Content recommendation method and apparatus, computer device, and storage medium

Abstract

The present disclosure provides a content recommendation method, device, computer equipment and storage medium. The content recommendation method comprises: obtaining scene data of a target recommendation scene in a target application, input data of a plurality of feature domains of a plurality of to-be-recommended element objects, and domain label data of each feature domain; extracting scene features from the scene data, and extracting input features from the plurality of input data, and generating scene weights of each to-be-recommended element object based on the scene features and the plurality of input features; extracting domain label features from the plurality of domain label data, and generating target object features of each to-be-recommended element object based on the scene weights, the plurality of input features and the plurality of domain label features; determining recommendation prediction data of each to-be-recommended element object according to the target object features, and recommending the plurality of to-be-recommended element objects according to the recommendation prediction data. The embodiment of the present disclosure can improve the accuracy of recommended content.

Description

Technical Field

[0001] This disclosure relates to the field of content recommendation technology, and in particular to a content recommendation method, apparatus, computer device, and storage medium. Background Technology

[0002] In the internet age, internet applications typically need to push content (such as short videos and articles) to platform users to meet their content needs. Content recommendation is a process that uses algorithms to analyze user preferences and other relevant information to predict content that users might be interested in and then recommends that content to them. Content recommendation algorithms can recommend videos, articles, and other content that users might be interested in, thereby increasing user engagement with the application.

[0003] Currently, since each application may have multiple recommendation scenarios, and these scenarios typically require individually designed customized models to adapt to the differences in data distribution and feature space under different scenarios, related technologies usually employ Embedding Personalized Network (EPNET) models to achieve multi-scenario, multi-task content recommendation.

[0004] However, the current EPNET model does not recommend content to users accurately enough in various application scenarios. Summary of the Invention

[0005] This disclosure provides a content recommendation method, apparatus, computer device, and storage medium, which can improve the accuracy of recommended content.

[0006] According to one aspect of this disclosure, a content recommendation method is provided, comprising:

[0007] Acquire scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain;

[0008] Scene features are obtained by extracting features from the scene data, and multiple input features are obtained by extracting features from multiple input data. Scene weights are generated for each element object to be recommended based on the scene features and the multiple input features.

[0009] Multiple domain label features are obtained by extracting features from multiple domain label data, and target object features corresponding to each recommended element object are generated by fusing the scene weight, the multiple input features and the multiple domain label features.

[0010] Based on the characteristics of the target object, the recommendation prediction data for each of the proposed element objects is determined, and recommendations are made to the plurality of proposed element objects based on the recommendation prediction data.

[0011] According to one aspect of this disclosure, a content recommendation apparatus is provided, comprising:

[0012] The acquisition unit is used to acquire scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain.

[0013] The weight generation unit is used to extract features from the scene data to obtain scene features, and to extract features from multiple input data to obtain multiple input features, and to generate scene weights corresponding to each of the recommended element objects based on the scene features and the multiple input features.

[0014] The fusion unit is used to extract features from multiple domain label data to obtain multiple domain label features, and to fuse the scene weights, multiple input features and multiple domain label features to generate target object features corresponding to each element object to be recommended.

[0015] The recommendation unit is used to determine the recommendation prediction data for each element object to be recommended based on the characteristics of the target object, and to recommend the plurality of element objects to be recommended based on the recommendation prediction data.

[0016] Optionally, in some embodiments, the fusion unit is further configured to:

[0017] The input features and the domain label features corresponding to each feature domain are fused to obtain the first fused feature corresponding to each feature domain.

[0018] Based on the scene weights, each of the first fusion features is weighted to obtain the first weighted feature corresponding to each feature domain.

[0019] The target object feature corresponding to each element object to be recommended is generated by fusing multiple first weighted features.

[0020] Optionally, in some embodiments, the fusion unit is further configured to:

[0021] The first weighted feature and the domain label feature corresponding to each feature domain are fused to obtain the second fused feature corresponding to each feature domain.

[0022] The target object feature corresponding to each element object to be recommended is generated by fusing multiple second fusion features.

[0023] Optionally, in some embodiments, the fusion unit is further configured to:

[0024] Obtain the average preference value of each feature domain in the target recommendation scenario, and obtain the target preference value of the account object for each feature domain;

[0025] The feature preference weight of each feature domain is determined based on the average preference value and the target preference value.

[0026] Based on the multiple feature preference weights, feature fusion is performed on multiple second fusion features to obtain the target object features corresponding to each element object to be recommended.

[0027] Optionally, in some embodiments, the fusion unit is further configured to:

[0028] Each input feature is weighted based on the scene weights to obtain the weighted input features corresponding to each feature domain.

[0029] The weighted input features and the domain label features corresponding to each feature domain are fused to obtain the third fused feature corresponding to each feature domain.

[0030] The third fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

[0031] Optionally, in some embodiments, the weight generation unit is further configured to:

[0032] The scene features and the multiple input features are fused to obtain the scene fusion feature corresponding to each element object to be recommended.

[0033] The first gating unit based on the preset neural network model performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended.

[0034] Optionally, in some embodiments, the weight generation unit is further configured to:

[0035] The scene fusion features are subjected to linear activation processing to obtain scene activation features;

[0036] The scene activation features are subjected to nonlinear activation processing to obtain initial weights;

[0037] The initial weights are scaled based on preset scaling parameters to obtain the scene weights corresponding to each element object to be recommended.

[0038] Optionally, in some embodiments, the recommendation unit is further configured to:

[0039] The recommendation prediction sub-model based on the preset neural network model performs recommendation prediction on the features of the target object to obtain the initial prediction data of each element object to be recommended.

[0040] The initial prediction data is normalized to determine the recommendation prediction data for each of the elements to be recommended.

[0041] Optionally, in some embodiments, the recommendation prediction sub-model includes multiple second gating units and multiple neural network layers corresponding to the second gating units, and the multiple neural network layers are hierarchically connected.

[0042] The recommendation unit is also used for:

[0043] The target object features are simultaneously input into multiple second gating units to obtain the gating weight of each second gating unit;

[0044] The output features of multiple neural network layers are weighted based on multiple gating weights, and the weighted features are used as input features of the next neural network layer. The initial prediction data for each element object to be recommended is determined based on the features output by the last neural network layer.

[0045] Optionally, in some embodiments, the content recommendation apparatus provided in this disclosure further includes: a pooling unit, the pooling unit being used for:

[0046] The dimension of the target object's feature is determined to be greater than a first threshold;

[0047] The target object features are pooled so that the dimension of the pooled target object features is equal to the first threshold.

[0048] Optionally, in some embodiments, the recommendation unit is further configured to:

[0049] The multiple elements to be recommended are sorted according to the recommendation prediction data to obtain the sorting result;

[0050] Based on the ranking results, recommendations are made to the plurality of element objects to be recommended.

[0051] A third aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the content recommendation method as described in the first aspect.

[0052] This fourth aspect of the disclosure provides a computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the content recommendation method as described in the first aspect.

[0053] The fifth aspect of this disclosure provides a computer program product comprising a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the content recommendation method as described in the first aspect.

[0054] The content recommendation method provided in this disclosure includes: acquiring scene data of a target recommendation scenario in a target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain; further, extracting scene features from the scene data to obtain scene features, and extracting multiple input features from the multiple input data to obtain multiple input features, and generating scene weights corresponding to each element object to be recommended based on the scene features and the multiple input features; further, extracting multiple domain label features from the multiple domain label data to obtain multiple domain label features, and fusing the scene weights, multiple input features, and multiple domain label features to generate target object features corresponding to each element object to be recommended; further, determining recommendation prediction data for each element object to be recommended based on the target object features, and recommending multiple element objects to be recommended based on the recommendation prediction data.

[0055] In this embodiment of the disclosure, when performing content recommendation and extracting features from the elements to be recommended, not only is the content data of multiple feature domains corresponding to each element object extracted, but also the domain label data of each feature domain. This avoids the loss of inter-domain difference information in the element features during content recommendation. In other words, this embodiment of the disclosure, by introducing domain label features, deeply considers the impact of differences between feature domains on the recommendation scenario, which can improve the representation of differences between feature domains and thus improve the accuracy of recommended content.

[0056] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objectives and other advantages of this disclosure may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0057] The accompanying drawings are provided to further understand the technical solutions of this disclosure and constitute a part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure.

