Data processing method and device, equipment and storage medium

Abstract

The embodiment of the invention discloses a data processing method and device, equipment and a storage medium. The data processing method and device can be applied to various scenes such as artificial intelligence and intelligent traffic. The data processing method comprises the following steps: through a pre-training generation model, according to a first historical interaction feature of a business object for pushed historical media resources, predicting to obtain a preference media resource list; through a service identification model, according to a second resource feature of a historical media resource in the preference media resource list, extracting a key resource feature used for reflecting the preference of the service object for the target media resource from the first resource feature of the to-be-pushed target media resource; and according to the key resource feature, the first historical interaction feature and the first resource feature, obtaining a resource interaction tag of the business object for the target media resource, and according to the resource interaction tag, pushing the target media resource to the business object. By adopting the method and the device, the pushing accuracy of the target media resource can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data processing method, apparatus, device, and storage medium. Background Art

[0002] With the development of Internet technology and the ever-increasing scale of network data, a large number of media applications (such as gaming, video, audio, and advertising applications) have emerged one after another. The media resources within these applications (such as gaming, video, audio, and advertising resources) have also continued to grow. The needs of business entities (such as users) are also becoming increasingly diverse and personalized. To increase the fun and usage of media applications, media resources within these applications can be pushed to business entities. In practice, it has been found that different business entities often have different media resource requirements. If media resources are pushed uniformly to all business entities, some business entities may find it difficult to obtain the media resources they are interested in, resulting in low push accuracy. Summary of the Invention

[0003] The embodiments of the present application provide a data processing method, apparatus, device, and storage medium, which can improve the push accuracy of target media resources.

[0004] An embodiment of the present application provides a data processing method, including:

[0005] Acquire a first historical interaction feature of the business object for the pushed historical media resource and a first resource feature of the target media resource to be pushed;

[0006] A pre-trained model is generated, and a preferred media resource list is predicted based on the first historical interaction feature; the preferred media resource list includes historical media resources preferred by the business object;

[0007] Extracting, from the first resource feature, a key resource feature reflecting the preference of the business object for the target media resource based on the second resource feature of the historical media resource in the preferred media resource list using the business identification model;

[0008] A resource interaction tag of the business object for the target media resource is obtained according to the key resource feature, the first historical interaction feature, and the first resource feature, and the target media resource is pushed to the business object according to the resource interaction tag.

[0009] An embodiment of the present application provides a data processing device, including:

[0010] A first acquisition module is configured to acquire a first historical interaction feature of a business object for a pushed historical media resource and a first resource feature of a target media resource to be pushed;

[0011] A first prediction module, configured to predict a list of preferred media resources according to first historical interaction features through a pre-trained generation model; the list of preferred media resources includes historical media resources preferred by a business object.

[0012] A first extraction module, configured to extract, from first resource features, key resource features for reflecting the preference of the business object for a target media resource according to second resource features of historical media resources in the list of preferred media resources through a service recognition model.

[0013] A push module, configured to obtain a resource interaction label of the business object for the target media resource according to the key resource features, the first historical interaction features, and the first resource features, and push the target media resource to the business object according to the resource interaction label.

[0014] On the one hand, an embodiment of the present application provides a computer-readable storage medium storing a computer program, which is suitable for being loaded and executed by a processor so that a computer device having the processor executes the method provided by the embodiment of the present application.

[0015] On the one hand, an embodiment of the present application provides a computer program product or a computer program, the computer program product or the computer program including 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 the processor executes the computer instructions so that the computer device executes the method provided by the embodiment of the present application.

[0016] In the embodiment of the present application, through a pre-trained generation model, according to first historical interaction features of a business object for historical media resources, a list of preferred media resources including historical media resources preferred by the business object is predicted. Since the pre-trained generation model is a large model with a large number of model parameters and a relatively deep number of model layers, it can accurately depict the interest of the business object in media resources, and at the same time can accurately extract the sequential information of the first historical interaction features in time, reduce the loss of sequential information, and improve the prediction accuracy of the list of preferred media resources. Further, the list of preferred media resources is used as an additional feature to input into the service recognition model, and the service recognition model extracts key resource features for reflecting the preference of the business object for the target media resource from the first resource features. Then, according to the key resource features, the first historical interaction features, and the first resource features, a resource interaction label of the business object for the target media resource is obtained, and the target media resource is pushed to the business object according to the resource interaction label. In this way, by the service recognition model outputting the resource interaction label in real time, it can not only ensure that the media resource interest of the business object is accurately depicted, improve the accuracy of pushing the target media resource, but also ensure the real-time performance of pushing the target media resource. Brief Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 It is a schematic structural diagram of a data processing system provided by an embodiment of the present application;

