Information recommendation method and device, computer device, storage medium and program product

Abstract

The application relates to an information recommendation method and device, computer equipment, a storage medium and a computer program product. The method can be applied to the fields of artificial intelligence, intelligent transportation and cloud computing, and comprises the following steps: acquiring fine-grained features and coarse-grained features of a media to be recommended; performing feature generalization processing on the coarse-grained features to obtain a coarse-grained feature vector; performing feature strengthening processing on the fine-grained features and the coarse-grained features to obtain a fine-grained feature vector; weighting and fusing the fine-grained feature vector and the coarse-grained feature vector to obtain a fusion vector; selecting a target object from candidate objects based on the fusion vector and an object feature vector of the candidate objects, and pushing recommendation information of the media to be recommended to the target object. The method can improve the accuracy of recommendation and the conversion rate of media information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an information recommendation method, apparatus, computer device, storage medium, and computer program product. Background Technology

[0002] With the continuous development of internet technology, users can obtain various media information to be recommended from the internet, such as advertisements, and then display this information on the media page. For media providers, it is necessary to accurately recommend the developed media information to users with relevant needs in order to improve product usage.

[0003] Since developed media information is typically used by only a small number of users, traditional recommendation schemes both recommend existing and interesting media information to users, and also push new media information as an add-on to improve the conversion rate of new media information. However, in the above recommendation scheme, users may be interested in existing media information, but not necessarily in new media information, resulting in low accuracy in recommending items to users and a low conversion rate for media information. Summary of the Invention

[0004] Therefore, it is necessary to provide an information recommendation method, apparatus, computer equipment, computer-readable storage medium, and computer program product to address the aforementioned technical problems, thereby improving the accuracy of recommendations and the conversion rate of media information.

[0005] Firstly, this application provides an information recommendation method. The method includes:

[0006] Obtain the fine-grained and coarse-grained features of the media to be recommended;

[0007] The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector;

[0008] The fine-grained and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector;

[0009] The fine-grained representation vector and the coarse-grained representation vector are weighted and fused to obtain a fused vector;

[0010] Based on the fusion vector and the object representation vector of the candidate objects, a target object is selected from the candidate objects, and recommendation information of the media to be recommended is pushed to the target object.

[0011] Secondly, this application also provides an information recommendation device. The device includes:

[0012] The acquisition module is used to acquire fine-grained and coarse-grained features of the media to be recommended;

[0013] The first processing module is used to perform feature generalization processing on the coarse-grained features to obtain a coarse-grained representation vector.

[0014] The second processing module is used to perform feature enhancement processing on the fine-grained features and coarse-grained features to obtain a fine-grained representation vector;

[0015] The fusion module is used to weightedly fuse the fine-grained representation vector and the coarse-grained representation vector to obtain a fused vector;

[0016] The recommendation module is used to select a target object from the candidate objects based on the fusion vector and the object representation vector of the candidate objects, and push the recommendation information of the media to be recommended to the target object.

[0017] In one embodiment, the acquisition module is further configured to acquire at least two media features corresponding to the media to be recommended; determine the granularity entropy of each feature among the at least two media features; and perform granularity division on each feature among the at least two media features based on the granularity entropy to obtain the fine-grained features and coarse-grained features of the media to be recommended.

[0018] In one embodiment, the acquisition module is further configured to: determine the number of interactive objects corresponding to different feature values ​​for each feature among the at least two media features; determine the first conversion rate for each feature among the at least two media features when taking different feature values ​​based on the number of interactive objects and the total number of objects in the media to be recommended; and determine the granular entropy of each feature among the at least two media features based on the conversion rate when taking different feature values ​​and the corresponding number of objects.

[0019] In one embodiment, the acquisition module is further configured to: determine the number of object interactions corresponding to different feature values ​​for each feature among the at least two media features; determine the second conversion rate for each feature among the at least two media features when taking different feature values ​​based on the number of object interactions and the total number of interactions for the media to be recommended; and determine the granular entropy of each feature among the at least two media features based on the second conversion rate when taking different feature values ​​and the corresponding number of object interactions.

[0020] In one embodiment, the acquisition module is further configured to acquire at least two media features corresponding to the media to be recommended; determine a coarse-to-fine granularity ratio, and divide each feature among the at least two media features into fine-grained features and coarse-grained features according to the coarse-to-fine granularity ratio to obtain fine-grained features and coarse-grained features of the media to be recommended; or, in response to a granularity division request, divide each feature among the at least two media features into fine-grained features and coarse-grained features.

[0021] In one embodiment, the first processing module is further configured to concatenate the coarse-grained features belonging to the same media to be recommended, respectively, to obtain a first concatenated feature for each of the media to be recommended; and sequentially input each of the first concatenated features into the coarse-grained network of the recommendation model, so that the multilayer perceptron in the coarse-grained network performs feature processing on the input first concatenated features to obtain a coarse-grained representation vector.

[0022] In one embodiment, the second processing module is further configured to concatenate the fine-grained features and the coarse-grained features to obtain a second concatenated feature; and to sequentially input the second concatenated feature into the fine-grained network of the recommendation model, so that the multilayer perceptron in the fine-grained network performs feature processing on the input second concatenated feature to obtain a fine-grained representation vector.

[0023] In one embodiment, the device further includes:

[0024] The first determining module is used to determine the modulus or information entropy of the feature embedding vector of the media to be recommended; and to determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector respectively based on the fine-grained features, the coarse-grained features and the modulus; or, to determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector respectively based on the fine-grained features, the coarse-grained features and the information entropy.

[0025] The fusion module is further configured to perform weighted fusion of the fine-grained representation vector and the coarse-grained representation vector based on the weight parameters.

[0026] In one embodiment, the recommendation module is further configured to determine a first score value of the candidate object based on the fusion vector and the object representation vector of the candidate object; the first score value is used to reflect the conversion rate of the candidate object after receiving the media to be recommended; and a target object is selected from the candidate objects according to the first score value.

[0027] In one embodiment, the coarse-grained representation vector, the fine-grained representation vector, and the fused vector are obtained through a machine learning model; the apparatus further includes:

[0028] The acquisition module is also used to acquire the training fine-grained features and training coarse-grained features of the media samples;

[0029] The first processing module is further configured to perform feature generalization processing on the training coarse-grained features through the first network branch of the machine learning model to obtain the training coarse-grained representation vector.

[0030] The second processing module is further configured to perform feature enhancement processing on the training fine-grained features and the training coarse-grained features through the second network branch of the machine learning model to obtain the training fine-grained representation vector;

[0031] The training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused through the fusion network of the machine learning model to obtain the training fusion vector;

[0032] The fusion module is further configured to determine a second score value for the candidate object sample based on the training fusion vector and the training object representation vector of the candidate object sample;

[0033] An optimization module is used to optimize the parameters of the machine learning model based on the loss between the second score value and the score label.

[0034] In one embodiment, the device further includes:

[0035] The second determining module is used to determine a third score value of the candidate object sample based on the training coarse-grained representation vector and the training object representation vector of the candidate object sample; and to determine a fourth score value of the candidate object sample based on the training fine-grained representation vector and the training object representation vector of the candidate object sample.

[0036] The optimization module is further configured to optimize the parameters of the machine learning model based on the loss between the second score value and the score label, the loss between the third score value and the score label, and the loss between the fourth score value and the score label.

[0037] In one embodiment, the device further includes:

[0038] The third determining module is used to determine the training degree of the media sample; the training degree is used to represent the cumulative number of iterations in which the media sample participates in training; based on the training degree, the fine-grained training features and the coarse-grained training features, the weight coefficients corresponding to the fine-grained training representation vector and the coarse-grained training representation vector are determined respectively.

[0039] The fusion module is further configured to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight coefficients.

[0040] In one embodiment, the device further includes:

[0041] The fourth determining module is used to determine the magnitude or information entropy of the feature embedding vector of the media sample; and to determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively based on the training fine-grained features, the training coarse-grained features, and the magnitude; or, to determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively based on the training fine-grained features, the training coarse-grained features, and the information entropy.

[0042] The fusion module is further configured to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight coefficients.

[0043] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0044] Obtain the fine-grained and coarse-grained features of the media to be recommended;

[0045] The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector;

[0046] The fine-grained and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector;

[0047] The fine-grained representation vector and the coarse-grained representation vector are weighted and fused to obtain a fused vector;

[0048] Based on the fusion vector and the object representation vector of the candidate objects, a target object is selected from the candidate objects, and recommendation information of the media to be recommended is pushed to the target object.

