers to each expert network in the three expert networks.

[0088] It should be noted that the information involved in this application (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals (including but not limited to signals transmitted between user terminals and other devices, etc.) are all fully authorized by the user or relevant parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the object features and resource features involved in this application are obtained with the full knowledge and authorization of the user.

[0089] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0090] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, pre-trained models, operating / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be fine-tuned and widely applied to downstream tasks across various AI disciplines. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0091] Natural language processing (NLP) is a key area in computer science and artificial intelligence. It studies various theories and methods that enable effective communication between humans and computers using natural language. Natural language processing involves natural language, the language we use everyday, and is closely related to linguistics. It also involves key model training techniques in computer science, mathematics, and other artificial intelligence fields. Pre-trained models (PTMs) are derived from large language models in the NLP field. Pre-trained models, also known as cornerstone models or large models, refer to deep neural networks (DNNs) with large parameters. These models are trained on massive amounts of unlabeled data. Leveraging the function approximation capabilities of large-parameter DNNs, PTMs extract common features from the data. Through fine-tuning, efficient parameter fine-tuning (PEFT), and prompt-tuning techniques, they are then adapted for downstream tasks. Therefore, pre-trained models can achieve ideal results in few-shot or zero-shot scenarios. PTM is divided into language models, visual models, speech models, multimodal models, etc. according to the data modality processed. Examples of language models include ELMO (Embeddings from Language Model), BERT (Bidirectional Encoder Representations from Transformers), and GPT (Generative Pre-trained Transformer). A large language model is a large-scale deep learning model that typically uses an autoregressive loss as a training objective. This model predicts the next word in a given context, thereby learning to generate grammatically correct and semantically coherent text to understand and generate human language. By learning from large amounts of text data, large language models can understand the complex patterns and contextual relationships of language, thereby generating coherent and relevant text. A key feature of large language models is their scale, typically containing billions or even trillions of parameters. This enables them to capture and simulate the rich diversity and complexity of human language. Large language models have demonstrated excellent performance in various language tasks, including but not limited to text generation, text understanding, machine translation, and sentiment analysis. Multimodal models are models that represent features from two or more data modalities.Pre-trained models are important tools for producing artificial intelligence generated content (AIGC) and can also serve as a universal interface for connecting multiple task-specific models. After fine-tuning, large language models can be widely applied to downstream tasks. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0092] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning by demonstration. Pretrained models are the latest development in deep learning, integrating these techniques.

[0093] The solution provided in the embodiment of the present application implements a multimedia resource recommendation method based on artificial intelligence machine learning technology.

[0094] Multimedia resource recommendation algorithms are widely used in daily life, such as short video recommendations, product recommendations, and song recommendations. Accurately recommending these multimedia resources to appropriate users is a key challenge facing multimedia resource recommendation algorithms. In these recommendation scenarios, the scale of recommended multimedia resources is enormous, often numbering in the millions or even tens of millions. Furthermore, recommendation requirements are diverse, such as those involving different delivery targets and media placements. Therefore, the recommendation tasks modeled by recommendation models must also be diverse. Therefore, multi-task joint modeling is currently the mainstream approach for recommendation algorithms. This involves training a recommendation model that shares multiple tasks to improve modeling efficiency. However, due to the differences between recommendation tasks in different recommendation scenarios, multiple recommendation tasks may interfere with each other, resulting in the recommendation model being unable to effectively perceive the deviations between different recommendation tasks. Multi-task joint modeling is prone to negative transfer and seesaw phenomena. Negative transfer reduces the overall recommendation effectiveness of the recommendation model, while seesaw phenomena occur when the recommendation model performs well on some tasks but poorly on others.

[0095] Furthermore, the sparsity of training samples is a significant challenge in multimedia resource recommendation algorithms. For example, in product recommendation scenarios, after a product is recommended to a user, the percentage of users who provide feedback on the product is relatively small. This lack of user feedback makes it difficult for the multimedia resource recommendation algorithm to fully learn relevant features. Consequently, recommendation results are poor for cold-start multimedia resources or long-tail multimedia resources. Cold-start multimedia resources refer to newly created multimedia resources, while long-tail multimedia resources refer to multimedia resources with low conversion rates.

[0096] In summary, multimedia resource recommendation algorithms face two challenges: one is the difficulty in modeling task deviations in multi-task joint modeling, and the other is the poor recommendation effect of cold-start or long-tail multimedia resources.

[0097] In response to the first challenge mentioned above, the multimedia resource recommendation method proposed in the embodiment of the present application sets expert networks of different granularities for the three levels of scene granularity, task granularity and global granularity, and further integrates the features extracted by expert networks of different granularities to effectively understand the commonalities and differences between different scenes and different tasks, thereby modeling a recommendation model with higher accuracy. In response to the second challenge mentioned above, the multimedia resource recommendation proposed in the embodiment of the present application sets up a memory-biased expert network and a generalization-biased expert network on the basis of the main structure of multiple expert networks of different granularities. The memory-biased expert network learns features related to feedback behavior, and the generalization-biased expert network learns features unrelated to feedback behavior, thereby decoupling the learning process of memory-biased features and generalization-biased features. Since the generalization-biased expert network is conducive to modeling cold-start or long-tail multimedia resources, it can solve or alleviate the problem of poor recommendation effect of cold-start or long-tail multimedia resources to a certain extent.

[0098] Therefore, the embodiments of the present application propose a method for recommending multimedia resources based on a multi-scenario multi-task hybrid expert network (MGA-MMEN, Memorization and Generalization Adaptive Perception) . This method is universal and can be integrated into recommendation scenarios of multi-scenario multi-task joint modeling, such as multi-task conversion prediction scenarios in advertising recommendation systems, content recommendation scenarios in video recommendation scenarios, and playback duration estimation scenarios. The detailed implementation process of this method can be found in the description of each embodiment below.

[0099] The multimedia resource recommendation method provided in the embodiment of the present application can be executed by a computer device. Optionally, the computer device is a terminal or a server. Optionally, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal is a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, smart voice interaction device, smart home appliance and car terminal, etc., but is not limited to this.

[0100] In some embodiments, the computer program involved in the embodiments of the present application can be deployed and executed on a computer device, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected through a communication network. Multiple computer devices distributed at multiple locations and interconnected through a communication network can constitute a blockchain system.

[0101] In some embodiments, Figure 1 This is a schematic diagram of an implementation environment provided by an embodiment of the present application. Figure 1 , the implementation environment includes a terminal 101 and a server 102. The terminal 101 and the server 102 are connected via a wireless or wired network.

[0102] Among them, the server 102 stores the object characteristics of any object and the resource characteristics of any multimedia resource. The terminal 101 sends a recommendation request for a certain object to the server 102. After receiving the recommendation request, the server 102 generates a target recommendation task for any multimedia resource to be recommended. The target recommendation task is to predict whether to recommend the multimedia resource to the object. The server 102 uses the object characteristics of the object and the resource characteristics of the multimedia resource as the task characteristics of the target recommendation task, and determines the prediction result based on the task characteristics. The prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object. Based on the prediction result, it can be determined whether to recommend the multimedia resource to the object. When the prediction result indicates that the multimedia resource is recommended to the object, the server 102 sends the multimedia resource to the terminal 101, and the terminal 101 displays the multimedia resource based on the object.

[0103] Figure 2 This is a flowchart of a multimedia resource recommendation method provided by an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 2, the method comprising:

[0104] 201. The computer device obtains task features of a target recommendation task, where the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any recommendation scenario among multiple recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0105] The multimedia resource recommendation method in the embodiment of the present application can be applied to any recommendation task in any recommendation scenario among multiple recommendation scenarios. Each recommendation scenario corresponds to at least one recommendation task. The multiple recommendation scenarios and the at least one recommendation task in each recommendation scenario can be predefined recommendation scenarios or recommendation tasks.

[0106] For example, various recommendation scenarios include video recommendation, product recommendation, song recommendation, image and text recommendation, or advertisement recommendation. Taking the video recommendation scenario as an example, the various recommendation tasks in this video recommendation scenario include the same-city video recommendation task, the homepage video recommendation task, and the search-term-based video recommendation task.

[0107] The target recommendation task is any recommendation task in any recommendation scenario, and the target recommendation task is to predict whether to recommend a multimedia resource to an object. The object can be an account, etc. The multimedia resource can also vary in different recommendation scenarios, for example, the multimedia resource can be a video resource, a product resource, a song resource, a graphic resource, or an advertisement resource, etc., which is not limited in the present embodiment.

[0108] For the target recommendation task, the computer device obtains the object features of the object in the target recommendation task and the resource features of the multimedia resources. The object features and the resource features constitute the task features of the target recommendation task, and subsequently predicts whether to recommend the multimedia resources to the object based on the task features.

[0109] 202. The computer device extracts the task features through the scene expert network in the recommendation model to obtain scene granularity features.

[0110] In an embodiment of the present application, the recommendation model is a multi-task joint modeling recommendation model, that is, the recommendation task can be used to predict any recommendation task in any recommendation scenario, and is not limited to processing a single recommendation task in a single scenario.

[0111] The recommendation model includes a scenario expert network, which is used to extract scene-level features. A computer device inputs the task features of the target recommendation task into the scenario expert network, and outputs scene-level features. These features reflect the characteristics of the target recommendation task at the scene level and can distinguish between recommendation tasks in different recommendation scenarios.