[0058] Figure 1 A system architecture diagram for the content recommendation method used in embodiments of this disclosure;

[0059] Figure 2 A flowchart illustrating the content recommendation method provided in this disclosure;

[0060] Figure 3 A schematic diagram of the structure of the first gating unit provided in an embodiment of this disclosure;

[0061] Figure 4 A flowchart illustrating the process of determining the target object features corresponding to the element object to be recommended, provided in an embodiment of this disclosure;

[0062] Figure 5A A schematic diagram of an algorithm structure for determining the features of a target object provided in an embodiment of this disclosure;

[0063] Figure 5B A schematic diagram of an algorithm structure for determining a second fusion feature provided in an embodiment of this disclosure;

[0064] Figure 6 A schematic diagram of another algorithm structure for determining the features of a target object provided in this embodiment of the disclosure;

[0065] Figure 7 A schematic diagram of an algorithm structure for determining a third fusion feature provided in an embodiment of this disclosure;

[0066] Figure 8 A schematic diagram of the model structure of the recommendation prediction sub-model provided in this embodiment of the disclosure;

[0067] Figure 9 A specific flowchart illustrating the content recommendation method provided in this embodiment of the disclosure;

[0068] Figure 10 A schematic diagram of the structure of the content recommendation device provided in the embodiments of this disclosure;

[0069] Figure 11 This is a terminal structure diagram for implementing various methods according to an embodiment of the present disclosure;

[0070] Figure 12 This is a server structure diagram illustrating the implementation of various methods according to an embodiment of the present disclosure. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this disclosure.

[0072] Before providing a further detailed description of the embodiments of this disclosure, the terms and concepts used in these embodiments are explained, and they are subject to the following interpretations:

[0073] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, 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 attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0074] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, pre-trained model technology, operating / interactive systems, and mechatronics. Among these, pre-trained models, also known as large-scale models or foundational models, can be widely applied to downstream tasks across various AI fields after fine-tuning. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0075] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and pre-trained learning. Pre-trained models are the latest development in deep learning, integrating all of these techniques.

[0076] Neural network models are computational models inspired by biological neural networks, used in machine learning and artificial intelligence. They consist of a hierarchical structure of multiple neurons, each connected to neurons in the next layer. These connections have weights, and through these weights and activation functions, the neural network can learn complex patterns and relationships in the input data. Neural network models are frequently used in tasks such as image recognition, natural language processing, and prediction.

[0077] Recommendation refers to a method where a system or platform proactively sends information, content, or services to users without requiring explicit requests or searches from them. Recommendation methods typically use users' interests, preferences, and historical behavior to target specific user data and provide personalized content.

[0078] Nonlinear activation function (Sigmoid function): also known as a logistic function or sigmoid function. The Sigmoid function maps an input value to an output value between 0 and 1, meaning the output value ranges from 0 to 1. When the input value approaches negative infinity, the output value approaches 0; when the input value approaches positive infinity, the output value approaches 1.

[0079] The normalization exponent (softmax function) is used to compress a K-dimensional vector containing arbitrary real numbers into another K-dimensional vector, such that each element is in the range (0,1) and the sum of all elements is 1.

[0080] Embedded vector: refers to the data form obtained by embedding the original low-dimensional data manifold into the high-dimensional space where the original data resides.

[0081] In the internet age, internet applications typically need to push content (such as short videos and articles) to platform users to meet their content needs. Content recommendation is a process that uses algorithms to analyze user preferences and other relevant information to predict content that users may be interested in and recommend that content to them. Content recommendation algorithms can recommend videos, articles, and other content that users might be interested in, thereby increasing user engagement with the application. In the field of content recommendation, element objects can represent the specific content recommended to the user, such as videos and articles. Context-based content recommendation refers to the process of recommending specific content to users in different application scenarios or user interaction environments. Since current recommendation systems may have multiple recommendation scenarios for each business, it is usually necessary to use customized development and optimization models for specific scenarios to adapt to the differences in data distribution and feature space under different scenarios.

[0082] Currently, in content recommendation systems, related technologies typically employ Embedding Personalized Networks (EPNETs) to design feature embedding layers for multiple scenarios. This is achieved by introducing Gate Neural Units (Gate NUs) within EPNETs to select personalized embedding layers. EPNETs, ​​by introducing a small number of parameters and achieving rapid convergence, enable the selection of personalized embedding layers and incorporate scene-specific features into the model. This allows for the adaptation and balancing of feature importance across multiple domains, thereby improving the performance of the recommendation system and enabling a single model to serve all scenarios (One Serve All).

[0083] However, while EPNET can achieve personalized embedding layer selection by introducing Gate NU, Gate NU applies the same weight to each original feature. This means that all vectors output from the entire embedding layer undergo the same weight adjustment, causing the generated embedding vectors to easily lose some inter-domain differentiation—that is, the differences between feature domains are lost. Therefore, when this embedding vector is applied to subsequent recommendation models, it can easily lead to inaccurate content recommendations for users.

[0084] To address the issue of inaccurate content recommendations for users in the aforementioned scenarios, this disclosure provides a content recommendation method aimed at improving the accuracy of recommended content.

[0085] System architecture and scenario description of the embodiments disclosed herein

[0086] Figure 1 This is a system architecture diagram for the content recommendation method applied according to embodiments of the present disclosure. It includes a terminal 110, an Internet 120, a gateway 130, a server 140, etc.

[0087] Terminal 110 includes various device forms such as desktop computers, laptops, PDAs (personal digital assistants), mobile phones, vehicle terminals, home theater terminals, dedicated terminals, intelligent voice interaction devices, smart home appliances, or aircraft. Furthermore, terminal 110 can be a single device or a collection of multiple devices. For example, multiple devices can be connected via a local area network, sharing a single display device to work collaboratively, collectively forming terminal 110. Terminal 110 can communicate with the Internet 120 via wired or wireless means to exchange data.

[0088] Server 140 refers to a computer system capable of providing certain services to terminal 110. Compared to ordinary terminal 110, server 140 has higher requirements in terms of stability, security, and performance. Server 140 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines). In this embodiment, server 140 specifically provides storage functionality; that is, server 140 can be a database node in a distributed database, and this disclosure includes a cluster of multiple servers 140 forming a distributed database.

[0089] Gateway 130, also known as an internetwork connector or protocol converter, implements network interconnection at the transport layer and acts as a translator. It is a computer system or device that translates between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateway 130 also provides filtering and security functions. Messages sent from terminal 110 to server 140 are processed and forwarded through gateway 130. Messages sent from server 140 to terminal 110 are also processed and forwarded through gateway 130. In this embodiment, terminal 110 sends a content recommendation request to server 140 through gateway 130, and server 140 returns a content recommendation result to terminal 110 via gateway 130.

[0090] The content recommendation method provided in this embodiment can be implemented in terminal 110, server 140, or partly in terminal 110 and partly in server 140.

[0091] When the content recommendation method provided in this embodiment is implemented in terminal 110, in response to the content recommendation task performed by the target object on terminal 110, terminal 110 can first obtain scene data of the target recommendation scene in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain; then terminal 110 can perform feature extraction on the scene data to obtain scene features, and perform feature extraction on the multiple input data to obtain multiple input features, and generate scene weights corresponding to each element object to be recommended based on the scene features and the multiple input features; further, terminal 110 can perform feature extraction on the multiple domain label data to obtain multiple domain label features, and fuse the scene weights, multiple input features, and multiple domain label features to generate target object features corresponding to each element object to be recommended; further, terminal 110 can determine the recommendation prediction data of each element object to be recommended based on the target object features, and recommend multiple element objects to be recommended based on the recommendation prediction data.

[0092] When the content recommendation method provided in this embodiment is implemented in server 140, in response to the content recommendation task performed by the target object on server 140, server 140 can first obtain scene data of the target recommendation scene in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain; then server 140 can perform feature extraction on scene data to obtain scene features, and perform feature extraction on multiple input data to obtain multiple input features, and generate scene weights corresponding to each element object to be recommended based on scene features and multiple input features; further, server 140 can perform feature extraction on multiple domain label data to obtain multiple domain label features, and fuse scene weights, multiple input features, and multiple domain label features to generate target object features corresponding to each element object to be recommended; further, server 140 can determine recommendation prediction data for each element object to be recommended based on target object features, and recommend multiple element objects to be recommended based on recommendation prediction data.

