Information recommendation method and related apparatus

Abstract

The application relates to the technical field of artificial intelligence, and provides an information recommendation method and related devices to improve the accuracy of video recommendation. The method comprises the following steps: in response to target media content triggered by a target object, obtaining attribute information of the target object and content information of the target media content; performing feature conversion on the attribute information and the content information respectively to obtain attribute features of the attribute information and original features of the content information; obtaining enhanced features of the content information based on stored clustering features of each reference clustering center, wherein each reference clustering center is determined based on content information of each historical media content triggered by each sample object; obtaining object features of the target object based on the attribute features, the original features and the enhanced features, and determining recommendation information based on the object features.

Description

Technical Field

[0001] This application relates to the field of computer technology, and provides an information recommendation method and related apparatus. Background Technology

[0002] With the continuous development of computer technology, various applications are emerging. To meet the needs of different users, these applications use recommendation systems to achieve personalized content recommendations. For example, when a target user clicks on a video shared by another user, recommended videos related to that video are presented.

[0003] In related technologies, the corresponding object feature representation and content feature representation are usually obtained based on the attribute information of the sample objects contained in each training data and the content information of the content to be recommended. Then, the recommendation model is trained based on each object feature representation and content feature representation.

[0004] However, since recommendation models need to fit parameters using training samples, and the target videos to be shared are affected by the frequency of operations and are difficult to cover in the training dataset, there is a problem of sample sparsity. Therefore, it is difficult to learn an effective vector representation when training the model, resulting in low accuracy of content recall. Summary of the Invention

[0005] This application provides an information recommendation method and related apparatus to improve effective vector representation and enhance the accuracy of content recall.

[0006] In a first aspect, embodiments of this application provide an information recommendation method, including:

[0007] In response to target media content triggered by a target object, obtain the attribute information of the target object and the content information of the target media content;

[0008] The attribute information and the content information are respectively subjected to feature transformation to obtain the attribute features of the attribute information and the original features of the content information;

[0009] Enhanced features of the content information are obtained based on the clustering features of each stored reference cluster center; wherein, each reference cluster center is determined based on the content information of each historical media content triggered by each sample object;

[0010] Based on the attribute features, the original features, and the enhanced features, the object features of the target object are obtained, and recommendation information is determined based on the object features.

[0011] Secondly, embodiments of this application provide an information recommendation device, comprising:

[0012] An information acquisition unit is used to acquire attribute information of the target object and content information of the target media content in response to target media content triggered by the target object.

[0013] The feature acquisition unit is used to perform feature transformation on the attribute information and the content information respectively to obtain the attribute features of the attribute information and the original features of the content information;

[0014] A clustering enhancement unit is used to obtain enhanced features of the content information based on the clustering features of each of the stored reference clustering centers; wherein, each of the reference clustering centers is determined based on the content information of each historical media content triggered by each sample object;

[0015] The content recommendation unit is used to obtain the object features of the target object based on the attribute features, the original features, and the enhanced features, and to determine recommendation information based on the object features.

[0016] As one possible implementation, when determining the cluster centers in the current iteration based on the cluster centers in the previous iteration, the training unit is specifically used for:

[0017] Based on the cluster centers in the previous iteration, determine the target erasure information and target addition information in the current iteration;

[0018] Based on the target erasure information and target addition information in the current iteration, the cluster centers in the previous iteration are updated to obtain the cluster centers in the current iteration.

[0019] As one possible implementation, when determining the target erasure information and target addition information in the current iteration based on the cluster centers in the previous iteration, the training unit is specifically used for:

[0020] Based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data, the update weight in the current iteration is determined.

[0021] Based on the updated weights and the initial erasure information set, the target erasure information in the current iteration is determined;

[0022] Based on the updated weights and the initial increment information set, the target increment information in the current iteration is determined.

[0023] As one possible implementation, when determining the update weights in the current iteration based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data, the training unit is specifically used for:

[0024] Calculate the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data;

[0025] Based on the calculated similarities, the updated sub-weights corresponding to each cluster center in the previous iteration are obtained; each sub-weight is determined according to the proportion of the corresponding similarity in the total similarity.

[0026] Based on the obtained update sub-weights, the update weights in the current iteration are obtained.

[0027] As one possible implementation, the recommendation model includes a clustering enhancement layer and a feature representation layer; then, when obtaining the enhanced features of the content information based on the clustering features of each stored reference cluster center, the clustering enhancement unit is specifically used for:

[0028] The original features of the content information are input into the clustering enhancement layer to obtain the enhanced features of the content information. The clustering enhancement layer stores the clustering features of each of the reference cluster centers.

[0029] When obtaining the object features of the target object based on the attribute features, the original features, and the enhanced features, the content recommendation unit is specifically used for:

[0030] The attribute features, the original features, and the enhanced features are input into the feature representation layer to obtain the object features of the target object.

[0031] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described information recommendation method.

[0032] Fourthly, embodiments of this application provide a computer-readable storage medium including a computer program, which, when run on an electronic device, causes the electronic device to perform the steps of the information recommendation method described above.

[0033] Fifthly, embodiments of this application provide a computer program product, the program product including a computer program stored in a computer-readable storage medium, wherein a processor of an electronic device reads from the computer-readable storage medium and executes the computer program, causing the electronic device to perform the steps of the above-described information recommendation method.

[0034] In this embodiment, in response to target media content triggered by a target object, the attribute information of the target object and the content information of the target media content are obtained; feature transformation is performed on the attribute information and the content information respectively to obtain the attribute features of the attribute information and the original features of the content information; based on the clustering features of each stored reference cluster center, the enhanced features of the content information are obtained, wherein each reference cluster center is determined based on the content information of each historical media content triggered by each sample object; based on the attribute features, the original features and the enhanced features, the object features of the target object are obtained, and recommendation information is determined based on the object features.

[0035] In this way, by utilizing the content information of historical media content in the sample data, data augmentation is performed on the target media content. Since the feature representations of media content with similar content also have similarities, the obtained augmented features can effectively strengthen the features of the target media content, thereby effectively solving the data sparsity problem, improving the accuracy of content recall, and thus improving the accuracy of relevant media content recommendation.

[0036] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0038] Figure 1 This is a schematic diagram illustrating an application scenario provided in the embodiments of this application;

[0039] Figure 2A This is a flowchart illustrating a model training method provided in an embodiment of this application;

[0040] Figure 2B This is a flowchart illustrating an iterative training method provided in an embodiment of this application;

[0041] Figure 3 This is a schematic diagram of an operation interface provided in an embodiment of this application;

[0042] Figure 4A This is a schematic diagram of a sample pair provided in an embodiment of this application;

[0043] Figure 4B This is a schematic diagram of a negative sample provided in an embodiment of this application;

[0044] Figure 5 This is a network diagram of a recommendation model provided in the embodiments of this application;

[0045] Figure 6 This is a schematic diagram of a clustering enhancement layer provided in an embodiment of this application;

[0046] Figure 7 This is a schematic diagram of a process for obtaining the cluster centers in the current iteration, provided in an embodiment of this application.