[0112] 203. The computer device extracts the task features through the task expert network in the recommendation model to obtain the task granularity features.

[0113] The recommendation model also includes a task expert network, which is used to extract task-granularity features. The computer device inputs the task features of the target recommendation task into the task expert network, and outputs task-granularity features from the task expert network. These task-granularity features reflect the characteristics of the target recommendation task at the task granularity, enabling differentiation between different recommendation tasks within the same recommendation scenario.

[0114] 204. The computer device extracts the task features through the global expert network in the recommendation model to obtain global granular features.

[0115] The recommendation model also includes a global expert network, which is used to extract features at a global granularity. The computer device inputs the task features of the target recommendation task into the global expert network to obtain the global granularity features output by the global expert network. The global granularity features are used to reflect the characteristics of the target recommendation task at a global granularity. The global granularity features can be understood as being able to control the overall characteristics of the recommendation task. When two different recommendation tasks belong to the same recommendation task in the same recommendation scenario, the global granularity features can distinguish between the two different recommendation tasks. For example, if both recommendation tasks belong to the same-city recommendation task in the video recommendation scenario, but the objects and multimedia resources involved in the two recommendation tasks are different, then the two recommendation tasks are also different.

[0116] 205. The computer device determines a prediction result based on the scene granularity feature, the task granularity feature, and the global granularity feature. The prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0117] After obtaining the scene-level, task-level, and global-level features of the target recommendation task through the recommendation model, the computer device combines these three levels of features to determine a prediction result. Subsequently, based on the prediction result, the computer device may determine whether to recommend the multimedia resource to the subject. For example, if the interaction situation represented by the prediction result meets the expected interaction situation, the multimedia resource is recommended to the subject. If the interaction situation represented by the prediction result does not meet the expected interaction situation, the multimedia resource is not recommended to the subject.

[0118] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0119] Figure 3 This is a schematic diagram of the structure of a recommendation model provided in an embodiment of the present application. Figure 3 As shown, the recommendation model includes a scene expert network, a task expert network, and a global expert network. The scene expert network is used to extract scene-level features, the task expert network is used to extract task-level features, and the global expert network is used to extract global-level features. In one possible implementation, the task features of the target recommendation task include memory-biased features and generalization-biased features. The scene expert network, task expert network, and global expert network are further divided based on memory bias and generalization bias, as described below.

[0120] In one possible implementation, Figure 3As shown, the scene expert network includes a scene generalization expert network and a scene memory expert network, the task expert network includes a task memory expert network, and the global expert network includes a global generalization expert network. The input of the scene generalization expert network is a generalization bias feature, which is used to extract scene granularity generalization features. The input of the scene memory expert network is a memory bias feature, which is used to extract scene granularity memory features. The input of the task memory expert network is a memory bias feature, which is used to extract task granularity memory features. The input of the global generalization expert network is a generalization bias feature, which is used to extract global granularity generalization features.

[0121] In one possible implementation, Figure 3 As shown, the number of scene memory expert networks is x, the number of scene generalization expert networks is y, the number of task memory expert networks is n, and the number of global generalization expert networks is k. These expert networks can have the same network structure but different network parameters.

[0122] In one possible implementation, Figure 3 As shown, the recommendation model also includes a global granularity fusion network, a first scene granularity fusion network, a second scene granularity fusion network and a task granularity fusion network. The global granularity fusion network is connected to k global generalization expert networks to fuse the global granularity generalization features output by the k global generalization expert networks. The first scene granularity fusion network is connected to x scene memory expert networks to fuse the scene granularity memory features output by x scene memory expert networks. The second scene granularity fusion network is connected to y scene generalization expert networks to fuse the scene granularity generalization features output by y scene generalization expert networks. The task granularity fusion network is connected to n task memory expert networks to fuse the task granularity memory features output by n task memory expert networks. Among them, the global granularity fusion network, the first scene granularity fusion network, the second scene granularity fusion network and the task granularity fusion network belong to the first-level fusion network in the recommendation model.

[0123] In one possible implementation, Figure 3 As shown, the recommendation model also includes a generalization fusion network and a memory fusion network. The generalization fusion network is connected to the global granularity fusion network and the second scene granularity fusion network to fuse the global granularity generalization features output by the global granularity fusion network and the scene granularity generalization features output by the second scene granularity fusion network. The memory fusion network is connected to the first scene granularity fusion network and the task granularity fusion network to fuse the scene granularity memory features output by the first scene granularity fusion network and the task granularity memory features output by the task granularity fusion network. The generalization fusion network and the memory fusion network belong to the second-level fusion network in the recommendation model.

[0124] In one possible implementation, Figure 3 As shown, the recommendation model also includes a target fusion network, which is connected to the generalization fusion network and the memory fusion network to fuse the generalization fusion features output by the generalization fusion network with the memory fusion features output by the memory fusion network. The target fusion network is the third-level fusion network in the recommendation model.

[0125] In one possible implementation, Figure 3 As shown, the recommendation model also includes an output network and a feature mapping network. The output network is connected to the target fusion network, and the feature mapping network is connected to the output network. The output network is used to process the target fusion features output by the target fusion network to obtain target output features. The feature mapping network is used to process the target output features output by the output network to obtain a prediction result. The prediction result output by the feature mapping network is the prediction result obtained by the recommendation model based on the task features of the target recommendation task.

[0126] The following embodiments are combined Figure 3 The recommendation model shown in the figure provides a detailed introduction to the implementation process of the multimedia resource recommendation method.

[0127] In some embodiments, as Figure 3 As shown in Figure 2, the scene expert network includes the scene memory expert network and the scene generalization expert network, the task expert network is the task memory expert network, and the global expert network is the global generalization expert network. The implementation process of the multimedia resource recommendation method is as follows: Figure 4 An embodiment of Figure 4 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 4 , the method comprising:

[0128] 401. The computer device obtains task features of a target recommendation task, where the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any recommendation scenario among multiple recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0129] In one possible implementation, a computer device obtains object data of an object in a target recommendation task and resource data of a multimedia resource; performs embedded feature extraction on the object data to obtain object features, and performs embedded feature extraction on the resource data to obtain resource features.

[0130] In one possible implementation, the object data includes the object's interactions with multimedia resources within a preset time period before the current time point, including clicks, playbacks, and the like. The multimedia resources may be recommended to the object. The object data also includes the object's attribute information, which refers to information provided when registering the object, such as the object's location or age.

[0131] In one possible implementation, the resource data includes a multimedia resource identifier, a material identifier within the multimedia resource, and the type of target audience to which the multimedia is delivered. The resource data also includes information about the target audience's interaction with the multimedia resource within a preset time period prior to the current time point, including clicks and playbacks. The target audience may be the target audience to whom the multimedia resource was recommended. The resource data also includes a description of the multimedia resource's attributes, such as the type or industry to which the multimedia resource belongs.

[0132] It should be noted that the object data of the above objects and the resource data of the multimedia resources are all obtained with the full knowledge and authorization of the user.

[0133] 402. The computer device obtains memory bias features and generalization bias features from task features. The memory bias features refer to features that are affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features that are not affected by the interaction between the object and the multimedia resource.

[0134] Among them, the interaction between the object and the multimedia resource can be called the object's feedback behavior to the multimedia resource. The memory bias feature can be understood as a feature related to the object's feedback behavior, that is, it will be affected by the interaction between the object and the multimedia resource. The generalization bias feature can be understood as a feature unrelated to the object's feedback behavior, that is, it will not be affected by the interaction between the object and the multimedia resource.

[0135] After obtaining the task features of the target recommendation task, the computer device divides the task features into memory bias features and generalization bias features. It should be noted that task features are composed of object features and resource features. Object features include memory bias features and generalization bias features, while resource features also include memory bias features and generalization bias features.

[0136] In one possible implementation, for the object features in the task features, the computer device determines the object features related to the feedback behavior of the object as memory bias features. For example, the memory bias features in the object features include features extracted based on the object's interaction with multimedia resources within a preset time period before the current time point. The computer device determines the object features unrelated to the object's feedback behavior as generalization bias features. For example, the generalization bias features in the object features include features extracted based on the object's attribute information, as well as keyword features or type features of the multimedia resources searched by the object.

[0137] In one possible implementation, for the resource features in the task features, the computer device determines the resource features related to the feedback behavior of the object as memory bias features. For example, the memory bias features in the resource features include features extracted based on the interaction between the object and the multimedia resource within a preset time period before the current time point, as well as features extracted based on the multimedia resource identifier, the material identifier in the multimedia resource, and the type of object to which the multimedia is delivered. The computer device determines the resource features that are not related to the feedback behavior of the object as generalization bias features. For example, the generalization bias features in the resource features include features extracted based on the attribute description of the multimedia resource (such as the type or industry to which the multimedia resource belongs, etc.), etc.

[0138] 403. The computer device extracts the memory bias feature through the scene memory expert network in the recommendation model to obtain the scene granularity memory feature.

[0139] like Figure 3 As shown, the recommendation model includes a scene memory expert network, which is used to extract memory bias features. The computer device inputs the memory bias features into the scene memory expert network to obtain the scene granularity memory features output by the scene memory expert network. For example, the memory bias features input by the scene memory expert network are denoted as E mem , the scene granularity memory feature output by the scene memory expert network is recorded as

[0140] Among them, the scene memory expert network belongs to the memory-biased expert network. The memory-biased expert network utilizes the memory-biased features related to the feedback behavior in the recommendation scenario, and focuses on learning the potential association semantics between different objects, objects and multimedia resources, and multimedia resources and multimedia resources.