[0093] When the content recommendation method provided in this embodiment is partially implemented in terminal 110 and partially implemented in server 140, in response to the content recommendation task performed by the target object on terminal 110, terminal 110 can first obtain scene data of the target recommendation scene in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain; then terminal 110 can send the scene data, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain to server 140; furthermore, server 140 can perform feature extraction on the scene data to obtain scene features, and perform feature extraction on multiple element objects to obtain scene features. The input data undergoes feature extraction to obtain multiple input features, and scene weights are generated for each element object to be recommended based on scene features and multiple input features. Then, server 140 can extract features from multiple domain label data to obtain multiple domain label features, and generate target object features for each element object to be recommended based on scene weights, multiple input features, and multiple domain label features. Further, server 140 can determine recommendation prediction data for each element object to be recommended based on the target object features, and send the recommendation prediction data to terminal 110. Finally, terminal 110 can recommend multiple element objects to be recommended based on the recommendation prediction data.

[0094] The content recommendation method provided in this disclosure can be applied to content recommendation tasks in various Internet applications, specifically in scenarios such as content recommendation tasks in video push applications and content recommendation tasks in article push applications.

[0095] For example, when the content recommendation method provided in this embodiment is applied to the content recommendation task of a video push application, the following steps can be taken: First, scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain can be obtained. Then, scene features are extracted from the scene data to obtain scene features, and multiple input features are extracted from the multiple input data to obtain multiple input features. Scene weights corresponding to each element object to be recommended are generated based on the scene features and the multiple input features. Further, multiple domain label features are extracted from the multiple domain label data to obtain multiple domain label features, and target object features corresponding to each element object to be recommended are generated by fusing the scene weights, multiple input features, and multiple domain label features. Finally, recommendation prediction data for each element object to be recommended is determined based on the target object features, and multiple element objects to be recommended are recommended based on the recommendation prediction data. At this time, each element object to be recommended is equivalent to a video to be pushed in the video push application. This embodiment can perform video recommendation based on the determined recommendation prediction data of each element object to be recommended. In this way, the differential representation between feature domains can be improved, thereby improving the accuracy of recommended content.

[0096] The above examples do not limit the scope of protection in this case.

[0097] General Description of Embodiments in this Disclosure

[0098] According to one embodiment of this disclosure, a content recommendation method is provided. For example... Figure 2 The diagram shown is a flowchart illustrating a content recommendation method provided in this disclosure. This method can be executed by a computer device, specifically a terminal or a server. The content recommendation method may include:

[0099] Step 210: Obtain scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain.

[0100] As disclosed above, the content recommendation method provided in this embodiment can be applied to content recommendation tasks in various Internet applications, recommending more accurate content to account objects logged into the target application. The target application refers to the application platform for which content is recommended to the account object, and is not limited thereto. The account object refers to a user logged into the target application who wants to view the recommended content. Here, a recommended element object indicates a content object to be recommended in the target application, which can be a product, video, article, service, etc. A recommended element object can contain multiple feature domains, and each feature domain indicates the attribute features associated with the recommended element object, such as the attribute information of the account object, the element object attributes of the recommended element object (such as type, duration, popularity, author, etc.), etc. Therefore, the input data for each feature domain indicates the specific information of the corresponding feature domain, i.e., the specific information of the aforementioned attribute features. Domain label data is used to characterize the information for classifying and labeling the feature domain. For example, when the recommended element object is a recommended video, the corresponding feature domain can include video type (such as action, comedy, science fiction, etc.), video author information, user interest tags, etc.

[0101] Note that when obtaining information about multiple elements to be recommended from a target audience, the consent of each element must be obtained beforehand. Furthermore, the collection, use, and processing of information about multiple elements to be recommended will comply with relevant laws, regulations, and standards. When obtaining consent from an element, individual permission or consent can be obtained through pop-up windows or redirection to a confirmation page.

[0102] The target application can encompass multiple business scenarios, and the different scenarios in which the account object is located will all affect the final recommended content. The target recommendation scenario indicates the specific scenario in which the recommended element object will be presented, and can include business scenario information and the current scenario information of the account object. Scenario data refers to the specific data corresponding to the target recommendation scenario. Typically, account objects may view similar content in the same scenario. For example, in a work scenario, an account object may be more likely to view work-related push content; at home, it may be more likely to view push content related to its own preferences. Therefore, scenario data can be used for content prediction, which helps improve the effectiveness of push processing. Thus, scenario data can include geographical area data, current input day type data (weekday, weekend, or holiday), weather data, etc. Based on this, embodiments of this disclosure can determine scenario data by collecting the account object's behavior logs on the application platform, such as clicks, browsing, and dwell time, as well as scenario context information, such as time, location, and device type.

[0103] It should be noted that when obtaining the context data of the account object, the consent of the account object must be obtained in advance, similar to the above, so it will not be repeated.

[0104] In the above embodiments, this disclosure introduces scenario data to better understand the needs and preferences of account objects in specific scenarios. By introducing input data for multiple feature domains corresponding to each element object to be recommended, as well as domain label data for each feature domain, the impact of differences between feature domains on the recommendation scenario can be considered in depth, improving the representation of differences between feature domains and thus improving the accuracy of recommended content.

[0105] Step 220: Extract scene features from scene data and extract multiple input features from multiple input data, and generate scene weights for each element object to be recommended based on scene features and multiple input features.

[0106] Feature extraction refers to the process of extracting features from raw data that can effectively represent data characteristics and related information. Its purpose is to transform the raw data into a set of feature vectors to facilitate subsequent analysis, modeling, and decision-making. Scene features are features extracted from scene data and can be in vector form. Input features are features extracted from the corresponding input data and can also be in vector form. Specifically, in this embodiment, scene data and each input data point can be input into the feature embedding layer of a preset neural network model to obtain corresponding scene features and input features. Scene weights are used to characterize the degree of matching between the corresponding element to be recommended and the target recommendation scene, thereby providing a basis for recommendation ranking.

[0107] It should be noted that a vector is an array of values ​​in different dimensions, representing a point in a multidimensional space. The line segment between this point and the origin in the multidimensional coordinate system has magnitude and direction, which are the magnitude and direction of the vector. Each value in the vector is the point value projected onto the corresponding coordinate axis in the multidimensional coordinate system, i.e., a vector element. Vector elements can be numerical values ​​or symbols, etc. The vector elements in the feature vectors mentioned in this disclosure, such as the vectors of scene features, input features, and domain label features (described in subsequent embodiments, and will not be repeated here), can be numerical values ​​and are not limited thereto.

[0108] Taking scene features as an example, if the scene data is character-based, it is input into the embedding layer to obtain scene features in vector form. If the scene data is numerical, it is mapped to numerical values ​​to obtain scene data in vector form. The specific mapping process for input features, domain label features, and other features can be found in the section on scene features and will not be elaborated further.

[0109] In this embodiment, features can be vectorized according to the characteristics of different types of features, ensuring the accuracy of the feature vectorization results.

[0110] It should be noted that the preset neural network model refers to a model that processes scene data, input data of multiple feature domains corresponding to multiple elements to be recommended, and domain label data of each feature domain to obtain recommendation prediction data for each element to be recommended. The preset neural network model used in the embodiments of this disclosure can be a model built based on network structures such as Embedding Personalized Network (EPNET), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), and is not specifically limited here.

[0111] In some embodiments, the step of generating a scene weight corresponding to each element object to be recommended based on scene features and multiple input features in this disclosure may specifically include:

[0112] The scene features and multiple input features are fused to obtain the scene fusion features corresponding to each element object to be recommended.

[0113] The first gating unit, based on a pre-defined neural network model, performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended.

[0114] Feature fusion refers to combining features from different sources or types to improve model performance and accuracy. Feature concatenation is one method of feature fusion, which involves directly concatenating feature vectors from different sources to form a longer feature vector. For example, in neural networks, feature concatenation can stack multiple feature maps along the channel dimension. This concatenation method allows the network to learn features from different spatial locations and fuse them at the same level. Based on this, embodiments of this disclosure can first perform feature fusion on scene features and multiple input features to obtain scene fusion features corresponding to each element object to be recommended. The obtained scene fusion features are used to characterize the result of fusing scene features and multiple input features of each element object to be recommended. Further, feature processing can be performed on the scene fusion features based on the first gating unit of a preset neural network model to obtain scene weights corresponding to each element object to be recommended.

[0115] In some embodiments, the step of performing feature processing on scene fusion features based on the first gating unit of a preset neural network model to obtain the scene weight corresponding to each element object to be recommended in this disclosure may specifically include:

[0116] Linear activation processing is applied to the scene fusion features to obtain scene activation features;

[0117] The scene activation features are subjected to non-linear activation processing to obtain initial weights;

[0118] The initial weights are scaled based on preset scaling parameters to obtain the scene weights corresponding to each element object to be recommended.