[0049] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0050] Obtain the fine-grained and coarse-grained features of the media to be recommended;

[0051] The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector;

[0052] The fine-grained and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector;

[0053] The fine-grained representation vector and the coarse-grained representation vector are weighted and fused to obtain a fused vector;

[0054] Based on the fusion vector and the object representation vector of the candidate objects, a target object is selected from the candidate objects, and recommendation information of the media to be recommended is pushed to the target object.

[0055] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0056] Obtain the fine-grained and coarse-grained features of the media to be recommended;

[0057] The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector;

[0058] The fine-grained and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector;

[0059] The fine-grained representation vector and the coarse-grained representation vector are weighted and fused to obtain a fused vector;

[0060] Based on the fusion vector and the object representation vector of the candidate objects, a target object is selected from the candidate objects, and recommendation information of the media to be recommended is pushed to the target object.

[0061] The aforementioned information recommendation methods, devices, computer equipment, storage media, and computer program products divide the features of the media to be recommended into fine-grained features and coarse-grained features of varying degrees of granularity. The coarse-grained features undergo feature generalization processing to obtain a coarse-grained representation vector with strong media generalization ability. Meanwhile, the fine-grained and coarse-grained features undergo feature enhancement processing to obtain a fine-grained representation vector that captures the media's posterior information. Then, the fine-grained and coarse-grained representation vectors are weighted and fused, effectively balancing generalization and memorization. Therefore, when selecting target objects for recommendation based on the fused vector and the object representation vector of candidate objects, the recommendation accuracy of the media to be recommended can be effectively improved, which is beneficial to increasing the conversion rate of the media to be recommended. Attached Figure Description

[0062] Figure 1 This is a diagram illustrating the application environment of an information recommendation method in one embodiment.

[0063] Figure 2 This is a flowchart illustrating an information recommendation method in one embodiment;

[0064] Figure 3 This is a schematic diagram of a page with ad granularity division in one embodiment;

[0065] Figure 4 This is a schematic diagram of the structure of a machine learning model in one embodiment;

[0066] Figure 5 This is a schematic diagram of the feature extraction network structure in one embodiment;

[0067] Figure 6 This is a schematic diagram of a page displaying images and generation requirements in one embodiment;

[0068] Figure 7 This is a schematic diagram of a page displaying recommendation information in one embodiment;

[0069] Figure 8 This is a schematic diagram of the process of training a machine learning model in one embodiment;

[0070] Figure 9 This is a schematic diagram illustrating the training of a machine learning model in one embodiment;

[0071] Figure 10 This is a schematic diagram illustrating the training of a machine learning model in another embodiment;

[0072] Figure 11 This is a structural block diagram of an information recommendation device in one embodiment;

[0073] Figure 12 This is a structural block diagram of the information recommendation device in another embodiment;

[0074] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0076] It should be noted that in the following description, the terms "first, second, and third" are used only to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0077] Before describing the embodiments of this application, the technology involved in this application will be explained in detail below:

[0078] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0079] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, autonomous driving, natural language processing, and machine learning / deep learning.

[0080] The information recommendation method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0081] Server 104 can implement the information recommendation method of this application, sending the recommendation information of the media to be recommended to the terminal 102 corresponding to the target audience (such as a group interested in the media to be recommended), so that the terminal 102 can display the recommendation information of the media to be recommended, thereby increasing the click-through rate of the media to be recommended. For example, server 104 can push the advertising information of XX new energy vehicle to users interested in new energy vehicles, thereby increasing the click-through rate of the advertising information.

[0082] The terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, IoT device, or portable wearable device. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices, etc.

[0083] Server 104 can be a standalone physical server or a service node in a blockchain system. These service nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on top of the Transmission Control Protocol (TCP). Furthermore, server 104 can also be a server cluster composed of multiple physical servers, and can be a cloud server providing 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 (CDNs), and big data and artificial intelligence platforms.

[0084] Terminal 102 and server 104 can be connected via Bluetooth, USB (Universal Serial Bus) or network, etc., and this application does not impose any restrictions.

[0085] In one embodiment, such as Figure 2 As shown, an information recommendation method is provided, which can be derived from... Figure 1 The method is executed by a server or terminal, or by a server and terminal working together. Figure 1 Taking the server execution in [the context of the example] as an example, the steps include:

[0086] S202, obtain the fine-grained and coarse-grained features of the media to be recommended.

[0087] The media to be recommended can be at least one of the following: advertisements, videos, music, or live streams. The advertisements can be video ads, image ads, audio ads, text ads, or graphic ads targeting hardware products, software, or other products (such as consumer goods). The software can be applications or mini-programs in various fields, specifically game applications, social applications, shopping applications, video applications, and generative applications. Generative applications can generate corresponding articles, lyrics, images, animations, or videos based on given information. It should be noted that the media to be recommended can be cold-start media, such as newly launched media; alternatively, it can be hot-start media, such as mature media already on the market.

[0088] When using fixed feature values, if the number of media to be recommended covered by media feature 'a' is less than the number of media to be recommended covered by media feature 'b', then media feature 'a' is a fine-grained feature, and media feature 'b' is a coarse-grained feature. In other words, with fixed feature values, the number of media to be recommended covered by a fine-grained feature is less than the number covered by a coarse-grained feature. For example, regarding ad identifiers (IDs) and ad types, assuming the ad ID is "X brand car promotion poster" and the ad type is "new energy vehicle category," then the number of ads covered by "X brand car promotion poster" is significantly less than the number of ads covered by "new energy vehicle category." Therefore, "X brand car promotion poster" is a fine-grained feature, and "new energy vehicle category" is a coarse-grained feature. Here, ad type can also be called ad category.

[0089] In one embodiment, the server obtains the media features of the media to be recommended, and performs granular segmentation on the media features to obtain fine-grained features and coarse-grained features of the media to be recommended.

[0090] Specifically, the server obtains at least two media features corresponding to the media to be recommended; determines the coarse-to-fine granularity ratio, and divides each feature in the at least two media features into fine-grained features and coarse-grained features according to the coarse-to-fine granularity ratio to obtain the fine-grained features and coarse-grained features of the media to be recommended; or, in response to the granularity division request, divides each feature in the media features into fine-grained features and coarse-grained features.

[0091] The coarse-grained to fine-grained ratio can be the ratio between the number of coarse-grained features and the number of fine-grained features. This ratio can be an empirical value, derived from multiple experiments. For example, a ratio of 1:1 between coarse-grained and fine-grained features indicates that coarse-grained features account for 50% of the total features, and fine-grained features account for 50% of the total features.

[0092] After obtaining the coarse-to-fine granularity ratio, the server can automatically classify the media features of the media to be recommended into fine-grained features and coarse-grained features according to this ratio. For example, if the ratio between the number of coarse-grained features and the number of fine-grained features is 1:1, it means that coarse-grained features account for 50% of the total number of features, and fine-grained features account for 50% of the total number of features. In this case, the server can automatically classify 50% of the media features of the media to be recommended as fine-grained features and the other 50% as coarse-grained features.

[0093] Alternatively, the server can granularize media features according to the user's actual needs. The specific granularization steps include: the user can trigger a granularity granularization request through the client, and the client sends the granularity granularization request to the server. The granularity granularization request can carry feature identifiers and corresponding categories; the server divides the corresponding features of at least two media features into fine-grained features and coarse-grained features according to the feature identifiers and corresponding categories.

[0094] For example, such as Figure 3 As shown, on the client's granularity classification page, the user selects the ad identifier and advertiser identifier to be classified as fine-grained features, and selects the ad type and the industry to be classified as coarse-grained features. After clicking confirm, the client generates a granularity classification request based on the ad identifier, advertiser identifier, fine-grained features, ad type, industry to be classified as coarse-grained features, and sends the granularity classification request to the server. After receiving the granularity classification request, the server classifies the ad identifier and advertiser identifier as fine-grained features, and classifies the ad type and the industry to be classified as coarse-grained features.

[0095] In another embodiment, the server obtains at least two media features corresponding to the media to be recommended; determines the granularity entropy of each feature among the at least two media features; and divides each feature among the at least two media features into fine-grained features and coarse-grained features based on the granularity entropy to obtain the fine-grained features and coarse-grained features of the media to be recommended.