[0019] Figure 2 It is a schematic diagram of a media resource push method provided by an embodiment of the present application;

[0020] Figure 3 It is a schematic flowchart of a data processing method provided by an embodiment of the present application;

[0021] Figure 4 It is a schematic diagram of the training process of a pre-trained generation model provided by an embodiment of the present application;

[0022] Figure 5 It is a schematic diagram of predicting resource interaction tags provided by an embodiment of the present application;

[0023] Figure 6 It is a schematic flowchart of a data processing method provided by an embodiment of the present application;

[0024] Figure 7 It is a schematic diagram of pushing game props provided by an embodiment of the present application;

[0025] Figure 8 It is a schematic diagram of pushing video data provided by an embodiment of the present application;

[0026] Figure 9 It is a schematic structural diagram of a data processing device provided by an embodiment of the present application;

[0027] Figure 10 It is a schematic diagram of a computer device provided by an embodiment of the present application. Detailed Description of the Embodiments

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0029] This application relates to the field of artificial intelligence technology. Specifically, in the embodiments of this application, a pre-trained generation model can be used to predict a list of preferred media resources for a business object, improving the prediction accuracy and efficiency of the list of preferred media resources. Through a business recognition model, key resource features reflecting the preferences of the business object for the target media resource can be extracted, and then the resource interaction label of the business object for the target media resource can be identified, improving the prediction accuracy and efficiency of the resource interaction label.

[0030] Among them, Artificial Intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, enabling machines to have the functions of perception, reasoning, and decision-making. Artificial intelligence technology is an interdisciplinary subject with a wide range of fields, including both hardware-level technologies and software-level technologies. Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and intelligent transportation.

[0031] Specifically, this application specifically relates to machine learning under artificial intelligence technology. Machine Learning (ML) is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning.

[0032] This application also relates to the technical field of large models (PTM, Pre-training model). A pre-training model, also known as a foundation model or a large model, refers to a deep neural network (DNN) with a large number of parameters. It is trained on a large amount of unlabeled data, and uses the function approximation ability of the large-parameter DNN to extract common features from the data. Through techniques such as fine tune, parameter-efficient fine-tuning (PEFT), and prompt-tuning, it is applicable to downstream tasks. Therefore, the pre-training model can achieve ideal results in few-shot or zero-shot scenarios. PTM can be classified into language models (ELMO, BERT, GPT), vision models (swin-transformer, ViT, V-MOE), speech models (VALL-E), multi-modal models (ViBERT, CLIP, Flamingo, Gato), etc. according to the data modalities processed. Among them, a multi-modal model refers to a model that establishes feature representations of two or more data modalities. The pre-training model is an important tool for outputting artificial intelligence-generated content (AIGC) and can also be used as a general interface connecting multiple specific task models. Specifically, the pre-training generation model in this application can be a large model. Through the pre-training generation model, a list of preferred media resources for a business object can be predicted, improving the prediction accuracy and efficiency of the list of preferred media resources.

[0033] Please refer to Figure 1 , Figure 1 which is a schematic structural diagram of a data processing system provided by an embodiment of this application. As Figure 1 shown, the data processing system may include a server 10 and a cluster of terminal devices. The cluster of terminal devices may include one or more terminal devices, and the number of terminal devices will not be limited here. As Figure 1 shown, it may specifically include terminal device 100a, terminal device 100b, terminal device 100c, …, terminal device 100n. As Figure 1 shown, terminal device 100a, terminal device 100b, terminal device 100c, …, terminal device 100n may be respectively connected to the above-mentioned server 10 through a network, so that each terminal device can interact with the server 10 through this network connection for data. Of course, terminal device 100a, terminal device 100b, terminal device 100c, …, terminal device 100n can communicate with each other through a direct network connection, that is, point-to-point communication can be achieved between each terminal device; that is to say, when data needs to be exchanged between every two terminal devices, one terminal device (i.e., the sending terminal device) can directly send the data to another terminal device (i.e., the receiving terminal).

[0034] Among them, each terminal device in the terminal device cluster may include: smart phones, tablet computers, laptop computers, desktop computers, intelligent voice interaction devices, smart home appliances (such as smart TVs), wearable devices, vehicle-mounted terminals, and other intelligent terminals with data processing capabilities. It should be understood that, as Figure 1 shown, each terminal device in the terminal device cluster may be installed with a media application having the function of processing media resources. When the media application runs on each terminal device, it can respectively perform data interaction with the above Figure 1 shown server 10. For example, the media application may specifically include game applications, video applications, audio applications, and advertising applications. For ease of understanding, embodiments of the present application may select one terminal device as the target terminal device among Figure 1 shown multiple terminal devices. For example, embodiments of the present application may use Figure 1 shown terminal device 100a as the target terminal device. The target terminal device may be installed with a media application having the function of processing media resources. At this time, the target terminal device can achieve data interaction with the server 10 through the media application.