[0119] The first gating unit in this embodiment may include a two-layer structure. The first layer is a linear activation layer built based on a rectified linear unit (ReLU), used for cross-features and various prior information, which helps alleviate the gradient vanishing problem and allows the network to be trained at a deeper level. The second layer is a non-linear activation layer built based on the sigmoid function, used to generate a gating score (i.e., the initial weights in this embodiment), which can convert the linear input into a non-linear output, i.e., output a value between 0 and 1, providing the neural network with the ability to model non-linearity. Further, this gating score can be multiplied by a hyperparameter (i.e., a preset scaling parameter in this embodiment) to amplify the signal, which can be denoted as γ, and γ can take values ​​such as 2, 3, etc., without limitation. For example, as... Figure 3 The diagram shown is a structural schematic of the first gating unit provided in this disclosure. The first gating unit 310 includes two neural layers: a ReLU layer and a Sigmoid layer. The ReLU layer can first perform linear activation processing on the scene fusion features to obtain scene activation features. The process of calculating the scene activation features corresponding to each element object to be recommended can be as shown in Formula 1:

[0120] h = Relu(xW1 + b1) (Formula 1)

[0121] In Formula 1, h refers to the scene activation feature corresponding to each element object to be recommended, x refers to the scene fusion feature corresponding to each input element object to be recommended, W1 refers to the weight matrix of the ReLU layer, which is used to multiply with the input scene fusion feature to perform a linear transformation; b1 refers to the bias vector of the ReLU layer, which is used to add the result after multiplying with the weight matrix W1 to adjust the output of the activation function.

[0122] Furthermore, the Sigmoid layer of the first gating unit can perform non-linear activation processing on the scene activation features to obtain initial weights. The process of calculating the initial weights corresponding to each element object to be recommended can be shown in Equation 2:

[0123] δ=γ*Sigmoid(hW2+b2),δ∈[0,γ] (Formula 2)

[0124] In Formula 2, δ refers to the initial weight corresponding to each element object to be recommended, h refers to the scene activation feature corresponding to each element object to be recommended, γ refers to the preset scaling parameter, W2 refers to the weight matrix of the Sigmoid layer, which is used to multiply with the input scene activation feature to perform a linear transformation; b2 refers to the bias vector of the Sigmoid layer, which is used to add the result after multiplying with the weight matrix W2 to adjust the output of the activation function.

[0125] Furthermore, such as Figure 3 As shown, when the preset scaling parameter is 2, the initial weights can be scaled based on the preset scaling parameter to obtain the scene weights corresponding to each element object to be recommended. Element-wise Product refers to the method of combining the transformed embedded feature vector with the original embedded feature vector through element-wise product to generate the final personalized embedding output.

[0126] In the above embodiments, this disclosure can perform embedding transformations on scene data and multiple input data respectively, and fuse the transformed scene features with the multiple input features to input the fused scene features into a first gating unit for personalized transformation. This allows for adaptation to the feature importance of different users in multiple domains. Furthermore, the first gating unit can use domain-specific personalized prior information (i.e., combined with scene information) as input to the Gate NU, outputting a personalized gating vector to perform personalized transformations of the original embedding, thereby improving the accuracy of recommended content.

[0127] Step 230: Extract features from multiple domain label data to obtain multiple domain label features, and generate target object features corresponding to each element object to be recommended by fusing scene weights, multiple input features, and multiple domain label features.

[0128] Among them, domain label features refer to the features extracted from domain label data, which can be in vector form. Target object features refer to the features that are the result of fusing scene weights, multiple input features, and multiple domain label features. This avoids the problem of related technologies using only scene weights and multiple input features to apply the same weights to different feature domains, thus losing the differences between feature domains. It also avoids the impact of this loss of feature domain differences on recommendation accuracy.

[0129] In some embodiments, such as Figure 4 The diagram illustrates a flowchart of a process for determining the target object features corresponding to an element object to be recommended, as provided in this disclosure. Specifically, the step in this embodiment of the disclosure, which involves fusing scene weights, multiple input features, and multiple domain label features to generate the target object features corresponding to each element object to be recommended, may include:

[0130] Step 410: Perform feature fusion on the input features and domain label features corresponding to each feature domain to obtain the first fused feature corresponding to each feature domain.

[0131] Here, the first fusion feature refers to the feature obtained by fusing the input feature and the domain label feature corresponding to each feature domain. Thus, each element object to be recommended can include multiple corresponding first fusion features, and the number of first fusion features is the same as the number of feature domains. This embodiment of the disclosure fuses the input feature and the domain label feature corresponding to each feature domain. The domain label feature in this case is equivalent to a shared embedding vector representing shared public knowledge within the domain, which can solve the problem of identical weighting coefficients for each feature domain when the model handles personalized weights for each feature domain.

[0132] Step 420: Weight each first fusion feature based on scene weights to obtain the first weighted feature corresponding to each feature domain.

[0133] Here, the first weighted feature refers to the feature vector obtained by weighting each first fusion feature based on scene weights. Thus, each element object to be recommended can include multiple corresponding first weighted features, and the number of first weighted features is the same as the number of feature domains. The scene weights can be applied to each first fusion feature using element-wise multiplication to generate a personalized embedding output. Therefore, the process of calculating the first weighted feature corresponding to each feature domain can be shown in Equation 3:

[0134]

[0135] In formula 3, h f1 It refers to the first weighted feature corresponding to each feature domain. This refers to the input features corresponding to each feature domain. This refers to the domain label features of each feature domain (i.e., feature slot id embedding before EPNet). δ refers to the first fused feature corresponding to each feature domain, and δ refers to the scene weight corresponding to each element object to be recommended. Wherein, when the preset neural network model is a model built based on EPNET, then... Used to characterize domain label features combined prior to EPNET. Thus, This approach can be responsible for sharing common knowledge within the feature domain. During the backpropagation of the neural network gradient, only the embedding of the corresponding value is updated for different values ​​in the feature domain. Since different values ​​in the same feature domain also belong to the domain label data, different feature values ​​share the same domain label feature. Therefore, during the gradient update phase, gradients are shared and updated jointly, extracting common knowledge within the feature domain. This modeling approach is also better suited for features containing cold start information, such as business label data (item ID).

[0136] It's important to note that domain label features essentially share a common embedding vector of the same feature domain type for different feature values ​​belonging to the same feature domain. For example, a feature domain might characterize the time an account user spends on a target application each time they open it, such as 10 minutes or 30 minutes. To ensure that the embedding vector corresponding to this feature domain includes not only the embedding representations corresponding to specific feature values ​​like 10 minutes and 30 minutes, but also a shared representation of the "time spent on the target application" slot itself, the feature corresponding to "10 minutes" would be: "10 minutes" = "10 minutes" embedding vector + "time spent on the target application" category feature. In this way, both "10 minutes" and "30 minutes" can share a common knowledge or common representation belonging to the "time spent on the target application" slot (i.e., the feature domain).

[0137] Step 430: Generate the target object feature corresponding to each element object to be recommended by fusing multiple first weighted features.

[0138] Here, the target object feature refers to the feature vector generated by fusing the first weighted features corresponding to multiple feature domains. Thus, each element object to be recommended can correspond to one target object feature, and the number of target object features is the same as the number of element objects to be recommended. In some embodiments, such as... Figure 5A As shown, in this embodiment of the present disclosure, feature fusion can be performed on the input feature 510 and the domain label feature 520 corresponding to each feature domain to obtain the first fused feature corresponding to each feature domain. Then, based on the scene weight output by the first gating unit, each first fused feature is weighted to obtain the first weighted feature 530 corresponding to each feature domain. Afterwards, feature fusion is directly performed on multiple first weighted features 530 (e.g., ...). Figure 5A The n features (where n is the number of feature domains corresponding to each element object to be recommended, and is a positive integer) are used to generate 540 target object features for each element object to be recommended. Figure 5A The dashed line with an arrow in the middle indicates "FF W / O BP", which can be interpreted as "freight forwarder that does not use the backpropagation algorithm". Figure 5AThe solid line with an arrow indicates that "FF W / BP" can be interpreted as "freight forwarder with backpropagation." Here, "freight forwarder" refers to a stage in the data transmission or processing process that is given the ability to backpropagate in order to optimize or adjust its performance.

[0139] In steps 410 to 430 above, the embodiments of this disclosure can generate target object features corresponding to each element object to be recommended by introducing domain label data of each feature domain. This can deeply consider the impact of differences between feature domains on the recommendation scenario, improve the differential representation between feature domains, and thus improve the accuracy of recommended content.

[0140] In other embodiments, the step of generating the target object feature corresponding to each element object to be recommended based on the fusion of multiple first weighted features in this disclosure may specifically include:

[0141] For each feature domain, the first weighted feature and the domain label feature are fused to obtain the second fused feature for each feature domain.