[0047] Figure 8 This is a schematic diagram of a process for obtaining enhanced features provided in an embodiment of this application;

[0048] Figure 9 This is a flowchart illustrating an information recommendation method provided in an embodiment of this application;

[0049] Figure 10 This is a schematic diagram of another user interface provided in the embodiments of this application;

[0050] Figure 11 This is a logical schematic diagram of an information recommendation method provided in an embodiment of this application;

[0051] Figure 12 This is a schematic diagram of the structure of an information device provided in an embodiment of this application;

[0052] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0054] Target object: The target object refers to the object that presents the recommended content after the target video is triggered. In the embodiments of this application, the target object can be a user or the account logged in by a user.

[0055] Media content: Content generally includes various media forms such as text, sound, and images. The media used include text, pictures, photos, sound, animation, and video, as well as interactive functions provided by the program. In the embodiments of this application, it specifically refers to information and products that can be used to recommend to users, such as news, text and images, videos, and items.

[0056] Representation learning: In the field of deep learning, representation learning can refer to a set of techniques for learning a feature, such as transforming raw data into a form that can be learned by machines and used for effective development, removing invalid or redundant information from the raw data, and refining the effective information to form features.

[0057] Clustering: The clustering algorithm used for clustering processing may include, but is not limited to: K-means clustering algorithm, mean shift clustering algorithm, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm, Expectation-maximization (EM) clustering algorithm using Gaussian Mixture Model (GMM), and hierarchical clustering algorithm. This application does not limit the type of clustering algorithm used.

[0058] Recommendation system: Based on the target audience's interests and behaviors, recommend information and products that the target audience may be interested in.

[0059] Recall: Due to limitations in computing power and online system latency, recommendation systems typically consist of two phases: recall and ranking. The recall phase involves quickly selecting candidate content relevant to the user's interests from the entire media content library. For example, from a pool of millions of media content, thousands of content pieces might be selected that meet the system's objectives and computing power constraints. The ranking phase involves scoring the recalled media content and selecting the top k pieces based on their scores as the recommendation results. For example, from thousands of content pieces, the top k pieces might be selected for display.

[0060] This application relates to artificial intelligence (AI) and machine learning technologies.

[0061] 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.

[0062] Artificial intelligence (AI) studies the design principles and implementation methods of various intelligent machines, enabling them to perceive, reason, and make decisions. AI technology mainly includes computer vision, natural language processing, machine learning / deep learning, autonomous driving, and intelligent transportation. With the research and advancement of AI technology, it is being researched and applied in multiple fields, such as smart homes, intelligent customer service, virtual assistants, smart speakers, intelligent marketing, autonomous driving, robotics, and smart healthcare. It is believed that with further technological development, AI will be applied in even more fields and play an increasingly important role.

[0063] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. 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. Compared to data mining, which focuses on finding patterns in large datasets, machine learning emphasizes algorithm design, enabling computers to automatically "learn" patterns from data and use these patterns to predict unknown data.

[0064] 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, and inductive learning.

[0065] The recommendation model in this application embodiment is trained using machine learning or deep learning techniques. Based on the training method of the recommendation model in this application embodiment, recommendations for media content such as videos, images, and news can be made.

[0066] This application's embodiments involve two stages: model training and model application of the recommendation model. In the model training stage, the recommendation model is trained using machine learning techniques. Specifically, the training sample set provided in this application's embodiments is used to train the recommendation model. Through iterative training, the trained recommendation model is output. In the model training stage, content recommendations are performed based on the trained recommendation model. Detailed process for model training and application is described below.

[0067] With the continuous development of computer technology, various applications are emerging. To meet the needs of different users, these applications use recommendation systems to achieve personalized content recommendations. For example, when a target user clicks on a video shared by another user, recommended videos related to that video are presented.

[0068] In related technologies, the corresponding object feature representation and content feature representation are usually obtained based on the attribute information of the sample objects contained in each training sample and the content information of the content to be recommended. Then, the recommendation model is trained based on each object feature representation and content feature representation.

[0069] However, recommendation models require parameter fitting using training samples. Since the target videos for sharing are affected by the frequency of user interactions, they are difficult to cover in the training dataset, resulting in sample sparsity. Therefore, during model training, it is difficult to learn effective vector representations, leading to lower accuracy in content recall. For example, the popularity of target videos is polarized. If the target video is a popular video—meaning it receives a high frequency of clicks, comments, and likes within a certain timeframe—the training is more thorough due to the larger sample size of popular videos, typically without fitting issues. However, if the target video is a less popular video, its smaller or even homogenous audience results in very few training samples, making it difficult to ensure coverage in the training dataset and thus hindering the learning of effective vector representations.

[0070] In this embodiment, in response to target media content triggered by a target object, the attribute information of the target object and the content information of the target media content are obtained; feature transformation is performed on the attribute information and the content information respectively to obtain the attribute features of the attribute information and the original features of the content information; based on the clustering features of each stored reference cluster center, the enhanced features of the content information are obtained, wherein each reference cluster center is determined based on the content information of each historical media content triggered by each sample object; based on the attribute features, the original features and the enhanced features, the object features of the target object are obtained, and recommendation information is determined based on the object features.

[0071] In this way, by utilizing the content information of historical media content in the sample data to augment the target media content, and since the feature representations of similar media content also have similarities, the resulting augmented features can effectively strengthen the features of target media content that has not been sufficiently trained. This effectively solves the data sparsity problem, thereby improving the distribution accuracy and recall accuracy of the recommendation system, and ultimately improving the accuracy of relevant media content recommendations. Especially in scenarios involving related content recommendations based on sharing, it can increase users' viewing time and willingness to share, thereby enhancing the user experience.

[0072] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0073] See Figure 1The diagram illustrates an application scenario according to an embodiment of this application. The application scenario diagram includes a terminal device 110 and a server 120. The terminal device 110 and the server 120 can communicate via a communication network. The communication network can be a wired network or a wireless network.

[0074] Among them, terminal device 110 is a computer device used by the user, including but not limited to personal computers, mobile phones, tablets, laptops, e-book readers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc.

[0075] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0076] It should be noted that, Figure 1 The example shown is for illustration only; in reality, there is no limit to the number of terminal devices and servers.

[0077] In this embodiment of the application, the terminal device 110 may be equipped with an information recommendation client, which is used to recommend media content. The information recommendation client may be software that recommends media content, such as social software, instant messaging software, short video software, shopping software, etc., or it may be a mini program, webpage, etc., without specific limitations.

[0078] It should be noted that the information recommendation client in this application embodiment can also refer to various information recommendation applications that can be used in the car, such as education, news, travel, audiobooks, and advertisements. Correspondingly, media content can refer to news, books, guides, etc. related to education and travel, or advertisements, information flow messages, etc., without specific limitations.

[0079] Server 120 can be an information recommendation server corresponding to an information recommendation client. The information recommendation server provides media content materials to the information recommendation client; for example, the set of media content to be recalled in this embodiment can be stored on the information recommendation server side. Alternatively, the set of media content to be recalled can also be located locally on the information recommendation client. Furthermore, the information recommendation server can also be used for recommending media content, without specific limitations.

