A model training method, a propagation resource recommendation method, and related devices

Abstract

The embodiment of the application provides a model training method, a propagation resource recommendation method and related equipment, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and Internet of Vehicles; the embodiment of the application can obtain a propagation resource sample set; a main propagation resource tower of a to-be-trained resource recommendation model is used to perform feature extraction on the propagation resource sample to obtain resource sample attribute features corresponding to the propagation resource sample; an auxiliary propagation resource tower of the to-be-trained resource recommendation model is used to perform feature extraction on a propagation resource reference sample to obtain resource reference sample attribute features corresponding to the propagation resource reference sample; the to-be-trained resource recommendation model is trained according to the resource sample attribute features and the resource reference sample attribute features to obtain a trained resource recommendation model. The embodiment of the application can improve the accuracy of the trained resource recommendation model in predicting any to-be-recommended propagation resource.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a model training method, a method for recommending propagation resources, and related equipment. The related equipment includes a model training device, a propagation resource recommendation device, a computer device, and a computer-readable storage medium. Background Technology

[0002] With the development of internet technology, recommendation algorithms are being used more and more widely in the field of information dissemination resources, such as advertising. In this field, recommendation algorithms suggest information dissemination resources that users prefer.

[0003] Currently, recommendation algorithms used in the field of communication resources are generally recommendation neural network models. However, when predicting scores for communication resources, recommendation neural network models can only accurately score a subset of communication resources with specific attribute information, making it impossible to accurately predict the score of any single communication resource.

[0004] In summary, existing technologies suffer from the problem of low accuracy in predictions made by recommendation neural network models. Summary of the Invention

[0005] This application provides a model training method, a propagation resource recommendation method, and related equipment, which can improve the accuracy of the trained resource recommendation model in predicting any propagation resource to be recommended.

[0006] A model training method, comprising:

[0007] Obtain a set of dissemination resource samples, which includes dissemination resource samples and dissemination resource reference samples that are related to the dissemination resource samples in terms of dissemination resource attribute information. The relationship indicates that the dissemination resource samples and the dissemination resource reference samples have the same dissemination resource attribute information.

[0008] The main propagation resource tower of the resource recommendation model to be trained is used to extract features from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples.

[0009] Using the auxiliary propagation resource tower of the resource recommendation model to be trained, feature extraction is performed on the propagation resource reference sample to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0010] Based on the attribute characteristics of resource samples and resource reference samples, calculate the attribute correlation loss between the propagated resource samples and the propagated resource reference samples in terms of propagated resource attribute information;

[0011] The resource recommendation model to be trained is trained using attribute correlation loss to obtain the trained resource recommendation model.

[0012] Accordingly, embodiments of this application provide a model training apparatus, including:

[0013] The first acquisition unit can be used to acquire a set of propagation resource samples. The set of propagation resource samples includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. The relationship indicates that the propagation resource samples and the propagation resource reference samples have the same propagation resource attribute information.

[0014] The first extraction unit can be used to extract features from the propagation resource samples using the main propagation resource tower of the resource recommendation model to be trained, and obtain the resource sample attribute features corresponding to the propagation resource samples.

[0015] The second extraction unit can be used to extract features from the propagation resource reference sample using the auxiliary propagation resource tower of the resource recommendation model to be trained, and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0016] The calculation unit can be used to calculate the attribute correlation loss between the propagated resource sample and the propagated resource reference sample in terms of propagated resource attribute information, based on the attribute characteristics of the resource sample and the attribute characteristics of the resource reference sample.

[0017] The training unit can be used to train the resource recommendation model to be trained based on the attribute correlation loss, so as to obtain the trained resource recommendation model.

[0018] In some embodiments, the propagation resource sample set further includes object preference samples corresponding to the propagation resource samples; the model training device further includes a third extraction unit, which can be specifically used to extract features from the object preference samples using the object pyramid of the resource recommendation model to be trained, so as to obtain the preference sample features corresponding to the object preference samples.

[0019] Accordingly, the computing unit can be used to obtain the preference loss of the object preference sample based on the preference sample characteristics.

[0020] Accordingly, the training unit can be used to train the resource recommendation model to be trained based on attribute correlation loss and preference loss, so as to obtain the trained resource recommendation model.

[0021] In some embodiments, the preference loss includes preference correlation loss; the calculation unit can be specifically used to calculate the preference correlation loss between object preference samples and propagation resource samples based on preference sample characteristics and resource sample attribute characteristics.

[0022] Accordingly, the training unit can be used to train the resource recommendation model to be trained based on attribute correlation loss and preference correlation loss, so as to obtain the trained resource recommendation model.

[0023] In some embodiments, the preference loss further includes preference feature loss; the training unit may also be used to obtain label information corresponding to object preference samples; calculate the preference feature loss between preference sample features and label information; and train the resource recommendation model to be trained based on attribute correlation loss, preference correlation loss and preference feature loss to obtain the trained resource recommendation model.

[0024] In some embodiments, the propagation resource attribute information of the propagation resource sample includes coarse-grained resource attribute information and fine-grained resource attribute information, and the propagation resource attribute information of the propagation resource reference sample includes coarse-grained resource attribute information; the first extraction unit can be used to use the main propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information and the fine-grained resource attribute information of the propagation resource sample, so as to obtain the resource sample attribute features corresponding to the propagation resource sample; the second extraction unit can be used to use the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information of the propagation resource reference sample, so as to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0025] In some embodiments, the first acquisition unit may be used to acquire the propagation resource attribute information of the propagation resource sample; perform propagation resource attribute information weakening processing on the propagation resource sample according to the propagation resource attribute information to obtain the propagation resource reference sample corresponding to the propagation resource sample; and generate a propagation resource sample set according to the propagation resource sample and the propagation resource reference sample corresponding to the propagation resource sample.

[0026] In some embodiments, the first acquisition unit may be used to extract attribute feature information of propagation resource attribute information; and based on the attribute feature information, perform propagation resource attribute information weakening processing on the propagation resource sample to obtain a propagation resource reference sample corresponding to the propagation resource sample.

[0027] In some embodiments, the first acquisition unit can be used to identify the information type of the propagation resource attribute information based on the attribute feature information to obtain the attribute information type of the propagation resource attribute information; if the attribute information type is the target attribute information type, the propagation resource attribute information of the target attribute information type is removed from the propagation resource sample to obtain the propagation resource reference sample corresponding to the propagation resource sample.

[0028] Furthermore, embodiments of this application also provide a method for recommending propagated resources, including: a post-trained resource recommendation model as described above, wherein the post-trained resource recommendation model includes at least two post-trained propagation resource towers; the method includes:

[0029] Obtain the resources to be recommended for dissemination, and obtain the creation information of the resources to be recommended for dissemination;

[0030] Based on the information created by the propagation resources, select the target propagation resource tower corresponding to the propagation resource to be recommended from at least two propagation resource towers of the post-trained resource recommendation model.

[0031] The target training and propagation resource tower are used to extract features of the resources to be recommended and propagated, thereby obtaining the propagation resource attribute features of the resources to be recommended and propagated.

[0032] Based on the characteristics of the dissemination resources, determine the recommendation parameters for the dissemination resources to be recommended;

[0033] Based on the recommendation parameters, resources to be recommended for dissemination are suggested.

[0034] Accordingly, embodiments of this application provide a propagation resource recommendation device, such as the aforementioned post-training resource recommendation model, which includes at least two post-training propagation resource towers; including:

[0035] The second acquisition unit can be used to acquire resources to be recommended for dissemination, as well as to acquire the creation information of the resources to be recommended for dissemination.

[0036] The filtering unit can be used to filter out the target post-trained propagation resource tower corresponding to the propagation resource to be recommended from at least two post-trained propagation resource towers of the post-trained resource recommendation model based on the information created by the propagation resource;

[0037] The third extraction unit can be used to extract features from the propagation resource tower after target training to obtain the propagation resource attribute features of the propagation resource to be recommended.

[0038] The determining unit can be used to determine the recommendation parameters of the dissemination resources to be recommended based on the attribute characteristics of the dissemination resources;

[0039] The recommendation unit can be used to recommend resources to be promoted based on recommendation parameters.

[0040] In some embodiments, the post-training propagation resource tower includes a post-training main propagation resource tower and a post-training auxiliary propagation resource tower; the filtering unit can be used to determine the post-training main propagation resource tower as the target post-training propagation resource tower corresponding to the propagation resource to be recommended if the propagation resource creation information conforms to a preset rule; and to determine the post-training auxiliary propagation resource tower as the target post-training propagation resource tower corresponding to the propagation resource to be recommended if the propagation resource creation information does not conform to the preset rule.

[0041] In some embodiments, the post-trained resource recommendation model further includes a post-trained object tower; an extraction unit, which can be used to obtain object preference information of the recommended object for the resource to be recommended and propagated, wherein the object preference information represents the preference attribute information of the recommended object for the resource to be recommended and propagated; and to extract features from the object preference information using the post-trained object tower to obtain the object preference features corresponding to the object preference information.

[0042] Accordingly, the determination unit can be used to map the preference sample features and the propagation resource attribute features to the recommendation parameters, thereby obtaining the recommendation parameters of the propagation resource to be recommended.

[0043] Furthermore, embodiments of this application also provide a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute any of the model training methods and any of the resource recommendation methods provided in embodiments of this application.

[0044] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program adapted for loading by a processor to execute any of the model training methods and any of the resource recommendation methods provided in embodiments of this application.

[0045] Furthermore, this application also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the model training methods provided in this application.

[0046] Furthermore, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements any of the model training methods and any of the resource recommendation methods provided in embodiments of this application.