[0035] Among them, as Figure 1 shown, the server 10 may be a device that provides background services for applications in the terminal device. The server 10 may be an independent physical server, or may also be a server cluster or distributed system composed of multiple physical servers, or may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0036] It should be understood that a data processing system based on Figure 1 is applicable to the push scenario for media resources. It can be understood that the media resources in embodiments of the present application may be game resources (such as game props, game virtual characters, game scene maps, etc.), video resources (such as video data, video tasks, video benefits, etc.), advertising resources (such as product advertisements, economic advertisements, etc.). The historical media resources that have been pushed may refer to the media resources that have been pushed within a historical time period. Taking game props as an example of media resources, the game props that have been pushed may refer to the game props that have been pushed to users within a historical time period. Among them, the business object may refer to the user in the media application that provides the media resources.

[0037] Among them, the first historical interaction feature may include the historical interaction features of the business object for historical media resources, which may include resource usage features, resource click features, resource purchase features, resource sharing features, etc., and can be extracted from resource interaction data such as the resource usage data, resource click data, resource purchase data, and resource sharing data of the business object for historical media resources. For example, taking the media resource as a game prop, the first historical interaction feature of the historical game prop may include the prop usage feature, prop click feature, prop purchase feature, prop sharing feature, etc. of the business object for the historical game prop.

[0038] The target media resource to be pushed can refer to the media resource pushed to the business object in a future time period, such as an optimized media resource or a newly released media resource. Taking the target media resource as the target game prop in a game application as an example, the target game prop can refer to a newly released game prop or a game prop obtained by optimizing the previously pushed historical game prop. In specific implementation, the target game prop can be pushed to the business object to increase the return rate and activity of the business object in the game application. At the same time, if the target game prop can be accurately pushed, it can avoid pushing the target game prop to invalid business objects (i.e., game objects that do not need the target game prop), thus greatly saving the pushing cost of the game prop.

[0039] The first resource feature of the target media resource may include one or more of the resource features such as the media resource ID, media resource name, media resource type, media resource usage, media resource style, and media resource value of the target media resource. For example, taking the target media resource as the target game prop in a game application as an example, the first resource feature of the target game prop may include one or more of the prop ID, prop name, prop user, prop price, prop usage, prop style, prop type, etc.

[0040] The pre-trained generation model can be a large model, a deep neural network with a large number of model parameters, and has a deep number of model layers (such as the number of model layers is greater than the layer threshold, such as the number of model layers is greater than 12 layers). It is pre-trained in a self-supervised manner through the interaction features of a large number of unlabeled sample media resources and is used to predict the list of preferred media resources of the business object. Since the pre-trained generation model has a large number of model parameters and a deep number of model layers, it can accurately extract the sequential information of the behavior (i.e., historical interaction features) of the business object in time, can reduce the loss of sequential information, and thus can well depict the interest of the business object in media resources, and then accurately predict the list of preferred media resources of the business object. It can be understood that in this application, the pre-trained generation model can be used to predict the media resources preferred by the business object, reduce the loss of sequential information, and improve the accuracy and efficiency of predicting the list of preferred media resources.

[0041] Optionally, the pre-trained generative model can be any one of models such as the GPT (Generative Pre-trained Transformer) model, the BART (Bidirectional and Auto-Regressive Transformers) model, and the T5 model. The GPT model is a generative natural language processing (NLP) model based on the Transformer (a neural network model that learns context and thus meaning by detecting relationships in sequential data) architecture.

[0042] Among them, the GPT model is an autoregressive language model. In this application, the GPT model can predict the preferred media resource of the business object at the next moment through the first historical interaction feature of the historical media resources, that is, predict the media resource that the business object will interact with. The principle of the GPT model mainly relies on the Transformer architecture, especially the self-attention mechanism in the Transformer architecture, to capture the dependencies between features at different positions in the input first historical interaction feature. In this way, GPT can handle long-distance dependencies and complex semantic structures, thus achieving high-quality natural language generation and understanding. In summary, GPT is a generative, pre-trained, autoregressive NLP model that uses the Transformer architecture to learn the probability distribution of a large number of media resources.