[0080] The information recommendation method in this embodiment can be executed by the server or the terminal device alone, or by both the server and the terminal device. For example, a target object triggers target media content by logging into an information recommendation client installed on the terminal device. The terminal device responds to the target media content triggered by the target object, sending relevant information about the target object and the target media content to the server. Based on the relevant information about the target object and the target media content, the server obtains the attribute features of the target object, the original features of the target media content, and the enhanced features of the target media content. Based on the attribute features, original features, and enhanced features, the server obtains the object features of the target object. Then, based on the object features of the target object and the recall content features of each media content to be recalled, the server determines the target recall media content from each media content to be recalled and returns the target recall media content to the terminal device. The terminal device displays the target recall media content to the target object. The server also obtains the object features of the target object, the context features related to the content recommendation request, and the content features corresponding to each content to be recommended. Based on these features, the server determines the final recommendation order of each content to be recommended for the target object and returns it to the terminal device. The terminal device then displays the recommendation results to the target object according to the determined recommendation order. It should be noted that the execution process of one of the information recommendation methods listed above is only an example and is not specifically limited here.

[0081] In this embodiment, during the model training phase, the recommendation model can be deployed on either the terminal device 110 or the server 120 for training. During the model application phase, after the recommendation model is trained using the training method described in this embodiment, it can be directly deployed on the server 120 or the terminal device 110. Optionally, generally, the recommendation model is directly deployed on the server 120 for both model training and application.

[0082] The information recommendation method provided in this application embodiment can be applied to various application scenarios that include media content recommendation tasks, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving. The training samples used in different scenarios are different, and they will not be listed one by one here.

[0083] The information recommendation method provided by the exemplary embodiments of this application will be described below with reference to the accompanying drawings and the application scenarios described above. It should be noted that the application scenarios described above are only shown to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way in this respect.

[0084] See Figure 2A The diagram shown is a flowchart illustrating a training method for a recommendation model provided in this embodiment. This method is applied to an electronic device, which can be a terminal device or a server. The specific process is as follows:

[0085] S201. Based on the sample dataset, obtain the initial cluster centers. Each sample dataset contains the content information of a historical media content triggered by the sample object.

[0086] In this embodiment, the historical media content can be news, images and text, videos, products, etc. Content information includes, but is not limited to, media content attribute information such as author and tags. The sample data may also include a content identifier (ID) for the media content. The content information of the media content can be obtained through the content identifier.

[0087] Taking videos as an example, the content information is the video information itself, which includes video attributes such as the author and tags. Video tags can include categories like entertainment videos, documentaries, and movie clips. Similarly, taking products as an example, the content information is the product information, which includes the product provider and tags. Product tags can include categories like entertainment videos, documentaries, and movie clips.

[0088] Each sample data set contains content information about a historical media content triggered by the sample object, as well as attribute information about the sample object. Attribute information includes basic object attributes, object behavior sequences, etc., with basic attributes including, but not limited to, age, gender, and city. For example, the sample data may contain an object identifier for the sample object, through which the sample object's attribute information can be obtained.

[0089] In this embodiment of the application, the scenario where media content related to a specific media content is presented after a media content is triggered by an object is referred to as a related content recommendation scenario. For example, the media content triggered by the object may be shared by other objects.

[0090] In related content recommendation scenarios, a piece of historical media content triggered by a sample object can be called the triggering content, and media content related to that historical media content can be called associated content. Taking related video recommendation as an example, a piece of historical video triggered by a sample object can also be called the triggering video, and videos related to that historical video can be called associated videos.

[0091] Clearly, related content needs to be relevant to the triggering content and be able to capture the object's interests and preferences. Related content recommendation is a content recommendation scenario with relevance constraints. In related content recommendation, each sample data point contains, in addition to the triggering content's information, the sample object's attribute information and related content's information. For example, the sample data can contain related content identifiers, through which the attribute information of the related content can be obtained. The following explanation uses the related content recommendation scenario as an example.

[0092] In this embodiment, the sample dataset contains sample data that can be divided into positive samples and negative samples. Positive samples can be determined based on the historical behavior of the sample object. For example, if the historical behavior indicates that the sample object triggered a historical media content and then triggered a set operation on the presented related content, then the historical behavior is taken as a positive sample. The operation type of the set operation includes, but is not limited to, clicking, resuming playback, liking, and forwarding.

[0093] For example, see related video recommendations. Figure 3 As shown, this is a schematic diagram of a possible user interface provided in an embodiment of this application. When object A clicks on video A, the following is displayed: Figure 3 The interface shown includes a video display area and a video recommendation area. The video display area is used to play the trigger video, and the video recommendation area is used to display recommended related videos. The trigger video is video A, and the related videos are videos B1, B2, B3, etc. It should be noted that... Figure 3 This example only uses the case where the video recommendation area is located to the right of the video presentation area. In actual application, there are no restrictions on the relative position between the video recommendation area and the video presentation area.

[0094] Negative samples can be collected through, but are not limited to, the following two methods:

[0095] The first method: perform negative sampling within a batch based on positive samples.

[0096] Specifically, positive samples can be divided into at least one batch, with each batch containing ns positive samples. Each positive sample constitutes a sample pair (object-side information and recall-side information). The object-side information includes the attribute information of the sample object and the content information of the triggering content, while the recall-side information is the content information of the associated content. Obviously, each batch contains ns sample pairs. For each batch, any sample pair is taken as the target sample pair, and the object-side information contained in the target sample pair is combined with the recall-side information contained in any sample pair from other sample pairs in that batch to obtain a negative sample. In this way, for each batch, based on ns sample pairs, corresponding ns negative samples can be obtained.

[0097] For example, see Figure 4AAs shown, each batch contains ns sample pairs, namely sample pair 1, sample pair 2, ..., sample pair ns. The object-side information contained in sample pair 1 includes the attribute information of sample object 1 and the content information of trigger content 1. The recall-side information contained in sample pair 1 is the content information of associated content 1. The object-side information contained in sample pair 2 includes the attribute information of sample object 2 and the content information of trigger content 2. The recall-side information contained in sample pair 2 is the content information of associated content 2. Similarly, the object-side information contained in sample pair ns includes the attribute information of sample object ns and the content information of trigger content ns. The recall-side information contained in sample pair ns is the content information of associated content ns.

[0098] See Figure 4B As shown, for a batch, the object-side information contained in a sample pair is combined with the recall-side information contained in any other sample pair in the batch to obtain negative sample 1, negative sample 2, ..., negative sample ns. Specifically, the object-side information in negative sample 1 includes the attribute information of sample object 1 and the content information of trigger content 1; the recall-side information in sample pair 1 is the content information of associated content ns. Similarly, the object-side information in negative sample 2 includes the attribute information of sample object 2 and the content information of trigger content 2; the recall-side information in negative sample 2 is the content information of associated content 1. Likewise, the object-side information in negative sample ns includes the attribute information of sample object ns and the content information of trigger content ns; the recall-side information in negative sample ns is the content information of associated content 2.