[0047] This application embodiment can obtain a set of propagation resource samples, which includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. Using the main propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples. Using the auxiliary propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource reference samples to obtain the resource reference sample attribute features corresponding to the propagation resource reference samples. Based on the resource sample attribute features and the resource reference sample attribute features, the attribute correlation loss between the propagation resource samples and the propagation resource reference samples in terms of propagation resource attribute information is calculated. Based on the attribute correlation loss, the resource recommendation model to be trained is trained to obtain the trained resource recommendation model. Since the embodiments of this application can employ a main propagation resource tower and an auxiliary propagation resource tower of the resource recommendation model to be trained, and extract features from propagation resource samples and propagation resource reference samples that have a correlation in propagation resource attributes, attribute correlation loss can be calculated based on the extracted resource sample attribute features and resource reference sample attribute features. This allows the resource reference sample attribute features of the auxiliary propagation resource tower to fit the resource sample attribute features of the main propagation resource tower as closely as possible based on the attribute correlation loss. Consequently, the dependence of the auxiliary propagation resource tower on propagation resource attribute information in the trained resource recommendation model can be weakened, thereby improving the accuracy of the trained resource recommendation model in predicting any propagation resource to be recommended. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of a scenario for the model training method provided in an embodiment of this application;

[0050] Figure 2 This is a schematic flowchart of the model training method provided in the embodiments of this application;

[0051] Figure 3 This is a schematic diagram illustrating the acquisition and propagation of a sample set of resources provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the preference loss for obtaining object preference samples provided in an embodiment of this application;

[0053] Figure 5 This is a flowchart illustrating the method for recommending dissemination resources provided in an embodiment of this application;

[0054] Figure 6 Two flowcharts illustrating the method for recommending dissemination resources provided in this application embodiment;

[0055] Figure 7 This is a schematic diagram of the resource recommendation model to be trained provided in an embodiment of this application;

[0056] Figure 8 These are two schematic diagrams illustrating the resource recommendation model to be trained provided in the embodiments of this application;

[0057] Figure 9 These are three schematic diagrams illustrating the resource recommendation model to be trained provided in the embodiments of this application;

[0058] Figure 10 This is a schematic diagram of the structure of the model training device provided in the embodiments of this application;

[0059] Figure 11 These are two schematic diagrams illustrating the structure of the dissemination resource recommendation device provided in the embodiments of this application;

[0060] Figure 12 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] This application provides a model training method, apparatus, computer device, and computer-readable storage medium. The model training apparatus can be integrated into the computer device, which may be a server or a terminal, etc.

[0063] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or 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 network (CDN) acceleration services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. The embodiments of this application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0064] This application relates to Artificial Intelligence (AI), which is the theory, method, technology, and application system for simulating, extending, and expanding human intelligence using digital computers or machines controlled by digital computers to perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines 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 perception, reasoning, and decision-making capabilities.

[0065] 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, natural language processing, and machine learning / deep learning.

[0066] For example, see Figure 1Taking the integration of the model training device into a computer device as an example, the computer device can acquire a set of propagation resource samples, which includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. Using the main propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples. Using the auxiliary propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource reference samples to obtain the resource reference sample attribute features corresponding to the propagation resource reference samples. Based on the resource sample attribute features and the resource reference sample attribute features, the attribute correlation loss between the propagation resource samples and the propagation resource reference samples in terms of propagation resource attribute information is calculated. Based on the attribute correlation loss, the resource recommendation model to be trained is trained to obtain the trained resource recommendation model.

[0067] The presentation form of the dissemination resource sample can be any one of images, audio and video, audio, and text, or a combination of at least two of them; specifically, the dissemination resource sample can be an advertising sample.

[0068] The presentation form of the communication resource reference sample can be any one of images, audio and video, audio, and text, or a combination of at least two of them; specifically, the communication resource reference sample can be an advertising sample.

[0069] Both the main propagation resource tower and the auxiliary propagation resource tower can be towers constructed from at least one layer of a neural network. The main propagation resource tower can be a tower used for feature extraction of propagation resource samples; the auxiliary propagation resource tower can be a tower used for feature extraction of propagation resource reference samples.

[0070] The propagation resource attribute information can refer to the attribute information of the propagation resource sample and the attribute information of the propagation resource reference sample. Specifically, the propagation resource attribute information can include at least one of coarse-grained resource attribute information and fine-grained resource attribute information.

[0071] Coarse-grained resource attribute information can refer to the coarse-grained attribute information of a propagation resource sample or a propagation resource reference sample. Specifically, for example, coarse-grained resource attribute information can include the category, color, size, etc. of the propagation resource sample, or the category, color, size, etc. of the propagation resource reference sample.

[0072] Fine-grained resource attribute information can refer to the fine-grained attribute information of a propagation resource sample or a propagation resource reference sample; specifically, for example, fine-grained resource attribute information may include the unique identifier, name, etc. of the propagation resource sample, or the unique identifier, name, etc. of the propagation resource sample.

[0073] Here, resource sample attribute features can refer to the features corresponding to the propagation resource attribute information of the propagated resource sample. Resource reference sample attribute features can refer to the features corresponding to the propagation resource attribute information of the propagation resource reference sample.

[0074] The association relationship indicates that the communication resource sample and the communication resource reference sample have the same communication resource attribute information. Specifically, when the communication resource attribute information of the communication resource reference sample and the communication resource sample have some of the same information types, this situation can be said to indicate that the communication resource reference sample and the communication resource sample have an association relationship in terms of communication resource attribute information. For example, the information types include the aforementioned coarse-grained resource attribute information and coarse-grained resource attribute information.

[0075] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0076] This embodiment will be described from the perspective of a model training device, which can be integrated into a computer device, such as a server or a terminal. The terminal can include tablet computers, laptops, personal computers (PCs), wearable devices, virtual reality devices, or other smart devices that can acquire data.

[0077] like Figure 2 As shown, the specific process of this model training method is as follows: steps S201 to S205:

[0078] S201. Obtain a sample set of propagation resources.

[0079] The set of communication resource samples includes communication resource samples and communication resource reference samples that are related to the communication resource samples in terms of communication resource attribute information. The relationship indicates that the communication resource samples and communication resource reference samples have the same communication resource attribute information.

[0080] In some application embodiments, such as Figure 3 As shown, the method for obtaining the propagation resource sample set can be as described in steps S301 to S303:

[0081] S301. Obtain the propagation resource attribute information of the propagation resource sample.

[0082] In one example, propagating resource attribute information could mean propagating the attribute information carried by the resource sample.

[0083] In one example, the propagation resource attribute information may be information bound to the propagation resource sample and stored in the cloud or locally. Based on this, embodiments of this application can obtain the propagation resource attribute information of the propagation resource sample from the cloud or locally.

[0084] S302. Based on the propagation resource attribute information, the propagation resource sample is weakened to obtain the propagation resource reference sample corresponding to the propagation resource sample.

[0085] In one example, the method for weakening the propagation resource attribute information of a propagation resource sample based on the propagation resource attribute information to obtain the propagation resource reference sample can be as follows: extract the attribute feature information of the propagation resource attribute information; and weaken the propagation resource attribute information of the propagation resource sample based on the attribute feature information to obtain the propagation resource reference sample.

[0086] In one example, attribute feature information can be information carried by the propagation resource attribute information itself.

[0087] In one example, attribute feature information can be obtained by extracting features from the propagation resource attribute information. Specifically, a feature extraction model can be used to extract features from the propagation resource attribute information to obtain the attribute feature information.

[0088] Among them, attribute feature information can be information that characterizes the attributes of propagated resources. Specifically, the representation form of attribute feature information can be a string. For example, attribute feature information can be a string with a pre-defined format. Alternatively, attribute feature information can refer to information that satisfies preset candidate rules. For example, when attribute feature information matches preset attribute feature information, this situation can be said to satisfy the preset candidate rules; or, for example, when the keyword of attribute feature information is the same as the preset keyword, this situation can be said to satisfy the preset candidate rules.

[0089] In one example, the method for weakening the propagation resource attribute information of the propagation resource sample based on attribute feature information to obtain the propagation resource reference sample can be as follows: based on attribute feature information, the propagation resource attribute information is identified by information type to obtain the attribute information type of the propagation resource attribute information; if the attribute information type is the target attribute information type, the propagation resource attribute information of the target attribute information type is removed from the propagation resource sample to obtain the propagation resource reference sample corresponding to the propagation resource sample.

[0090] In one example, the method for identifying the attribute type of the propagation resource attribute information based on attribute feature information and obtaining the attribute information type of the propagation resource attribute information can be as follows: obtain a mapping relationship set, which includes the mapping relationship between preset attribute feature information and preset attribute information type; determine the target preset attribute feature information corresponding to the attribute feature information from the mapping relationship set; determine the target preset attribute information type corresponding to the target preset attribute feature information based on the mapping relationship set; and determine the target preset attribute information type as the attribute information type of the propagation resource attribute information.

[0091] The information types can include coarse-grained resource attribute information and fine-grained resource attribute information. The target attribute information type can be fine-grained resource attribute information. In this embodiment, fine-grained resource attribute information can be removed from the propagation resource sample to obtain a propagation resource reference sample corresponding to the propagation resource sample.

[0092] For example, the propagation resource attribute information of a propagation resource sample includes coarse-grained resource attribute information such as the category, color, and size of the propagation resource sample, and fine-grained resource attribute information such as the unique identifier of the propagation resource sample. In this embodiment, the fine-grained resource attribute information can be removed from the propagation resource sample to obtain a propagation resource reference sample corresponding to the propagation resource sample. In this case, the propagation resource attribute information of the propagation resource reference sample includes the fine-grained resource attribute information.

[0093] S303. Generate a set of propagation resource samples based on the propagation resource samples and the corresponding propagation resource reference samples.

[0094] In this embodiment, a dissemination resource sample and a corresponding dissemination resource reference sample can be bound together to construct a dissemination resource sample set. For example, taking a lipstick advertisement as a dissemination resource sample, the dissemination resource sample has the color of the lipstick, the category of the lipstick, and a unique identifier for the lipstick advertisement. Correspondingly, the dissemination resource reference sample has the color of the lipstick and the category of the lipstick, but does not have a unique identifier for the lipstick advertisement.

[0095] In some embodiments of the application, the propagation resource sample set can be pre-set and stored in a database. Based on this, the embodiments of the present application can directly extract the propagation resource sample set from the database.

[0096] S202. Using the main propagation resource tower of the resource recommendation model to be trained, feature extraction is performed on the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples.

[0097] In some application embodiments, the method of using the main propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource samples and obtain the resource sample attribute features corresponding to the propagation resource samples can be as follows: using the main propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information and the fine-grained resource attribute information of the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples.

[0098] In one example, the main propagation resource tower of the resource recommendation model to be trained is used to extract features from the coarse-grained and fine-grained resource attribute information of the propagated resource samples. The method to obtain the resource sample attribute features corresponding to the propagated resource samples can be as follows: the main propagation resource tower of the resource recommendation model to be trained is used to extract features from the coarse-grained resource attribute information of the propagated resource samples to obtain the first coarse-grained resource attribute features; the main propagation resource tower of the resource recommendation model to be trained is used to extract features from the fine-grained resource attribute information of the propagated resource samples to obtain the fine-grained resource attribute features; based on the first coarse-grained resource attribute features and the fine-grained resource attribute features, the resource sample attribute features corresponding to the propagated resource samples are determined.