[0043] Among them, the BART model is an Auto-Encoding language representation model. The BERT model represents bidirectional encoder representations from Transformer. The BERT model aims to pre-train deep bidirectional representations by jointly conditioning the left and right contexts in all layers to generate deep bidirectional language representations. The T5 model is a sequence-to-sequence (Seq2Seq) model based on the Transformer structure. Its main feature is the ability to unify multiple NLP tasks in a single framework for training. The T5 model consists of two parts: an encoder and a decoder. The encoder is a multi-layer Transformer network responsible for converting the input natural language text into a low-dimensional context vector representation. The decoder also uses a multi-layer Transformer network to convert the context vector generated by the encoder into the target natural language text.

[0044] This application can generate a pre-trained model to predict a list of preferred media resources for a business object based on the first historical interaction features of the business object. The list of preferred media resources includes historical media resources preferred by the business object. It can be understood that the list of preferred media resources includes historical media resources that the business object may be interested in. The list of preferred media resources can be partially ordered, and the prediction probability of the historical media resources earlier in the list of preferred media resources is greater. This means that the probability that the business object is interested in the historical media resources earlier in the list of preferred media resources is greater.

[0045] Since the pre-trained generation model is a large model, it consumes a large amount of computing resources and takes a long time to calculate. Therefore, the list of preferred media resources of the business object can be updated periodically (such as on a daily or weekly basis) through the pre-trained model. The business recognition model can use the list of preferred media resources periodically output by the pre-trained generation as additional features, and combine the first historical interaction features of the business object and the first resource features of the target media resource to predict the resource interaction label of the business object for the target media resource. The resource interaction label can refer to the purchase probability, click probability, share probability, usage probability, etc. of the business object for the target media resource.

[0046] Specifically, the business recognition model can be used to predict the resource interaction label of the business object for the target media resource. The business recognition model can be an end-to-end lightweight model with fewer model parameters and strong real-time model calculation, which can quickly output the resource interaction label. In this way, the real-time performance of pushing media resources to the business object can be ensured, and the pushing efficiency of pushing media resources to the business object can be improved. Optionally, the business recognition model can be composed of any one of the DIN model, DeepFM model, FM model, FFM (Field-aware Factorization Machine) model, and a cross-attention layer based on the cross-attention mechanism.

[0047] The DIN model is a deep learning model that captures the points of interest of business objects through the Attention Mechanism to improve the model's expressive ability and prediction effect. The design of the DIN model takes into account the diversity of user interests and their changes over time, as well as the differences in the impact of different products on users' clicks on advertisements. The DeepFM model consists of two parts: the FM model part (Factorization Machine part) and the DNN model part (Neural Network part). The FM model part can be used to extract low-order features, and the DNN model part can be used to extract high-order features. Moreover, the DeepFM model has a relatively fast training speed. Low-order features can be understood as linear feature combinations, and high-order features can be understood as features formed after multiple linear-nonlinear combination operations, which are highly abstract features. The FM model extracts feature combinations through the inner product of latent variables for each dimension of features, theoretically solving the problem of extracting low-order and high-order combined features. Compared with the FM model, the FFM model introduces the concept of feature field awareness. For each feature, different latent vector features need to be learned for different cross-domains, which can better reflect the differences in features in different domains.

[0048] In this application, the list of preferred media resources of the business object output by the pre-trained generation model is used as an additional feature and input into the business recognition model. The business recognition model can extract, from the first resource features, the key resource features that reflect the preference of the business object for the target media resource according to the second resource features of the media resources in the list of preferred media resources, that is, the key resource features perceived for the target media resource. The business recognition model can obtain the resource interaction label of the business object for the target media resource based on the key resource features, the first historical interaction features, and the first resource features.

[0049] Since the pre-trained generation model has a large number of model parameters and deep model layers, and can accurately extract the sequential information of the behavior of the business object in time, it can reduce the loss of sequential information, well characterize the media resource interests of the business object, and thus can accurately predict the list of preferred media resources of the business object. Therefore, using the list of preferred media resources output by the pre-trained generation model as an additional feature and inputting it into the business recognition model, and the resource interaction label is output in real time through the business recognition model. In this way, it can not only ensure that the media resource interests of the business object are well characterized and fully utilized, improving the accuracy of media resource push, but also ensure the real-time nature of media resource push.

[0050] As Figure 2 shown, Figure 2 is a schematic diagram of a media resource push method provided by an embodiment of this application. As Figure 2 shown, the terminal device 20a can be the aboveFigure 1 Any terminal device in the terminal device cluster, such as terminal device 20a, can be the terminal device 100b in the above Figure 1 The terminal device 20a is installed with a media application. The server 20c can be the server 10 in the above Figure 1 The business object 20b can be a user in the media application installed on the terminal device 20a. The multimedia application installed on the terminal device 20a can provide media resources for the business object 20b, that is, the business object 20b can interact through the media resources provided by the media application. For example, the media application can be a game application, the media resources can be game props in the game application, and the business object 20b can perform interactions such as purchasing game props, using game props, and sharing game props.