[0096] Among them, considering that the typical characteristics of fine-grained features are a large number of candidate feature values ​​and large frequency differences, granularity entropy is used as the basis for distinguishing between coarse and fine-grained features.

[0097] In one embodiment, the step of determining the granular entropy of each feature among at least two media features may specifically include: the server determining the number of interactive objects corresponding to different feature values ​​for each feature among the at least two media features; determining a first conversion rate for each feature with different feature values ​​based on the number of interactive objects and the total number of objects in the media to be recommended; and determining the granular entropy of each feature among the at least two media features based on the conversion rate for different feature values ​​and the corresponding number of objects. Alternatively, the server determines the number of object interactions corresponding to different feature values ​​for each feature among the at least two media features; determines a second conversion rate for each feature with different feature values ​​based on the number of object interactions and the total number of interactions in the media to be recommended; and determines the granular entropy of each feature among the at least two media features based on the second conversion rate for different feature values ​​and the corresponding number of object interactions.

[0098] The number of interactive objects can be the number of recommended media outlets that have generated conversions. For example, in an ad recommendation scenario, the number of interactive objects could be the number of clicked ads. The number of object interactions can be the number of interactions or the duration of interactions corresponding to the recommended media outlets that have generated conversions. For example, in a video recommendation scenario, the number of object interactions could be the number of video clicks, comments, likes, shares, favorites, and viewing time. The total number of interactions for recommended media outlets can be the total number of interactions for all media outlets.

[0099] For a media feature i with K possible values, the conversion rate for each feature value k is calculated as follows:

[0100]

[0101] in, n represents the conversion rate. i＝k∧label＝1 The number of interactive objects corresponds to the media feature i taking the value k, and N is the total number of objects in the media to be recommended. For example, in an advertising recommendation scenario, n i＝k∧label＝1 This can represent the number of ads that convert when media feature i is the ad type. For example, in a video recommendation scenario, n... i＝k∧label＝1 This can represent the number of videos that are clicked or watched when media feature i is a video type (such as comedy, drama, or science fiction).

[0102] Or, n i＝k∧label＝1 The value of media feature i represents the number of object interactions when i is k, and N is the total number of interactions for the media to be recommended. For example, in an advertising recommendation scenario, n i＝k∧label＝1 This can represent the number of interactions when media feature i is an ad type (such as new energy vehicle ads), such as the number of ad clicks or viewing duration. For example, in a video recommendation scenario, n... i＝k∧label＝1 This can represent the number of interactions when media feature i is a video type (such as comedy, drama, or science fiction), such as at least one of the following: number of video clicks, viewing time, number of comments, number of likes, number of shares, or number of favorites.

[0103] After calculating the conversion rate of each feature at different feature values ​​k, the granular entropy can be calculated using the following formula:

[0104]

[0105] Among them, the smaller the granularity entropy, the more concentrated the number of interactive objects or the number of object interactions are on a few feature values, and the more the media features are biased towards fine granularity. In practice, granularity entropy < 0.001 can be selected as the criterion for judging coarse and fine granularity.

[0106] S204 performs feature generalization processing on the coarse-grained features to obtain the coarse-grained representation vector.

[0107] Since coarse-grained features cover a large number of media to be recommended when taking fixed feature values, the resulting coarse-grained representation vector after feature generalization processing can be a feature vector with strong generalization ability, that is, it can have strong new media generalization ability, such as strong new advertising generalization ability.

[0108] In one embodiment, the server concatenates the coarse-grained features belonging to the same media to be recommended to obtain the first concatenated feature of each media to be recommended. Then, the server uses a coarse-grained network to perform feature generalization processing on the coarse-grained features to obtain a coarse-grained representation vector.

[0109] Specifically, the server concatenates the coarse-grained features belonging to the same media to be recommended, obtaining the first concatenated feature for each media to be recommended. Each first concatenated feature is then sequentially input into the coarse-grained network of the recommendation model, so that the multilayer perceptron in the coarse-grained network processes the input first concatenated features to obtain a coarse-grained representation vector. For example, each layer in the multilayer perceptron performs linear processing on the features input to that layer based on the network parameters, and performs nonlinear processing on the result to finally obtain the coarse-grained representation vector. For details, refer to the following calculation formulas (1) and (2).

[0110] Before concatenation, the coarse-grained features of the media to be recommended can be vectorized to obtain the embedded representations corresponding to the coarse-grained features. Then, the embedded representations corresponding to the coarse-grained features of the same media to be recommended are concatenated to obtain the first concatenated feature. It should be noted that this embedded representation can also be called an embedded representation, specifically a feature embedding vector. Therefore, for ease of description, in this embodiment, the embedded representation corresponding to the coarse-grained feature can be simply referred to as coarse-grained embedding.

[0111] For example, in an ad recommendation scenario, the embedding representations corresponding to the coarse-grained features of an ad are used. Perform splicing, and then obtain the first splicing feature e coarse ,in:

[0112]

[0113] It should be noted that, It can be the embedded representation of the i-th coarse-grained feature, e coarse It can be a concatenation of all coarse-grained feature embeddings. This is the splicing symbol. After obtaining the first splicing feature e... coarse Next, the first splicing feature e coarse Input into a coarse-grained network, such as Figure 4 As shown. Among them, the first splicing feature ecoarse This refers to coarse-grained embedding after splicing.

[0114] The first splicing feature e is obtained by using a multilayer perceptron in a coarse-grained network. coarse Feature generalization is performed to obtain a coarse-grained representation vector. The function expression for the generalization process is as follows:

[0115] r coarse =f coarse (e coarse (1)

[0116] In the above formula, r coarse For coarse-grained representation vectors, the network function f coarse It can be represented as:

[0117] r (l+1) =ReLU(W (l) r (l) +b (l) (2)

[0118] Among them, W (l) and b (l) r is the learnable parameter of the l-th layer. (0) =e coarse It should be noted that the ReLU function mentioned above is only one type of processing function; other functions, such as the sigmoid function and the tanh function, can also be used for processing.

[0119] S206 performs feature enhancement processing on fine-grained and coarse-grained features to obtain fine-grained representation vectors.

[0120] Although fine-grained features cover fewer media to be recommended when taking fixed feature values, they have high accuracy. By concatenating fine-grained features and coarse-grained features and performing feature enhancement processing, a fine-grained representation vector can be obtained. This fine-grained representation vector captures the posterior information of old media, such as the posterior information of old advertisements.

[0121] In one embodiment, the server concatenates fine-grained and coarse-grained features to obtain a second concatenated feature. This second concatenated feature is then sequentially input into the fine-grained network of the recommendation model, allowing a multilayer perceptron within the fine-grained network to process the input second concatenated feature and obtain a fine-grained representation vector. For example, each layer of the multilayer perceptron performs linear processing on the input features based on network parameters and then performs nonlinear processing on the result to obtain the fine-grained representation vector.

[0122] Before concatenation, the fine-grained features of the media to be recommended can be vectorized to obtain the embedded representations corresponding to the fine-grained features. Then, the embedded representations corresponding to the fine-grained features of the same media to be recommended are concatenated to obtain the first concatenated feature. This embedded representation can also be called a feature embedding vector. For ease of description, the embedded representation corresponding to the fine-grained features can be simply referred to as fine-grained embedding.

[0123] For the extraction process of fine-grained representation vectors, please refer to S204 above.

[0124] S208, weighted and fused the fine-grained representation vector and the coarse-grained representation vector to obtain the fused vector.

[0125] Among them, the fusion vector can refer to the representation vector of the media to be recommended, such as the representation vector of advertisements, video representation vectors, music representation vectors, and live room representation vectors.

[0126] In one embodiment, the server can utilize a lightweight gating network to adaptively weight and fuse fine-grained representation vectors and coarse-grained representation vectors to obtain a fused vector.

[0127] For example, a gating network determines the weight parameters corresponding to fine-grained and coarse-grained representation vectors, and then uses these weight parameters to perform a weighted fusion of the two vectors to obtain a fused vector. The weight parameters for the fine-grained and coarse-grained representation vectors are different.

[0128] Specifically, the server determines the magnitude or information entropy of the feature embedding vector of the media to be recommended; based on the fine-grained features, coarse-grained features, and magnitude, it determines the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively; or, based on the fine-grained features, coarse-grained features, and information entropy, it determines the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively. S208 may specifically include: the server performing weighted fusion of the fine-grained representation vector and the coarse-grained representation vector based on the weight parameters.