[0099] In one example, the method to determine the resource sample attribute features corresponding to the propagation resource sample based on the first coarse-grained resource attribute features and the fine-grained resource attribute features can be as follows: the first coarse-grained resource attribute features and the fine-grained resource attribute features are weighted to obtain the resource sample attribute features corresponding to the propagation resource sample.

[0100] In one example, the method to determine the resource sample attribute features corresponding to the propagation resource sample based on the first coarse-grained resource attribute features and the fine-grained resource attribute features can be as follows: the first coarse-grained resource attribute features and the fine-grained resource attribute features are concatenated to obtain the resource sample attribute features corresponding to the propagation resource sample.

[0101] S203. Using the auxiliary propagation resource tower of the resource recommendation model to be trained, feature extraction is performed on the propagation resource reference sample to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0102] In some application embodiments, the method of using the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource reference sample and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample can be as follows: using the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information of the propagation resource reference sample and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0103] S204. Based on the attribute characteristics of the resource sample and the attribute characteristics of the resource reference sample, calculate the attribute correlation loss between the propagated resource sample and the propagated resource reference sample in terms of propagated resource attribute information.

[0104] In some application embodiments, the method for calculating the attribute correlation loss between the propagation resource sample and the propagation resource reference sample on the propagation resource attribute information based on the resource sample attribute characteristics and the resource reference sample attribute characteristics can be as follows: using an attribute correlation loss function, the attribute correlation loss between the propagation resource sample and the propagation resource reference sample on the propagation resource attribute information can be calculated based on the resource sample attribute characteristics and the resource reference sample attribute characteristics.

[0105] Among them, the attribute correlation loss function can be an unsupervised loss function. Specifically, the attribute correlation loss function can be InfoNCE (Info Noise Contrastive Estimation loss), or it can be the cross-entropy loss function.

[0106] S205. Based on the attribute correlation loss, train the resource recommendation model to be trained to obtain the trained resource recommendation model.

[0107] In this embodiment, the model parameters of the resource recommendation model to be trained are converged based on the attribute correlation loss to obtain the trained resource recommendation model.

[0108] In some embodiments, the propagation resource sample set also includes object preference samples corresponding to the propagation resource samples; based on this, such as Figure 4 As shown, before step S205 in this embodiment of the application, the method further obtains the preference loss of the object preference sample. Obtaining the preference loss of the object preference sample includes steps S401 to S402:

[0109] S401. Use the object pyramid of the resource recommendation model to extract features from the object preference samples to obtain the preference sample features corresponding to the object preference samples.

[0110] The object tower can be a tower constructed from at least one layer of a neural network. Specifically, the object tower can be a tower used for feature extraction from object preference samples.

[0111] S402. Based on the characteristics of the preference samples, obtain the preference loss of the object preference samples.

[0112] In some embodiments of the application, preference loss may include, but is not limited to, at least one of preference correlation loss and preference feature loss.

[0113] Based on steps S401 to S402, step S205 trains the resource recommendation model to be trained according to the attribute correlation loss, and obtains the trained resource recommendation model in the following way: train the resource recommendation model to be trained according to the attribute correlation loss and the preference loss, and obtain the trained resource recommendation model.

[0114] Based on the foregoing, in some application embodiments, preference loss includes preference relevance loss; accordingly, the preference loss of object preference samples can be obtained by calculating the preference relevance loss between object preference samples and propagation resource samples based on preference sample characteristics and resource sample attribute characteristics.

[0115] In one example, the method to calculate the preference correlation loss between object preference samples and propagated resource samples based on preference sample features and resource sample attribute features can be as follows: using a preference correlation loss function, the preference correlation loss between object preference samples and propagated resource samples can be calculated based on preference sample features and resource sample attribute features.

[0116] Among them, the preference correlation loss can be an unsupervised loss function. Specifically, the preference correlation loss can be InfoNCE (Info Noise Contrastive Estimation loss), and the attribute correlation loss function can also be the cross-entropy loss function.

[0117] Based on the above, the method to train the resource recommendation model to be trained and obtain the trained resource recommendation model according to the attribute correlation loss and preference loss can be as follows: Train the resource recommendation model to be trained according to the attribute correlation loss and preference correlation loss to obtain the trained resource recommendation model.

[0118] In one example, the resource recommendation model to be trained is trained based on attribute relevance loss and preference relevance loss to obtain the trained resource recommendation model. This can be done by: obtaining the first weights of attribute relevance loss and preference relevance loss respectively; weighting the attribute relevance loss and preference relevance loss according to the first weights to obtain the weighted relevance loss; and training the resource recommendation model to be trained based on the weighted relevance loss to obtain the trained resource recommendation model.

[0119] Based on the foregoing, in some application embodiments, the preference loss also includes preference feature loss. Based on this, the method for training the resource recommendation model to be trained according to the attribute correlation loss and preference loss to obtain the trained resource recommendation model can be as follows: obtain the label information corresponding to the object preference sample; calculate the preference feature loss between the preference sample features and the label information; train the resource recommendation model to be trained according to the attribute correlation loss, preference correlation loss and preference feature loss to obtain the trained resource recommendation model.

[0120] The label information can be pre-set information used for supervised learning of the resource recommendation model to be trained. The label information can be information about the preferences of object preference samples for the propagated resources.

[0121] In one example, the preference feature loss between preference sample features and label information can be calculated by using a preference feature loss function.

[0122] The preference feature loss can be a mean squared error loss function, a logarithmic loss function, or something similar.

[0123] In one example, the resource recommendation model to be trained is trained based on attribute relevance loss, preference relevance loss, and preference feature loss to obtain the trained resource recommendation model. This can be achieved by: obtaining the second weights of attribute relevance loss, preference relevance loss, and preference feature loss respectively; weighting the attribute relevance loss, preference relevance loss, and preference feature loss according to the second weights to obtain the weighted loss; and training the resource recommendation model to be trained based on the weighted loss to obtain the trained resource recommendation model.

[0124] This application embodiment can obtain a set of propagation resource samples, which includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. Using the main propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples. Using the auxiliary propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource reference samples to obtain the resource reference sample attribute features corresponding to the propagation resource reference samples. Based on the resource sample attribute features and the resource reference sample attribute features, the attribute correlation loss between the propagation resource samples and the propagation resource reference samples in terms of propagation resource attribute information is calculated. Based on the attribute correlation loss, the resource recommendation model to be trained is trained to obtain the trained resource recommendation model. Since the embodiments of this application can employ a main propagation resource tower and an auxiliary propagation resource tower of the resource recommendation model to be trained, and extract features from propagation resource samples and propagation resource reference samples that have a correlation in propagation resource attributes, attribute correlation loss can be calculated based on the extracted resource sample attribute features and resource reference sample attribute features. This allows the resource reference sample attribute features of the auxiliary propagation resource tower to fit the resource sample attribute features of the main propagation resource tower as closely as possible based on the attribute correlation loss. Consequently, the dependence of the auxiliary propagation resource tower on propagation resource attribute information in the trained resource recommendation model can be weakened, thereby improving the accuracy of the trained resource recommendation model in predicting any propagation resource to be recommended.

[0125] Based on the method described in the above embodiments, the following examples will provide further detailed explanations.

[0126] This embodiment will be described from the perspective of a dissemination resource recommendation device, which can be integrated into a computer device, such as a server or a terminal. The terminal can include tablet computers, laptops, personal computers (PCs), wearable devices, virtual reality devices, or other smart devices that can acquire data.

[0127] like Figure 5 As shown, the specific process of this resource recommendation method is illustrated in steps S501 to S505:

[0128] S501. Obtain the resources to be recommended for dissemination, and obtain the creation information of the resources to be recommended for dissemination.

[0129] In some embodiments, the propagation resource creation information can be information carried by the propagation resource to be recommended. The propagation resource creation information can be the creation time information of the propagation resource to be recommended, or it can be the creation location information of the propagation resource to be recommended. Specifically, in this embodiment, it is preferred that the propagation resource creation information be the creation time information of the propagation resource to be recommended.

[0130] S502. Based on the propagation resource creation information, select the target post-training propagation resource tower corresponding to the propagation resource to be recommended from at least two post-training propagation resource towers of the post-training resource recommendation model.

[0131] Among them, the post-training resource recommendation model is the same as the aforementioned post-training resource recommendation model.

[0132] The post-training propagation resource tower can be a tower constructed from at least one layer of a neural network. The propagation resource towers in the resource recommendation model to be trained, such as the main propagation resource tower, auxiliary propagation resource tower, and object tower, are trained to obtain the post-training propagation resource tower.

[0133] In some embodiments, the post-training propagation resource tower includes a primary post-training propagation resource tower and a secondary post-training propagation resource tower. Based on this, the method for selecting the target post-training propagation resource tower corresponding to the resource to be recommended from at least two post-training propagation resource towers of the post-training resource recommendation model, according to the propagation resource creation information, can be as follows: if the propagation resource creation information conforms to a preset rule, then the primary post-training propagation resource tower is determined as the target post-training propagation resource tower corresponding to the resource to be recommended; if the propagation resource creation information does not conform to the preset rule, then the secondary post-training propagation resource tower is determined as the target post-training propagation resource tower corresponding to the resource to be recommended.

[0134] In one example, when the propagation resource creation information matches the preset propagation resource creation information, this situation can be said to be that the propagation resource creation information conforms to the preset rules; when the propagation resource creation information does not match the preset propagation resource creation information, this situation can be said to be that the propagation resource creation information does not conform to the preset rules.

[0135] In one example, the creation information of the propagation resource is the creation time information of the propagation resource to be recommended. When the creation time information is greater than or equal to a preset time threshold, the creation information of the propagation resource can be said to conform to the preset rules. When the creation time information is less than the preset time threshold, the creation information of the propagation resource can be said to not conform to the preset rules.

[0136] S503. After training the target, use the propagation resource tower to extract features of the resources to be recommended and obtain the propagation resource attribute features of the resources to be recommended.