[0051] To improve the activity and interest of the business object 20b in the media application, the terminal device 20a can push the media resources that the business object 20b is interested in to the business object. Specifically, the terminal device 20a can obtain the historical interaction data of the business object 20b for the historical media resources that have been pushed, such as media resource click data, media resource purchase data, media resource usage data, media resource sharing data, etc. Further, the terminal device 20a can extract features from the historical interaction data for the historical media resources to obtain the first historical interaction feature for the historical media resources, and send the first historical interaction feature to the server 20c. Of course, the terminal device 20a can directly send the historical interaction data for the historical media resources to the server 20c.

[0052] Among them, the terminal device 20a can determine the target media resource that needs to be pushed currently, and obtain the first resource feature of the target media resource to be pushed. The first resource feature can include one or more of the resource features such as the media resource ID, media resource name, media resource type, media resource usage, media resource style, and media resource value of the target media resource. Further, the terminal device 20a can send the first resource feature of the target media resource to the server 20c. Of course, the server 20c can maintain a set of push media resources, and the set of push media resources includes one or more media resources to be pushed, and the target media resource belongs to any one of the media resources in the set of push media resources. The set of push media resources can be uploaded and managed by a media resource management object. At the same time, the server 20c can store the resource features corresponding to each media resource in the set of push media resources.

[0053] Further, the server 20c can input the first historical interaction feature of the service object 20b for the historical media resource into the pre-trained generation model. The pre-trained generation model is a large model with a large number of model parameters and a deep number of model layers, which can accurately characterize the interest of the service object in the media resource. For example, the pre-trained generation model can be a GPT model. Through the pre-trained model, according to the first historical interaction feature, a list of preferred media resources of the service object 20b is predicted periodically (such as on a daily or weekly basis). The list of preferred media resources includes the historical media resources that the service object 20b prefers (i.e., is interested in), that is, which historical media resources the service object 20b is interested in. Since the pre-trained generation model is a GPT model with a large number of model parameters and deep model layers, it can accurately extract the sequential information of the behavior of the service object in time, reduce the loss of sequential information, and thus can well characterize the interest of the service object in the media resource, and then accurately predict the list of preferred media resources of the service object.

[0054] The server 20c inputs the list of preferred media resources of the service object 20b as an additional feature of the service recognition model into the service recognition model. The service recognition model has fewer model parameters and strong model calculation real-time performance, and can quickly output a model of resource interaction labels. The service recognition model can include a cross-attention layer (i.e., a module based on the cross-attention mechanism), and a service sub-model. The service sub-model can be any one of a DIN model, a DeepFM model, an FM model, and an FFM (Field-aware Factorization Machine) model. Specifically, through the cross-attention layer in the service recognition model, attention cross-processing is performed on the first historical interaction feature and the first resource feature, and key resource features for reflecting the preference of the service object 20b for the target media resource are extracted from the first resource feature. It can be understood that the key resource features can be used to reflect some perceptual information of the service object 20b for the target media resource. For example, the key resource features can be used to reflect whether the service object 20b is interested in the target media resource, the degree of interest, and which aspects of the target media resource the service object is interested in, etc.

[0055] Further, the server 20c can, through the service sub-model in the service recognition model, identify the resource interaction label of the service object for the target media resource according to the key resource feature, the first historical interaction feature, and the first resource feature. The service sub-model can be any one of the DIN model, the DeepFM model, the FM model, and the FFM (Field-aware Factorization Machine) model. The server 20c can return the resource interaction label of the service object for the target media resource to the terminal device 20a, and the terminal device 20a can determine whether to push the target media resource to the service object 20b according to the resource interaction label. Specifically, if the terminal device 20a determines that the target media resource meets the push condition according to the resource interaction label, the target media resource is pushed to the service object 20b. For example, the resource interaction label can be the purchase probability. When the purchase probability of the service object 20b for the target media resource is greater than or equal to the purchase probability threshold, it is determined that the target media resource meets the push condition.

[0056] As Figure 2 shown, the target media resource can be a virtual pet puppy. When the target media resource meets the push condition, the terminal device 20a can display the target media resource 20e in the resource display interface 20d provided in the media application market associated with the media resource. At the same time, the resource display interface 20d also displays a purchase control. When the service object 20b needs to purchase the target media resource 20e, a trigger operation for the purchase control can be implemented, and the terminal device 20a can execute the purchase operation for the target media resource in response to the trigger operation for the purchase control.