[0099] In this embodiment of the application, when obtaining each initial cluster center based on the sample dataset, the initial cluster center can be obtained based on the attribute information of the triggering content contained in each sample data in the sample dataset.

[0100] The second method involves obtaining negative samples through global negative sampling. Specifically, negative samples are randomly sampled from the material library, which contains at least the attribute information of the sample object, the content information of the triggering content, and the content information of the associated content.

[0101] In this embodiment of the application, each sample data annotation may have a label value, which is used to indicate whether the sample data is a positive sample or a negative sample.

[0102] S202. Based on the sample dataset and each initial cluster center, the recommendation model is iteratively trained to obtain the trained recommendation model, and the cluster centers of the last iteration are used as reference cluster centers.

[0103] See Figure 5 As shown, it is a schematic diagram of the network structure of a possible recommendation model provided in the embodiments of this application.

[0104] The recommendation model consists of a user tower on the left and an item tower on the right. The user tower is used to obtain object features based on the input object-side information, which includes the object's attributes and the content information of the media content triggered by that object. The item tower is used to obtain recall content features based on the input recall-side information, which includes, but is not limited to, the content information of the media content to be recalled. During model training, the content information of the media content to be recalled can also be understood as the content information of related content.

[0105] The user tower includes a clustering enhancement layer, an embedding layer, and a multilayer perceptron (MLP). The embedding layer takes object-side information as input, including content and attribute information, and outputs attribute features and original features. The clustering enhancement layer takes the original features of the content information as input and outputs enhanced features of the content information. The MLP takes attribute features, original features, and enhanced features as input and outputs object features. Specifically, in this embodiment, after obtaining the content and attribute information, the embedding layer obtains attribute features and original features. The original features are then input to the clustering enhancement layer, which outputs enhanced features. The attribute features, original features, and enhanced features are then concatenated to obtain concatenated features. Finally, the concatenated features are input to the MLP, which outputs object features.

[0106] The item tower includes an embedding layer and a multilayer perceptron. The embedding layer takes recall information as input and outputs features corresponding to that recall information. The multilayer perceptron takes the features corresponding to the recall information as input and outputs recall content features. Specifically, in this embodiment, after obtaining item-side information, the embedding layer obtains features corresponding to that information. These features are then concatenated to obtain concatenated features. Finally, the concatenated features are input to the multilayer perceptron, which outputs recall content features.

[0107] In this embodiment, the embedding layer can be used to perform embedding processing on sparse features. Specifically, it converts high-dimensional categorical variables into low-dimensional dense representations. This method randomly initializes a lookup table, where each row represents a value of the categorical variable, the number of rows is the number of values ​​for the categorical variable, and the number of columns is the dimension of the low-dimensional dense vector. When calculating the feature embedding, all categorical embeddings are searched and then summed using weighted averages. For continuous features, processing is required based on the feature's value range and distribution to ensure that the processed value range remains as close as possible to [0,1]. Common processing methods include normalization and taking the logarithm.

[0108] It should be noted that, in the embodiments of this application, both the user tower and the item tower can be implemented by, but are not limited to, deep neural networks (DNN).

[0109] Because the dual-tower model incorporates both recall and object-side information, it allows for the integration of richer information beyond object IDs and recalled content IDs. Furthermore, the dual-tower model is well-suited for online recommendation services and is relatively simple to operate. Therefore, this embodiment only uses the dual-tower model as an example for illustration. However, other network structures can be used in practical applications. Similarly, when using other network structures, the features output from the embedding layer can be input into the clustering enhancement layer to obtain the enhancement vector.

[0110] In this embodiment of the application, during each iteration, for a sample data, based on the cluster centers of the previous iteration, the cluster centers of the current iteration are determined, and based on the clustering features of each cluster center of the current iteration, the enhanced features of the content information contained in the sample data are determined. Then, based on the determined enhanced features, the prediction result is obtained, and the model loss is determined based on the prediction result, and the model parameters are adjusted based on the model loss.

[0111] The following is combined with Figure 5 The recommended model shown provides a detailed explanation of the model training process. For specifics, please refer to [link / reference]. Figure 2B As shown, when executing S202, the following steps may be used, but are not limited to:

[0112] S2021. Obtain sample data x. Sample data x can be any sample data from the sample dataset.

[0113] S2022. Based on the cluster centers of the previous iteration, determine the cluster centers of the current iteration.

[0114] In the implementation of this application, the process of determining the cluster centers of the current iteration based on the cluster centers of the previous iteration can be called a write operation, and the process of determining the enhanced features of the content information contained in the sample data x based on the clustering features of each cluster center of the current iteration can be called a read operation.

[0115] See Figure 6 As shown in this embodiment, the clustering enhancement layer adopts a memory network structure. The memory network contains two memory vectors, which are used to store cluster centers and clustering features of the cluster centers, respectively. The memory vector used to store cluster centers is called the root feed attr memory emb, abbreviated as RAE. The memory vector used to store the clustering features of the cluster centers is called the root feed cluster id emb, abbreviated as CIE. In this embodiment, clustering features refer to vectors that express the id embedding of the corresponding cluster center. In this embodiment, M(1), M(1), ..., M(m) represent the m cluster centers stored in the RAE, and E(1), E(1), ..., E(m) represent the clustering features of the m cluster centers stored in the CIE.

[0116] During model training, for the clustering enhancement module, the cluster centers stored in the RAE can be updated first through write operations. Then, based on the updated cluster centers in the RAE and the corresponding clustering features stored in the CIE, the corresponding enhancement features can be obtained through read operations. This allows the recommendation network to provide the current output based on historical inputs, and further enables data augmentation of the target video using the attribute features of historical inputs during the model application phase. The clustering enhancement module can also include a controller to implement the read and write functions of the RAE.

[0117] Considering that updates to cluster centers include both data addition and data erasure, this embodiment utilizes target erasure information and target addition information to achieve fast and accurate updates to cluster centers. For details, please refer to... Figure 7 As shown in the embodiments of this application, when executing S2022, the following steps may be adopted, but are not limited to:

[0118] S20221. Based on the cluster centers in the previous iteration, determine the target erasure information and target addition information in the current iteration.

[0119] Specifically, when executing S20221, the following steps may be used, but are not limited to:

[0120] Step 1: Based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data x, determine the update weight in the current iteration.

[0121] Specifically, when performing step one, the following steps may be used, but are not limited to:

[0122] Step 1 (1) Calculate the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data x.

[0123] The similarity between each cluster center and the original features of the content information contained in the sample data x can be calculated using, but is not limited to, the inner product cosine distance.

[0124] For example, the similarity between each cluster center and the sample data x can be calculated using formula (1):

[0125]

[0126] Where R represents the original features containing the content information of sample data x, and M t-1 (i) represents the i-th cluster center in the previous iteration, K(R,M) t-1 (i) represents the cluster center M. t-1 (i) The similarity between the original features R and the content information contained in the sample data x, where m is the number of cluster centers.