[0137] In some application embodiments, when the target post-training propagation resource tower is the post-training master propagation resource tower, the post-training master propagation resource tower can be used to extract features of the propagation resource to be recommended, thereby obtaining the propagation resource attribute features of the propagation resource to be recommended.

[0138] In some application embodiments, when the target post-training propagation resource tower is a post-training auxiliary propagation resource tower, the post-training auxiliary propagation resource tower can be used to extract features of the propagation resource to be recommended, thereby obtaining the propagation resource attribute features of the propagation resource to be recommended.

[0139] In one example, when the target post-training propagation resource tower is a post-training auxiliary propagation resource tower, the post-training auxiliary propagation resource tower is used to extract features from the propagation resource to be recommended, and the resulting propagation resource attribute features can specifically include: obtaining the propagation resource attribute information of the propagation resource to be recommended; weakening the propagation resource to be recommended based on the propagation resource attribute information to obtain the weakened propagation resource; extracting features from the weakened propagation resource using the post-training auxiliary propagation resource tower to obtain the weakened propagation resource features; and using the weakened propagation resource features as the propagation resource attribute features of the propagation resource to be recommended.

[0140] In one example, the propagation resource to be recommended is weakened based on the propagation resource attribute information. The specific method for obtaining the weakened propagation resource can be found in the explanation above, which states that "the propagation resource sample is weakened based on the propagation resource attribute information to obtain the propagation resource reference sample corresponding to the propagation resource sample". It will not be repeated here.

[0141] S504. Based on the characteristics of the dissemination resources, determine the recommendation parameters for the dissemination resources to be recommended.

[0142] The recommendation parameters can be parameters that characterize the resource to be recommended. Specifically, recommendation parameters can be recommendation scores, recommendation probabilities, and so on.

[0143] In some application embodiments, the post-trained resource recommendation model further includes a post-trained object tower; before step S504, the method further includes: obtaining object preference information of the recommended objects targeted by the resource to be recommended and propagated; and using the post-trained object tower to extract features from the object preference information to obtain object preference features corresponding to the object preference information.

[0144] Among them, object preference information represents the preference attribute information of the recommended object for the resource to be recommended and disseminated.

[0145] Based on the above, the method to determine the recommendation parameters of the propagation resources to be recommended according to the characteristics of propagation resource attributes can be as follows: map the preference sample characteristics and the propagation resource attribute characteristics to obtain the recommendation parameters of the propagation resources to be recommended.

[0146] In one example, a mapping function can be obtained; based on the mapping function, the recommendation parameters are mapped to the preference sample features and the propagation resource attribute features to obtain the recommendation parameters of the propagation resource to be recommended.

[0147] In one example, the similarity between preference sample features and propagation resource attribute features can be calculated to obtain the recommendation parameters for the propagation resource to be recommended.

[0148] S505. Based on the recommendation parameters, recommend resources to be recommended for dissemination.

[0149] In some application embodiments, the method for recommending resources to be promoted based on recommendation parameters can be as follows: sorting the resources according to the size of the recommendation parameters to obtain sorted recommendation parameters; based on the sorted recommendation parameters, selecting a preset number of target resources to be promoted from the resources to be promoted; and recommending the target resources to be promoted.

[0150] In some application embodiments, the method for recommending resources to be promoted based on recommendation parameters can be as follows: obtaining preset parameters; weighting the preset parameters and recommendation parameters to obtain weighted parameters; and recommending resources to be promoted based on the weighted parameters.

[0151] In one example, the method for recommending resources to be promoted based on the weighted parameters can be as follows: sort the resources to be promoted based on the magnitude of the weighted parameters to obtain sorted resources to be promoted; extract a preset number of target sorted resources to be promoted from the sorted resources; and recommend the target sorted resources to be promoted.

[0152] As can be seen from the above, the embodiments of this application can obtain the resources to be recommended for dissemination, as well as the dissemination resource creation information of the resources to be recommended for dissemination; based on the dissemination resource creation information, a target dissemination resource tower corresponding to the resources to be recommended is selected from at least two dissemination resource towers of the trained resource recommendation model; the target dissemination resource tower is used to extract features from the resources to be recommended for dissemination, thereby obtaining the dissemination resource attribute features of the resources to be recommended for dissemination; the recommendation parameters of the resources to be recommended for dissemination are determined based on the dissemination resource attribute features; and the resources to be recommended for dissemination are recommended based on the recommendation parameters. Since the embodiments of this application can select the target dissemination resource tower corresponding to the resources to be recommended for dissemination from at least two dissemination resource towers of the trained resource recommendation model based on the dissemination resource creation information, the target dissemination resource tower can be used to extract features from the resources to be recommended for dissemination, thereby obtaining the dissemination resource attribute features of the resources to be recommended for dissemination. Therefore, the recommendation parameters of any resources to be recommended for dissemination can be accurately predicted based on the dissemination resource attribute features, thereby accurately recommending the resources to be recommended for dissemination based on the recommendation parameters.

[0153] In this embodiment, the model training device and the dissemination resource recommendation device are specifically integrated into a computer device. This embodiment uses dissemination resource samples as advertising samples, dissemination resource reference samples as advertising reference samples, and the resource to be recommended as the advertisement to be recommended as an example. The computer device is a server. This embodiment can be applied to the recall and coarse-ranking stages of an advertising system. This embodiment can improve the accuracy of predicting any advertisement in the advertising system, such as new and old advertisements, thereby enhancing the recommendation effect of the advertising system; for users, it can improve the user experience.

[0154] like Figure 6 As shown, a method for recommending dissemination resources, the specific process of which is as follows: steps S601 to S614:

[0155] S601, Computer equipment acquires a set of dissemination resource samples.

[0156] The set of communication resource samples includes communication resource samples, communication resource reference samples that are related to the communication resource samples in terms of communication resource attribute information, and object preference samples corresponding to the communication resource samples. The relationship indicates that the communication resource samples and communication resource reference samples have the same communication resource attribute information.

[0157] In some embodiments, the method for obtaining the propagation resource sample set may be as follows: obtaining the propagation resource attribute information of the propagation resource sample; performing a weakening process on the propagation resource attribute information of the propagation resource sample based on the propagation resource attribute information to obtain the propagation resource reference sample corresponding to the propagation resource sample; and generating the propagation resource sample set based on the propagation resource sample and the propagation resource reference sample corresponding to the propagation resource sample.

[0158] In one example, propagating resource attribute information could mean propagating the attribute information carried by the resource sample.

[0159] In one example, the method for weakening the propagation resource attribute information of a propagation resource sample based on the propagation resource attribute information to obtain the propagation resource reference sample can be as follows: extract the attribute feature information of the propagation resource attribute information; and weaken the propagation resource attribute information of the propagation resource sample based on the attribute feature information to obtain the propagation resource reference sample.

[0160] In one example, the method for weakening the propagation resource attribute information of the propagation resource sample based on attribute feature information to obtain the propagation resource reference sample can be as follows: based on attribute feature information, the propagation resource attribute information is identified by information type to obtain the attribute information type of the propagation resource attribute information; if the attribute information type is the target attribute information type, the propagation resource attribute information of the target attribute information type is removed from the propagation resource sample to obtain the propagation resource reference sample corresponding to the propagation resource sample.

[0161] The information types can include coarse-grained resource attribute information and fine-grained resource attribute information. The target attribute information type can be fine-grained resource attribute information. In this embodiment, fine-grained resource attribute information can be removed from the propagation resource sample to obtain a propagation resource reference sample corresponding to the propagation resource sample.

[0162] In this embodiment, the propagation resource sample set can be input into the resource recommendation model to be trained in batches, including propagation resource samples, propagation resource reference samples, and object preference samples.

[0163] This application embodiment can also employ a memory bank to store target samples previously trained by the resource recommendation model, thereby expanding the number of samples in the propagation resource sample set based on the target samples and increasing the comparative learning scope of the resource recommendation model. Specifically, this application embodiment can configure the memory bank according to the memory usage constraints, speed, accuracy, and other requirements of the resource recommendation model during training.

[0164] S602. The computer equipment uses the main propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource samples and obtain the resource sample attribute features corresponding to the propagation resource samples.

[0165] Among them, the resource recommendation model to be trained in the embodiments of this application is, for example... Figure 7As shown. The resource recommendation model to be trained includes an object tower, a main propagation resource tower, and an auxiliary propagation resource tower.

[0166] The object tower can be a tower used to process object preference samples; the object tower includes a first feature extraction layer and a first feature output layer. The first feature extraction layer can be a neural network layer used to extract features from the object preference samples to obtain initial preference sample features; the first feature output layer can be a neural network layer used to process the initial preference sample features from the first feature extraction layer to obtain object preference features.

[0167] The main propagation resource tower can be a tower used to process propagation resource samples; the main propagation resource tower includes a second feature extraction layer and a second feature output layer. The second feature extraction layer can be a neural network layer used to extract features from the propagation resource samples to obtain initial resource sample attribute features; the second feature output layer can be a neural network layer used to process the initial resource sample attribute features from the second feature extraction layer to obtain resource sample attribute features.

[0168] The auxiliary propagation resource tower can be a tower used to process propagation resource reference samples; the auxiliary propagation resource tower includes a third feature extraction layer and a third feature output layer. The third feature extraction layer can be a neural network layer used to extract features from the propagation resource reference samples to obtain initial resource reference sample attribute features; the third feature output layer can be a neural network layer used to process the resource reference sample attribute features from the third feature extraction layer to obtain resource reference sample attribute features.

[0169] Based on step S602, in some application embodiments, the method of using the main propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource samples and obtain the resource sample attribute features corresponding to the propagation resource samples can be as follows: using the main propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information and the fine-grained resource attribute information of the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples.

[0170] Specifically, the method for extracting features from the coarse-grained and fine-grained resource attribute information of the propagated resource samples using the main propagation resource tower of the resource recommendation model to obtain the resource sample attribute features corresponding to the propagated resource samples can be as follows: First, the first coarse-grained resource attribute features are obtained by extracting features from the coarse-grained resource attribute information of the propagated resource samples using the main propagation resource tower of the resource recommendation model; second, the fine-grained resource attribute features are obtained by extracting features from the fine-grained resource attribute information of the propagated resource samples using the main propagation resource tower of the resource recommendation model; and third, the resource sample attribute features corresponding to the propagated resource samples are determined based on the first coarse-grained resource attribute features and the fine-grained resource attribute features.