[0057] It can be seen that in the implementation of this application, through the pre-trained generation model with more model parameters and deeper model layers, the sequential information and other effective information of the behavior of the service object in time are accurately extracted, the loss of sequential information can be reduced, the interest of the service object in the media resource can be well characterized, and then the list of preferred media resources of the service object can be accurately predicted. Further, the list of preferred media resources is added as an additional feature to the service recognition model, and the cross-attention mechanism is used to extract the key resource features (such as effective resource features) existing in the first resource feature. In this way, not only can the interest of the service object in the media resource be well characterized and fully utilized, but also the accuracy of media resource push can be improved, and the real-time performance of media resource push can be ensured.

[0058] Further, please refer to Figure 3 , Figure 3 which is a schematic flowchart of a data processing method provided by an embodiment of this application. As Figure 3 shown, this method can be performed by Figure 1Any terminal device in can execute, or can be executed by Figure 1 server 10 in, or can also be jointly executed by the Figure 1 terminal device and the server in. The device used to execute this data processing method in this application can be collectively referred to as a computer device. Among them, this data processing method can include but is not limited to the following steps:

[0059] S101, Obtain the first historical interaction feature of the business object for the pushed historical media resources, and the first resource feature of the target media resources to be pushed.

[0060] Specifically, in order to improve the return rate and activity of the business object in the media application that provides media resources, media resources can be pushed to the business object. At the same time, if the media resources can be accurately pushed to the business object, it is possible to avoid pushing media resources to invalid business objects (that is, game objects that do not need media resources), thereby greatly saving the pushing cost of media resources. Specifically, the computer device can obtain the historical interaction data of the business object for the pushed historical media resources, extract features from the historical interaction data, and obtain the first historical interaction feature (that is, the historical interaction feature of the business object for the pushed historical media resources). The first historical interaction feature can include one or more of the resource usage feature, resource click feature, resource purchase feature, and resource sharing feature for the historical media resources, etc.

[0061] At the same time, the computer device can also obtain the target media resources to be pushed. The target media resources can refer to the media resources pushed to the business object in the future time period, such as optimized media resources or newly released media resources. Taking the target game props in the game application as an example of the target media resources, the target game props can refer to newly released game props, or can refer to the game props obtained by optimizing the pushed historical game props. The computer device can obtain the first resource feature (that is, the resource feature of the target media resources). The first resource feature can include one or more of the resource features such as media resource ID, media resource name, media resource type, media resource usage, media resource style, and media resource value. For example, taking the target game props in the game application as an example of the target media resources, the first resource feature of the target game props can include one or more of the prop ID, prop name, prop user, prop price, prop usage, prop style, and prop type.

[0062] S102, Through the pre-trained generation model, predict the list of preferred media resources according to the first historical interaction feature.

[0063] Specifically, the computer device can call a pre-trained generation model. Through this pre-trained generation model, according to the first historical interaction feature, a list of preferred media resources can be predicted. The list of preferred media resources includes the historical media resources preferred by the business object. For example, taking the media resource as a game prop, the list of preferred media resources can be a list of preferred game props, and the list of preferred game props can be the game props that the business object likes (or is interested in). The pre-trained generation model can be a large model, a deep neural network with a large number of model parameters, and has a deep number of model layers. It is self-supervised pre-trained through the interaction features of a large number of unlabeled sample media resources. For example, the pre-trained generation model can refer to any one of models such as the GPT model, the BART model, and the T5 model. Since the pre-trained generation model has a large number of model parameters and a deep number of model layers, it can accurately extract the sequential information and other feature information of the business object's behavior in time, can reduce the loss of sequential information, and thus can well depict the business object's interest in media resources, and then accurately predict the list of preferred media resources of the business object.

[0064] Optionally, the pre-trained generation model can be the GPT model. The pre-trained generation model can include M transformation encoders (i.e., Transformer encoders). Transformer is a sequence-to-sequence model based on the attention mechanism, mainly used for machine translation tasks and can perform parallel computing. When the pre-trained generation model is the GPT model, the specific way for the computer device to predict the list of preferred media resources according to the first historical interaction feature through the pre-trained generation model can include: calling N self-attention layers in the first transformation encoder to perform attention feature extraction on the first historical interaction feature to obtain N-dimensional attention features; the first transformation encoder is the transformation encoder with the first arrangement order among the M transformation encoders, and N is a positive integer. Calling the neural network layer in the first transformation encoder to perform deep feature extraction on the N-dimensional attention features to obtain resource preference features. Determining the list of preferred media resources preferred by the business object according to the resource preference features and the remaining transformation encoders; the remaining transformation encoders are the transformation encoders other than the first transformation encoder among the M transformation encoders.