[0127] For example, assuming there are 10 cluster centers, i.e., m is 10, K(R,M) can be calculated using formula (1). t-1 (1)), K(R,M t-1 (2)), ..., K(R,M) t-1 (10)), where M t-1 (1) M t-1 (2), ..., M t-1 (10) are all 100-dimensional vectors.

[0128] Step 1 (2): Based on the calculated similarities, obtain the updated sub-weights corresponding to each cluster center in the previous iteration. Each sub-weight is determined based on the proportion of its corresponding similarity within the total similarity.

[0129] For example, the updated sub-weights corresponding to each cluster center can be calculated using formula (2):

[0130]

[0131] Among them, w w(i) represents the updated sub-weight corresponding to the i-th cluster center in the previous iteration, K(R, M) t-1 (i) represents the cluster center M. t-1 (i) The similarity between the original features R containing the content information in the sample data x, where m is the number of cluster centers in the previous iteration.

[0132] For example, based on K(R,M) t-1 (1)), K(R,M t-1 (2)), ..., K(R,M) t-1 (10)), M is calculated using formula (2). t-1 (1) Corresponding w w (1) M t-1 (2) The corresponding w w (2), ..., and obtain M t-1 (10) corresponding to w w (10).

[0133] Step 1 (3): Based on the obtained update sub-weights, obtain the update weights in the current iteration.

[0134] Assuming there are m cluster centers, for each cluster center, we can obtain the corresponding updated sub-weights, i.e., w. w (1), w w (2), ..., w w (m), based on the obtained updated sub-weights, an m-dimensional weight vector {w} is obtained. w (1), w w (2), ..., w w (m)}, and the m-dimensional weight vector {w w (1), w w (2), ..., w w (m)} serves as the update weight w in the current iteration. w .

[0135] For example, based on w w (1) w w (2), ..., w w (10) Obtain the updated weight w in the current iteration. w w w Given a 10-dimensional vector, including w w (1) w w (2), ..., w w (10).

[0136] Step 2: Based on the updated weights and the initial erase information set, determine the target erase information in the current iteration, and based on the updated weights and the initial increment information set, determine the target increment information in the current iteration.

[0137] In this embodiment, both the initial erasure information and the initial addition information can be randomly generated d-dimensional vectors. Here, d represents the length of the cluster centers, i.e., M. t-1 The length of (i).

[0138] For example, the target erasure information in the current iteration can be calculated using formula (3):

[0139]

[0140] Among them, w w This indicates updating the weights, e t This indicates the initial erase information.

[0141] The target addition information in the current iteration can be calculated using formula (4):

[0142]

[0143] Among them, w w This indicates updating the weights, a t This indicates the initial addition of information.

[0144] For example, suppose w w Let a be a 1×10 vector. t e t If both are vectors of size 100×1, then A t E t Both are 10×100 vectors.

[0145] When performing the write operation, although w in formulas (3) and (4) w The values ​​of the objects being calculated may differ, meaning the initial erasure information and the initial increment information may be different. Therefore, the target erasure information and the target iteration information may also be different.

[0146] By using the above implementation method, the target erasure information and target addition information in the current iteration can be determined by using the update weights in the same current iteration, which improves the efficiency of cluster center update and can effectively control the written content through update weights.

[0147] S20222. Based on the target erasure information and target addition information in the current iteration, update each cluster center in the previous iteration to obtain each cluster center in the current iteration.

[0148] For example, the cluster centers in the current iteration can be calculated using formula (5):

[0149] M t = (1-E t )⊙M t-1 +A t Formula (5)

[0150] Among them, A t E represents the added information in the current iteration. t M represents the erasure information in the current iteration. t-1 M represents the cluster centers in the previous iteration. t This represents the cluster centers in the current iteration. M t Let M be the memory matrix at time t, with size m×d, where m is the number of cluster centers, d is the dimension of the vector representation of the cluster centers, and M... t (i) is M t The data vector of the i-th unit.

[0151] S2023. Based on the clustering characteristics of each cluster center in the current iteration, determine the enhanced features of the content information contained in the sample data x.

[0152] In this embodiment of the application, the original features are input into the clustering enhancement layer. In the clustering enhancement layer, the enhanced features of the content information contained in the sample data x are determined by using the clustering features of each cluster center in the current iteration.

[0153] Since media content with similar content also has similar feature representations, in this embodiment of the application, as a possible implementation method, the cluster center most similar to the original feature is determined from each cluster center by the similarity between the original feature and the cluster center, and then the cluster feature corresponding to the determined cluster center is used as the enhanced feature of the content information contained in the sample data x.

[0154] As another possible implementation, to further improve the accuracy of the enhanced features, in the embodiments of this application, see [reference needed]. Figure 8 As shown, when executing S2023, the following steps may be used, but are not limited to:

[0155] S20231. Obtain each cluster center in the current iteration, and based on the similarity between the original features and each cluster center in the current iteration, obtain the sub-weights corresponding to each cluster center in the current iteration.

[0156] For example, the sub-weights corresponding to each cluster center in the current iteration can be calculated using formula (6):

[0157]

[0158] Where w(i) represents the sub-weight corresponding to the i-th cluster center in the current iteration, K(R, M) t (i) represents the cluster center M. t (i) Similarity between the original features R and the content information contained in the sample data x.

[0159] For example, suppose that M is calculated using formula (6). t (1) The corresponding w(1) is 0.1, M t (2) The corresponding w(2) are 0.2, ..., and we obtain M t The w(10) corresponding to (10) is 0.1.

[0160] S20232. Based on the obtained sub-weights and the clustering features of each cluster center in the current iteration, the enhanced features of the content information are obtained.

[0161] Specifically, during execution S20232, the clustering features of each cluster center in the current iteration are obtained; based on the obtained sub-weights, the obtained clustering features are weighted and summed to obtain the enhanced features of the content information. The clustering features corresponding to each reference cluster center can be obtained through embedding.

[0162] For example, the enhanced features of content information can be calculated using formula (7):

[0163]

[0164] Where r represents the enhanced features of content information, w(i) represents the sub-weight corresponding to the i-th cluster center in the current iteration, and E(i) represents the cluster center M. t (i) Corresponding clustering features.

[0165] For example, suppose that w(1), w(2), ..., w(10) are 0.1, 0.2, ..., 0.1 respectively. Then, by formula (7), r = 0.1E(1) + 0.2E(2) + ... + 0.1E(10).

[0166] By utilizing the similarity between the original features and the cluster centers, and based on the cluster features of each cluster center, the cluster features most similar to the original features are used as enhancement features, making the sample enhancement features more accurate, thereby improving the accuracy of feature representation and thus enhancing recall precision.

[0167] S2024. Based on the determined enhancement features, the prediction results are obtained.

[0168] In this embodiment, based on the attribute features, original features, and enhanced features of sample data x, object features of sample objects in sample data x are obtained. Based on the recall information of sample data x, recall content features are acquired, and the similarity between object features and recall content features is calculated. The similarity between object features and recall content features can be represented using, but is not limited to, cosine similarity.

[0169] S2025. Based on the prediction results, determine the model loss.

[0170] In this embodiment of the application, the model loss may be, but is not limited to, cross-entropy loss. For example, the cross-entropy loss is determined based on the similarity between the calculated object features and the recalled content features, and the determined cross-entropy loss is used as the model loss.