[0171] Furthermore, specifically, the method for extracting features from the coarse-grained resource attribute information of the propagation resource samples using the main propagation resource tower of the resource recommendation model to obtain the first coarse-grained resource attribute features can be as follows: the second feature extraction layer of the main propagation resource tower in the resource recommendation model to be trained is used to extract features from the coarse-grained resource attribute information of the propagation resource samples to obtain the first initial coarse-grained resource attribute features; the second feature output layer of the main propagation resource tower in the resource recommendation model to be trained is used to extract features from the first initial coarse-grained resource attribute features of the propagation resource samples to obtain the first coarse-grained resource attribute features.

[0172] Furthermore, specifically, the method for extracting fine-grained resource attribute information from the propagation resource samples using the main propagation resource tower of the resource recommendation model to obtain fine-grained resource attribute features can be as follows: the second feature extraction layer of the main propagation resource tower in the resource recommendation model to be trained is used to extract features from the fine-grained resource attribute information of the propagation resource samples to obtain initial fine-grained resource attribute features; the second feature output layer of the main propagation resource tower in the resource recommendation model to be trained is used to extract features from the initial fine-grained resource attribute features of the propagation resource samples to obtain fine-grained resource attribute features.

[0173] S603. The computer equipment uses the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource reference samples and obtain the resource reference sample attribute features corresponding to the propagation resource reference samples.

[0174] In some application embodiments, the method of using the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the propagation resource reference sample and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample can be as follows: using the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information of the propagation resource reference sample and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0175] Furthermore, specifically, the method for extracting features from the coarse-grained resource attribute information of the propagation resource reference sample using the auxiliary propagation resource tower of the resource recommendation model to be trained, and obtaining the resource reference sample attribute features corresponding to the propagation resource reference sample, can be as follows: the third feature extraction layer of the auxiliary propagation resource tower in the resource recommendation model to be trained is used to extract features from the coarse-grained resource attribute information of the propagation resource reference sample to obtain the initial resource reference sample attribute features corresponding to the propagation resource reference sample; the third feature output layer of the auxiliary propagation resource tower in the resource recommendation model to be trained is used to extract features from the initial resource reference sample attribute features to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0176] S604. The computer equipment uses the object pyramid of the resource recommendation model to be trained to extract features from the object preference samples and obtain the preference sample features corresponding to the object preference samples.

[0177] In some application embodiments, the method of using the object pyramid of the resource recommendation model to extract features from the object preference samples to obtain the preference sample features corresponding to the object preference samples can be as follows: using the first feature extraction layer of the object pyramid in the resource recommendation model to extract features from the object preference samples to obtain the initial preference sample features corresponding to the object preference samples; using the first feature output layer of the object pyramid in the resource recommendation model to extract features from the initial preference sample features to obtain the preference sample features corresponding to the object preference samples.

[0178] In the embodiments of this application, multiple training tasks can be used to train the resource recommendation model to be trained. For example, Figure 7 As shown, the training task may include a main training task, a first auxiliary training task, and a second auxiliary training task.

[0179] The main training task is a supervised task, which primarily uses object preference samples and label information to enable the resource recommendation model to learn the preferences of the objects corresponding to the object preference samples for the disseminated resources. Objects can refer to users, and object preference samples can refer to users' user preference samples, which may include, for example, age, gender, location, preferred disseminated resource type, and historical behavior related to the disseminated resources. For details regarding the main training task, please refer to the preference feature loss section below.

[0180] The first auxiliary training task is an unsupervised task, which mainly uses object preference samples and propagated resource samples to enable the resource recommendation model to learn the correlation between the object preference samples and the propagated resource samples. For details on the first auxiliary training task, please refer to the preference correlation loss below.

[0181] The second auxiliary training task is an unsupervised task. By propagating resource samples and reference samples, the resource recommendation model to be trained learns the correlation between the propagated resource samples and the reference samples, thereby enhancing the accuracy of the auxiliary propagation resource pyramid in predicting the resources to be recommended. For details on the second auxiliary training task, please refer to the attribute correlation loss section below.

[0182] S605. The computer equipment calculates the attribute correlation loss between the propagated resource sample and the propagated resource reference sample in terms of propagated resource attribute information, based on the attribute characteristics of the resource sample and the attribute characteristics of the resource reference sample.

[0183] In some application embodiments, the method for calculating the attribute correlation loss between the propagation resource sample and the propagation resource reference sample on the propagation resource attribute information based on the resource sample attribute characteristics and the resource reference sample attribute characteristics can be as follows: using an attribute correlation loss function, the attribute correlation loss between the propagation resource sample and the propagation resource reference sample on the propagation resource attribute information can be calculated based on the resource sample attribute characteristics and the resource reference sample attribute characteristics.

[0184] Specifically, as mentioned above, the attribute correlation loss is the loss corresponding to the second auxiliary training task. For example... Figure 8 As shown, the propagation resource sample is processed sequentially through the second feature extraction layer and the second feature output layer to obtain the resource sample attribute features corresponding to the propagation resource sample. The propagation resource reference sample is processed sequentially through the third feature extraction layer and the third feature output layer to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0185] It can be understood here that, in order to weaken the dependence of the auxiliary dissemination resource tower on strong memory features such as fine-grained resource attribute information and unique advertising identifiers, the dissemination resource reference sample can be a sample obtained by weakening the dissemination resource sample. Specifically, the dissemination resource sample includes all dissemination resource attribute information, such as coarse-grained resource attribute information and fine-grained resource attribute information, while the dissemination resource reference sample includes some dissemination resource attribute information, such as coarse-grained resource attribute information. Based on this, the attribute features of the resource sample have the features corresponding to all dissemination resource attribute information, while the attribute features of the resource reference sample have the features corresponding to some dissemination resource attribute information.

[0186] Then, based on the resource sample attribute features and resource reference sample attribute features, unsupervised learning is performed using InfoNCE (InfoNoise Contrastive Estimation loss) to make the resource sample attribute features and the corresponding resource reference sample attribute features as similar as possible, and to make the resource sample attribute features and non-corresponding resource reference sample attribute features as dissimilar as possible. This weakens the dependence of the post-trained auxiliary propagation resource tower in the post-trained resource recommendation model on the fine-grained resource attribute information in the propagated resource attribute information, thereby improving the accuracy of the post-trained auxiliary propagation resource tower in predicting the resources to be recommended in the post-trained resource recommendation model.

[0187] What can be understood here is that when a propagation resource sample is input into the main propagation resource tower, a propagation resource reference sample corresponding to the propagation resource sample will be input into the auxiliary propagation resource tower, thus obtaining the resource reference sample attribute features corresponding to the resource sample attribute features; when a propagation resource reference sample that does not correspond to the propagation resource sample is input into the auxiliary propagation resource tower, the resource reference sample attribute features that do not correspond to the resource sample attribute features can be obtained.

[0188] Here, it can be understood that when a communication resource reference sample is a sample obtained by weakening a communication resource sample, and the communication resource attribute information of the communication resource reference sample is contained within the communication resource attribute information of the communication resource sample, this situation can be called a sample corresponding to the communication resource sample, that is, a positive sample of the communication resource reference sample. For example, taking an advertisement where both the communication resource sample and the communication resource reference sample are the same as an advertisement, the difference between the communication resource sample and the communication resource reference sample is that the communication resource sample has the advertisement's unique identifier, while the communication resource reference sample does not; apart from this, all other communication resource attribute information of the communication resource sample and the communication resource reference sample is the same. In this case, the communication resource reference sample can be called a sample corresponding to the communication resource sample.

[0189] When the propagation resource attribute information of the propagation resource reference sample is not included in the propagation resource attribute information of the propagation resource sample, this situation can be called a propagation resource reference sample that does not correspond to the propagation resource sample, that is, a negative sample of the propagation resource reference sample. For example, if the propagation resource sample is a lipstick advertisement and the propagation resource reference sample is a yogurt advertisement, this situation can be called a propagation resource reference sample that does not correspond to the propagation resource sample.

[0190] For details regarding InfoNCE, please refer to formula (1):

[0191]

[0192] Among them, L q InfoNCE represents τ; τ represents the temperature coefficient, which can be a preset hyperparameter. In this embodiment, for attribute correlation loss, q in formula (1) represents the resource sample attribute feature; k + k represents the resource reference sample attribute feature corresponding to the resource sample attribute feature; i This represents the attribute features of the resource reference sample corresponding to the i-th propagation resource reference sample.

[0193] S606. The computer device obtains the preference loss of the object preference sample based on the preference sample characteristics.

[0194] In some embodiments, the preference loss may include preference relevance loss and preference feature loss. As previously mentioned, the preference relevance loss is the loss corresponding to the first auxiliary training task, and the preference feature loss is the loss corresponding to the main training task.

[0195] I. Regarding the loss of preference relevance:

[0196] In some application embodiments, the specific way to obtain the preference loss of object preference samples based on preference sample characteristics can be: to calculate the preference correlation loss between object preference samples and propagation resource samples based on preference sample characteristics and resource sample attribute characteristics.

[0197] Specifically, such as Figure 9 As shown, the object preference sample is processed sequentially through the first feature extraction layer and the second feature output layer to obtain the preference sample features corresponding to the object preference sample. The propagation resource sample is processed sequentially through the second feature extraction layer and the second feature output layer to obtain the resource sample attribute features corresponding to the propagation resource sample.

[0198] Then, based on preference sample features and resource sample attribute features, unsupervised learning is performed using InfoNCE (Info Noise Contrastive Estimation loss) to ensure that preference sample features and their corresponding resource sample attribute features are as similar as possible, and preference sample features and their non-corresponding resource sample attribute features are as dissimilar as possible. This strengthens the resource recommendation model's ability to learn the correlation between object preference samples and propagated resource samples. Through cross-comparison learning between the object tower and the main propagation resource tower, the model enhances the discriminative power of different objects' preferences on different propagation resource samples, thereby improving the prediction accuracy of the trained resource recommendation model.

[0199] What can be understood here is that when an object preference sample is input into the object tower, a corresponding propagation resource sample will be input into the main propagation resource tower, thus obtaining the resource sample attribute features corresponding to the preference sample features; when a propagation resource sample that does not correspond to the object preference sample is input into the main propagation resource tower, the resource sample attribute features that do not correspond to the preference sample features can be obtained.