[0141] In one possible implementation, Figure 5 This is a schematic diagram of the structure of a memory expert network and a generalized expert network provided by an embodiment of the present application. The structure of the scene memory expert network is as follows: Figure 5As shown in (1), the scene memory expert network includes a feedforward neural network layer and a layer normalization network layer. The feedforward neural network layer includes a Dense layer (dense layer) and a Relu layer (activation layer). The processing process of the scene memory expert network can be expressed by the following formula:

[0142]

[0143] H m =LN(F m )

[0144] Among them, E mem Indicates memory bias characteristics, and are the network parameters of the feedforward neural network layer in the scene memory expert network, is the output of the Dense layer, Relu(·) represents the activation layer, and F m represents the output of the activation layer, LN(·) represents the layer normalization network layer, H m Represents scene-granular memory features.

[0145] 404. The computer device extracts the generalization bias feature through the scene generalization expert network in the recommendation model to obtain the scene granularity generalization feature, which includes the scene granularity memory feature and the scene granularity generalization feature.

[0146] like Figure 3 As shown, the recommendation model includes a scene generalization expert network, which is used to extract generalization bias features. The computer device inputs the generalization bias features into the scene generalization expert network to obtain the scene granularity generalization features output by the scene generalization expert network. For example, the generalization bias features input by the scene generalization expert network are denoted as E gen , the scene granularity generalization feature output by the scene generalization expert network is recorded as

[0147] Among them, the scenario-generalizing expert network is a generalization-biased expert network. This generalization-biased expert network utilizes generalization-biased features that are unrelated to feedback behavior in recommendation scenarios, focusing on learning semantic similarities between different objects, objects and multimedia resources, and multimedia resources. For cold-start or long-tail multimedia resources, due to the very low conversion volume, the memory-biased expert network cannot learn sufficient effective features. In this case, the generalization-biased expert network can transfer features from other multimedia resources with the same attributes that are not cold-start or long-tail, effectively assisting in learning the features of these cold-start or long-tail multimedia resources.

[0148] In one possible implementation, Figure 5This is a schematic diagram of the structure of a memory expert network and a generalized expert network provided by an embodiment of the present application. The structure of the generalized expert network in this scenario is as follows Figure 5 As shown in (2), the scene generalization expert network includes a feedforward neural network layer and a layer normalization network layer. The feedforward neural network layer includes a Dense layer (dense layer) and a Relu layer (activation layer). The processing process of the scene generalization expert network can be expressed by the following formula:

[0149]

[0150] H g =LN(N g )

[0151] Among them, E gen represents the generalization bias feature, and are the network parameters of the feedforward neural network layer in the scene generalization expert network, is the output of the Dense layer, Relu(·) represents the activation layer, N g represents the output of the activation layer, LN(·) represents the layer normalization network layer, H g Represents scene-level generalization features.

[0152] It should be noted that if Figure 5 As shown in FIG, the network structures of the scene memory expert network and the scene generalization expert network can be the same, but the network parameters of the scene memory expert network and the scene generalization expert network are different, and the input and output of the network are also different.

[0153] 405. The computer device extracts the memory bias feature through the task memory expert network in the recommendation model to obtain the task granularity memory feature, and the task granularity feature includes the task granularity memory feature.

[0154] like Figure 3 As shown, the recommendation model includes a task memory expert network, which is used to extract memory bias features. The computer device inputs the memory bias features into the task memory expert network to obtain the task granularity memory features output by the task memory expert network. For example, the memory bias features input by the task memory expert network are denoted as E mem , the task granularity memory feature output by the task memory expert network is recorded as

[0155] The task memory expert network is a memory-biased expert network. Its network structure and processing are similar to those of the scene memory expert network and will not be further elaborated here. It should be noted that the network structures of the task memory expert network and the scene memory expert network can be the same, but their network parameters are different, and the network inputs and outputs are also different.

[0156] 406. The computer device extracts the generalization bias feature through the global generalization expert network in the recommendation model to obtain the global granularity generalization feature, and the global granularity feature includes the global granularity generalization feature.

[0157] like Figure 3 As shown, the recommendation model includes a global generalization expert network, which is used to extract generalization bias features. The computer device inputs the generalization bias features into the global generalization expert network to obtain the global granularity generalization features output by the global generalization expert network. For example, the generalization bias features input by the global generalization expert network are denoted as E gen , the global granularity generalization feature output by the global generalization expert network is recorded as

[0158] The global generalized expert network is a generalized biased expert network. Its network structure and processing are similar to those of the scene generalized expert network and will not be further elaborated here. It should be noted that the network structures of the global generalized expert network and the scene generalized expert network can be the same, but their network parameters are different, and the network inputs and outputs are also different.

[0159] 407. The computer device determines a prediction result based on the scene granularity feature, the task granularity feature, and the global granularity feature. The prediction result is used to represent the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0160] After the computer device obtains the scene granularity features, task granularity features and global granularity features of the target recommendation task through the recommendation model, it combines the features of these three granularities to determine the prediction result, and then determines whether to recommend the multimedia resource to the object based on the prediction result. The detailed implementation process of step 407 can be found in the following Figure 7 Steps 701-704 of the embodiment, or the following Figure 10 Steps 1007-1015 in the embodiment are not described here for the time being.

[0161] In one possible implementation, the prediction result is a predicted probability, which represents the probability that the subject will interact with the multimedia resource when the multimedia resource is recommended to the subject. After the computer device determines the prediction result based on the scene-level granularity features, the task-level granularity features, and the global granularity features, the method further includes: recommending the multimedia resource to the subject if the predicted probability is greater than a preset probability threshold.

[0162] In multimedia resource recommendation using multi-scenario, multi-task joint modeling, the recommendation model needs to simultaneously account for multiple recommendation scenarios and multiple recommendation tasks within each recommendation scenario. For example, there will be varying degrees of commonality and diversity between different recommendation scenarios, between different recommendation tasks within the same recommendation scenario, and between different recommendation tasks within different recommendation scenarios. This makes task deviation modeling in multi-task joint modeling difficult. The recommendation model in the embodiment of the present application utilizes a multi-expert network, setting up expert networks at three levels: scenario granularity, task granularity, and global granularity, which facilitates learning the commonalities and differences between various recommendation tasks.

[0163] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0164] In some embodiments, as Figure 3 As shown in the figure, the number of scene expert networks in the recommendation model is m, the number of scene granularity features is m, and the network parameters of each scene expert network are different; the number of task expert networks is n, the number of task granularity features is n, and the network parameters of each task expert network are different; the number of global expert networks is k, the number of global granularity features is k, and the network parameters of each global expert network are different; where m, n, and k are all positive integers. The implementation process of the multimedia resource recommendation method can be seen in the following Figure 6 Example of . Figure 6 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 6 , the method comprising:

[0165] 601. The computer device obtains task features of a target recommendation task, where the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any recommendation scenario among multiple recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0166] The process of step 601 is the same as that of step 401 above, and will not be repeated here.

[0167] 602. The computer device extracts the task features through the m scene expert networks in the recommendation model to obtain m scene granular features.

[0168] In an embodiment of the present application, the recommendation model includes m scenario expert networks, each with different network parameters. Optionally, each scenario expert network has the same network structure. The scenario expert network is a scenario-granular expert network, which is used to learn the commonalities between different recommendation tasks in the same recommendation scenario. Optionally, the number of scenario expert networks is equal to the number of recommendation scenarios, with one scenario expert network corresponding to one recommendation scenario. The scenario expert network corresponding to any recommendation scenario is used to learn the commonalities between different recommendation tasks in that recommendation scenario.

[0169] The computer device inputs the task features into m scene expert networks respectively, obtains the scene granularity features output by each scene expert network, and thus obtains m scene granularity features.

[0170] In one possible implementation, Figure 3 As shown, the m scene expert networks include x scene memory expert networks and y scene generalization expert networks. Task features include memory bias features and generalization bias features. Memory bias features refer to features affected by the interaction between the object and the multimedia resources, and generalization bias features refer to features not affected by the interaction between the object and the multimedia resources. Then, step 602 includes: the computer device extracts features of the memory bias features through the x scene memory expert networks to obtain x scene granularity memory features respectively; and extracts features of the generalization bias features through the y scene generalization expert networks to obtain y scene granularity generalization features respectively. The scene granularity features include x scene granularity memory features and y scene granularity generalization features.

[0171] Alternatively, the processing of x scene memory expert networks can be expressed by the following formula:

[0172]

[0173]

[0174] in, and is the network parameter of the x-th scene memory expert network, E mem Indicates memory bias characteristics, represents the x-th scene granularity memory feature, Represents a set of x scene-granularity memory features.

[0175] Alternatively, the processing of y scene bias expert networks can be expressed by the following formula:

[0176]

[0177]

[0178] in, and is the network parameter of the generalized expert network for the yth scenario, E gen represents the generalization bias feature, represents the y-th scene granularity generalization feature, Represents a set of y scene-granularity generalized features.

[0179] 603. The computer device extracts the task features through the n task expert networks in the recommendation model to obtain n task granularity features.