[0171] S2026. Adjust model parameters based on model loss.

[0172] S2027. Determine whether the model convergence condition is met. If yes, execute S2028; otherwise, return to execute S2021.

[0173] In this embodiment, the model convergence condition may include the number of iterations reaching an iteration threshold, the model loss of the current iteration meeting a first preset loss range, or the model loss of n consecutive iterations meeting a second preset loss range, but is not limited to these.

[0174] S2028, Output the recommended model after training.

[0175] In this embodiment of the application, after the recommendation model is trained, the data stored in ARE and CIE is also updated. During the model application stage, the enhanced features of the content information can be determined using the data stored in ARE and CIE.

[0176] See Figure 9 The diagram shown is a flowchart of an information recommendation method provided in this application embodiment. This method is applied to an electronic device, which can be a terminal device or a server. The specific process is as follows:

[0177] S901. In response to the target media content triggered by the target object, obtain the attribute information of the target object and the content information of the target media content.

[0178] S902. Perform feature transformation on the attribute information and content information respectively to obtain the attribute features of the attribute information and the original features of the content information.

[0179] In this embodiment, attribute information and content information can be input into the embedding layer of the user tower for feature transformation to obtain the attribute features of the attribute information and the original features of the content information. See S2022 for details, which will not be repeated here.

[0180] S903. Based on the clustering features of each reference clustering center stored in the database, obtain enhanced features of the content information; wherein, each reference clustering center is determined based on the content information of each historical media content triggered by each sample object.

[0181] In this embodiment of the application, when obtaining the enhanced features of content information based on the clustering features of each of the stored reference clustering centers, the original features of the content information can be input into the clustering enhancement layer to obtain the enhanced features of the content information.

[0182] Specifically, when executing S903, the following steps may be used, but are not limited to:

[0183] First, obtain the stored reference cluster centers, and based on the similarity between the original features and the stored reference cluster centers, obtain the sub-weights corresponding to each stored reference cluster center.

[0184] For example, the sub-weights corresponding to each reference cluster center can be calculated using formula (8):

[0185]

[0186] Among them, w c (i) represents the sub-weight corresponding to the i-th reference cluster center, K(R, M) t (i) represents the i-th reference cluster center M. c (i) The similarity between the original features R of the content information of the target media content and the target media content.

[0187] Secondly, based on the obtained sub-weights and the clustering features of each reference cluster center, enhanced features of content information are obtained.

[0188] Specifically, the clustering features of each stored reference cluster center are obtained; based on the obtained sub-weights, the obtained clustering features are weighted and summed to obtain the enhanced features of the content information. The clustering features corresponding to each reference cluster center can be obtained through embedding.

[0189] For example, the enhanced features of content information can be calculated using formula (7):

[0190]

[0191] Where r represents the enhanced features of content information, wc (i) represents the i-th reference cluster center M c (i) corresponds to the sub-weight, E c (i) represents the i-th reference cluster center M c (i) Corresponding clustering features.

[0192] S904. Based on attribute features, original features, and enhanced features, obtain the object features of the target object, and determine the recommendation information based on the object features.

[0193] In this embodiment, attribute features, original features, and enhanced features are input into a feature representation layer to obtain the object features of the target object. The feature representation layer is a user-based MLP. Specifically, the attribute features, original features, and enhanced features are concatenated to obtain concatenated features, which are then input into the feature representation layer to obtain the object features of the target object.

[0194] Determining recommendation information based on object features includes: identifying target recall media content from the target recall media content based on the similarity between the object features of the target object and the recall content features of each media content to be recalled. The number of target recall media content can be one or more.

[0195] The recall content features for each media content to be recalled can be obtained in advance based on the item tower in the recommendation model, or they can be obtained based on the item tower in the recommendation model after the target object triggers the target media content; there are no restrictions on this. The similarity between the object features and the recall content features can be represented by, but is not limited to, cosine similarity.

[0196] For example, suppose there are 1,000 media contents to be recalled. Based on the similarity between the object characteristics of the target object and the content characteristics of each media contents to be recalled, 10 target media contents to be recalled are identified from each media contents to be recalled.

[0197] The present application will now be described with reference to several specific embodiments.

[0198] Example 1: Video recommendation scenario.

[0199] See Figure 10 As shown, when the target object clicks on video A in the operation interface, the terminal device responds to the target object's click operation on the target video by obtaining the object ID of the target object and the video ID of the target video. It then retrieves the attribute information of the target object based on the object ID and the content information of the target video based on the video ID. The attribute information of the target object includes its basic object attributes and object behavior sequence, while the content information of the target video includes video attribute information such as author and tags.

[0200] See Figure 11 As shown, the object's basic attributes, object behavior sequence, target video ID, and video attribute information are input into the embedding layer of the user tower for feature transformation, resulting in features corresponding to the object's basic attributes, the object behavior sequence, the target video ID, and the original features corresponding to the target video's video attribute information.

[0201] Then, using the reference cluster centers stored in ARE and their respective clustering features based on the reference cluster centers stored in CIE, the enhanced features corresponding to the video attribute information of the target video are obtained. Then, the attribute features, original features, and enhanced features are processed to obtain the object features of the target object.

[0202] Subsequently, based on the similarity between the object features of the target object to be recalled and the recall content features of each pre-generated video to be recalled, the target recall video is determined from each video to be recalled.

[0203] Example 2: Product recommendation scenario.

[0204] For example, when the target object clicks on product B in the operation interface, the terminal device responds to the target object's click operation on the target video by obtaining the object ID of the target object and the product ID of the target product. It then obtains the attribute information of the target object based on its object ID and the content information of the target video based on its product ID. The attribute information of the target object includes its basic object attributes and object behavior sequence, while the content information of the target video is the product attribute information.

[0205] The object's basic attributes, object behavior sequence, target product ID, and product attribute information are input into the user tower's embedding layer for feature transformation, resulting in features corresponding to the object's basic attributes, object behavior sequence, target video ID, and original features corresponding to the product attribute information.

[0206] Then, using the reference cluster centers stored in ARE and their respective clustering features based on the reference cluster centers stored in CIE, the enhanced features corresponding to the product attribute information are obtained. Then, the attribute features, original features, and enhanced features are processed to obtain the object features of the target object.

[0207] Subsequently, based on the similarity between the object characteristics of the target object to be recalled and the recall content characteristics of each pre-generated product to be recalled, the target product to be recalled is determined from each product to be recalled.

[0208] Example 3: News recommendation scenario.

[0209] For example, when the target object clicks on news item B in the interface, the terminal device responds to the target object's click operation on the target video by obtaining the object ID of the target object and the news ID of the target news item. It then retrieves the attribute information of the target object based on its object ID and the content information of the target video based on the news ID. The attribute information of the target object includes its basic object attributes and object behavior sequence, while the content information of the target video is its news attribute information.