[0200] It can be understood here that when the propagation resource sample is a sample that has been processed by the object corresponding to the object preference sample, such as by clicking or browsing, this case can be called the propagation resource sample corresponding to the object preference sample, that is, the propagation resource sample is a positive sample of the object preference sample.

[0201] It can be understood here that when the propagation resource sample is a sample of the object corresponding to the object preference sample that has not undergone any operation processing, such as clicking or browsing, this situation can be called a propagation resource sample that does not correspond to the object preference sample, that is, the propagation resource sample is a negative sample of the object preference sample.

[0202] InfoNCE can be found in the aforementioned formula (1). Based on formula (1), in this embodiment of the application, for the preference relevance loss, q in formula (1) represents the preference sample feature; k + k represents the resource sample attribute features corresponding to the preference sample features. i This represents the resource sample attribute feature corresponding to the i-th propagation resource sample.

[0203] II. Regarding the loss based on preference features:

[0204] In some application embodiments, the specific way to obtain the preference loss of an object preference sample based on the preference sample features can be: obtaining the label information corresponding to the object preference sample; calculating the preference feature loss between the preference sample features and the label information.

[0205] Specifically, such as Figure 8 As shown, the object preference sample is processed sequentially through the first feature extraction layer and the second feature output layer to obtain the preference sample features corresponding to the object preference sample. This embodiment of the application can use preference feature loss to enable the recommendation model to be trained to learn the object's preference for propagation resources corresponding to the object preference sample.

[0206] S607. The computer equipment trains the resource recommendation model to be trained based on attribute correlation loss and preference loss, and obtains the trained resource recommendation model.

[0207] In one embodiment of the application, the resource recommendation model to be trained is trained based on attribute correlation loss and preference loss to obtain the trained resource recommendation model.

[0208] S608. The computer device acquires the resources to be recommended for dissemination, and acquires the creation information of the dissemination resources to be recommended for dissemination.

[0209] In some embodiments of the application, the propagation resource creation information can be information carried by the propagation resource to be recommended. The propagation resource creation information can be the creation time information of the propagation resource to be recommended.

[0210] S609. The computer device, based on the propagation resource creation information, selects the target post-training propagation resource tower corresponding to the propagation resource to be recommended from at least two post-training propagation resource towers of the post-training resource recommendation model.

[0211] The post-training resource recommendation model includes a post-training main propagation resource tower, a post-training auxiliary propagation resource tower, and an object tower.

[0212] In some embodiments, the post-training propagation resource tower includes a primary post-training propagation resource tower and a secondary post-training propagation resource tower. Based on this, the method for selecting the target post-training propagation resource tower corresponding to the resource to be recommended from at least two post-training propagation resource towers of the post-training resource recommendation model, according to the propagation resource creation information, can be as follows: if the propagation resource creation information conforms to a preset rule, then the primary post-training propagation resource tower is determined as the target post-training propagation resource tower corresponding to the resource to be recommended; if the propagation resource creation information does not conform to the preset rule, then the secondary post-training propagation resource tower is determined as the target post-training propagation resource tower corresponding to the resource to be recommended.

[0213] What can be understood here is that, in the application phase of the post-trained resource recommendation model, the resource to be recommended needs to determine which tower, the main post-trained propagation resource tower or the auxiliary post-trained propagation resource tower, to enter based on the propagation resource creation information.

[0214] In one example, the creation information of the propagation resource is the creation time information of the propagation resource to be recommended. When the creation time information is greater than or equal to a preset time threshold, the creation information of the propagation resource can be said to conform to the preset rules. When the creation time information is less than the preset time threshold, the creation information of the propagation resource can be said to not conform to the preset rules.

[0215] S610. The computer equipment uses the target training and propagation resource tower to extract features of the propagation resource to be recommended, and obtains the propagation resource attribute features of the propagation resource to be recommended.

[0216] In some application embodiments, when the target post-training propagation resource tower is the post-training master propagation resource tower, the post-training master propagation resource tower can be used to extract features of the propagation resource to be recommended, thereby obtaining the propagation resource attribute features of the propagation resource to be recommended.

[0217] The process of extracting features from the propagation resource to be recommended using the trained master propagation resource tower to obtain the propagation resource attribute features can be found in the aforementioned processing process of the propagation resource samples by the master propagation resource tower, and will not be repeated here.

[0218] In some application embodiments, when the target post-training propagation resource tower is a post-training auxiliary propagation resource tower, the post-training auxiliary propagation resource tower can be used to extract features of the propagation resource to be recommended, thereby obtaining the propagation resource attribute features of the propagation resource to be recommended.

[0219] The process of using a post-trained auxiliary propagation resource tower to extract features from the resources to be recommended and obtain the propagation resource attribute features can be found in the aforementioned process of the auxiliary propagation resource tower processing the propagation resource reference samples, and will not be repeated here.

[0220] In one example, when the target post-training propagation resource tower is a post-training auxiliary propagation resource tower, the post-training auxiliary propagation resource tower is used to extract features from the propagation resource to be recommended, and the resulting propagation resource attribute features can specifically include: obtaining the propagation resource attribute information of the propagation resource to be recommended; weakening the propagation resource to be recommended based on the propagation resource attribute information to obtain the weakened propagation resource; extracting features from the weakened propagation resource using the post-training auxiliary propagation resource tower to obtain the weakened propagation resource features; and using the weakened propagation resource features as the propagation resource attribute features of the propagation resource to be recommended.

[0221] S611. The computer device obtains the object preference information of the recommended objects for the resources to be recommended and disseminated.

[0222] Among them, object preference information can be information obtained in real time.

[0223] S612. The computer equipment uses the trained object pyramid to extract features from the object preference information, and obtains the object preference features corresponding to the object preference information.

[0224] Among them, the object preference information is extracted by using the trained object tower to obtain the object preference features corresponding to the object preference information. For details, please refer to the above-mentioned object tower processing process for object preference samples, which will not be repeated here.

[0225] S613. The computer equipment performs mapping processing on the preference sample features and the propagation resource attribute features to obtain the recommendation parameters of the propagation resource to be recommended.

[0226] In one embodiment of the application, the similarity between preference sample features and propagation resource attribute features can be calculated to obtain recommendation parameters for the propagation resource to be recommended.

[0227] Specifically, the inner product of the preference sample features and the propagation resource attribute features can be calculated to obtain the recommendation parameters of the propagation resource to be recommended.

[0228] S614. The computer equipment recommends resources to be disseminated based on the recommended parameters.

[0229] In some application embodiments, the method for recommending resources to be promoted based on recommendation parameters can be as follows: sorting the resources according to the size of the recommendation parameters to obtain sorted recommendation parameters; based on the sorted recommendation parameters, selecting a preset number of target resources to be promoted from the resources to be promoted; and recommending the target resources to be promoted.

[0230] Steps S602, S603, and S604 can be parallel steps.

[0231] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0232] Based on the above, it can be seen that the embodiments of this application physically split the propagation resource tower into at least two towers, such as a main propagation resource tower and an auxiliary propagation resource tower. Using contrastive learning, the auxiliary propagation resource tower, while removing strong memory attribute information such as fine-grained resource attribute information, tries to fit the resource sample attribute features output by the main propagation resource tower as closely as possible. This improves the scoring ability of the trained main propagation resource tower for new advertisements in the trained resource recommendation model without affecting the prediction effect of the trained main propagation resource tower for old advertisements. The embodiments of this application also introduce cross-contrast learning between the object tower and the main propagation resource tower to strengthen the discriminative power of different objects' preferences on different propagation resource samples, thereby improving the performance of the advertisement recommendation algorithm.

[0233] This application embodiment can obtain a set of propagation resource samples, which includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. Using the main propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples. Using the auxiliary propagation resource tower of the resource recommendation model to be trained, features are extracted from the propagation resource reference samples to obtain the resource reference sample attribute features corresponding to the propagation resource reference samples. Based on the resource sample attribute features and the resource reference sample attribute features, the attribute correlation loss between the propagation resource samples and the propagation resource reference samples in terms of propagation resource attribute information is calculated. Based on the attribute correlation loss, the resource recommendation model to be trained is trained to obtain the trained resource recommendation model. Since the embodiments of this application can employ a main propagation resource tower and an auxiliary propagation resource tower of the resource recommendation model to be trained, and extract features from propagation resource samples and propagation resource reference samples that have a correlation in propagation resource attributes, attribute correlation loss can be calculated based on the extracted resource sample attribute features and resource reference sample attribute features. This allows the resource reference sample attribute features of the auxiliary propagation resource tower to fit the resource sample attribute features of the main propagation resource tower as closely as possible based on the attribute correlation loss. Consequently, the dependence of the auxiliary propagation resource tower on propagation resource attribute information in the trained resource recommendation model can be weakened, thereby improving the accuracy of the trained resource recommendation model in predicting any propagation resource to be recommended.

[0234] To better implement the above methods, this application also provides a model training device that can be integrated into a computer device, such as a server or terminal. The terminal may include a tablet computer, a laptop computer, and / or a personal computer.

[0235] For example, such as Figure 10 As shown, the model training device may include a first acquisition unit 301, a first extraction unit 302, a second extraction unit 303, a calculation unit 304, a training unit 305, and a third extraction unit 306, as follows:

[0236] (1) First acquisition unit;

[0237] The first acquisition unit can be used to acquire a set of propagation resource samples. The set of propagation resource samples includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. The relationship indicates that the propagation resource samples and the propagation resource reference samples have the same propagation resource attribute information.

[0238] In some embodiments, the first acquisition unit may be used to acquire the propagation resource attribute information of the propagation resource sample; perform propagation resource attribute information weakening processing on the propagation resource sample according to the propagation resource attribute information to obtain the propagation resource reference sample corresponding to the propagation resource sample; and generate a propagation resource sample set according to the propagation resource sample and the propagation resource reference sample corresponding to the propagation resource sample.

[0239] In some embodiments, the first acquisition unit may be used to extract attribute feature information of propagation resource attribute information; and based on the attribute feature information, perform propagation resource attribute information weakening processing on the propagation resource sample to obtain a propagation resource reference sample corresponding to the propagation resource sample.

[0240] In some embodiments, the first acquisition unit can be used to identify the information type of the propagation resource attribute information based on the attribute feature information to obtain the attribute information type of the propagation resource attribute information; if the attribute information type is the target attribute information type, the propagation resource attribute information of the target attribute information type is removed from the propagation resource sample to obtain the propagation resource reference sample corresponding to the propagation resource sample.