[0065] Specifically, when the pre-trained generation model is a GPT model, the pre-trained generation model includes M transformation encoders. Each transformation encoder includes N self-attention layers (i.e., self-attention) and a neural network layer (such as a feed-forward neural network layer, feed-forward neural network layer). Each self-attention layer can include three linear transformations (i.e., Linear Transformation), namely query, key, and value. The computer device can call the N self-attention layers in the first transformation encoder in the pre-trained generation model to extract attention features from the first historical interaction features, obtaining N-dimensional attention features. It can be understood that the M transformation encoders included in the pre-trained generation model are arranged in sequence, and each transformation encoder has its own arrangement order. The first transformation encoder is the transformation encoder with the first arrangement order among the M transformation encoders. The N self-attention layers in each transformation encoder can be used to extract information in different-dimensional semantic spaces. One self-attention layer is used to extract one-dimensional attention features, that is, one self-attention layer is used to extract information in one-dimensional semantic space. Since the GPT model is a sequence model, it can accurately extract the sequential information of the first historical interaction features in time, and thus can accurately predict the list of preferred media resources.

[0066] Further, the computer device can perform deep feature extraction on the N-dimensional attention features through the neural network layer in the first transformation encoder to obtain a resource preference feature (i.e., the hidden state processed by the first transformation encoder). The computer device can determine the list of preferred media resources preferred by the business object from the candidate media resource set according to the resource preference feature and the remaining transformation encoders. The remaining transformation encoders are the transformation encoders other than the first transformation encoder among the M transformation encoders. Among them, the candidate media resource set can refer to what the pre-trained generation model learns from the historical interaction data of a large number of business objects for historical media resources. Specifically, the computer device can input the resource preference feature into the next transformation encoder of the first transformation encoder (i.e., the transformation encoder with the second arrangement order among the M transformation encoders), and the N self-attention layers and neural network layer in the next transformation encoder continue to perform feature extraction on the resource preference feature, and so on, until the last transformation encoder of the pre-trained generation model finishes processing. Among them, after the last transformation encoder of the pre-trained generation model, a linear transformation layer and an activation function can also be connected to map the output of the last transformation encoder to the probability distribution of the predicted labels.

[0067] Optionally, the pre-trained generation model can be pre-trained. The training process of the pre-trained generation model can include: obtaining second historical interaction features corresponding to the first sample media resources interacted by the first sample object in the first historical time period, and the second sample media resources interacted by the first sample object in the second historical time period. The first historical time period is before the second historical time period, and the first historical time period is adjacent to the second historical time period. Through the initial pre-trained generation model, according to the second historical interaction features, a first predicted media resource list is predicted; the first predicted media resource list includes the first sample media resources preferred by the first sample object. According to the second sample media resources and the first predicted media resource list, the initial pre-trained generation model is trained until the initial pre-trained generation model meets the convergence condition, and the pre-trained generation model is obtained.

[0068] Specifically, the computer device can obtain the second historical interaction features corresponding to the first sample media resources interacted by the first sample object in the first historical time period. The first sample object can be an object in the media application that has generated interaction data with the media resources. The first historical time period is before the second historical time period, and the first historical time period is adjacent to the second historical time period, that is, the second historical time period is the next time period of the first historical time period. The first sample media resources can also be the media resources with which the first sample object has generated interaction data in the first historical time period, and the second historical interaction features are the historical interaction features of the second sample object for the first sample media resources. The second historical interaction features can also include resource usage features, resource click features, resource purchase features, and resource sharing features for the first sample media resources, etc. The computer device can also obtain the second sample media resources interacted by the first sample object in the second time period, that is, the second sample media resources are the media resources with which the first sample object has generated interaction data in the second time period. It can be understood that the pre-trained generation model can obtain the sample media resources interacted by the first sample object in each time period and the corresponding historical interaction features for unsupervised method training, without using the supervised method of manual annotation, which can reduce the training cost and improve the training efficiency of the pre-trained generation model.

[0069] Furthermore, the computer device can generate a model through initial training and predict a first predicted media resource list based on the second historical interaction features. The first predicted media resource list includes first sample media resources preferred by the first sample object. The number of first sample objects can be one or more. When the number of first sample objects is multiple, the number of first sample media resources is also multiple, and the number of second sample media resources is also multiple. Therefore, the first sample media resources included in the first predicted media resource list can be extracted from the interaction data of multiple first sample objects with respect to the first sample media resources. Specifically, the initial pre-training generation model can also be any one of models such as the GPT model, the BART model, and the T5 model. When the initial pre-training generation model is the GPT model, the specific processing process of the initial pre-training generation model can refer to the specific process of the above pre-training generation model, which will not be elaborated in the embodiments of this application.