[0180] In this embodiment of the present application, the recommendation model includes n task expert networks, each with different network parameters. Optionally, each task expert network has the same network structure. The task expert network is a task-granular expert network, which is used to learn the differences between different recommendation tasks in the same recommendation scenario.

[0181] The computer device inputs the task features into n task expert networks respectively, obtains the task granularity features output by each task expert network, and thus obtains n task granularity features.

[0182] In one possible implementation, Figure 3 As shown, the n task expert networks are n task memory expert networks, and the task features include memory bias features, which are features affected by the interaction between the object and the multimedia resource. Step 603 then includes: extracting the memory bias features using the n task memory expert networks to obtain n task-granularity memory features, each of which includes n task-granularity memory features.

[0183] Alternatively, the processing of n task memory expert networks can be expressed by the following formula:

[0184]

[0185]

[0186] in, and is the network parameter of the nth task memory expert network, E mem Indicates memory bias characteristics, represents the nth task granularity memory feature, Represents a set of n task-granularity memory features.

[0187] 604. The computer device extracts the task features through the k global expert networks in the recommendation model to obtain k global granular features.

[0188] In this embodiment of the present application, the recommendation model includes k global expert networks, each with different network parameters. Optionally, each global expert network has the same network structure. The global expert network is a global-granularity expert network that does not distinguish between recommendation scenarios and recommendation tasks and is used to learn commonalities across recommendation scenarios and tasks.

[0189] The computer device inputs the task features into k global expert networks respectively, obtains the global granular features output by each global expert network, and thus obtains k global granular features.

[0190] In one possible implementation, Figure 3 As shown, the k global expert networks are k global generalization expert networks. Task features include generalization bias features, which are features that are not affected by the interaction between the object and the multimedia resource. Step 604 then includes: extracting the generalization bias features using the k global generalization expert networks to obtain k global granularity generalization features, which include k global granularity generalization features.

[0191] Alternatively, the processing of k global generalization expert networks can be expressed by the following formula:

[0192]

[0193]

[0194] in, and is the network parameter of the kth global generalization expert network, E gen represents the generalization bias feature, represents the kth global granularity generalization feature, Represents a set of k global granularity generalization features.

[0195] 605. The computer device determines a prediction result based on the m scene granularity features, the n task granularity features, and the k global granularity features. The prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0196] After the computer device obtains the scene granularity features, task granularity features and global granularity features of the target recommendation task through the recommendation model, it combines the features of these three granularities to determine the prediction result, and then determines whether to recommend the multimedia resource to the object based on the prediction result. The detailed implementation process of step 407 can be found in the following Figure 7 Steps 701-704 of the embodiment, or the following Figure 10 Steps 1007-1015 in the embodiment are not described here for the time being.

[0197] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0198] In some embodiments, the recommendation model further includes a fusion network. The computer device determines a prediction result based on the scene granularity features, the task granularity features, and the global granularity features, including: fusing the scene granularity features, the task granularity features, and the global granularity features through the fusion network in the recommendation model to obtain a target fusion feature; and generating a prediction result based on the target fusion feature.

[0199] In Figure 4 Based on the embodiment of the example, the scene granularity features include scene granularity memory features and scene granularity generalization features, the task granularity features include task granularity memory features, and the global granularity features include global granularity generalization features. The fusion network includes a memory fusion network, a generalization fusion network, and a target fusion network. The fusion process is described below. Figure 7 Example of . Figure 7 This is a flow chart of a fusion method provided in an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 7 , the method comprising:

[0200] 701. The computer device fuses the scene-granularity memory features and the task-granularity memory features through the memory fusion network in the recommendation model to obtain the memory fusion features.

[0201] like Figure 3 As shown, the recommendation model includes a memory fusion network, which is used to fuse memory-granular features. After obtaining scene-granular memory features and task-granular memory features, the computer device inputs the scene-granular memory features and task-granular memory features into the memory fusion network to obtain the memory fusion features output by the memory fusion network.

[0202] In one possible implementation, task features include memory-biased features and generalization-biased features. Memory-biased features are features affected by the interaction between an object and a multimedia resource, while generalization-biased features are features not affected by the interaction between an object and a multimedia resource. Step 701 includes: a computer device, using a memory fusion network, normalizes the concatenation of the memory-biased features and the generalization-biased features to obtain a first memory weight and a second memory weight; and, using the memory fusion network, weightedly fuses the scene-granularity memory features and the task-granularity memory features based on the first memory weight and the second memory weight to obtain a memory fusion feature.

[0203] Optionally, the memory fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate weights, and the gating part is implemented based on an adaptive perception gating mechanism, and the weighted fusion part is used to perform weighted summation. Figure 8 This is a schematic diagram of the structure of a memory fusion network provided in an embodiment of the present application. Figure 8 As shown, the gating part of the memory fusion network includes a Dense layer and a normalization layer, the Dense layer and the normalization layer are connected, and the weighted fusion part of the memory fusion network includes a multiplication operator and an addition operator. Optionally, the processing process of the memory fusion network can be expressed by the following formula:

[0204]

[0205]

[0206] in, and is the network parameter of the gate control part in the memory fusion network, E mem Indicates the memory bias characteristic, E gen Represents the generalization bias feature, E mem ||Egen represents the concatenation of memory bias features and generalization bias features, softmax(·) represents the normalization layer, Represents the first memory weight, 1- represents the second memory weight, represents the scene granularity memory feature, represents the task-granularity memory feature, Represents memory fusion characteristics.

[0207] 702. The computer device fuses the scene granularity generalized features and the global granularity generalized features through the generalized fusion network in the recommendation model to obtain a generalized fusion feature.

[0208] like Figure 3 As shown, the recommendation model includes a generalization fusion network, which is used to fuse features of generalized granularity. After obtaining scene-granularity generalization features and global-granularity generalization features, the computer device inputs the scene-granularity generalization features and global-granularity generalization features into the generalization fusion network to obtain the generalization fusion features output by the generalization fusion network.

[0209] In one possible implementation, the task features include memory-biased features and generalization-biased features. Memory-biased features are features that are affected by the interaction between the object and the multimedia resource, while generalization-biased features are features that are not affected by the interaction between the object and the multimedia resource. Step 702 includes: the computer device, using a generalization fusion network, normalizes the concatenation of the memory-biased features and the generalization-biased features to obtain a first generalization weight and a second generalization weight; and, using the generalization fusion network, weightedly fuses the scene-granularity generalization features and the global-granularity generalization features based on the first generalization weight and the second generalization weight to obtain a generalization-fused feature.

[0210] Optionally, the generalized fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate the weight, and the weighted fusion part is used to perform weighted summation. The network structure of the generalized fusion network is similar to that of the memory fusion network and will not be described in detail here. Optionally, the processing process of the generalized fusion network can be expressed by the following formula:

[0211]

[0212]

[0213] in, and is the network parameter of the gating part in the generalized fusion network, E mem Indicates the memory bias characteristic, E gen Represents the generalization bias feature, E mem ||E genrepresents the concatenation of memory bias features and generalization bias features, softmax(·) represents the normalization layer, represents the first generalization weight, represents the second generalization weight, represents the scene granularity generalization feature, represents the global granularity generalization feature, Represents the generalized fusion feature.

[0214] 703. The computer device fuses the memory fusion feature and the generalization fusion feature through the target fusion network in the recommendation model to obtain the target fusion feature.

[0215] like Figure 3 As shown, the recommendation model includes a target fusion network, which is used to fuse features at the memory granularity and features at the generalization granularity. After obtaining the memory fusion features and the generalization fusion features, the computer device inputs the memory fusion features and the generalization fusion features into the target fusion network, and obtains the target fusion features output by the target fusion network.

[0216] In one possible implementation, step 703 includes: the computer device obtains memory biased resource features from multimedia resource features, where the memory biased resource features refer to multimedia resource features affected by the interaction between the object and the multimedia resource; normalizes the memory biased resource features through a target fusion network to obtain a first target weight and a second target weight; and weightedly fuses the memory fusion features and the generalization fusion features based on the first target weight and the second target weight through the target fusion network to obtain a target fusion feature.

[0217] Among them, the memory-biased resource feature refers to the feature of the multimedia resource feature that is related to the feedback behavior in the recommendation scenario, that is, the multimedia resource feature that is affected by the interaction between the object and the multimedia resource. Optionally, the memory-biased resource feature includes features related to the conversion status of the multimedia resource, such as the conversion rate of the multimedia resource, the exposure rate and click volume, whether it has passed the cold start period, etc. The first target weight and the second target weight are determined based on the memory-biased resource feature, which can control the proportion of memory fusion features and generalization fusion features in the weighted fusion process. The expected effect is that for cold-start or long-tail multimedia resources, the focus is on retaining generalization-biased features, and conversely, for popular or non-long-tail multimedia resources, the focus is on retaining memory-biased features.

[0218] Optionally, the target fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate weights, and the weighted fusion part is used to perform weighted summation. Figure 9 This is a schematic diagram of the structure of a target fusion network provided in an embodiment of the present application. Figure 9As shown, the gating part of the target fusion network includes a Dense layer and a normalization layer, the Dense layer and the normalization layer are connected, and the weighted fusion part of the target fusion network includes a multiplication operator and an addition operator. Optionally, the processing process of the target fusion network can be expressed by the following formula:

[0219]

[0220]

[0221] in, and is the network parameter of the gated part in the target fusion network, E stat represents the memory bias resource feature, softmax(·) represents the normalization layer, θ s_t represents the first target weight, 1-θ s_t represents the second target weight, Represents the memory fusion feature, Represents the generalized fusion feature.