[0129] In this context, both the magnitude and information entropy of the feature embedding vector can be used to represent the convergence of the media to be recommended. For example, in the scenario of ad recommendation, the convergence of the ad can be represented by the magnitude or information entropy of the ad ID feature embedding.

[0130] After obtaining the embedded representation of the convergence of the media to be recommended, the embedded representations of fine-grained features, coarse-grained features, and the embedded representation of the convergence of the media to be recommended are concatenated. The concatenated result is then input into a gating network, such as... Figure 4 As shown, the corresponding weight parameters can be obtained by performing feature processing on the splicing result through this gating network.

[0131] The weight parameters can be calculated using the following formula:

[0132]

[0133] Among them, W g e is a learnable parameter in a gated neural network. pop It is an embedding representation describing convergence, such as the embedding representation of advertising convergence, w fuse It is the output of the gating network.

[0134] After calculating the weight parameters of the fine-grained and coarse-grained representation vectors, these weight parameters are used to perform a weighted fusion of the two vectors to obtain a fused vector. The specific fusion expression is as follows:

[0135]

[0136] Where, r ad For the fusion vector, r fine For fine-grained representation vectors, The weight parameters are for the fine-grained representation vector. These are the weight parameters for the coarse-grained representation vector.

[0137] S210: Based on the fusion vector and the object representation vector of the candidate objects, select the target object from the candidate objects and push the recommendation information of the recommended media to the target object.

[0138] Here, candidate objects can refer to potential audience groups. Object representation vectors can refer to the feature vectors of candidate objects. Target objects can refer to individuals selected from the candidate objects who may be interested in the recommended media.

[0139] Recommendation information can be media data, key information, or key segments of the media to be recommended; in addition, the recommendation information can also be other types of information generated using media data or key information of media data, such as text-type advertising information, or recommendation information of corresponding data types (such as video type, animation type, or audio type) generated by using a multimodal large model to utilize the preferences of the target audience.

[0140] The aforementioned media data can refer to all information about the media to be recommended; key information can be key content of the media to be recommended, such as keywords, key sentences, or key images; key segments can be popular segments of the media to be recommended.

[0141] For example, in advertising recommendation scenarios, the recommendation information can be the key content of the advertisement (or the entire advertisement content). For instance, if the target audience is interested in discounted brand clothing, the recommendation can display images of the brand clothing and the corresponding discount in a visually appealing way. In video recommendation scenarios, the recommendation information can be a highlight of the video, such as the segment with the most comments or a clip featuring a popular celebrity. This video could be a TV series, movie, game video, sports video, or entertainment video. If the video is a live stream, the recommendation information can be the live feed. In live streaming room recommendation scenarios, the recommendation information can also be the live feed. In music recommendation scenarios, the recommendation information can be a climax of the music, such as the clips most commonly used in short videos.

[0142] In one embodiment, the extraction of object representation vectors may include the following steps: the server can vectorize the object information of candidate objects to obtain object embeddings, and then extract features from the object embeddings of candidate objects through a feature extraction network to obtain object representation vectors of candidate objects, such as... Figure 5 As shown.

[0143] The object information can be at least one of the following: the object identifier of the candidate object, media interaction information, or interest information.

[0144] In one embodiment, obtaining recommendation information may include the following steps: the server acquires media data of the media to be recommended; if the media data is text data, recommendation information corresponding to the data type of interest to the target object can be generated based on the media data or key information in the media data. Alternatively, the server acquires recommendation information corresponding to the data type specified by the media provider from a database; this recommendation information is generated by the server based on a business request initiated by the client.

[0145] For example, if the target audience prefers video content, then video recommendations can be generated; if the target audience prefers image content, then image recommendations can be generated; and if the target audience prefers audio content, then audio recommendations can be generated. Furthermore, a multimodal, large-scale model can be used to generate recommendations for the corresponding categories.

[0146] In another embodiment, the specific steps for obtaining recommendation information may include: if the media data is image data, the server obtains recommendation information corresponding to the image type from the information database. This recommendation information is generated by the server based on the business request initiated by the client.

[0147] For example, such as Figure 6As shown, media providers can input the required car image and corresponding demand information for advertising on the session page. A business request will then be generated based on this information. Upon receiving the business request, the server can generate video and graphic ads related to the car, storing them in a database. This allows for the delivery of video or graphic ads to the target audience based on their preferences when needed. It should be noted that the server can also generate audio ads based on the actual business request and the car image and demand information.

[0148] In one embodiment, the server determines a first score value for a candidate object based on the fusion vector and the object representation vector of the candidate object; and selects a target object from the candidate objects based on the first score value.

[0149] The first score can be a predicted score, reflecting the conversion rate of the candidate after receiving the recommended media. The candidate can include all people using media applications (such as video and music), all people using social applications, and all people using applications associated with the media application.

[0150] When the target audience is a user of a media application, the server can push recommended media information to the target audience's media application account. For example, while the target audience is watching a video, the server can push video advertisements for food or electronic products that the target audience may be interested in, or game videos that the target audience may be interested in, or live streams that the target audience may be interested in, to the target audience's video account.

[0151] When the target audience is a user of a social media application, the server can push recommended media information to that user's social media account. For example, when the target audience enters a relevant module (such as a video channel) of the social media application, the server can use the user's social media account to push video ads for food or electronic products that the user might be interested in, or game videos that the user might be interested in, or live streams that the user might be interested in. Alternatively, when the target audience enters the user's profile or community of the social media application, the server can use the user's social media account to push richly illustrated advertisements, such as... Figure 7 As shown.

[0152] In the above embodiments, the features of the media to be recommended are divided into fine-grained features and coarse-grained features of different degrees of coarseness. The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector with strong media generalization ability. The fine-grained features and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector that can capture the media's posterior information. Then, the fine-grained representation vector and the coarse-grained representation vector are weighted and fused to effectively balance generalization and memory. Therefore, when selecting target objects for recommendation based on the fused vector and the object representation vector of the candidate objects, the recommendation accuracy of the media to be recommended can be effectively improved, which is conducive to improving the conversion rate of the media to be recommended.

[0153] In one embodiment, the coarse-grained representation vector, fine-grained representation vector, and fused vector described above are obtained through a machine learning model; such as Figure 8 As shown, the method also includes:

[0154] S802, obtain the training fine-grained features and training coarse-grained features of the media samples.

[0155] The media samples can be media used for model training, including at least one of advertisements, videos, music, or live streams. These media samples can be cold-start media, such as newly launched media; alternatively, they can be hot-start media, such as mature media already on the market.

[0156] The training fine-grained features and training coarse-grained features refer to the fine-grained and coarse-grained features corresponding to the media samples used in the training phase, respectively. With fixed feature values, the number of media samples covered by the training fine-grained features is less than the number covered by the training coarse-grained features.

[0157] S804 performs feature generalization processing on the training coarse-grained features through the first network branch of the machine learning model to obtain the training coarse-grained representation vector.

[0158] The first network branch can be a coarse-grained network in a machine learning model. This machine learning model can be an Automatic Fusion Network (AutoFuse), as shown in [reference needed]. Figure 4 .

[0159] The training coarse-grained representation vector can be the coarse-grained representation vector obtained by performing feature generalization processing on the training coarse-grained features during the training phase. Since the training coarse-grained features cover a large number of media samples when taking fixed feature values, the training coarse-grained representation vector obtained after feature generalization processing can be a feature vector with strong generalization ability, that is, it can have strong new media generalization ability, and thus strong new advertising generalization ability.

[0160] S806 uses the second network branch of the machine learning model to perform feature enhancement processing on the training fine-grained features and the training coarse-grained features, thereby obtaining the training fine-grained representation vector.

[0161] The second network branch can refer to the fine-grained network in a machine learning model.

[0162] Although the number of media samples covered by the fine-grained training features is small when taking fixed feature values, the coverage is highly accurate. By concatenating the fine-grained training features and the coarse-grained training features and performing feature enhancement processing, a fine-grained training representation vector can be obtained. This fine-grained training representation vector captures the posterior information of old media, such as the posterior information of old advertisements.

[0163] S808 uses a fusion network of machine learning models to weightedly fuse the training fine-grained representation vector and the training coarse-grained representation vector to obtain the training fusion vector.

[0164] Among them, the fusion vector can refer to the representation vector of the media sample, such as the advertising representation vector, video representation vector, music representation vector, and live room representation vector.