[0241] (2) First extraction unit;

[0242] The first extraction unit can be used to extract features from the propagation resource samples using the main propagation resource tower of the resource recommendation model to be trained, and obtain the resource sample attribute features corresponding to the propagation resource samples.

[0243] In some embodiments, the propagation resource attribute information of the propagation resource sample includes coarse-grained resource attribute information and fine-grained resource attribute information; the first extraction unit can be used to use the main propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information and the fine-grained resource attribute information of the propagation resource sample, so as to obtain the resource sample attribute features corresponding to the propagation resource sample.

[0244] (3) Second extraction unit;

[0245] The second extraction unit can be used to extract features from the propagation resource reference sample using the auxiliary propagation resource tower of the resource recommendation model to be trained, and obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0246] In some embodiments, the propagation resource attribute information of the propagation resource reference sample includes coarse-grained resource attribute information; the second extraction unit can be used to use the auxiliary propagation resource tower of the resource recommendation model to be trained to extract features from the coarse-grained resource attribute information of the propagation resource reference sample, so as to obtain the resource reference sample attribute features corresponding to the propagation resource reference sample.

[0247] (4) Calculation unit;

[0248] The calculation unit can be used to calculate the attribute correlation loss between the propagated resource sample and the propagated resource reference sample in terms of propagated resource attribute information, based on the attribute characteristics of the resource sample and the attribute characteristics of the resource reference sample.

[0249] In some embodiments, the computing unit can be used to obtain the preference loss of an object preference sample based on the preference sample characteristics.

[0250] In some embodiments, preference loss includes preference correlation loss; the calculation unit can be used to calculate the preference correlation loss between object preference samples and propagation resource samples based on preference sample characteristics and resource sample attribute characteristics.

[0251] (5) Training unit;

[0252] The training unit can be used to train the resource recommendation model to be trained based on the attribute correlation loss, so as to obtain the trained resource recommendation model.

[0253] In some embodiments, the training unit can be used to train the resource recommendation model to be trained based on attribute correlation loss and preference loss, so as to obtain the trained resource recommendation model.

[0254] In some embodiments, the training unit can be used to train the resource recommendation model to be trained based on attribute correlation loss and preference correlation loss, so as to obtain the trained resource recommendation model.

[0255] In some embodiments, the preference loss further includes preference feature loss; a training unit, which can be used to obtain label information corresponding to object preference samples; calculate the preference feature loss between preference sample features and label information; and train the resource recommendation model to be trained based on attribute correlation loss, preference correlation loss and preference feature loss to obtain the trained resource recommendation model.

[0256] (6) Third extraction unit;

[0257] The third extraction unit can be used to extract features from object preference samples using the object pyramid of the resource recommendation model to be trained, and obtain the preference sample features corresponding to the object preference samples.

[0258] As can be seen from the above, the first acquisition unit of this application embodiment can be used to acquire a set of propagation resource samples, which includes propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information; the first extraction unit can be used to extract features from the propagation resource samples using the main propagation resource tower of the resource recommendation model to be trained, and obtain the resource sample attribute features corresponding to the propagation resource samples; the second extraction unit can be used to extract features from the propagation resource reference samples using the auxiliary propagation resource tower of the resource recommendation model to be trained, and obtain the resource reference sample attribute features corresponding to the propagation resource reference samples; the calculation unit can be used to calculate the attribute correlation loss between the propagation resource samples and the propagation resource reference samples in terms of propagation resource attribute information based on the resource sample attribute features and the resource reference sample attribute features; the training unit can be used to train the resource recommendation model to be trained based on the attribute correlation loss, and obtain the trained resource recommendation model. Since the embodiments of this application can employ a main propagation resource tower and an auxiliary propagation resource tower of the resource recommendation model to be trained, and extract features from propagation resource samples and propagation resource reference samples that have a correlation in propagation resource attributes, attribute correlation loss can be calculated based on the extracted resource sample attribute features and resource reference sample attribute features. This allows the resource reference sample attribute features of the auxiliary propagation resource tower to fit the resource sample attribute features of the main propagation resource tower as closely as possible based on the attribute correlation loss. Consequently, the dependence of the auxiliary propagation resource tower on propagation resource attribute information in the trained resource recommendation model can be weakened, thereby improving the accuracy of the trained resource recommendation model in predicting any propagation resource to be recommended.

[0259] To better implement the above methods, embodiments of this application also provide a propagation resource recommendation device, including the aforementioned post-trained resource recommendation model, which includes at least two post-trained propagation resource towers. This propagation resource recommendation device can be integrated into a computer device, such as a server or terminal, which may include a tablet computer, laptop computer, and / or personal computer.

[0260] For example, such as Figure 11 As shown, the model training device may include a second acquisition unit 501, a filtering unit 502, a fourth extraction unit 503, a determination unit 504, and a recommendation unit 505, as follows:

[0261] (1) Second acquisition unit;

[0262] The second acquisition unit can be used to acquire resources to be recommended for dissemination, as well as information on the creation of those resources.

[0263] (2) Filtering unit;

[0264] The filtering unit can be used to filter out the target post-trained propagation resource tower corresponding to the propagation resource to be recommended from at least two post-trained propagation resource towers of the post-trained resource recommendation model based on the propagation resource creation information.

[0265] In some embodiments, the post-training propagation resource tower includes a post-training main propagation resource tower and a post-training auxiliary propagation resource tower; the filtering unit can be used to determine the post-training main propagation resource tower as the target post-training propagation resource tower corresponding to the propagation resource to be recommended if the propagation resource creation information conforms to a preset rule; and to determine the post-training auxiliary propagation resource tower as the target post-training propagation resource tower corresponding to the propagation resource to be recommended if the propagation resource creation information does not conform to the preset rule.

[0266] (3) Fourth extraction unit;

[0267] The fourth extraction unit can be used to extract features from the propagation resource tower after target training to obtain the propagation resource attribute features of the propagation resource to be recommended.

[0268] In some embodiments, the fourth extraction unit can be used to obtain object preference information of the recommended object targeted by the resource to be recommended and propagated, wherein the object preference information represents the preference attribute information of the recommended object for the resource to be recommended and propagated; and to extract features from the object preference information using a trained object pyramid to obtain the object preference features corresponding to the object preference information.

[0269] (4) Determine the unit;

[0270] The determination unit can be used to determine the recommendation parameters of the resources to be recommended based on the attribute characteristics of the resources.

[0271] In some embodiments, the post-trained resource recommendation model further includes a post-trained object tower; a determination unit, which can be used to map the preference sample features and the propagation resource attribute features to the recommendation parameters to obtain the recommendation parameters of the resources to be recommended for propagation.

[0272] (5) Recommended Units;

[0273] The recommendation unit can be used to recommend resources to be promoted based on recommendation parameters.

[0274] As can be seen from the above, the second acquisition unit in this application embodiment can be used to acquire the resource to be recommended for propagation and to acquire the propagation resource creation information of the resource to be recommended for propagation; the filtering unit can be used to filter out the target propagation resource tower corresponding to the resource to be recommended from at least two propagation resource towers of the trained resource recommendation model based on the propagation resource creation information; the fourth extraction unit can be used to extract features of the resource to be recommended for propagation using the target propagation resource tower to obtain the propagation resource attribute features of the resource to be recommended for propagation; the determination unit can be used to determine the recommendation parameters of the resource to be recommended for propagation based on the propagation resource attribute features; and the recommendation unit can be used to recommend the resource to be recommended for propagation based on the recommendation parameters. Since the embodiments of this application can create information based on propagation resources, and select a target propagation resource tower corresponding to the propagation resource to be recommended from at least two propagation resource towers of the post-trained resource recommendation model, the target propagation resource tower can be used to extract features of the propagation resource to be recommended, thereby obtaining the propagation resource attribute features of the propagation resource to be recommended. Based on the propagation resource attribute features, the recommendation parameters of any propagation resource to be recommended can be accurately predicted, and the propagation resource to be recommended can be accurately recommended according to the recommendation parameters.

[0275] This application also provides a computer device, such as... Figure 12 As shown, it illustrates a structural schematic diagram of the computer device involved in the embodiments of this application, specifically:

[0276] The computer device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 12 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0277] The processor 401 is the control center of the computer device, connecting various parts of the computer device through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 402, and by calling data stored in the memory 402. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and computer programs, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0278] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, computer programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0279] The computer device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0280] The computer device may also include an input unit 404, which can be used to receive input digital or character information communication, and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0281] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the computer device loads the executable files corresponding to the processes of one or more computer programs into the memory 402 according to the following instructions, and the processor 401 runs the computer programs stored in the memory 402 to realize various functions, as follows:

[0282] A set of propagation resource samples is obtained, including propagation resource samples and propagation resource reference samples that are related to the propagation resource samples in terms of propagation resource attribute information. The related relationship indicates that the propagation resource samples and propagation resource reference samples have the same propagation resource attribute information. The main propagation resource tower of the resource recommendation model to be trained is used to extract features from the propagation resource samples to obtain the resource sample attribute features corresponding to the propagation resource samples. The auxiliary propagation resource tower of the resource recommendation model to be trained is used to extract features from the propagation resource reference samples to obtain the resource reference sample attribute features corresponding to the propagation resource reference samples. Based on the resource sample attribute features and the resource reference sample attribute features, the attribute correlation loss between the propagation resource samples and the resource reference samples in terms of propagation resource attribute information is calculated. Based on the attribute correlation loss, the resource recommendation model to be trained is trained to obtain the trained resource recommendation model.

[0283] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0284] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by a computer program, or by a computer program controlling related hardware. The computer program can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0285] Therefore, embodiments of this application provide a computer-readable storage medium storing a computer program that can be loaded by a processor to execute any of the model training methods and any of the resource recommendation methods provided in embodiments of this application.

[0286] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0287] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0288] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the model training methods provided in the embodiments of this application, the beneficial effects that any of the model training methods and any of the propagation resource recommendation methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0289] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above embodiments.