[0210] The basic attributes of the object, the sequence of object behavior, the target news ID, and the news attribute information are input into the embedding layer of the user tower for feature transformation, resulting in the features corresponding to the basic attributes of the object, the sequence of object behavior, the target video ID, and the original features corresponding to the news attribute information.

[0211] Then, using the reference cluster centers stored in ARE and their respective clustering features based on the reference cluster centers stored in CIE, enhanced features corresponding to the news attribute information are obtained. Then, the attribute features, original features, and enhanced features are processed to obtain the object features of the target object.

[0212] Subsequently, based on the similarity between the object characteristics of the target object to be recalled and the recall content characteristics of each pre-generated news item to be recalled, the target news item to be recalled is determined from each news item to be recalled.

[0213] Based on the same inventive concept, embodiments of this application provide an information recommendation device. For example... Figure 12 As shown, this is a structural schematic diagram of the information recommendation device 1200, which may include:

[0214] Information acquisition unit 1201 is used to acquire attribute information of the target object and content information of the target media content in response to target media content triggered by the target object;

[0215] The feature acquisition unit 1202 is used to perform feature transformation on the attribute information and the content information respectively to obtain the attribute features of the attribute information and the original features of the content information;

[0216] The clustering enhancement unit 1203 is used to obtain enhanced features of the content information based on the clustering features of each of the stored reference clustering centers; wherein, each of the reference clustering centers is determined based on the content information of each historical media content triggered by each sample object;

[0217] The content recommendation unit 1204 is used to obtain the object features of the target object based on the attribute features, the original features and the enhanced features, and to determine recommendation information based on the object features.

[0218] As one possible implementation, when obtaining the enhanced features of the content information based on the clustering features of each stored reference clustering center, the clustering enhancement unit 1203 is specifically used for:

[0219] Obtain the stored reference cluster centers, and based on the similarity between the original features and the obtained reference cluster centers, obtain the sub-weights corresponding to each reference cluster center;

[0220] Based on the obtained sub-weights and the clustering features of each reference cluster center, the enhanced features of the content information are obtained.

[0221] As one possible implementation, when obtaining the enhanced features of the content information based on the obtained sub-weights and the clustering features of each reference cluster center, the clustering enhancement unit 1203 is specifically used for:

[0222] Obtain the clustering features of each of the stored reference cluster centers;

[0223] Based on the obtained sub-weights, the acquired clustering features are weighted and summed to obtain the enhanced features of the content information.

[0224] As one possible implementation, the device further includes a training unit 1205, which is used for:

[0225] Based on the sample dataset, we obtain the initial cluster centers. Each sample dataset contains content information of a historical media content triggered by a sample object.

[0226] Based on the sample dataset and the initial cluster centers, the recommendation model is iteratively trained to obtain the trained recommendation model, and the cluster centers of the last iteration are used as the reference cluster centers. During each iteration, the following operations are performed:

[0227] For a given sample data, based on the cluster centers obtained in the previous iteration, determine the cluster centers for the current iteration, and based on the clustering features of each cluster center in the current iteration, determine the enhanced features of the content information contained in the sample data.

[0228] Based on the determined enhancement features, a prediction result is obtained, and the model loss is determined based on the prediction result, and the model parameters are adjusted based on the model loss.

[0229] As one possible implementation, when determining the cluster centers in the current iteration based on the cluster centers in the previous iteration, the training unit 1205 is specifically used for:

[0230] Based on the cluster centers in the previous iteration, determine the target erasure information and target addition information in the current iteration;

[0231] Based on the target erasure information and target addition information in the current iteration, the cluster centers in the previous iteration are updated to obtain the cluster centers in the current iteration.

[0232] As one possible implementation, when determining the target erasure information and target addition information in the current iteration based on the cluster centers in the previous iteration, the training unit 1205 is specifically used for:

[0233] Based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data, the update weight in the current iteration is determined.

[0234] Based on the updated weights and the initial erasure information set, the target erasure information in the current iteration is determined;

[0235] Based on the updated weights and the initial increment information set, the target increment information in the current iteration is determined.

[0236] As one possible implementation, when determining the update weights in the current iteration based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data, the training unit 1205 is specifically used for:

[0237] Calculate the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data;

[0238] Based on the calculated similarities, the updated sub-weights corresponding to each cluster center in the previous iteration are obtained; each sub-weight is determined according to the proportion of the corresponding similarity in the total similarity.

[0239] Based on the obtained update sub-weights, the update weights in the current iteration are obtained.

[0240] As one possible implementation, the recommendation model includes a clustering enhancement layer and a feature representation layer; then, when obtaining the enhanced features of the content information based on the clustering features of each stored reference cluster center, the clustering enhancement unit 1203 is specifically used for:

[0241] The original features of the content information are input into the clustering enhancement layer to obtain the enhanced features of the content information. The clustering enhancement layer stores the clustering features of each of the reference cluster centers.

[0242] When obtaining the object features of the target object based on the attribute features, the original features, and the enhanced features, the content recommendation unit 1204 is specifically used for:

[0243] The attribute features, the original features, and the enhanced features are input into the feature representation layer to obtain the object features of the target object.

[0244] For ease of description, the above sections are divided into modules (or units) according to their functions and described separately. Of course, in implementing this application, the functions of each module (or unit) can be implemented in one or more software or hardware components.

[0245] Regarding the apparatus in the above embodiments, the specific manner in which each unit executes the request has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0246] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0247] Based on the same inventive concept, embodiments of this application also provide an electronic device. In one embodiment, the electronic device can be a server or a terminal device. See also... Figure 13 As shown, it is a schematic diagram of a possible electronic device provided in an embodiment of this application. Figure 13 In the electronic device 1300, there are: processor 1310 and memory 1320.

[0248] The memory 1320 stores a computer program that can be executed by the processor 1310. The processor 1310 can execute the steps of the information recommendation method described above by executing the instructions stored in the memory 1320.

[0249] Memory 1320 may be volatile memory, such as random-access memory (RAM); memory 1320 may also be non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1320 may be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1320 may also be a combination of the above-described memories.

[0250] Processor 1310 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 1310 implements the above-described recommended information method when executing computer programs stored in memory 1320.

[0251] In some embodiments, the processor 1310 and the memory 1320 may be implemented on the same chip, while in other embodiments they may be implemented on separate chips.

[0252] This application embodiment does not limit the specific connection medium between the processor 1310 and the memory 1320. This application embodiment takes the connection between the processor 1310 and the memory 1320 via a bus as an example. Figure 13 The diagram uses thick lines to describe the connections between other components; these are merely illustrative and not intended to be limiting. Buses can be categorized as address buses, data buses, control buses, etc. For ease of description, Figure 13 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0253] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the steps of the aforementioned information recommendation method. In some possible implementations, various aspects of the information recommendation method provided in this application can also be implemented as a program product including a computer program. When the program product is run on an electronic device, the computer program causes the electronic device to perform the steps of the aforementioned information recommendation method. For example, the electronic device can perform actions such as... Figure 2A , Figure 2B or Figure 9 The steps are shown in the figure.

[0254] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, RAM, ROM, erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0255] The program product of the embodiments of this application may be a CD-ROM and include a computer program, and may run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium may be any tangible medium that contains or stores a computer program that may be used by or in conjunction with a command execution system, apparatus, or device.