[0165] In one embodiment, the server determines the training degree of the media sample; the training degree is used to represent the cumulative number of iterations in which the media sample participates in training; based on the training degree, fine-grained training features, and coarse-grained training features, the server determines the weight coefficients corresponding to the fine-grained training representation vector and the coarse-grained training representation vector, respectively; then, through the fusion network of the machine learning model, the fine-grained training representation vector and the coarse-grained training representation vector are weighted and fused based on the weight coefficients.

[0166] In another embodiment, the server determines the magnitude or information entropy of the feature embedding vector of the media sample; based on the training fine-grained features, training coarse-grained features, and magnitude, it determines the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector, respectively; or, based on the training fine-grained features, training coarse-grained features, and information entropy, it determines the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector, respectively; then, through the fusion network of the machine learning model, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused based on the weight coefficients.

[0167] The specific implementation steps of S802 to S808 above can be found by referring to Figure 2 S202 to S208 in the embodiment.

[0168] S810, based on the training fusion vector and the training object representation vector of the candidate object sample, determine the second score value of the candidate object sample.

[0169] The second score can be a score predicted during the training phase, used to reflect the conversion rate of candidate object samples after receiving media samples. Candidate object samples can include all people using media applications (such as video and music), all people using social applications, and all people using applications associated with the media application.

[0170] S812 optimizes the parameters of the machine learning model based on the loss between the second score value and the score label.

[0171] The score label can refer to the conversion status of the candidate object sample to the media sample, such as whether the recommended video, advertisement, live room or music was clicked. If it was clicked, the score label is 1, and if it was not clicked, the score label is 0.

[0172] In one embodiment, the server uses the main loss function to calculate the loss between the second score and the score label, obtaining the corresponding loss. Then, it uses this loss to optimize the parameters of the machine learning model, specifically optimizing the parameters of the fine-grained network, coarse-grained network, and gating network within the machine learning model. (See reference...) Figure 9 Furthermore, this loss can be used to optimize the parameters of the feature extraction network. The structure of the feature extraction network can be found by referring to... Figure 5 .

[0173] To ensure each network branch is trained more thoroughly and the model achieves better results, an auxiliary loss function can be introduced during model training. The specific steps include: the server determines the third score value of a candidate object sample based on the training coarse-grained representation vector and the training object representation vector of the candidate object sample; it determines the fourth score value of the candidate object sample based on the training fine-grained representation vector and the training object representation vector of the candidate object sample; the coarse-grained loss function is used to determine the loss between the third score value and the score label, and the fine-grained loss function is used to determine the loss between the fourth score value and the score label. Then, the machine learning model parameters are optimized based on these three losses. (See reference...) Figure 10 This allows the gated network to allocate weights more reasonably between the two network branches based on media convergence, thereby enabling each network branch of the machine learning model to be trained more fully.

[0174] In the above embodiments, the features of media samples are divided into training fine-grained features and training coarse-grained features of different degrees of coarseness. The training coarse-grained features are subjected to feature generalization processing to obtain a training coarse-grained representation vector with strong media generalization ability. The training fine-grained features and training coarse-grained features are subjected to feature enhancement processing to obtain a training fine-grained representation vector that can capture the media posterior information. Then, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused to effectively balance generalization and memorization. Therefore, the second score value is calculated using the training object representation vector and the training fused vector of the candidate object sample. The parameters of the machine learning model are optimized based on the loss between the second score value and the score label, so that the model can learn media features that balance generalization and memorization, thereby improving the generalization and accuracy of the model.

[0175] To better understand the technical solution of this application, it is explained here in the context of an advertising recommendation scenario, as follows:

[0176] The machine learning model used in the advertising recommendation scenario is AutoFuse, such as... Figure 4 As shown, the AutoFuse mainly consists of four parts:

[0177] Feature access layer: Divides the various features of the advertisement (referred to as advertisement features) into coarse-grained features and fine-grained features according to granularity entropy;

[0178] Prediction network layer: a coarse-grained network and a fine-grained network constructed through an asymmetric sharing method. The input of the coarse-grained network is the embedded representation of the coarse-grained features, which yields a coarse-grained advertising representation vector. The input of the fine-grained network is the embedded representation of the coarse-grained features and the fine-grained features, respectively, which yields a fine-grained advertising representation vector.

[0179] Representation fusion layer: The gating network dynamically determines the weights of coarse-grained and fine-grained ad representation vectors according to the training stage of the ad, and completes the fusion between coarse-grained and fine-grained ad representation vectors, balancing generalization ability and memory ability.

[0180] Training strategy: Introduce multi-task training technology to ensure that the gradient is fully transmitted to both coarse-grained and fine-grained networks during the training process, while achieving efficient learning of generalization and memory capabilities.

[0181] (I) Feature Access Layer

[0182] The feature access layer can divide advertising features into coarse-grained features and fine-grained features. Coarse-grained features are an important basis for the generalization ability of new ads, while fine-grained features are an important support for the memory ability of old ads. By explicitly distinguishing these two types of features in the feature access layer, it is easy to construct coarse-grained and fine-grained ad representation vectors in the next layer through asymmetric sharing.

[0183] Since fine-grained features are typically characterized by a large number of candidate feature values ​​and significant frequency differences, granularity entropy can be used as the basis for distinguishing between coarse and fine granularity of feature groups. Specifically, for an advertising feature i with K possible values, the conversion rate for each feature value k can be expressed as:

[0184]

[0185] Where, n i＝k∧label＝1 The number of positive samples where ad feature i takes the value k, such as the number of ads that converted when ad feature is ad ID; N is the total number of samples, such as the total number of ads. Granularity entropy is defined as:

[0186]

[0187] The smaller the granularity entropy, the more concentrated the positive sample distribution is on a few feature values, and the more the features are biased towards fine granularity. In practice, a granularity entropy of <0.001 can be selected as the criterion for judging coarse-grained features.

[0188] (II) Network Layer Prediction

[0189] The prediction network layer is mainly constructed through asymmetric sharing, consisting of coarse-grained and fine-grained networks. This prediction network layer accepts feature embeddings as input and outputs coarse-grained and fine-grained advertising representation vectors.

[0190] In this context, the coarse-grained network accepts the embedded representation of coarse-grained features (referred to as coarse-grained embedding) as network input:

[0191]

[0192] In the above formula, It is the embedding representation of the i-th coarse-grained feature, e coarse It is a concatenation of all coarse-grained feature embeddings. The coarse-grained network extracts coarse-grained advertising representation vectors from the input coarse-grained embeddings using a multilayer perceptron.

[0193] r coarse =f coarse (e coarse )

[0194] In the above formula, the network function f coarse It can be represented as:

[0195] r (l+1) =ReLU(W (l) r (l) +b (l) )

[0196] Among them, W (l) and b (l) r is the learnable parameter of the l-th layer. (0) =e coarse .

[0197] Fine-grained networks have a similar structure to coarse-grained networks, the main difference being that fine-grained networks simultaneously accept embedded representations corresponding to both coarse-grained and fine-grained features as input.

[0198]

[0199] Among them, e fine It is a splicing of the embedded representations of all fine-grained and coarse-grained features.

[0200] (III) Characterization Fusion Layer

[0201] Coarse-grained ad representation vectors exhibit strong generalization ability for new ads, while fine-grained ad representation vectors capture posterior information from older ads. To combine these two approaches, this application employs a lightweight gating network to adaptively fuse the representations.

[0202]

[0203] Among them, W g It is a learnable parameter, e pop It is an embedded representation describing the convergence of advertising, w fuse This is the output of the gating network. The final fused advertising representation is:

[0204]

[0205] AutoFuse uses the cumulative number of training iterations participated in by the advertisement (referred to as the training participation rate) as a representation of the advertisement's convergence. In practice, other candidate methods can also be used to represent the advertisement's convergence, such as using the modulus or information entropy embedded in the advertisement ID features.

[0206] (iv) Training Strategies

[0207] The ad representation obtained through AutoFuse can be flexibly applied to ranking models in most ad recommendation domains. Taking the classic dual-tower conversion rate model as an example, the final prediction result of the network can be represented as follows:

[0208]

[0209] Where, r u This is the representation vector output from the user side. Conversion rate prediction results are typically trained using cross-entropy loss.

[0210]

[0211] Experiments revealed that adding multiple loss functions allows each side of the AutoFuse network to be trained more fully, thereby enabling the gating network mechanism to function properly and more rationally distribute weights between the two sides of the network based on the convergence of the advertisement.