[0222] 704. The computer device generates a prediction result based on the target fusion features.

[0223] In one possible implementation, the recommendation model includes an output network and a feature mapping network. The computer device generates a prediction result based on the target fusion features, including: normalizing the target fusion features using the output network in the recommendation model to obtain target output features; and performing feature mapping on the target output features using the feature mapping network in the recommendation model to obtain the prediction result. Optionally, the output network includes a feedforward neural network layer and a normalization layer.

[0224] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0225] In Figure 6Taking the embodiment of as an example, the scene granularity features include scene granularity memory features and scene granularity generalization features, the task granularity features include task granularity memory features, and the global granularity features include global granularity generalization features. Moreover, the number of scene granularity memory features, scene granularity generalization features, task granularity memory features and global granularity generalization features is multiple, and the fusion network includes a first scene granularity fusion network, a second scene granularity fusion network, a task granularity fusion network, a global granularity fusion network, a memory fusion network, a generalization fusion network and a target fusion network. The implementation process of the multimedia resource recommendation method is described below. Figure 10 An embodiment of Figure 10 This is a flowchart of another multimedia resource recommendation method provided by an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 10 , the method comprising:

[0226] 1001. The computer device obtains task features of a target recommendation task, where the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any recommendation scenario among multiple recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0227] 1002. The computer device obtains memory bias features and generalization bias features from task features. The memory bias features refer to features that are affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features that are not affected by the interaction between the object and the multimedia resource.

[0228] The process of steps 1001-1002 is the same as that of steps 401-402, and will not be repeated here.

[0229] 1003. The computer device extracts the memory bias features through the x scene memory expert networks in the recommendation model to obtain x scene granularity memory features.

[0230] In the embodiment of this application, Figure 3 As shown, the scene expert network of the recommendation model includes x scene memory expert networks. The computer device inputs the memory bias features into the x scene memory expert networks respectively, obtains the scene granularity memory features output by each scene memory expert network, and thus obtains x scene granularity memory features.

[0231] 1004. The computer device extracts the generalization bias features through the y scene generalization expert networks in the recommendation model to obtain y scene granularity generalization features.

[0232] In the embodiment of this application, Figure 3As shown, the scene expert network of the recommendation model includes y scene generalization expert networks. The computer device inputs the generalization bias features into the y scene generalization expert networks respectively, obtains the scene granularity generalization features output by each scene generalization expert network, and thus obtains y scene granularity generalization features.

[0233] The process of steps 1003-1004 is the same as that of step 602 above, and will not be repeated here.

[0234] 1005. The computer device extracts the memory bias feature through the n task memory expert networks in the recommendation model to obtain n task granularity memory features.

[0235] In the embodiment of this application, Figure 3 As shown, the task expert network of the recommendation model includes n task memory expert networks. The computer device inputs the memory bias features into the n task memory expert networks respectively, obtains the task granularity memory features output by each task memory expert network, and thus obtains n task granularity memory features.

[0236] The process of step 1005 is the same as that of step 603 above, and will not be repeated here.

[0237] 1006. The computer device extracts the generalization bias features through the k global generalization expert networks in the recommendation model, and obtains k global granularity generalization features respectively.

[0238] In the embodiment of this application, Figure 3 As shown, the global expert network of the recommendation model includes k global generalization expert networks. The computer device inputs the generalization bias features into the k global generalization expert networks respectively, obtains the global granularity generalization features output by each global generalization expert network, and thus obtains k global granularity generalization features.

[0239] The process of step 1006 is the same as that of step 604 above, and will not be repeated here.

[0240] 1007. The computer device fuses the x scene granularity memory features through the first scene granularity fusion network to obtain a fused scene granularity memory feature.

[0241] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes a first scene granularity fusion network, and the first scene granularity fusion feature is used to fuse scene granularity memory features. The computer device inputs x scene granularity memory features into the first scene granularity fusion network, and obtains the fused scene granularity memory features output by the first scene granularity fusion network.

[0242] In one possible implementation, step 1007 includes: for any scene-granularity memory feature, performing normalization based on the product of the scene-granularity memory feature and the transpose of the scene-granularity memory feature using a first scene-granularity fusion network to obtain a first scene weight for the scene-granularity memory feature; and performing weighted fusion of multiple scene-granularity memory features based on their first scene weights using the first scene-granularity fusion network to obtain a fused scene-granularity memory feature. The multiple scene-granularity memory features refer to the x scene-granularity memory features.

[0243] Optionally, the first scene granularity fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate weights, and the weighted fusion part is used to perform weighted summation. Figure 11 This is a structural diagram of a first scenario granularity fusion network provided by an embodiment of the present application, such as Figure 11 As shown, the gating part of the first scene granularity fusion network includes a self-attention layer and a normalization layer, which are connected to each other, and the weighted fusion part of the first scene granularity fusion network includes a multiplication operator and an addition operator. Optionally, the processing process of the first scene granularity fusion network can be expressed by the following formula:

[0244]

[0245]

[0246] in, represents the granular memory feature of the i-th scene, and are the network parameters of the gated part in the first scene granularity fusion network, softmax(·) represents the normalization layer, is the output of the self-attention layer, represents the first scene weight of the i-th scene granularity memory feature, Represents the fused scene granularity memory features.

[0247] 1008. The computer device fuses the y scene granularity generalization features through a second scene granularity fusion network to obtain a fused scene granularity generalization feature.

[0248] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes a second scene granularity fusion network, which is used to fuse scene granularity generalization features. The computer device inputs y scene granularity generalization features into the second scene granularity fusion network, and obtains the fused scene granularity generalization features output by the second scene granularity fusion network.

[0249] In one possible implementation, step 1008 includes: normalizing any scene-granularity generalization feature using a second scene-granularity fusion network based on the product of the scene-granularity generalization feature and the transpose of the scene-granularity generalization feature to obtain a second scene weight for the scene-granularity generalization feature; and performing weighted fusion of multiple scene-granularity generalization features using the second scene weights of the multiple scene-granularity generalization features using the second scene weights to obtain a fused scene-granularity generalization feature. The multiple scene-granularity generalization features refer to the y scene-granularity generalization features.

[0250] Optionally, the second scene granularity fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate the weight, and the weighted fusion part is used to perform weighted summation. Among them, the network structure and processing process of the second scene granularity fusion network are the same as the network structure and processing process of the first scene granularity fusion network, and will not be repeated here. It should be noted that the network structures of the first scene granularity fusion network and the second scene granularity fusion network can be the same, but the network parameters of the first scene granularity fusion network and the second scene granularity fusion network are different, and the input and output of the network are also different.

[0251] 1009. The computer device fuses n task granularity memory features through a task granularity fusion network to obtain a fused task granularity memory feature.

[0252] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes a task-granularity fusion network, which is used to fuse task-granularity memory features. The computer device inputs n task-granularity memory features into the task-granularity fusion network, and obtains the fused task-granularity memory features as output by the task-granularity fusion network.

[0253] In one possible implementation, step 1009 includes: for any task-granularity memory feature, normalizing the feature based on the product of the task-granularity memory feature and the transpose of the task-granularity memory feature using a task-granularity fusion network to obtain a task weight for the task-granularity memory feature; and weightedly fusing multiple task-granularity memory features based on their task weights using the task-granularity fusion network to obtain a fused task-granularity memory feature. The multiple task-granularity memory features refer to the n task-granularity memory features.

[0254] Optionally, the task granularity fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate the weights, and the weighted fusion part is used to perform weighted summation. Among them, the network structure and processing process of the task granularity fusion network are the same as the network structure and processing process of the first scene granularity fusion network, and will not be repeated here. It should be noted that the network structure of the first scene granularity fusion network and the task granularity fusion network can be the same, but the network parameters of the first scene granularity fusion network and the task granularity fusion network are different, and the input and output of the network are also different.

[0255] 1010. The computer device fuses k global granularity generalization features through a global granularity fusion network to obtain a fused global granularity generalization feature.

[0256] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes a global granularity fusion network, which is used to fuse task-granularity memory features. The computer device inputs k global granularity generalization features into the global granularity fusion network, and obtains the fused global granularity generalization features output by the global granularity fusion network.

[0257] In one possible implementation, step 1010 includes: for any global granularity generalization feature, normalizing it based on the product of the global granularity generalization feature and the transpose of the global granularity generalization feature through a global granularity fusion network to obtain a global weight of the global granularity generalization feature; and weightedly fusing multiple global granularity generalization features based on the global weights of the multiple global granularity generalization features through the global granularity fusion network to obtain a fused global granularity generalization feature.

[0258] Optionally, the global granularity fusion network includes a gating part and a weighted fusion part, the gating part is used to calculate the weight, and the weighted fusion part is used to perform weighted summation. Among them, the network structure and processing process of the global granularity fusion network are the same as the network structure and processing process of the first scene granularity fusion network, and will not be repeated here. It should be noted that the network structure of the first scene granularity fusion network and the global granularity fusion network can be the same, but the network parameters of the first scene granularity fusion network and the global granularity fusion network are different, and the input and output of the network are also different.