[0142] The target object features corresponding to each element object to be recommended are generated by fusing multiple second fusion features.

[0143] Among them, such as Figure 5B As shown, in this embodiment of the present disclosure, the first weighted feature 530 and the domain label feature 520 corresponding to each feature domain can be fused to obtain the second fused feature corresponding to the feature domain. Therefore, the process of calculating the second fused feature corresponding to each feature domain can be as shown in Formula 4:

[0144]

[0145] In formula 4, h f2 This refers to the second fusion feature corresponding to each feature domain. This refers to the input features corresponding to each feature domain. This refers to the domain label features of each feature domain. It also refers to the domain label features of each feature domain (i.e., feature slot id embedding after EPNet), and and The difference lies in the location of their application; δ refers to the scene weight corresponding to each element object to be recommended. Specifically, when the preset neural network model is a model built on EPNET, then... The domain label features used to characterize the combined features after EPNET enable differentiated processing of <feature domain-scene domain> pairs. Although the same scene domain receives the same scene weight across different feature domains, the introduction of domain label features adds differentiated representations of different <feature domain-scene domain> pairs while preserving the original model's personalized weights for each scene domain. Thus, this embodiment avoids the problem in related technologies where different feature domains are weighted using the same weights based solely on scene weights and the fusion of multiple input features, resulting in the loss of differences between feature domains. This also avoids the impact of this loss of feature domain differences on recommendation accuracy.

[0146] Furthermore, such as Figure 6 As shown, after obtaining the second fusion features corresponding to multiple feature domains, these features can be input into the recommendation prediction sub-model of a preset neural network model for prediction. Alternatively, this embodiment can perform feature fusion on the obtained second fusion features corresponding to multiple feature domains to obtain target object features 540 for each element to be recommended, and then input the target object features 540 into the recommendation prediction sub-model of the preset neural network model for prediction. Thus, this embodiment can introduce domain label features within the same feature domain to achieve shared common knowledge within the domain, while also relying on this shared embedding vector to achieve differentiation between different feature domains.

[0147] In other embodiments, the step of generating target object features corresponding to each element object to be recommended based on the fusion of multiple second fusion features in this disclosure may specifically include:

[0148] Obtain the average preference value of each feature domain in the target recommendation scenario, and obtain the target preference value of the account object for each feature domain;

[0149] The feature preference weights for each feature domain are determined based on the average preference value and the target preference value.

[0150] Based on multiple feature preference weights, multiple second fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

[0151] The average preference score refers to the average degree of correlation between the push preferences and push tendencies of multiple application users in the target application and the feature domain in the target recommendation scenario. The average preference score can be obtained by analyzing the behavioral data, statistical data, or other information of multiple application users. The average preference score can prevent the problem of insufficient content diversity in content recommendation processing, which only recommends content of the most interest to account users. The target preference score refers to the degree of correlation between the push preferences and push tendencies of the target account user in the target application and the feature domain in the target recommendation scenario. For example, if an account user frequently watches science fiction movies, then the science fiction feature domain will have a high target preference score.

[0152] Note that when obtaining statistical data or other information from application users, the consent of each application user must be obtained in advance. Furthermore, the collection, use, and processing of these user characteristics will comply with relevant laws, regulations, and standards. When obtaining consent from application users, individual permission or consent can be obtained through pop-up windows or redirection to confirmation pages.

[0153] Furthermore, based on the obtained average preference value and target preference value, the feature preference weight of the account object for each feature domain can be determined. Feature preference weight refers to the weight value determined by considering the account object's personalized preference for the feature domain and avoiding insufficient diversity in the distributed content. Thus, if the account object's target preference value for a feature domain is significantly higher than the average preference value, then the feature preference weight corresponding to that feature domain should be higher.

[0154] It should be noted that the step of determining the feature preference weight for each feature domain based on the average preference value and the target preference value may specifically include: determining a first score based on the average preference value, and determining a second score based on the target preference value; calculating a total score based on the first and second scores, and determining the feature preference weight based on the total score. The first and second scores can be determined using a pre-defined table of correspondence between preference values ​​and scores, and the two tables can be the same or different, without limitation. Then, the feature preference weight corresponding to the feature domain can be determined based on the table of correspondence between the total score and preference weights. Furthermore, this embodiment can use the average or weighted average to determine the total score, without limitation. Therefore, this embodiment can flexibly set different weights for the average preference value and the target preference value according to the needs of actual applications, improving the flexibility of determining feature preference weights and increasing processing efficiency.

[0155] Furthermore, in this embodiment of the present disclosure, the second fusion feature corresponding to each feature domain can be weighted and calculated based on the feature preference weight of each feature domain, and the weighted features of multiple feature domains can be concatenated to obtain the target object feature corresponding to each element object to be recommended.

[0156] In other embodiments, the step of generating the target object feature corresponding to each element object to be recommended based on scene weights, multiple input features, and multiple domain label features in this disclosure may further include:

[0157] Each input feature is weighted based on the scene weight to obtain the weighted input feature corresponding to each feature domain.

[0158] For each feature domain, the weighted input features and domain label features are fused to obtain the third fused feature for each feature domain.

[0159] Multiple third-party fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

[0160] Among them, such as Figure 7 As shown, since the improvement of this embodiment of the present disclosure lies in considering the domain label feature 520, each input feature 510 can also be weighted based on the scene weights output by the first gating unit to obtain the weighted input feature 710 corresponding to each feature domain. Then, the weighted input feature 710 and the domain label feature 520 corresponding to each feature domain are fused to obtain the third fused feature 720 corresponding to each feature domain. Further, the obtained third fused features 720 corresponding to multiple feature domains can be input into the recommendation prediction sub-model of the preset neural network model for prediction.

[0161] In other embodiments, the present disclosure embodiments may also perform feature fusion on the third fusion features corresponding to the multiple feature domains obtained to obtain the target object features corresponding to each element object to be recommended, and then input the target object features into the recommendation prediction sub-model of the preset neural network model for prediction.

[0162] In the above embodiments, this disclosure can employ different methods to fuse scene weights, multiple input features, and multiple domain label features, ensuring that the feature change ratios for each feature domain are not identical but varied. This avoids the problem in related technologies where different feature domains are weighted using the same weights based solely on the fusion of scene weights and multiple input features, thus losing the differences between feature domains and preventing the impact of this loss of difference on recommendation accuracy. Therefore, this disclosure can enhance the differentiated representation between feature domains, thereby improving the accuracy of recommended content.

[0163] In some embodiments, after step S230, the content recommendation method provided in this disclosure may further include:

[0164] The dimension of the target object's features is determined to be greater than the first threshold;

[0165] The target object features are pooled so that the dimension of the pooled target object features is equal to the first threshold.

[0166] Here, the first threshold refers to the numerical value used to determine whether the target object feature needs pooling based on its dimension. The dimension of the target object feature refers to the number of elements in the corresponding feature vector. If the dimension of the target object feature is greater than the first threshold, pooling is required; if the dimension is less than or equal to the first threshold, pooling is not required. For example, assuming the recommended element object corresponding to the target object feature is video A, after processing the scene data, input data of multiple feature domains, and domain label data of each feature domain corresponding to video A, the target object feature is obtained as "[70,90,80,10,40,50,30,20,80,90,100,50]", with a dimension of 12. If the first threshold is set to 8, then the target object feature with a dimension of 12 is determined to be a target object feature with a dimension greater than the first threshold.

[0167] Alternatively, the first threshold can also refer to a standard value for the dimension of the target object feature. In this embodiment, target object features with dimensions greater than the first threshold can be pooled to make the dimension of the pooled target object feature equal to the first threshold. For example, if the first threshold is preset to 8, it can be determined that the dimension of a target object feature with dimension 12 is greater than the first threshold. Thus, the target object feature with dimension 12 can be pooled into a new target object feature with dimension 8.

[0168] It should be noted that pooling is a dimensionality reduction technique that aggregates feature values ​​within a feature map (e.g., taking the maximum or average value) by sliding a pooling kernel (also called a window) across the map, thereby generating a new feature map. The size and type of the pooling kernel determine the degree of dimensionality reduction and the amount of feature information preserved. Pooling can reduce the size of the feature map and decrease computational cost. Furthermore, pooling techniques can include average pooling and max pooling, among others.

[0169] In the above embodiments, the present disclosure can use pooling operations to adjust the dimensions of the target object features, so that the target object features corresponding to each element object to be recommended have the same dimensions, effectively reducing the amount of computation, while increasing the robustness of the features, preventing overfitting, and thus improving the efficiency of content recommendation.