[0212] The final loss function is:

[0213] Loss = Loss final +Loss fine +Loss coarse

[0214] The loss functions for the two levels are as follows:

[0215]

[0216]

[0217] By implementing the above solution, the following technical effects can be achieved:

[0218] (1) Comparative Experiment

[0219] As shown in Table 1, Table 1 illustrates the effectiveness of applying the present invention to prediction tasks on two sets of real datasets. All experimental results consist of the mean AUC and its variance from three repeated experiments, with the best results shown in bold.

[0220] Table 1

[0221]

[0222] By simultaneously exploring both fine-grained and coarse-grained features, AutoFuse outperformed all comparison algorithms, achieving optimal results on both new and old ads, as well as the full set of ads, validating its feasibility and superiority in mitigating the cold start problem in large-scale scenarios. Compared to currently deployed DNN models, AutoFuse achieved AUC gains of 0.76% and 0.73% on new ads within the community and public account datasets. Furthermore, the results demonstrate that the model's predictions for old ads and the full set of ads are also enhanced, further proving the synergistic optimization effect of AutoFuse in predicting both new and old ads.

[0223] (2) Ablation test

[0224] To further understand the characteristics of the AutoFuse model, ablation experiments were conducted on the model's key design points. First, the impact of different input structures on prediction performance was tested experimentally, and the results are shown in Table 2.

[0225] Table 2

[0226] Input structure New Advertisement Decrease Old Advertisement Decrease All Ads Decrease Asymmetric input 0.8251 - 0.852 - 0.8478 - Symmetrical input 0.8186 -0.79％ 0.8513 -0.08％ 0.8432 -0.54％ Feature isolation 0.8107 -1.75％ 0.7864 -7.70％ 0.8023 -5.37％

[0227] Experiments show that symmetric sharing, compared to AutoFuse's asymmetric sharing, results in a 0.79% decrease in AUC for new ads, while the decrease is relatively lower for older ads (0.08%). This is mainly because new ads are more significantly affected by fine-grained features, while older ads, which are primarily characterized by fine-grained features, are less affected. Feature isolation exhibits a significant performance degradation across both new and older ads, particularly for older ads, demonstrating that coarse-grained features also play a crucial role for older ads.

[0228] Furthermore, the impact of different convergence representations on the selection efficiency of the gating network was experimentally tested. Several convergence representations were mainly compared:

[0229] (1) Multi-segment binning based on training degree (referred to as multi-segment binning): The convergence of the advertising vector is represented by mapping the training degree to multiple interval segments.

[0230] (2) Divide into two segments according to the training level (referred to as two-segment binning): the same as the previous method, but only two interval segments are distinguished;

[0231] (3) Non-convergence based on the degree of participation (abbreviated as non-convergence).

[0232] The corresponding results are shown in Table 3:

[0233] Table 3

[0234] Convergence characterization New Advertisement Decrease Old Advertisement Decrease All Ads Decrease Multi-stage bins 0.8251 - 0.852 - 0.8478 - Two-stage bucket 0.8187 -0.78％ 0.8453 -0.79％ 0.8441 -0.44％ none 0.8188 -0.76％ 0.8453 -0.79％ 0.8441 -0.44％

[0235] Experiments show that incorporating training degree representation can effectively improve the selection accuracy of gating networks.

[0236] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0237] Based on the same inventive concept, this application also provides an information recommendation apparatus for implementing the information recommendation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more information recommendation apparatus embodiments provided below can be found in the limitations of the information recommendation method described above, and will not be repeated here.

[0238] In one embodiment, such as Figure 11 As shown, an information recommendation device is provided, including: an acquisition module 1102, a first processing module 1104, a second processing module 1106, a fusion module 1108, and a recommendation module 1110, wherein:

[0239] The acquisition module 1102 is used to acquire the fine-grained features and coarse-grained features of the media to be recommended;

[0240] The first processing module 1104 is used to perform feature generalization processing on the coarse-grained features to obtain a coarse-grained representation vector.

[0241] The second processing module 1106 is used to perform feature enhancement processing on the fine-grained features and coarse-grained features to obtain a fine-grained representation vector.

[0242] The fusion module 1108 is used to weightedly fuse the fine-grained representation vector and the coarse-grained representation vector to obtain a fused vector;

[0243] The recommendation module 1110 is used to select a target object from the candidate objects based on the fusion vector and the object representation vector of the candidate objects, and push the recommendation information of the media to be recommended to the target object.

[0244] In one embodiment, the acquisition module 1102 is further configured to acquire at least two media features corresponding to the media to be recommended; determine the granularity entropy of each feature among the at least two media features; and divide each feature among the at least two media features into fine-grained features and coarse-grained features based on the granularity entropy to obtain the fine-grained features and coarse-grained features of the media to be recommended.

[0245] In one embodiment, the acquisition module 1102 is further configured to determine the number of interactive objects corresponding to different feature values ​​for each feature among at least two media features; determine the first conversion rate for each feature among at least two media features when taking different feature values ​​based on the number of interactive objects and the total number of objects in the media to be recommended; and determine the granular entropy of each feature among at least two media features based on the conversion rate when taking different feature values ​​and the corresponding number of objects.

[0246] In one embodiment, the acquisition module 1102 is further configured to: determine the number of object interactions corresponding to different feature values ​​for each feature among at least two media features; determine the second conversion rate for each feature among at least two media features when taking different feature values ​​based on the number of object interactions and the total number of interactions for the media to be recommended; and determine the granular entropy of each feature among at least two media features based on the second conversion rate when taking different feature values ​​and the corresponding number of object interactions.

[0247] In one embodiment, the acquisition module 1102 is further configured to acquire at least two media features corresponding to the media to be recommended; determine the coarse-fine granularity ratio, and divide each feature in the at least two media features into fine-grained features and coarse-grained features according to the coarse-fine granularity ratio to obtain the fine-grained features and coarse-grained features of the media to be recommended; or, in response to the granularity division request, divide each feature in the at least two media features into fine-grained features and coarse-grained features.

[0248] In one embodiment, the first processing module 1104 is further configured to concatenate the coarse-grained features belonging to the same media to be recommended, respectively, to obtain the first concatenated feature of each media to be recommended; and sequentially input each first concatenated feature into the coarse-grained network of the recommendation model, so that the multilayer perceptron in the coarse-grained network performs feature processing on the input first concatenated features to obtain a coarse-grained representation vector.

[0249] In one embodiment, the second processing module 1106 is further configured to concatenate the fine-grained features and the coarse-grained features to obtain a second concatenated feature; and to sequentially input the second concatenated feature into the fine-grained network of the recommendation model so that the fine-grained network performs linear processing on the second concatenated feature based on the network parameters, and performs nonlinear processing on the result to obtain a fine-grained representation vector.

[0250] In one embodiment, such as Figure 12 As shown, the device also includes:

[0251] The first determining module 1112 is used to determine the modulus or information entropy of the feature embedding vector of the media to be recommended; and to determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector respectively based on the fine-grained features, coarse-grained features and modulus; or, to determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector respectively based on the fine-grained features, coarse-grained features and information entropy.

[0252] The fusion module 1108 is also used to perform weighted fusion of fine-grained representation vectors and coarse-grained representation vectors based on weight parameters.

[0253] In one embodiment, the recommendation module 1110 is further configured to determine a first score value of a candidate object based on the fusion vector and the object representation vector of the candidate object; the first score value is used to reflect the conversion rate of the candidate object after receiving the media to be recommended; and the target object is selected from the candidate objects according to the first score value.

[0254] In the above embodiments, the features of the media to be recommended are divided into fine-grained features and coarse-grained features of different degrees of coarseness. The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector with strong media generalization ability. The fine-grained features and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector that can capture the media's posterior information. Then, the fine-grained representation vector and the coarse-grained representation vector are weighted and fused to effectively balance generalization and memory. Therefore, when selecting target objects for recommendation based on the fused vector and the object representation vector of the candidate objects, the recommendation accuracy of the media to be recommended can be effectively improved, which is conducive to improving the conversion rate of the media to be recommended.

[0255] In one embodiment, the coarse-grained representation vector, the fine-grained representation vector, and the fused vector are obtained through a machine learning model; such as Figure 12 As shown, the device also includes:

[0256] The acquisition module 1102 is also used to acquire the training fine-grained features and training coarse-grained features of the media samples;

[0257] The first processing module 1104 is also used to perform feature generalization processing on the training coarse-grained features through the first network branch of the machine learning model to obtain the training coarse-grained representation vector.