[0259] 1011. The computer device fuses the fused scene granularity memory features and the fused task granularity memory features through a memory fusion network to obtain a memory fusion feature.

[0260] 1012. The computer device fuses the fused scene granularity generalized features and the fused global granularity generalized features through a generalized fusion network to obtain a generalized fusion feature.

[0261] 1013. The computer device fuses the memory fusion feature and the generalization fusion feature through the target fusion network to obtain the target fusion feature.

[0262] The process of steps 1011-1013 is the same as that of steps 701-703 above, and will not be repeated here.

[0263] 1014. The computer device normalizes the target fusion features based on the output network in the recommendation model to obtain the target output features.

[0264] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes an output network, and the computer device inputs the target fusion features into the output network to obtain the target output features output by the output network.

[0265] In one possible implementation, the output network includes a feedforward neural network layer and a normalization layer. The feedforward neural network layer processes the target fusion features to obtain a processing result, and the normalization layer normalizes the processing result to obtain a target output feature.

[0266] 1015. The computer device performs feature mapping on the target output features through the feature mapping network in the recommendation model to obtain a prediction result.

[0267] In the embodiment of this application, Figure 3 As shown, the recommendation model also includes a feature mapping network, and the computer device inputs the target output feature into the feature mapping network to obtain the prediction result output by the feature mapping network.

[0268] In a possible implementation, the prediction result is a prediction probability, where the prediction probability is a probability that the object interacts with the multimedia resource when the multimedia resource is recommended to the object.

[0269] The core of the embodiment of the present application lies in a multi-level and multi-granularity hybrid expert network, which specifically includes feature decoupling of memory bias and generalization bias, expert networks of three granularities: task-exclusive, scene-shared, and global-shared, and a three-level fusion network based on an adaptive perception gating mechanism. On the one hand, expert networks of different granularities are set up for the three levels of global, scene, and task, and the expert networks of different granularities are cleverly integrated through an adaptive gating mechanism, which can efficiently mine and characterize the commonalities and individualities of tasks in different scenarios in the feature space, and effectively alleviate the problem of difficult task bias modeling in multi-task joint modeling. On the other hand, based on the main structure of the hybrid expert network, the expert network with memory bias and generalization bias is decoupled. Since the expert network with generalization bias is conducive to the modeling of cold start and long-tail multimedia resources, the problem of poor recommendation effect of cold start and long-tail multimedia resources is effectively alleviated. At the same time, the expert network with memory bias is used to strengthen the memory of recommendation preferences for other popular and non-long-tail multimedia resources, further enhancing the recommendation effect.

[0270] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0271] Figure 12 This is a flowchart of a training method for a recommendation model provided in an embodiment of the present application. The embodiment of the present application is executed by a computer device. Figure 12 , the method comprising:

[0272] 1201. The computer device obtains sample task features and sample labels of the sample recommendation task. The sample task features include sample object features of the sample object and sample resource features of the sample multimedia resource. The sample recommendation task belongs to any recommendation task in any recommendation scenario. The sample label is used to characterize the actual interaction between the sample object and the sample multimedia resource after the sample multimedia resource is recommended to the sample object.

[0273] The sample task features of the sample recommendation task are similar to the task features of the target recommendation task mentioned above, and will not be repeated here.

[0274] Optionally, the sample label is a sample probability. If the sample object interacts with the sample multimedia resource, the sample probability is equal to 1; if the sample object does not interact with the sample multimedia resource, the sample probability is equal to 0.

[0275] 1202. The computer device extracts the sample task features through the scene expert network in the recommendation model to obtain the sample scene granularity features.

[0276] 1203. The computer device extracts the sample task features through the task expert network in the recommendation model to obtain the sample task granularity features.

[0277] 1204. The computer device extracts the sample task features through the global expert network in the recommendation model to obtain the sample global granularity features.

[0278] 1205. The computer device determines a sample prediction result based on the sample scene granularity feature, the sample task granularity feature, and the sample global granularity feature. The sample prediction result is used to characterize the interaction between the sample object and the sample multimedia resource when the sample multimedia resource is recommended to the sample object.

[0279] The process of determining the sample prediction results in steps 1202-1205 is similar to the process of determining the prediction results of the target recommendation task in the above embodiments, and will not be repeated here.

[0280] 1206. The computer device trains the recommendation model based on the difference between the sample prediction results and the sample labels.

[0281] The sample label can reflect the actual interaction between the sample object and the sample multimedia resource, and the sample prediction result is the prediction result obtained based on the recommendation model, which can reflect the predicted interaction between the sample object and the sample multimedia resource. The smaller the difference between the sample prediction result and the sample label, the more accurate the sample prediction result, that is, the more accurate the recommendation model. Therefore, the computer device trains the recommendation model based on the difference between the sample prediction result and the sample label to reduce the difference between the sample prediction result obtained based on the trained recommendation model and the sample label, thereby improving the accuracy of the recommendation model.

[0282] The method provided in the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity, and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity, and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0283] Figure 13 This is a schematic diagram of the structure of a multimedia resource recommendation device provided by an embodiment of the present application. Figure 13 , the device comprises:

[0284] Feature acquisition module 1301, used to acquire task features of a target recommendation task, where the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any of multiple recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task.

[0285] The feature extraction module 1302 is used to extract the task features through the scene expert network in the recommendation model to obtain scene granularity features;

[0286] The feature extraction module 1302 is further used to extract the task features through the task expert network in the recommendation model to obtain task granularity features;

[0287] The feature extraction module 1302 is further used to extract the task features through the global expert network in the recommendation model to obtain global granular features;

[0288] The prediction module 1303 is used to determine a prediction result based on the scene granularity feature, the task granularity feature and the global granularity feature. The prediction result is used to represent the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

[0289] The multimedia resource recommendation device provided by the embodiment of the present application, for any recommendation task in any recommendation scenario, first obtains the task features of the recommendation task, uses the recommendation model to make predictions based on the task features, and sets three expert networks of different granularities for the three levels of scene granularity, task granularity and global granularity in the recommendation model. The three expert networks of different granularities respectively extract features of the task features to obtain features of three different granularities, and then predicts the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object based on the features of the three different granularities, so as to determine whether to recommend the multimedia resource to the object. Since the processing is performed at the scene granularity, task granularity and global granularity respectively, the commonalities and differences of different tasks in each scenario in the feature space can be efficiently mined, which can effectively alleviate the problem of poor recommendation effect of the recommendation model shared by multiple tasks, thereby improving the accuracy of multimedia resource recommendation.

[0290] Alternatively, see Figure 14 The scene expert network includes a scene memory expert network and a scene generalization expert network; the feature extraction module 1302 is used to:

[0291] Obtain memory bias features and generalization bias features from task features. Memory bias features refer to features affected by the interaction between the object and the multimedia resources, while generalization bias features refer to features not affected by the interaction between the object and the multimedia resources.

[0292] Through the scene memory expert network, the memory bias features are extracted to obtain the scene granularity memory features;

[0293] Through the scene generalization expert network, the generalization bias features are extracted to obtain the scene granularity generalization features, which include scene granularity memory features and scene granularity generalization features.

[0294] Alternatively, see Figure 14 The task expert network is a task memory expert network; the feature extraction module 1302 is used to:

[0295] Obtain memory bias features from task features. Memory bias features refer to features affected by the interaction between the subject and multimedia resources.

[0296] The memory bias features are extracted through the task memory expert network to obtain task granularity memory features, which include task granularity memory features.

[0297] Alternatively, see Figure 14 , the global expert network is a global generalized expert network; the feature extraction module 1302 is used to:

[0298] Obtain generalization bias features from task features. Generalization bias features refer to features that are not affected by the interaction between the object and the multimedia resources.

[0299] The generalization bias features are extracted through the global generalization expert network to obtain the global granularity generalization features, which include the global granularity generalization features.

[0300] Alternatively, see Figure 14 , the number of scene expert networks is m, the number of scene granularity features is m, and the network parameters of each scene expert network are different; the number of task expert networks is n, the number of task granularity features is n, and the network parameters of each task expert network are different; the number of global expert networks is k, the number of global granularity features is k, and the network parameters of each global expert network are different; among them, m, n and k are all positive integers.

[0301] Alternatively, see Figure 14 The recommendation model also includes a fusion network; a prediction module 1303, which is used to:

[0302] A fusion unit 1313 is configured to fuse the scene granularity features, the task granularity features, and the global granularity features through a fusion network in the recommendation model to obtain a target fusion feature;

[0303] The prediction unit 1323 is configured to generate a prediction result based on the target fusion feature.

[0304] Alternatively, see Figure 14 , the scene granularity features include scene granularity memory features and scene granularity generalization features, the task granularity features include task granularity memory features, and the global granularity features include global granularity generalization features; the fusion network includes memory fusion network, generalization fusion network and target fusion network;

[0305] The fusion unit 1313 is configured to:

[0306] Through the memory fusion network, the scene granularity memory features and task granularity memory features are fused to obtain the memory fusion features;

[0307] Through the generalization fusion network, the scene granularity generalization features and the global granularity generalization features are fused to obtain the generalization fusion features;

[0308] Through the target fusion network, the memory fusion features and the generalization fusion features are fused to obtain the target fusion features.