[0290] The foregoing has provided a detailed description of a model training method, a propagation resource recommendation method, a model training device, a propagation resource recommendation device, a computer device, and a computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A model training method, characterized in that, The method comprises the following steps: obtaining a set of propagation resource samples, the set of propagation resource samples comprising a propagation resource sample, a propagation resource reference sample having an association relationship with the propagation resource sample in propagation resource attribute information, and an object preference sample corresponding to the propagation resource sample, the association relationship indicating that the propagation resource sample and the propagation resource reference sample have the same propagation resource attribute information; wherein the propagation resource sample is any one of an image, an audio / video, an audio, and a text, or a combination of at least any two of them; the propagation resource reference sample is any one of an image, an audio / video, an audio, and a text, or a combination of at least any two of them; performing feature extraction on coarse-grained resource attribute information and fine-grained resource attribute information of the propagation resource sample by using a main propagation resource tower of a to-be-trained resource recommendation model to obtain resource sample attribute features corresponding to the propagation resource sample; performing feature extraction on coarse-grained resource attribute information of the propagation resource reference sample by using an auxiliary propagation resource tower of the to-be-trained resource recommendation model to obtain resource reference sample attribute features corresponding to the propagation resource reference sample; calculating attribute correlation loss between the propagation resource sample and the propagation resource reference sample in the propagation resource attribute information according to the resource sample attribute features and the resource reference sample attribute features; performing feature extraction on the object preference sample by using an object tower of the to-be-trained resource recommendation model to obtain preference sample features corresponding to the object preference sample; obtaining a preference loss of the object preference sample according to the preference sample features; wherein the preference loss comprises a preference correlation loss and a preference feature loss; training the to-be-trained resource recommendation model according to the attribute correlation loss, the preference correlation loss, and the preference feature loss to obtain a trained resource recommendation model.

2. The model training method of claim 1, wherein, The method comprises the following steps: calculating a preference correlation loss between the object preference sample and the propagation resource sample according to the preference sample features and the resource sample attribute features.

3. The model training method of claim 1, wherein, The preference feature loss is obtained in the following manner: obtaining label information corresponding to the object preference sample; calculating a preference feature loss between the preference sample features and the label information. 4.The model training method of any one of claims 1 to 3, wherein, The method comprises the following steps: obtaining propagation resource attribute information of the propagation resource sample; performing weakening processing of the propagation resource attribute information on the propagation resource sample according to the propagation resource attribute information to obtain a propagation resource reference sample corresponding to the propagation resource sample; generating the set of propagation resource samples according to the propagation resource sample and the propagation resource reference sample corresponding to the propagation resource sample.

5. The model training method of claim 4, wherein, The method comprises the following steps: extracting attribute feature information of the propagation resource attribute information; According to the attribute feature information, the propagation resource sample is subjected to weakening processing of the propagation resource attribute information, so as to obtain a propagation resource reference sample corresponding to the propagation resource sample.

6. The model training method of claim 5, wherein, The weakening processing of the propagation resource attribute information according to the attribute feature information comprises: According to the attribute feature information, the propagation resource attribute information is subjected to information type identification, so as to obtain an attribute information type of the propagation resource attribute information. If the attribute information type is a target attribute information type, the propagation resource attribute information of the target attribute information type is removed from the propagation resource sample, so as to obtain a propagation resource reference sample corresponding to the propagation resource sample.

7. A method of propagating resource recommendations, the method comprising: The method comprises: obtaining a to-be-recommended propagation resource and obtaining propagation resource creation information of the to-be-recommended propagation resource; According to the propagation resource creation information, a target trained propagation resource tower corresponding to the to-be-recommended propagation resource is selected from at least two trained propagation resource towers of the trained resource recommendation model, the trained propagation resource tower comprising a trained main propagation resource tower and a trained auxiliary propagation resource tower; if the propagation resource creation information meets a preset rule, the trained main propagation resource tower is determined as the target trained propagation resource tower corresponding to the to-be-recommended propagation resource; if the propagation resource creation information does not meet the preset rule, the trained auxiliary propagation resource tower is determined as the target trained propagation resource tower corresponding to the to-be-recommended propagation resource; The target trained propagation resource tower is used to extract features of the to-be-recommended propagation resource, so as to obtain propagation resource attribute features of the to-be-recommended propagation resource; According to the propagation resource attribute features, a recommendation parameter of the to-be-recommended propagation resource is determined; According to the recommendation parameter, the to-be-recommended propagation resource is recommended.

8. The method of propagating resource recommendations of claim 7, wherein, The trained resource recommendation model further comprises a trained object tower; before the recommendation parameter of the to-be-recommended propagation resource is determined according to the propagation resource attribute features, the method further comprises: object preference information of a recommended object to which the to-be-recommended propagation resource is directed is obtained; The trained object tower is used to extract features of the object preference information, so as to obtain object preference features corresponding to the object preference information; The recommendation parameter of the to-be-recommended propagation resource is determined according to the propagation resource attribute features, comprising: the preference sample features and the propagation resource attribute features are subjected to mapping processing of the recommendation parameter, so as to obtain the recommendation parameter of the to-be-recommended propagation resource.

9. A model training apparatus, comprising: comprises: The acquisition unit is configured to acquire a propagation resource sample set, the propagation resource sample set comprising a propagation resource sample, a propagation resource reference sample having an association relationship with the propagation resource sample in propagation resource attribute information, and an object preference sample corresponding to the propagation resource sample, the association relationship indicating that the propagation resource sample and the propagation resource reference sample have the same propagation resource attribute information; wherein the propagation resource sample is any one of an image, an audio / video, an audio, and a text, or a combination of at least any two thereof; and the propagation resource reference sample is any one of an image, an audio / video, an audio, and a text, or a combination of at least any two thereof. The first extraction unit is configured to perform feature extraction on coarse-grained resource attribute information and fine-grained resource attribute information of the propagation resource sample by using a main propagation resource tower of the to-be-trained resource recommendation model, to obtain resource sample attribute features corresponding to the propagation resource sample. The second extraction unit is configured to perform feature extraction on coarse-grained resource attribute information of the propagation resource reference sample by using an auxiliary propagation resource tower of the to-be-trained resource recommendation model, to obtain resource reference sample attribute features corresponding to the propagation resource reference sample. The third extraction unit is configured to perform feature extraction on the object preference sample by using an object tower of the to-be-trained resource recommendation model, to obtain preference sample features corresponding to the object preference sample. The calculation unit is configured to calculate attribute correlation loss between the propagation resource sample and the propagation resource reference sample in the propagation resource attribute information according to the resource sample attribute features and the resource reference sample attribute features, and to obtain preference loss of the object preference sample; wherein the preference loss comprises preference correlation loss and preference feature loss. The training unit is configured to train the to-be-trained resource recommendation model according to the attribute correlation loss, the preference correlation loss, and the preference feature loss, to obtain a trained resource recommendation model.

10. The apparatus of claim 9, wherein, The calculation unit is further configured to calculate preference correlation loss between the object preference sample and the propagation resource sample according to the preference sample features and the resource sample attribute features.

11. The apparatus of claim 9, wherein, The training unit is further configured to obtain label information corresponding to the object preference sample, and to calculate preference feature loss between the preference sample features and the label information.

12. The apparatus according to any one of claims 9 to 11, the first acquisition unit is configured to acquire propagation resource attribute information of the propagation resource sample; to perform weakening processing of the propagation resource attribute information on the propagation resource sample according to the propagation resource attribute information, to obtain a propagation resource reference sample corresponding to the propagation resource sample; and to generate the propagation resource sample set according to the propagation resource sample and the propagation resource reference sample corresponding to the propagation resource sample.

13. The apparatus of claim 12, wherein, The first acquisition unit is configured to extract attribute feature information of the propagation resource attribute information; to perform weakening processing of the propagation resource attribute information on the propagation resource sample according to the attribute feature information, to obtain a propagation resource reference sample corresponding to the propagation resource sample.

14. The apparatus of claim 13, wherein, The first obtaining unit is configured to perform information type identification on the propagation resource attribute information according to the attribute feature information, to obtain an attribute information type of the propagation resource attribute information; and if the attribute information type is a target attribute information type, the target attribute information type of the propagation resource attribute information is removed from the propagation resource sample, to obtain a propagation resource reference sample corresponding to the propagation resource sample.

15. A propagation resource recommendation device, comprising the trained resource recommendation model according to any one of claims 9 to 14, the trained resource recommendation model comprising at least two trained propagation resource towers; and comprising: The second obtaining unit is configured to obtain a to-be-recommended propagation resource and obtain propagation resource creation information of the to-be-recommended propagation resource. The screening unit is configured to screen a target trained propagation resource tower corresponding to the to-be-recommended propagation resource from the at least two trained propagation resource towers of the trained resource recommendation model according to the propagation resource creation information, the trained propagation resource tower comprising a trained main propagation resource tower and a trained auxiliary propagation resource tower; and comprising: if the propagation resource creation information meets a preset rule, the trained main propagation resource tower is determined as the target trained propagation resource tower corresponding to the to-be-recommended propagation resource; and if the propagation resource creation information does not meet the preset rule, the trained auxiliary propagation resource tower is determined as the target trained propagation resource tower corresponding to the to-be-recommended propagation resource. The third extracting unit is configured to perform feature extraction on the to-be-recommended propagation resource by using the target trained propagation resource tower, to obtain propagation resource attribute features of the to-be-recommended propagation resource. The determining unit is configured to determine a recommendation parameter of the to-be-recommended propagation resource according to the propagation resource attribute features. The recommendation unit is configured to recommend the to-be-recommended propagation resource according to the recommendation parameter.

16. The apparatus of claim 15, wherein, The trained resource recommendation model further comprises a trained object tower; and the extracting unit is configured to obtain object preference information of a recommendation object to which the to-be-recommended propagation resource is directed, the object preference information representing preference attribute information of the recommendation object to the to-be-recommended propagation resource. The trained object tower is used to perform feature extraction on the object preference information, to obtain object preference features corresponding to the object preference information. The determining unit is configured to perform mapping processing of the recommendation parameter on the preference sample features and the propagation resource attribute features, to obtain the recommendation parameter of the to-be-recommended propagation resource.

17. A computer device, comprising: The memory stores a computer program, and the processor is configured to run the computer program in the memory to execute the model training method in any one of claims 1 to 6 and the propagation resource recommendation method in any one of claims 7 to 8.

18. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded by the processor to execute the model training method in any one of claims 1 to 6 and the propagation resource recommendation method in any one of claims 7 to 8.

19. A computer program product, characterised in that, The computer program product stores a computer program, and the computer program is adapted to be loaded by the processor to execute the model training method in any one of claims 1 to 6 and the propagation resource recommendation method in any one of claims 7 to 8.

Images