[0258] The second processing module 1106 is also used to perform feature enhancement processing on the training fine-grained features and training coarse-grained features through the second network branch of the machine learning model to obtain the training fine-grained representation vector.

[0259] By using a fusion network of a machine learning model, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused to obtain a training fusion vector;

[0260] The fusion module 1108 is also used to determine the second score value of the candidate object sample based on the training fusion vector and the training object representation vector of the candidate object sample;

[0261] The optimization module 1114 is used to optimize the parameters of the machine learning model based on the loss between the second score value and the score label.

[0262] In one embodiment, such as Figure 12 As shown, the device also includes:

[0263] The second determining module 1116 is used to determine the third score value of the candidate object sample based on the training coarse-grained representation vector and the training object representation vector of the candidate object sample; and to determine the fourth score value of the candidate object sample based on the training fine-grained representation vector and the training object representation vector of the candidate object sample.

[0264] The optimization module 1114 is also used to optimize the parameters of the machine learning model based on the loss between the second score and the score label, the loss between the third score and the score label, and the loss between the fourth score and the score label.

[0265] In one embodiment, such as Figure 12 As shown, the device also includes:

[0266] The third determination module 1118 is used to determine the training degree of the media sample; the training degree is used to represent the cumulative number of iterations in which the media sample participates in training; based on the training degree, fine-grained training features and coarse-grained training features, the weight coefficients corresponding to the fine-grained training representation vector and the coarse-grained training representation vector are determined respectively.

[0267] The fusion module 1108 is also used to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight coefficients.

[0268] In one embodiment, such as Figure 12 As shown, the device also includes:

[0269] The fourth determining module 1120 is used to determine the magnitude or information entropy of the feature embedding vector of the media sample; based on the training fine-grained features, training coarse-grained features and magnitude, it determines the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively; or, based on the training fine-grained features, training coarse-grained features and information entropy, it determines the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively.

[0270] The fusion module 1108 is also used to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight coefficients.

[0271] In the above embodiments, the features of media samples are divided into training fine-grained features and training coarse-grained features of different degrees of coarseness. The training coarse-grained features are subjected to feature generalization processing to obtain a training coarse-grained representation vector with strong media generalization ability. The training fine-grained features and training coarse-grained features are subjected to feature enhancement processing to obtain a training fine-grained representation vector that can capture the media posterior information. Then, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused to effectively balance generalization and memorization. Therefore, the second score value is calculated using the training object representation vector and the training fused vector of the candidate object sample. The parameters of the machine learning model are optimized based on the loss between the second score value and the score label, so that the model can learn media features that balance generalization and memorization, thereby improving the generalization and accuracy of the model.

[0272] Each module in the aforementioned information recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0273] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores media data to be recommended. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an information recommendation method.

[0274] Those skilled in the art will understand that Figure 13The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0275] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the information recommendation method described above.

[0276] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the information recommendation method described above.

[0277] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the information recommendation method described above.

[0278] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0279] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0280] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0281] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An information recommendation method, characterized in that, The method includes: Obtaining fine-grained and coarse-grained features of the media to be recommended includes: determining the granularity entropy of each feature among at least two media features corresponding to the media to be recommended; dividing each feature among the at least two media features into fine-grained and coarse-grained features based on the granularity entropy to obtain the fine-grained and coarse-grained features of the media to be recommended; wherein, determining the granularity entropy of each feature among the at least two media features includes: determining the number of interactive objects corresponding to different feature values ​​for each feature among the at least two media features; determining the first conversion rate for each feature among the at least two media features when taking different feature values ​​based on the number of interactive objects and the total number of objects of the media to be recommended; determining the granularity entropy of each feature among the at least two media features based on the first conversion rate when taking different feature values ​​and the corresponding number of interactive objects; the smaller the granularity entropy, the more concentrated the distribution of the number of interactive objects is in a few feature values, and the more the at least two media features are biased towards fine-grained features; The coarse-grained features are subjected to feature generalization processing to obtain a coarse-grained representation vector; The fine-grained and coarse-grained features are subjected to feature enhancement processing to obtain a fine-grained representation vector; The fine-grained representation vector and the coarse-grained representation vector are weighted and fused to obtain a fused vector; Based on the fusion vector and the object representation vector of the candidate objects, a target object is selected from the candidate objects, and recommendation information of the media to be recommended is pushed to the target object.

2. The method according to claim 1, characterized in that, The method further includes: The object information of the candidate object is vectorized to obtain the object embedding; the object information includes at least one of the candidate object's object identifier, media interaction information, or interest information. The object embedding is used to extract features through a feature extraction network to obtain the object representation vector of the candidate object.

3. The method according to claim 1, characterized in that, The method further includes: If the media data of the media to be recommended is text data, recommendation information corresponding to the data type of interest to the target object is generated based on the media data or key information in the media data; or... The recommended information is obtained from the information database, which is generated based on the business request initiated by the client.

4. The method according to claim 1, characterized in that, The granularity entropy of each feature among the at least two media features corresponding to the media to be recommended includes: Among the at least two media features corresponding to the media to be recommended, determine the number of object interactions corresponding to different feature values ​​for each feature; Based on the number of interactions with the object and the total number of interactions with the media to be recommended, a second conversion rate is determined when each of the at least two media features takes different feature values; Based on the second conversion rate when taking different feature values ​​and the corresponding number of object interactions, the granularity entropy of each feature among the at least two media features is determined.

5. The method according to claim 1, characterized in that, The method further includes: Determine the coarse-to-fine granularity ratio, and divide each feature among the at least two media features into fine-grained and coarse-grained features according to the coarse-to-fine granularity ratio to obtain the fine-grained and coarse-grained features of the media to be recommended; or, In response to the granularity partitioning request, each of the at least two media features is divided into fine-grained features and coarse-grained features.

6. The method according to claim 1, characterized in that, The feature generalization processing of the coarse-grained features to obtain the coarse-grained representation vector includes: The coarse-grained features belonging to the same media to be recommended are spliced ​​together to obtain the first spliced ​​feature of each media to be recommended. Each of the first concatenated features is sequentially input into the coarse-grained network of the recommendation model, so that the multilayer perceptron in the coarse-grained network performs feature processing on the input first concatenated features to obtain a coarse-grained representation vector.

7. The method according to claim 1, characterized in that, The feature enhancement processing performed on the fine-grained features and coarse-grained features to obtain the fine-grained representation vector includes: The fine-grained features and the coarse-grained features are spliced ​​together to obtain a second spliced ​​feature; The second concatenated features are sequentially input into the fine-grained network of the recommendation model, so that the multilayer perceptron in the fine-grained network performs feature processing on the input second concatenated features to obtain a fine-grained representation vector.

8. The method according to claim 1, characterized in that, The method further includes: Determine the magnitude or information entropy of the feature embedding vector of the media to be recommended; Based on the fine-grained features, the coarse-grained features, and the modulus, determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively; or, based on the fine-grained features, the coarse-grained features, and the information entropy, determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively. The weighted fusion of the fine-grained representation vector and the coarse-grained representation vector includes: Based on the weight parameters, the fine-grained representation vector and the coarse-grained representation vector are weighted and fused.

9. The method according to claim 1, characterized in that, The step of selecting a target object from the candidate objects based on the fusion vector and the object representation vector of the candidate objects includes: Based on the fusion vector and the object representation vector of the candidate object, a first score value of the candidate object is determined; the first score value is used to reflect the conversion rate of the candidate object after receiving the media to be recommended. Among the candidate objects, the target object is selected based on the first score value.

10. The method according to any one of claims 1 to 9, characterized in that, The coarse-grained representation vector, the fine-grained representation vector, and the fused vector are obtained through a machine learning model; the method further includes: Obtain the fine-grained and coarse-grained training features of the media samples; The training coarse-grained features are generalized through the first network branch of the machine learning model to obtain the training coarse-grained representation vector. The second network branch of the machine learning model is used to perform feature enhancement processing on the training fine-grained features and the training coarse-grained features to obtain the training fine-grained representation vector. The training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused through the fusion network of the machine learning model to obtain the training fusion vector; Based on the training fusion vector and the training object representation vector of the candidate object sample, a second score value of the candidate object sample is determined; The parameters of the machine learning model are optimized based on the loss between the second score value and the score label.