[0170] Step 240: Determine the recommendation prediction data for each element object to be recommended based on the characteristics of the target object, and make recommendations for multiple element objects to be recommended based on the recommendation prediction data.

[0171] The recommendation prediction data represents the probability of recommending a corresponding element to an account in the current target recommendation scenario. This data also reflects the account's preference for the element (e.g., a recommended video) in the target recommendation scenario. Furthermore, multiple elements can be recommended based on the recommendation prediction data for each element, and the recommendation results can be pushed to the account's display interface.

[0172] In some embodiments of this disclosure, the step of determining the recommendation prediction data for each element object to be recommended based on the characteristics of the target object may specifically include:

[0173] The recommendation prediction sub-model based on the preset neural network model performs recommendation prediction on the features of the target object to obtain the initial prediction data of each element object to be recommended.

[0174] The initial prediction data is normalized to determine the recommended prediction data for each element object to be recommended.

[0175] The recommendation prediction sub-model refers to a model used to predict the degree to which a corresponding element object to be recommended will be recommended to an account object in the current target recommendation scenario. The recommendation prediction sub-model in this embodiment can employ a multilayer perceptron (MLP), a deep neural network tower (DNN tower), etc., without limitation. To map the initial prediction data of each element object to be recommended to the same numerical range, the initial prediction data can be normalized to ensure that the resulting recommendation prediction data falls within the range of 0 to 1, making different recommendation prediction data comparable and thus improving the accuracy of content recommendation.

[0176] In some embodiments, the step of performing recommendation prediction based on a recommendation prediction sub-model of a preset neural network model to obtain initial prediction data for each element object to be recommended in this disclosure may specifically include:

[0177] The target object features are simultaneously input into multiple second gating units to obtain the gating weight of each second gating unit;

[0178] The output features of multiple neural network layers are weighted based on multiple gating weights. The weighted features are then used as input features for the next neural network layer. The initial prediction data for each element to be recommended is determined based on the features output by the last neural network layer.

[0179] Specifically, when the recommendation prediction sub-model is constructed based on a DNN Tower structure, it can include multiple second gating units and multiple neural network layers corresponding to the second gating units, with hierarchical connections between the multiple neural network layers. For example... Figure 8 The diagram shown illustrates the model structure of the recommendation prediction sub-model provided in this embodiment. Based on this, this embodiment can simultaneously input the target object features 540 into multiple second gating units (such as...). Figure 8 The second gating units 1 to L (where L represents the number of second gating units) are used to obtain the gating weights output by each second gating unit. Then, the output features of the corresponding neural network layer are weighted based on each gating weight, and the weighted features are used as the input features of the next neural network layer. The initial prediction data for each element object to be recommended is determined based on the features output by the last neural network layer.

[0180] In some embodiments of this disclosure, the step of recommending multiple element objects to be recommended based on recommendation prediction data may specifically include:

[0181] The multiple elements to be recommended are sorted based on the recommendation prediction data to obtain the sorting results;

[0182] Recommendations are made based on the ranking results for multiple element objects to be recommended.

[0183] After obtaining the recommendation prediction data for each element to be recommended, multiple element objects can be sorted in descending order based on the recommendation prediction data to obtain the sorting result. Then, if there is no limit to the number of element objects to be recommended to be pushed to an account, the content can be pushed to the account in descending order of recommendation prediction data based on the sorting result. This means pushing target content to the target audience based on the weighted average probability value from highest to lowest.

[0184] This disclosure employs a method that fuses scene weights, multiple input features, and multiple domain label features to generate target object features for each element to be recommended. This avoids the problem in related technologies that rely solely on scene weights and the fusion of multiple input features, resulting in different feature domains being weighted with the same weights and losing the differences between feature domains. This also avoids the impact of this loss of feature domain differences on recommendation accuracy. Therefore, by introducing domain label features to deeply consider the impact of differences between feature domains in the recommendation scenario, this disclosure can improve the representation of differences between feature domains, thereby enhancing the accuracy of recommended content.

[0185] This disclosure provides a detailed description of embodiments in conjunction with specific application scenarios.

[0186] like Figure 9 The diagram shown is another flowchart illustrating the content recommendation method provided in this disclosure. This embodiment will describe each step of the content recommendation method in detail, focusing on the executing entity of each step. The method may specifically include the following steps:

[0187] Step 901: The computer device acquires scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain.

[0188] In this embodiment, each feature domain indicates the attribute features associated with the element object to be recommended, such as the attribute information of the account object, or the element object attributes of the element object to be recommended (e.g., type, duration, popularity, author, etc.). The input data of the feature domain is equivalent to a specific data. The domain label data of the feature domain is equivalent to a "category" data, which can be denoted as a slot id, used to explain the meaning of the feature domain itself of the corresponding type. The domain label data introduced in this embodiment is equivalent to a shared embedding vector of shared public knowledge within the domain, which can solve the problem that the weighting coefficients of each feature domain are exactly the same when the model processes the personalized weights of the feature domains.

[0189] Step 902: The computer device extracts features from scene data based on the feature embedding layer of a preset neural network model to obtain scene features, extracts features from multiple input data to obtain multiple input features, and extracts features from multiple domain label data to obtain multiple domain label features.

[0190] Scene features refer to features extracted from scene data, which can be in vector form. Input features refer to features extracted from the corresponding input data, which can also be in vector form. The process of feature extraction for the input data of each feature domain by the feature embedding layer can be seen in Equation 5 below, and the process of feature extraction for the domain label data (i.e., slot id) of each feature domain can be seen in Equation 6 below.

[0191] e f =embedding(v f ) (Formula 5)

[0192] In formula 5, v f This refers to the input data of the feature domain, e f This refers to the input features of the feature domain. Specifically, when extracting features from scene data, the input data can be replaced with scene data to obtain the corresponding scene features.

[0193]

[0194] In formula 6, id f This refers to the domain label data of the feature domain. This refers to the domain label features of the feature domain, and It can be from formula 4 above and

[0195] Step 903: The computer device performs feature fusion on scene features and multiple input features to obtain scene fusion features corresponding to each element object to be recommended, and performs feature processing on the scene fusion features based on the first gating unit of the preset neural network model to obtain the scene weight corresponding to each element object to be recommended.

[0196] In this embodiment, the first gating unit can be a structure built on Gate NU, used to inject personalized prior information (i.e., multiple input features corresponding to the element object to be recommended) into a preset neural network model. Thus, this embodiment can utilize... Figure 3 The first gating unit shown performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended.

[0197] Step 904: The computer device performs feature fusion on the input features and domain label features corresponding to each feature domain to obtain the first fused feature corresponding to each feature domain, and performs weighted processing on each first fused feature based on the scene weight to obtain the first weighted feature corresponding to each feature domain.

[0198] Among them, such as Figure 5AAs shown, this embodiment of the disclosure can introduce domain label features, which is equivalent to introducing shared common knowledge within each feature domain. Thus, during the gradient update phase, when gradients are shared and updated jointly, the common knowledge within the feature domain can be extracted. This allows for a deeper consideration of the impact of differences between feature domains on recommendation scenarios, improving the representation of differences between feature domains and thereby enhancing the accuracy of recommended content.

[0199] Step 905: The computer device performs feature fusion on the first weighted feature and the domain label feature corresponding to each feature domain to obtain the second fused feature corresponding to each feature domain, and generates the target object feature corresponding to each element object to be recommended based on the fusion of multiple second fused features.

[0200] This embodiment of the invention not only introduces domain label data within the feature domain, but also considers the domain label data again when fusing features from multiple feature domains, enabling differentiated processing of <feature domain-scene domain> pairs. By introducing domain label features, this embodiment can add differentiated representations of different <feature domain-scene domain> pairs while retaining the original model's personalized weights for each field. Thus, this embodiment avoids the problem in related technologies that rely solely on scene weights and the fusion of multiple input features, resulting in different feature domains being weighted with the same weights, thereby losing the differences between feature domains and preventing the impact of this loss of feature domain differences on recommendation accuracy.

[0201] In the feature vector embedding stage, this disclosure proposes three vectors: a feature value embedding vector (i.e., the input feature), and two feature domain slot ID embedding vectors (i.e., the feature domain slot ID embedding vectors). and These three vectors are structurally used at different levels to respectively share common knowledge within the feature domain and differentiate the <feature domain-scene domain> pairs between scene domains.

[0202] Step 906: The computer device performs recommendation prediction on the target object features based on the recommendation prediction sub-model of the preset neural network model, and determines the recommendation prediction data for each element object to be recommended.