[0256] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a computer program for use by or in conjunction with a command execution system, apparatus, or device.

[0257] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0258] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An information recommendation method, characterized in that, include: In response to target media content triggered by a target object, obtain the attribute information of the target object and the content information of the target media content; The target object refers to the user or the account the user logs into; The target media content refers to the media content triggered by the target object, and the media content refers to information or products used to recommend to the target object; the attribute information of the target object includes the target object's basic attributes and object behavior sequence; The attribute information and the content information are respectively subjected to feature transformation to obtain the attribute features of the attribute information and the original features of the content information; Based on the clustering features of each of the stored reference cluster centers, the enhanced features of the content information are obtained; Based on the attribute features, the original features, and the enhanced features, the object features of the target object are obtained, and recommendation information is determined based on the object features; The reference cluster centers are obtained as follows: Based on the sample dataset, the recommendation model is iteratively trained to obtain the trained recommendation model, and the cluster centers from the last iteration are used as the reference cluster centers. In each iteration, the following steps are performed: for a given sample data, the target erasure information and target addition information are updated based on the cluster centers obtained in the previous iteration, and the current iteration's cluster centers are obtained based on the updated target erasure information and target addition information; based on the clustering features of each cluster center in the current iteration, the enhancement features of the content information contained in the given sample data are determined; based on the determined enhancement features, a prediction result is obtained, and the model loss is determined based on the prediction result; and the model parameters are adjusted based on the model loss. The update weights used to update the target erasure information and the target addition information are determined based on the similarity between the original features of the content information contained in the sample data and each cluster center in the previous iteration; and each update sub-weight in the update weights is determined according to the proportion of the corresponding similarity in each similarity.

2. The method as described in claim 1, characterized in that, The enhanced features of the content information are obtained based on the clustering features of each reference clustering center stored in the database, including: Obtain the stored reference cluster centers, and based on the similarity between the original features and the obtained reference cluster centers, obtain the sub-weights corresponding to each reference cluster center; Based on the obtained sub-weights and the clustering features of each reference cluster center, the enhanced features of the content information are obtained.

3. The method as described in claim 2, characterized in that, The enhancement features of the content information obtained based on the acquired sub-weights and the clustering features of each reference cluster center include: Obtain the clustering features of each of the stored reference cluster centers; Based on the obtained sub-weights, the acquired clustering features are weighted and summed to obtain the enhanced features of the content information.

4. The method according to any one of claims 1 to 3, characterized in that, The process of updating the target erasure information and target addition information based on the cluster centers obtained in the previous iteration, and obtaining the cluster centers for the current iteration based on the updated target erasure information and target addition information, includes: Based on the cluster centers in the previous iteration, determine the target erasure information and target addition information in the current iteration; Based on the target erasure information and target addition information in the current iteration, the cluster centers in the previous iteration are updated to obtain the cluster centers in the current iteration.

5. The method as described in claim 4, characterized in that, The step of determining the target erasure information and target addition information in the current iteration based on the cluster centers in the previous iteration includes: Based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data, the update weight in the current iteration is determined. Based on the updated weights and the initial erasure information set, the target erasure information in the current iteration is determined; Based on the updated weights and the initial increment information set, the target increment information in the current iteration is determined.

6. The method as described in claim 5, characterized in that, The step of determining the update weight in the current iteration based on the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data includes: Calculate the similarity between each cluster center in the previous iteration and the original features of the content information contained in the sample data; Based on the calculated similarities, the updated sub-weights corresponding to each cluster center in the previous iteration are obtained; Based on the obtained update sub-weights, the update weights in the current iteration are obtained.

7. The method as described in claim 1, characterized in that, The recommendation model includes a clustering enhancement layer and a feature representation layer; therefore, obtaining enhanced features of the content information based on the clustering features of each stored reference cluster center includes: The original features of the content information are input into the clustering enhancement layer to obtain the enhanced features of the content information. The clustering enhancement layer stores the clustering features of each of the reference cluster centers. The process of obtaining the object features of the target object based on the attribute features, the original features, and the enhanced features includes: The attribute features, the original features, and the enhanced features are input into the feature representation layer to obtain the object features of the target object.

8. An information recommendation device, characterized in that, include: An information acquisition unit is used to acquire attribute information of the target object and content information of the target media content in response to target media content triggered by the target object. The target object refers to the user or the account the user logs into; The target media content refers to the media content triggered by the target object, and the media content refers to information or products used to recommend to the target object; the attribute information of the target object includes the target object's basic attributes and object behavior sequence; The feature acquisition unit is used to perform feature transformation on the attribute information and the content information respectively to obtain the attribute features of the attribute information and the original features of the content information; A clustering enhancement unit is used to obtain enhanced features of the content information based on the clustering features of each of the stored reference clustering centers; wherein, each of the reference clustering centers is determined based on the content information of each historical media content triggered by each sample object; The content recommendation unit is used to obtain the object features of the target object based on the attribute features, the original features, and the enhanced features, and to determine recommendation information based on the object features; The training unit is used to: iteratively train the recommendation model based on the sample dataset to obtain the trained recommendation model, and use the cluster centers of the last iteration as the reference cluster centers; wherein, in each iteration, the following are performed: for a sample data, update the target erasure information and target addition information based on the cluster centers obtained in the previous iteration, and obtain the cluster centers of the current iteration based on the updated target erasure information and target addition information; determine the enhancement features of the content information contained in the sample data based on the cluster features of the current iteration's cluster centers; obtain the prediction result based on the determined enhancement features, determine the model loss based on the prediction result, and adjust the model parameters based on the model loss; wherein the update weights used to update the target erasure information and the target addition information are determined based on the similarity between the original features of the content information contained in the sample data and the cluster centers of the current iteration; wherein each update sub-weight in the update weights is determined according to the proportion of the corresponding similarity in the various similarities.

9. The apparatus as claimed in claim 8, characterized in that, When obtaining enhanced features of the content information based on the clustering features of each stored reference clustering center, the clustering enhancement unit is specifically used for: Obtain the stored reference cluster centers, and based on the similarity between the original features and the obtained reference cluster centers, obtain the sub-weights corresponding to each reference cluster center; Based on the obtained sub-weights and the clustering features of each reference cluster center, the enhanced features of the content information are obtained.

10. The apparatus as claimed in claim 9, characterized in that, When obtaining the enhanced features of the content information based on the obtained sub-weights and the clustering features of each reference clustering center, the clustering enhancement unit is specifically used for: Obtain the clustering features of each of the stored reference cluster centers; Based on the obtained sub-weights, the acquired clustering features are weighted and summed to obtain the enhanced features of the content information.

11. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any of the methods described in claims 1 to 7.

12. A computer-readable storage medium, characterized in that, It includes a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described in claims 1 to 7.

13. A computer program product, characterized in that, It includes a computer program stored in a computer-readable storage medium, and a processor of an electronic device reads from and executes the computer program, causing the electronic device to perform the steps of any of the methods described in claims 1 to 7.

Images