11. The method according to claim 10, characterized in that, The method further includes: Based on the training coarse-grained representation vector and the training object representation vector of the candidate object sample, the third score value of the candidate object sample is determined. Based on the training fine-grained representation vector and the training object representation vector of the candidate object sample, the fourth score value of the candidate object sample is determined. The step of optimizing the parameters of the machine learning model based on the loss between the second score value and the score label includes: The machine learning model is optimized based on the loss between the second score and the score label, the loss between the third score and the score label, and the loss between the fourth score and the score label.

12. The method according to claim 10, characterized in that, The method further includes: Determine the training degree of the media sample; the training degree is used to represent the cumulative number of iterations in which the media sample participates in training; Based on the training parameters, the fine-grained training features, and the coarse-grained training features, determine the weight coefficients corresponding to the fine-grained training representation vector and the coarse-grained training representation vector, respectively. The weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector includes: Based on the weight coefficients, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused.

13. The method according to claim 10, characterized in that, The method further includes: Determine the magnitude or information entropy of the feature embedding vector of the media sample; Based on the training fine-grained features, the training coarse-grained features, and the modulus, determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector, respectively; or, based on the training fine-grained features, the training coarse-grained features, and the information entropy, determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector, respectively. The weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector includes: Based on the weight parameters, the training fine-grained representation vector and the training coarse-grained representation vector are weighted and fused.

14. An information recommendation device, characterized in that, The device includes: The acquisition module is used to acquire fine-grained features and coarse-grained features of the media to be recommended, including: determining the granularity entropy of each feature among at least two media features corresponding to the media to be recommended; dividing each feature among the at least two media features into fine-grained features based on the granularity entropy to obtain the fine-grained features and coarse-grained features of the media to be recommended; wherein, determining the granularity entropy of each feature among the at least two media features includes: determining the number of interactive objects corresponding to each feature taking different feature values ​​among the at least two media features; determining the first conversion rate when each feature takes different feature values ​​based on the number of interactive objects and the total number of objects of the media to be recommended; determining the granularity entropy of each feature among the at least two media features based on the first conversion rate when taking different feature values ​​and the corresponding number of interactive objects; the smaller the granularity entropy, the more concentrated the distribution of the number of interactive objects is in a few feature values, and the more the at least two media features are biased towards fine-grained features; The first processing module is used to perform feature generalization processing on the coarse-grained features to obtain a coarse-grained representation vector. The second processing module is used to perform feature enhancement processing on the fine-grained features and coarse-grained features to obtain a fine-grained representation vector; The fusion module is used to weightedly fuse the fine-grained representation vector and the coarse-grained representation vector to obtain a fused vector; The recommendation module is used to select a target object from the candidate objects based on the fusion vector and the object representation vector of the candidate objects, and push the recommendation information of the media to be recommended to the target object.

15. The apparatus according to claim 14, characterized in that, The acquisition module is further configured to vectorize the object information of the candidate object to obtain the object embedding; the object information includes at least one of the object identifier, media interaction information, or interest information of the candidate object; The object embedding is used to extract features through a feature extraction network to obtain the object representation vector of the candidate object.

16. The apparatus according to claim 14, characterized in that, The recommendation module is also used to generate recommendation information corresponding to the data type of interest of the target object based on the media data or key information in the media data if the media data of the media to be recommended is text data. Alternatively, recommendation information corresponding to the data type specified by the media provider can be obtained from the information database. The recommendation information is generated based on the business request initiated by the client.

17. The apparatus according to claim 14, characterized in that, The acquisition module is further configured to: determine the number of object interactions corresponding to different feature values ​​for each feature among the at least two media features corresponding to the media to be recommended; determine the second conversion rate for each feature among the at least two media features when taking different feature values ​​based on the number of object interactions and the total number of interactions for the media to be recommended; and determine the granularity entropy of each feature among the at least two media features based on the second conversion rate when taking different feature values ​​and the corresponding number of object interactions.

18. The apparatus according to claim 14, characterized in that, The acquisition module is further configured to determine the coarse-to-fine granularity ratio and divide each feature among the at least two media features into fine-grained features and coarse-grained features according to the coarse-to-fine granularity ratio to obtain the fine-grained features and coarse-grained features of the media to be recommended; or, in response to the granularity division request, divide each feature among the at least two media features into fine-grained features and coarse-grained features.

19. The apparatus according to claim 14, characterized in that, The first processing module is further configured to concatenate the coarse-grained features belonging to the same media to be recommended, respectively, to obtain a first concatenated feature for each of the media to be recommended; and to sequentially input each of the first concatenated features into the coarse-grained network of the recommendation model, so that the multilayer perceptron in the coarse-grained network performs feature processing on the input first concatenated features to obtain a coarse-grained representation vector.

20. The apparatus according to claim 14, characterized in that, The second processing module is further configured to concatenate the fine-grained features and the coarse-grained features to obtain a second concatenated feature; and to sequentially input the second concatenated feature into the fine-grained network of the recommendation model, so that the multilayer perceptron in the fine-grained network performs feature processing on the input second concatenated feature to obtain a fine-grained representation vector.

21. The apparatus according to claim 14, characterized in that, The device further includes: The first determining module is used to determine the magnitude or information entropy of the feature embedding vector of the media to be recommended; Based on the fine-grained features, the coarse-grained features, and the modulus, determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively; or, based on the fine-grained features, the coarse-grained features, and the information entropy, determine the weight parameters corresponding to the fine-grained representation vector and the coarse-grained representation vector, respectively. The fusion module is further configured to perform weighted fusion of the fine-grained representation vector and the coarse-grained representation vector based on the weight parameters.

22. The apparatus according to claim 14, characterized in that, The recommendation module is further configured to determine a first score value for the candidate object based on the fusion vector and the object representation vector of the candidate object; the first score value is used to reflect the conversion rate of the candidate object after receiving the media to be recommended. Among the candidate objects, the target object is selected based on the first score value.

23. The apparatus according to any one of claims 14 to 22, characterized in that, The coarse-grained representation vector, the fine-grained representation vector, and the fused vector are obtained through a machine learning model; the device further includes: The acquisition module is also used to acquire the training fine-grained features and training coarse-grained features of the media samples; The first processing module is further configured to perform feature generalization processing on the training coarse-grained features through the first network branch of the machine learning model to obtain the training coarse-grained representation vector. The second processing module is further configured to perform feature enhancement processing on the training fine-grained features and the training coarse-grained features through the second network branch of the machine learning model to obtain the training fine-grained representation vector; The fusion module is further configured to use the fusion network of the machine learning model to weightedly fuse the training fine-grained representation vector and the training coarse-grained representation vector to obtain a training fusion vector; and to determine a second score value of the candidate object sample based on the training fusion vector and the training object representation vector of the candidate object sample. An optimization module is used to optimize the parameters of the machine learning model based on the loss between the second score value and the score label.

24. The apparatus according to claim 23, characterized in that, The device further includes: The second determining module is used to determine a third score value of the candidate object sample based on the training coarse-grained representation vector and the training object representation vector of the candidate object sample; and to determine a fourth score value of the candidate object sample based on the training fine-grained representation vector and the training object representation vector of the candidate object sample. The optimization module is further configured to optimize the parameters of the machine learning model based on the loss between the second score value and the score label, the loss between the third score value and the score label, and the loss between the fourth score value and the score label.

25. The apparatus according to claim 23, characterized in that, The device further includes: The third determining module is used to determine the training degree of the media sample; the training degree is used to represent the cumulative number of iterations in which the media sample participates in training; based on the training degree, the fine-grained training features and the coarse-grained training features, the weight coefficients corresponding to the fine-grained training representation vector and the coarse-grained training representation vector are determined respectively. The fusion module is further configured to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight coefficients.

26. The apparatus according to claim 23, characterized in that, The device further includes: The fourth determining module is used to determine the magnitude or information entropy of the feature embedding vector of the media sample; and to determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively based on the training fine-grained features, the training coarse-grained features, and the magnitude; or, to determine the weight parameters corresponding to the training fine-grained representation vector and the training coarse-grained representation vector respectively based on the training fine-grained features, the training coarse-grained features, and the information entropy. The fusion module is further configured to perform weighted fusion of the training fine-grained representation vector and the training coarse-grained representation vector based on the weight parameters.

27. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 13.

28. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 13.

29. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 13.

Images