[0203] The recommendation prediction data is used to characterize the probability of recommending a corresponding element to an account in the current target recommendation scenario. This data can also reflect the account's preference for the element to be recommended (such as a recommended video) in the target recommendation scenario. For example, in a video recommendation scenario, the elements to be recommended include video V1 and video V2. After processing in steps 901 to 906, it can be determined that the recommendation prediction probability for video V1 is 85%, and the recommendation prediction probability for video V2 is 90%. Therefore, it can be indicated that the probability of recommending video V2 to the account is relatively high in the target recommendation scenario.

[0204] Step 907: The computer device sorts the multiple element objects to be recommended according to the recommendation prediction data, obtains the sorting results, and makes recommendations to the multiple element objects to be recommended based on the sorting results.

[0205] After obtaining the recommendation prediction data for each element to be recommended, multiple elements can be sorted in descending order based on the prediction data to obtain a ranking result. If there is no limit to the number of elements to be recommended that are pushed to an account, the content can be pushed to the account in descending order of the prediction data, based on the ranking result. This means pushing target content to the target audience based on the weighted average probability value from highest to lowest.

[0206] It should be noted that when comparing the content recommendation methods in related technologies with the content recommendation methods provided in this disclosure in video recommendation scenarios, this disclosure can conduct experiments on both the strong-target area and the differentiated processing area of ​​the non-targeted <feature domain-scene domain> pair on a personal computer (PC) of a target application. Table 1 below shows a comparison table illustrating the effects of related technologies and this disclosure in the focus image index, and the related technologies can use the BASE fine-ranking model (i.e., the control group). The first column in Table 1 indicates the specific items being compared, such as "Focus Image_Main Player_Average Views of Full-Length Videos_Featured Page", "Focus Image_Main Player_Average Playback Duration (Seconds)_Featured Page", etc. Furthermore, the average number of full-length video views per user (video views, vv) is used to measure the average number of videos watched by users on the featured page via the focus image and main player. The average full-length video playback duration per user measures the average duration of video viewing by users on the featured page via the focus image and main player, and can be expressed in seconds, minutes, or other time units. The second column in Table 1 represents the experimental results of the related technologies (i.e., the control group) in the corresponding matters. The third column in Table 1 represents the experimental results of the content recommendation method (i.e., the experimental group) of the present disclosure embodiments in the corresponding matters. The fourth column in Table 1 represents the percentage change between the experimental results of the present disclosure embodiments and the experimental results of the related technologies. That is, the percentage change can be obtained by subtracting the experimental results of the related technologies from the experimental results of the present disclosure embodiments and dividing the difference by the experimental results of the related technologies. "+" indicates a percentage increase, meaning that the method used in the present disclosure embodiments is more effective. Thus, the percentage change in specific data between the related technologies and the present disclosure embodiments can be characterized under the guidance of the focus image and the main player. Therefore, compared with the related technologies, the present disclosure embodiments show a significant increase of approximately 1.12% in the "Focus Image_Main Player_Average Full Movie VV_Featured Page" matter, and an increase of approximately +1.14% in the "Focus Image_Main Player_Average Full Movie Playback Duration (Seconds)_Featured Page" matter. Although there was an increase, this increase was not statistically significant, so the p-value was approximately 0.07, slightly higher than the commonly used significance level threshold of 0.05, meaning that there was not enough evidence to reject the null hypothesis (i.e., that the increase was due to random factors rather than design changes).

[0207]

[0208] Table 1

[0209] It should be noted that when comparing performance in video recommendation scenarios, this embodiment can also compare performance under metrics such as "following series" and "featured page" metrics. These experimental results demonstrate that when extracting features from elements to be recommended, this embodiment not only extracts the content data of multiple feature domains corresponding to each element object, but also extracts the domain label data of each feature domain. This avoids the loss of inter-domain difference information in element features during content recommendation. In other words, this embodiment introduces domain label features to deeply consider the impact of differences between feature domains on the recommendation scenario, which can improve the representation of differences between feature domains, thereby improving the accuracy of recommended content.

[0210] Description of apparatus and devices according to embodiments of this disclosure

[0211] It is understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this embodiment, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0212] Figure 10 A schematic diagram of the structure of a content recommendation device 1000 provided in an embodiment of this disclosure. The content recommendation device 1000 includes:

[0213] The acquisition unit 1001 is used to acquire scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain.

[0214] The weight generation unit 1002 is used to extract features from scene data to obtain scene features, and to extract features from multiple input data to obtain multiple input features, and to generate scene weights corresponding to each element object to be recommended based on scene features and multiple input features.

[0215] The fusion unit 1003 is used to extract features from multiple domain label data to obtain multiple domain label features, and to fuse them according to scene weights, multiple input features and multiple domain label features to generate target object features corresponding to each element object to be recommended.

[0216] The recommendation unit 1004 is used to determine the recommendation prediction data for each element object to be recommended based on the characteristics of the target object, and to recommend multiple element objects to be recommended based on the recommendation prediction data.

[0217] Optionally, in some embodiments, the fusion unit 1003 is further configured to:

[0218] For each feature domain, the input features and domain label features are fused to obtain the first fused feature for each feature domain.

[0219] Each first fusion feature is weighted based on the scene weight to obtain the first weighted feature corresponding to each feature domain.

[0220] The target object features for each element to be recommended are generated by fusing multiple first-weighted features.

[0221] Optionally, in some embodiments, the fusion unit 1003 is further configured to:

[0222] For each feature domain, the first weighted feature and the domain label feature are fused to obtain the second fused feature for each feature domain.

[0223] The target object features corresponding to each element object to be recommended are generated by fusing multiple second fusion features.

[0224] Optionally, in some embodiments, the fusion unit 1003 is further configured to:

[0225] Obtain the average preference value of each feature domain in the target recommendation scenario, and obtain the target preference value of the account object for each feature domain;

[0226] The feature preference weights for each feature domain are determined based on the average preference value and the target preference value.

[0227] Based on multiple feature preference weights, multiple second fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

[0228] Optionally, in some embodiments, the fusion unit 1003 is further configured to:

[0229] Each input feature is weighted based on the scene weight to obtain the weighted input feature corresponding to each feature domain.

[0230] For each feature domain, the weighted input features and domain label features are fused to obtain the third fused feature for each feature domain.

[0231] Multiple third-party fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

[0232] Optionally, in some embodiments, the weight generation unit 1002 is further configured to:

[0233] The scene features and multiple input features are fused to obtain the scene fusion features corresponding to each element object to be recommended.

[0234] The first gating unit, based on a pre-defined neural network model, performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended.

[0235] Optionally, in some embodiments, the weight generation unit 1002 is further configured to:

[0236] Linear activation processing is applied to the scene fusion features to obtain scene activation features;

[0237] The scene activation features are subjected to non-linear activation processing to obtain initial weights;

[0238] The initial weights are scaled based on preset scaling parameters to obtain the scene weights corresponding to each element object to be recommended.

[0239] Optionally, in some embodiments, the recommendation unit 1004 is further configured to:

[0240] The recommendation prediction sub-model based on the preset neural network model performs recommendation prediction on the features of the target object to obtain the initial prediction data of each element object to be recommended.

[0241] The initial prediction data is normalized to determine the recommended prediction data for each element object to be recommended.

[0242] Optionally, in some embodiments, the recommended prediction sub-model includes multiple second gating units and multiple neural network layers corresponding to the second gating units, and the multiple neural network layers are hierarchically connected.

[0243] Recommended unit 1004 is also used for:

[0244] The target object features are simultaneously input into multiple second gating units to obtain the gating weight of each second gating unit;

[0245] The output features of multiple neural network layers are weighted based on multiple gating weights. The weighted features are then used as input features for the next neural network layer. The initial prediction data for each element to be recommended is determined based on the features output by the last neural network layer.

[0246] Optionally, in some embodiments, the content recommendation apparatus provided in this disclosure further includes: a pooling unit (not shown), the pooling unit being used for:

[0247] The dimension of the target object's features is determined to be greater than the first threshold;

[0248] The target object features are pooled so that the dimension of the pooled target object features is equal to the first threshold.

[0249] Optionally, in some embodiments, the recommendation unit 1004 is further configured to:

[0250] The multiple elements to be recommended are sorted based on the recommendation prediction data to obtain the sorting results;

[0251] Recommendations are made based on the ranking results for multiple element objects to be recommended.