[0309] Alternatively, see Figure 14The task features include memory bias features and generalization bias features. The memory bias features refer to features affected by the interaction between the object and the multimedia resources, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resources. The fusion unit 1313 is used to:

[0310] Through the memory fusion network, the concatenation results of the memory bias feature and the generalization bias feature are normalized to obtain the first memory weight and the second memory weight;

[0311] Through the memory fusion network, the scene granularity memory features and the task granularity memory features are weightedly fused based on the first memory weight and the second memory weight to obtain the memory fusion features.

[0312] Alternatively, see Figure 14 The task features include memory bias features and generalization bias features. The memory bias features refer to features affected by the interaction between the object and the multimedia resources, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resources. The fusion unit 1313 is used to:

[0313] Through the generalization fusion network, the concatenation results of the memory bias feature and the generalization bias feature are normalized to obtain the first generalization weight and the second generalization weight;

[0314] The generalized fusion network is used to perform weighted fusion of the scene granularity generalized features and the global granularity generalized features based on the first generalization weight and the second generalization weight to obtain the generalized fusion features.

[0315] Alternatively, see Figure 14 , the fusion unit 1313 is used to:

[0316] Obtaining memory-biased resource features from multimedia resource features, where the memory-biased resource features refer to multimedia resource features affected by the interaction between the object and the multimedia resource;

[0317] Through the target fusion network, normalization is performed based on the memory bias resource characteristics to obtain the first target weight and the second target weight;

[0318] Through the target fusion network, the memory fusion feature and the generalization fusion feature are weightedly fused based on the first target weight and the second target weight to obtain the target fusion feature.

[0319] Alternatively, see Figure 14 The number of scene granularity memory features, scene granularity generalization features, task granularity memory features, and global granularity generalization features is multiple; the fusion network further includes a first scene granularity fusion network, a second scene granularity fusion network, a task granularity fusion network, and a global granularity fusion network; the fusion unit 1313 is further used to:

[0320] Through the first scene granularity fusion network, multiple scene granularity memory features are fused to obtain fused scene granularity memory features;

[0321] Through the second scene granularity fusion network, multiple scene granularity generalization features are fused to obtain the fused scene granularity generalization features;

[0322] Through the task granularity fusion network, multiple task granularity memory features are fused to obtain the fused task granularity memory features;

[0323] Through the global granularity fusion network, multiple global granularity generalization features are fused to obtain the fused global granularity generalization features;

[0324] Among them, the input of the memory fusion network is the fused scene granularity memory features and the fused task granularity memory features, and the input of the generalization fusion network is the fused scene granularity generalization features and the fused global granularity generalization features.

[0325] Alternatively, see Figure 14 , the fusion unit 1313 is used to:

[0326] For any scene granularity memory feature, normalize it based on the product of the scene granularity memory feature and the transpose of the scene granularity memory feature through the first scene granularity fusion network to obtain the first scene weight of the scene granularity memory feature;

[0327] Through the first scene granularity fusion network, multiple scene granularity memory features are weightedly fused based on the first scene weights of the multiple scene granularity memory features to obtain a fused scene granularity memory feature.

[0328] Alternatively, see Figure 14 The prediction result is a prediction probability, which is the probability that the object will interact with the multimedia resource when the multimedia resource is recommended to the object. After determining the prediction result based on the scene granularity features, the task granularity features, and the global granularity features, the device further includes a recommendation module 1304 for:

[0329] When the predicted probability is greater than a preset probability threshold, multimedia resources are recommended to the subject.

[0330] Alternatively, see Figure 14 The device further includes a model training module 1305, which is used to:

[0331] Obtaining sample task features and sample labels for the sample recommendation task. The sample task features include sample object features of the sample object and sample resource features of the sample multimedia resource. The sample recommendation task belongs to any recommendation task in any recommendation scenario. The sample label is used to characterize the actual interaction between the sample object and the sample multimedia resource after the sample multimedia resource is recommended to the sample object.

[0332] Through the scene expert network in the recommendation model, the sample task features are extracted to obtain the sample scene granularity features;

[0333] Through the task expert network in the recommendation model, the sample task features are extracted to obtain the sample task granularity features;

[0334] Through the global expert network in the recommendation model, the sample task features are extracted to obtain the global granularity features of the sample;

[0335] Determine a sample prediction result based on the sample scene granularity feature, the sample task granularity feature, and the sample global granularity feature. The sample prediction result is used to characterize the interaction between the sample object and the sample multimedia resource when the sample multimedia resource is recommended to the sample object.

[0336] The recommendation model is trained based on the difference between the sample prediction results and the sample labels.

[0337] It should be noted that the multimedia resource recommendation device provided in the above embodiment is merely an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multimedia resource recommendation device provided in the above embodiment and the multimedia resource recommendation method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0338] An embodiment of the present application also provides a computer device, which includes a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed in the multimedia resource recommendation method of the above embodiment.

[0339] Optionally, the computer device is provided as a terminal. Figure 15 FIG1 shows a schematic diagram of the structure of a terminal 1500 provided by an exemplary embodiment of the present application. The terminal 1500 includes: a processor 1501 and a memory 1502 .

[0340] The processor 1501 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1501 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field Programmable Gate Array), or PLA (Programmable Logic Array). The processor 1501 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 1501 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1501 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0341] Memory 1502 may include one or more computer-readable storage media, which may be non-transitory. Memory 1502 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in memory 1502 is used to store at least one computer program, which is used by processor 1501 to implement the multimedia resource recommendation method provided in the method embodiment of the present application.

[0342] In some embodiments, terminal 1500 may optionally include a peripheral device interface 1503 and at least one peripheral device. Processor 1501, memory 1502, and peripheral device interface 1503 may be connected via a bus or signal lines. Each peripheral device may be connected to peripheral device interface 1503 via a bus, signal lines, or circuit boards. Optionally, the peripheral device includes at least one of a radio frequency circuit 1504, a display screen 1505, a camera assembly 1506, an audio circuit 1507, and a power supply 1508.

[0343] The peripheral device interface 1503 can be used to connect at least one I / O (Input / Output)-related peripheral device to the processor 1501 and the memory 1502. In some embodiments, the processor 1501, the memory 1502, and the peripheral device interface 1503 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1501, the memory 1502, and the peripheral device interface 1503 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0344] RF circuit 1504 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. RF circuit 1504 communicates with communication networks and other communication devices via electromagnetic signals. RF circuit 1504 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. RF circuit 1504 may optionally include an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and the like. RF circuit 1504 may communicate with other devices via at least one wireless communication protocol. Such wireless communication protocols include, but are not limited to, metropolitan area networks (MANs), various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks (WLANs), and / or WiFi (Wireless Fidelity) networks. In some embodiments, RF circuit 1504 may also include circuitry related to NFC (Near Field Communication), although this application does not limit this.

[0345] The display screen 1505 is used to display a UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 1505 is a touch screen display, the display screen 1505 also has the ability to collect touch signals on the surface or above the surface of the display screen 1505. The touch signal can be input as a control signal to the processor 1501 for processing. In this case, the display screen 1505 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, there can be one display screen 1505, which is set on the front panel of the terminal 1500; in other embodiments, there can be at least two display screens 1505, which are respectively set on different surfaces of the terminal 1500 or in a folding design; in other embodiments, the display screen 1505 can be a flexible display screen, which is set on the curved surface or folding surface of the terminal 1500. Even more, the display screen 1505 can be set to a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 1505 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0346] The camera assembly 1506 is used to capture images or videos. Optionally, the camera assembly 1506 includes a front camera and a rear camera. The front camera is arranged on the front panel of the terminal 1500, and the rear camera is arranged on the back of the terminal 1500. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize panoramic shooting and VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 1506 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. The dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation under different color temperatures.

[0347] The audio circuit 1507 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals that are input into the processor 1501 for processing, or input into the radio frequency circuit 1504 to achieve voice communication. For the purpose of stereo sound collection or noise reduction, there may be multiple microphones, each located in different parts of the terminal 1500. The microphone may also be an array microphone or an omnidirectional collection microphone. The speaker is used to convert electrical signals from the processor 1501 or the radio frequency circuit 1504 into sound waves. The speaker may be a traditional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert electrical signals into sound waves audible to humans, but also convert electrical signals into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 1507 may also include a headphone jack.

[0348] Power supply 1508 is used to power various components in terminal 1500. Power supply 1508 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 1508 includes a rechargeable battery, the rechargeable battery can support wired charging or wireless charging. The rechargeable battery can also be used to support fast charging technology.

[0349] Those skilled in the art will understand that Figure 15 The structure shown in the figure does not constitute a limitation on the terminal 1500, and the terminal 1500 may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0350] Optionally, the computer device is provided as a server. Figure 16 This is a schematic diagram of the structure of a server provided in an embodiment of the present application. The server 1600 may have relatively large differences due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 1601 and one or more memories 1602, wherein the memory 1602 stores at least one computer program, and the at least one computer program is loaded and executed by the processor 1601 to implement the methods provided in the above-mentioned various method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server may also include other components for implementing device functions, which will not be described in detail here.

[0351] An embodiment of the present application further provides a computer-readable storage medium, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the operations performed by the multimedia resource recommendation method of the above embodiment.

[0352] An embodiment of the present application further provides a computer program product, including a computer program, which is loaded and executed by a processor to implement the operations performed by the multimedia resource recommendation method of the above embodiment.

[0353] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0354] The above description is merely an optional embodiment of the embodiments of the present application and is not intended to limit the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included in the scope of protection of the present application.