[0252] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0253] Reference Figure 11 , Figure 11 To implement the content recommendation method of this embodiment, the terminal 110 includes the following components: a radio frequency (RF) circuit 1110, a memory 1115, an input unit 1130, a display unit 1140, a sensor 1150, an audio circuit 1160, a wireless fidelity (WiFi) module 1170, a processor 1180, and a power supply 1190. Those skilled in the art will understand that... Figure 11 The terminal 110 structure shown does not constitute a limitation on a mobile phone or computer, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0254] The RF circuit 1110 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 1180; in addition, it transmits uplink data to the base station.

[0255] The memory 1115 can be used to store software programs and modules. The processor 1180 executes various terminal functions and data synchronization by running the software programs and modules stored in the memory 1115.

[0256] The input unit 1130 can be used to receive input numeric or character information, and to generate key signal inputs related to the settings and function control of the terminal. Specifically, the input unit 1130 may include a touch panel 1131 and other input devices 1132.

[0257] Display unit 1140 can be used to display input or provided information, as well as various menus of the terminal. Display unit 1140 may include display panel 1141.

[0258] Audio circuitry 1160, speaker 1161, and microphone 1162 provide an audio interface.

[0259] In this embodiment, the processor 1180 included in the terminal 110 can execute the content recommendation method of the previous embodiment.

[0260] The terminal 110 in this disclosure includes, but is not limited to, mobile phones, computers, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc.

[0261] Figure 12 This is a partial structural block diagram of a server 140 for implementing the content recommendation method of this disclosure. The server 140 can vary significantly due to different configurations or performance, and may include one or more central processing units (CPUs) 1222 (e.g., one or more processors) and memory 1232, and one or more storage media 1230 (e.g., one or more mass storage devices) for storing application programs 1242 or data 1244. The memory 1232 and storage media 1230 may be temporary or persistent storage. The program stored in the storage media 1230 may include one or more modules (not shown in the figure), each module including a series of instruction operations on the server 140. Furthermore, the CPU 1222 may be configured to communicate with the storage media 1230 and execute the series of instruction operations in the storage media 1230 on the server 140.

[0262] Server 140 may also include one or more power supplies 1226, one or more wired or wireless network interfaces 1250, one or more input / output interfaces 1258, and / or one or more operating systems 1241, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0263] The central processing unit 1222 in server 140 can be used to execute the content recommendation method of embodiments of this disclosure.

[0264] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the content recommendation methods of the foregoing embodiments.

[0265] This disclosure also provides a computer program product comprising a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the method described above.

[0266] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this disclosure and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “including,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0267] It should be understood that in this disclosure, "at least one item" means one or more, and "more than one" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0268] It should be understood that in the description of the embodiments of this disclosure, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0269] In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0270] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0271] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0272] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned computer-readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0273] It should also be understood that the various implementation methods provided in this disclosure can be combined arbitrarily to achieve different technical effects.

[0274] The above is a detailed description of the embodiments of this disclosure. However, this disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this disclosure. All such equivalent modifications or substitutions are included within the scope defined by the claims of this disclosure.

Claims

1. A content recommendation method, characterized in that, The method includes: Acquire scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain; Scene features are obtained by extracting features from the scene data, and multiple input features are obtained by extracting features from multiple input data. Scene weights are generated for each element object to be recommended based on the scene features and the multiple input features. Multiple domain label features are obtained by extracting features from multiple domain label data, and target object features corresponding to each recommended element object are generated by fusing the scene weight, the multiple input features and the multiple domain label features. Based on the characteristics of the target object, the recommendation prediction data for each of the proposed element objects is determined, and recommendations are made to the plurality of proposed element objects based on the recommendation prediction data.

2. The method according to claim 1, characterized in that, The step of fusing the scene weights, the multiple input features, and the multiple domain label features to generate the target object features corresponding to each element object to be recommended includes: The input features and the domain label features corresponding to each feature domain are fused to obtain the first fused feature corresponding to each feature domain. Based on the scene weights, each of the first fusion features is weighted to obtain the first weighted feature corresponding to each feature domain. The target object feature corresponding to each element object to be recommended is generated by fusing multiple first weighted features.

3. The method according to claim 2, characterized in that, The step of generating the target object feature corresponding to each element object to be recommended based on the fusion of multiple first weighted features includes: The first weighted feature and the domain label feature corresponding to each feature domain are fused to obtain the second fused feature corresponding to each feature domain. The target object feature corresponding to each element object to be recommended is generated by fusing multiple second fusion features.

4. The method according to claim 3, characterized in that, The step of generating the target object feature corresponding to each element object to be recommended based on multiple second fusion features includes: Obtain the average preference value of each feature domain in the target recommendation scenario, and obtain the target preference value of the account object for each feature domain; The feature preference weight of each feature domain is determined based on the average preference value and the target preference value. Based on the multiple feature preference weights, feature fusion is performed on multiple second fusion features to obtain the target object features corresponding to each element object to be recommended.

5. The method according to claim 1, characterized in that, The step of fusing the scene weights, the multiple input features, and the multiple domain label features to generate the target object features corresponding to each element object to be recommended further includes: Each input feature is weighted based on the scene weights to obtain the weighted input features corresponding to each feature domain. The weighted input features and the domain label features corresponding to each feature domain are fused to obtain the third fused feature corresponding to each feature domain. The third fusion features are fused to obtain the target object features corresponding to each element object to be recommended.

6. The method according to claim 1, characterized in that, The step of generating the scene weight corresponding to each element object to be recommended based on the scene features and the multiple input features includes: The scene features and the multiple input features are fused to obtain the scene fusion feature corresponding to each element object to be recommended. The first gating unit based on the preset neural network model performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended.

7. The method according to claim 6, characterized in that, The first gating unit based on the preset neural network model performs feature processing on the scene fusion features to obtain the scene weight corresponding to each element object to be recommended, including: The scene fusion features are subjected to linear activation processing to obtain scene activation features; The scene activation features are subjected to nonlinear activation processing to obtain initial weights; The initial weights are scaled based on preset scaling parameters to obtain the scene weights corresponding to each element object to be recommended.

8. The method according to claim 1, characterized in that, The step of determining the recommendation prediction data for each element object to be recommended based on the characteristics of the target object includes: The recommendation prediction sub-model based on the preset neural network model performs recommendation prediction on the features of the target object to obtain the initial prediction data of each element object to be recommended. The initial prediction data is normalized to determine the recommendation prediction data for each of the elements to be recommended.

9. The method according to claim 8, characterized in that, The recommendation prediction sub-model includes multiple second gating units and multiple neural network layers corresponding to the second gating units, and the multiple neural network layers are connected hierarchically. The recommendation prediction sub-model based on a preset neural network model performs recommendation prediction on the features of the target object to obtain initial prediction data for each element object to be recommended, including: The target object features are simultaneously input into multiple second gating units to obtain the gating weight of each second gating unit; The output features of multiple neural network layers are weighted based on multiple gating weights, and the weighted features are used as input features of the next neural network layer. The initial prediction data for each element object to be recommended is determined based on the features output by the last neural network layer.

10. The method according to claim 1, characterized in that, After generating the target object feature corresponding to each element object to be recommended by fusing the scene weights, the multiple input features, and the multiple domain label features, the method further includes: The dimension of the target object's feature is determined to be greater than a first threshold; The target object features are pooled so that the dimension of the pooled target object features is equal to the first threshold.

11. The method according to claim 1, characterized in that, The step of recommending the plurality of element objects to be recommended based on the recommendation prediction data includes: The multiple elements to be recommended are sorted according to the recommendation prediction data to obtain the sorting result; Based on the ranking results, recommendations are made to the plurality of element objects to be recommended.

12. A content recommendation device, characterized in that, The device includes: The acquisition unit is used to acquire scene data of the target recommendation scenario in the target application, input data of multiple feature domains corresponding to multiple element objects to be recommended, and domain label data of each feature domain. The weight generation unit is used to extract features from the scene data to obtain scene features, and to extract features from multiple input data to obtain multiple input features, and to generate scene weights corresponding to each of the recommended element objects based on the scene features and the multiple input features. The fusion unit is used to extract features from multiple domain label data to obtain multiple domain label features, and to fuse the scene weights, multiple input features and multiple domain label features to generate target object features corresponding to each element object to be recommended. The recommendation unit is used to determine the recommendation prediction data for each element object to be recommended based on the characteristics of the target object, and to recommend the plurality of element objects to be recommended based on the recommendation prediction data.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the content recommendation method according to any one of claims 1 to 11.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the content recommendation method according to any one of claims 1 to 11.

15. A computer program product comprising a computer program that is read and executed by a processor of a computer device, causing the computer device to perform the content recommendation method according to any one of claims 1 to 11.

Images