Claims

1. A multimedia resource recommendation method, characterized in that: The method comprises: Obtaining task features of a target recommendation task, the task features including object features of an object and resource features of a multimedia resource, the target recommendation task belonging to any one of a plurality of recommendation scenarios, and each recommendation scenario corresponding to at least one recommendation task; The task features are extracted through the scene expert network in the recommendation model to obtain scene granularity features; Extracting the task features through the task expert network in the recommendation model to obtain task granularity features; Extracting the task features through the global expert network in the recommendation model to obtain global granular features; A prediction result is determined based on the scene granularity feature, the task granularity feature, and the global granularity feature, where the prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

2. The method according to claim 1, characterized in that The scene expert network includes a scene memory expert network and a scene generalization expert network; the scene expert network in the recommendation model is used to extract the task features to obtain scene granularity features, including: Obtaining a memory bias feature and a generalization bias feature from the task features, wherein the memory bias feature refers to a feature affected by the interaction between the object and the multimedia resource, and the generalization bias feature refers to a feature not affected by the interaction between the object and the multimedia resource; Extracting the memory bias features through the scene memory expert network to obtain scene granularity memory features; The generalization bias feature is extracted through the scene generalization expert network to obtain a scene granularity generalization feature, where the scene granularity feature includes the scene granularity memory feature and the scene granularity generalization feature.

3. The method according to claim 1, characterized in that The task expert network is a task memory expert network; the task features are extracted through the task expert network in the recommendation model to obtain task granularity features, including: Acquire a memory bias feature from the task feature, wherein the memory bias feature refers to a feature affected by the interaction between the object and the multimedia resource; The memory bias feature is extracted through the task memory expert network to obtain a task granularity memory feature, wherein the task granularity feature includes the task granularity memory feature.

4. The method according to claim 1, wherein The global expert network is a global generalized expert network; the task features are extracted through the global expert network in the recommendation model to obtain global granular features, including: Obtaining a generalized bias feature from the task feature, wherein the generalized bias feature refers to a feature that is not affected by the interaction between the object and the multimedia resource; The generalization bias feature is extracted through the global generalization expert network to obtain a global granularity generalization feature, which includes the global granularity generalization feature.

5. The method according to claim 1, wherein The number of the scene expert networks is m, the number of the scene granularity features is m, and the network parameters of each scene expert network are different; the number of the task expert networks is n, the number of the task granularity features is n, and the network parameters of each task expert network are different; the number of the global expert networks is k, the number of the global granularity features is k, and the network parameters of each global expert network are different; wherein m, n and k are all positive integers.

6. The method according to claim 1, characterized in that The recommendation model further includes a fusion network; and determining a prediction result based on the scene granularity feature, the task granularity feature, and the global granularity feature includes: The scene granularity feature, the task granularity feature and the global granularity feature are fused through a fusion network in the recommendation model to obtain a target fusion feature; Based on the target fusion features, the prediction result is generated.

7. The method according to claim 6, characterized in that The scene granularity features include scene granularity memory features and scene granularity generalization features, the task granularity features include task granularity memory features, and the global granularity features include global granularity generalization features; the fusion network includes a memory fusion network, a generalization fusion network, and a target fusion network; The step of fusing the scene granularity features, the task granularity features, and the global granularity features through the fusion network in the recommendation model to obtain target fusion features includes: The scene granularity memory feature and the task granularity memory feature are fused through the memory fusion network to obtain a memory fusion feature; The scene granularity generalized features and the global granularity generalized features are fused through the generalized fusion network to obtain a generalized fusion feature; The memory fusion feature and the generalization fusion feature are fused through the target fusion network to obtain the target fusion feature.

8. The method according to claim 7, characterized in that The task features include memory bias features and generalization bias features, wherein the memory bias features refer to features affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resource; the memory fusion network is used to fuse the scene granularity memory features and the task granularity memory features to obtain memory fusion features, including: Normalizing the concatenation result of the memory bias feature and the generalization bias feature through the memory fusion network to obtain a first memory weight and a second memory weight; The memory fusion network performs weighted fusion on the scene granularity memory feature and the task granularity memory feature based on the first memory weight and the second memory weight to obtain the memory fusion feature.

9. The method according to claim 7, characterized in that The task features include memory bias features and generalization bias features, wherein the memory bias features refer to features affected by the interaction between the object and the multimedia resource, and the generalization bias features refer to features not affected by the interaction between the object and the multimedia resource; the generalization fusion network is used to fuse the scene granularity generalization features and the global granularity generalization features to obtain generalization fusion features, including: Normalizing the concatenation result of the memory bias feature and the generalization bias feature through the generalization fusion network to obtain a first generalization weight and a second generalization weight; The generalized fusion network performs weighted fusion on the scene granularity generalized features and the global granularity generalized features based on the first generalization weight and the second generalization weight to obtain the generalized fusion features.

10. The method according to claim 7, characterized in that The target fusion network is used to fuse the memory fusion feature and the generalization fusion feature to obtain the target fusion feature, including: Acquire a memory-biased resource feature from the multimedia resource features, wherein the memory-biased resource feature refers to a multimedia resource feature affected by an interaction between the object and the multimedia resource; Normalizing based on the memory bias resource characteristics through the target fusion network to obtain a first target weight and a second target weight; The target fusion network performs weighted fusion on the memory fusion feature and the generalization fusion feature based on the first target weight and the second target weight to obtain the target fusion feature.

11. The method according to claim 7, characterized in that The scene granularity memory feature, the scene granularity generalization feature, the task granularity memory feature, and the global granularity generalization feature are all multiple; the fusion network further includes a first scene granularity fusion network, a second scene granularity fusion network, a task granularity fusion network, and a global granularity fusion network; the method further includes: fusing multiple scene granularity memory features through the first scene granularity fusion network to obtain a fused scene granularity memory feature; fusing multiple scene granularity generalization features through the second scene granularity fusion network to obtain a fused scene granularity generalization feature; By using the task granularity fusion network, multiple task granularity memory features are fused to obtain fused task granularity memory features; By using the global granularity fusion network, multiple global granularity generalization features are fused to obtain a fused global granularity generalization feature; Among them, the input of the memory fusion network is the fused scene granularity memory features and the fused task granularity memory features, and the input of the generalization fusion network is the fused scene granularity generalization features and the fused global granularity generalization features.

12. The method according to claim 11, characterized in that The method of fusing multiple scene granularity memory features through the first scene granularity fusion network to obtain a fused scene granularity memory feature includes: For any scene granularity memory feature, normalizing the scene granularity memory feature by the first scene granularity fusion network based on the product of the scene granularity memory feature and the transpose of the scene granularity memory feature to obtain a first scene weight of the scene granularity memory feature; The first scene granularity fusion network performs weighted fusion on the multiple scene granularity memory features based on their first scene weights to obtain the fused scene granularity memory features.

13. The method according to any one of claims 1 to 12, characterized in that The prediction result is a prediction probability, which is a probability that the object will interact with the multimedia resource when the multimedia resource is recommended to the object; after determining the prediction result based on the scene granularity feature, the task granularity feature, and the global granularity feature, the method further includes: When the predicted probability is greater than a preset probability threshold, the multimedia resource is recommended to the object.

14. The method according to any one of claims 1 to 12, characterized in that The training process of the recommendation model includes: Obtaining sample task features and sample labels for a sample recommendation task, wherein the sample task features include sample object features of a sample object and sample resource features of a sample multimedia resource. The sample recommendation task belongs to any recommendation task in any recommendation scenario. The sample label is used to characterize the actual interaction between the sample object and the sample multimedia resource after the sample multimedia resource is recommended to the sample object. Extracting the sample task features through the scene expert network in the recommendation model to obtain sample scene granularity features; Extracting the sample task features through the task expert network in the recommendation model to obtain sample task granularity features; Extracting the sample task features through the global expert network in the recommendation model to obtain the sample global granularity features; Determining a sample prediction result based on the sample scene granularity feature, the sample task granularity feature, and the sample global granularity feature, wherein the sample prediction result is used to characterize an interaction between the sample object and the sample multimedia resource when the sample multimedia resource is recommended to the sample object; The recommendation model is trained based on the difference between the sample prediction result and the sample label.

15. A multimedia resource recommendation device, characterized in that: The device comprises: A feature acquisition module is used to acquire task features of a target recommendation task, wherein the task features include object features of an object and resource features of a multimedia resource. The target recommendation task belongs to any recommendation task in any of a plurality of recommendation scenarios, and each recommendation scenario corresponds to at least one recommendation task. A feature extraction module is used to extract the task features through the scene expert network in the recommendation model to obtain scene granularity features; The feature extraction module is further configured to extract the task features through the task expert network in the recommendation model to obtain task granularity features; The feature extraction module is further configured to extract the task features through the global expert network in the recommendation model to obtain global granularity features; A prediction module is used to determine a prediction result based on the scene granularity feature, the task granularity feature and the global granularity feature, wherein the prediction result is used to characterize the interaction between the object and the multimedia resource when the multimedia resource is recommended to the object.

16. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed by the multimedia resource recommendation method according to any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed by the multimedia resource recommendation method according to any one of claims 1 to 14.

18. A computer program product comprising a computer program, characterized in that The computer program is loaded and executed by a processor to implement the operations performed by the multimedia resource recommendation method according to any one of claims 1 to 14.

Images