[0070] Specifically, the initial pre-training generation model can determine the prediction probability corresponding to each first sample media resource in the first predicted media resource list. The first predicted media resource list is in a partial order. The higher the prediction probability of a first sample media resource, the higher the possibility that it is preferred (or interested) by the first sample object. The computer device can obtain the first sample media resource with the highest prediction probability from the first predicted media resource list as the predicted sample media resource. The computer device can obtain the cross-entropy loss between the second sample media resource and the predicted sample media resource, and adjust the model parameters of the initial pre-training generation model according to this cross-entropy loss. Among them, the computer device can obtain multiple cross-entropy losses based on the historical interaction features of the first sample object in multiple historical time periods and the sample media resources in their respective next historical time periods, and perform an average process on these cross-entropy losses to obtain the average cross-entropy loss. The model parameters of the initial pre-training generation model are adjusted according to this average cross-entropy loss until the initial pre-training generation model meets the convergence condition to obtain the pre-training generation model. The convergence condition can refer to that the number of model training times reaches the target number, or the model loss is less than or equal to the loss threshold.

[0071] As Figure 4 shown, Figure 4 is a schematic diagram of the training process of a pre-training generation model provided by an embodiment of this application. As Figure 4As shown, taking media resources as game props in a game application as an example, the computer device can obtain the historical game props that the first sample object interacted with in different historical time periods, such as game prop 1, game prop 2, game prop 3, …, and game prop J-1. The historical time periods corresponding to game prop 1, game prop 2, game prop 3, …, and game prop J-1 increase sequentially and are adjacent to each other pairwise. Specifically, the historical time period corresponding to game prop 1 is prior to the historical time period corresponding to game prop 2 and is adjacent to the historical time period corresponding to game prop 2 (i.e., the historical time period corresponding to game prop 2 is the next time period of the historical time period corresponding to game prop 1). Similarly, the historical time period corresponding to game prop 2 is prior to the historical time period corresponding to game prop 3 and is adjacent to the historical time period corresponding to game prop 3 (i.e., the historical time period corresponding to game prop 3 is the next time period of the historical time period corresponding to game prop 2). By analogy, the historical time period corresponding to game prop J-2 is prior to the historical time period corresponding to game prop J-1 and is adjacent to the historical time period corresponding to game prop J-1 (i.e., the historical time period corresponding to game prop J-1 is the next time period of the historical time period corresponding to game prop J-2).

[0072] The computer device can obtain the historical prop interaction features of the first sample object for the interacted historical game props, that is, the historical prop interaction features corresponding to game prop 1, game prop 2, game prop 3, …, and game prop J-1 respectively. The computer device can convert the historical prop interaction features corresponding to game prop 1, game prop 2, game prop 3, …, and game prop J-1 into dense interaction feature vectors through an embedding layer (i.e., the embedding layer). Taking game prop 1 as an example, the computer device can convert the historical prop interaction feature corresponding to game prop 1 into a dense vector through the embedding layer to obtain the dense interaction feature vector corresponding to game prop 1.

[0073] Then, the dense interaction feature vectors corresponding to game item 1, game item 2, game item 3, …, and game item J-1 are respectively input into the initial pre-trained generation model. Taking the GPT model structure as an example of the initial pre-trained generation model, the initial pre-trained generation model can adopt the autoregressive pre-training technology to predict the predicted game item that the first sample object will interact with in the next time period based on the game item in the previous time period (i.e., the actually interacted game item). The specific prediction process can refer to the content of step S102 above, and this will not be elaborated here in the embodiments of this application. For example, the initial pre-trained generation model can predict the predicted game item 2 that the first sample object may interact with based on the dense interaction feature vector corresponding to game item 1; predict the predicted game item 3 that the first sample object may interact with based on the dense interaction feature vector corresponding to game item 2; predict the predicted game item 4 that the first sample object may interact with based on the dense interaction feature vector corresponding to game item 3; and so on, until the predicted game item J that the first sample object may interact with is predicted based on the dense interaction feature vector corresponding to game item J-1, that is, predicted game item 2, predicted game item 2, predicted game item 4, …, predicted game item J.

[0074] The computer device can adopt the self-supervised pre-training technology to compare the predicted game item with the highest prediction probability (i.e., predicted game item 2, predicted game item 2, predicted game item 4, …, predicted game item J) predicted by the initial pre-trained generation model with the real behavior of the first sample object (i.e., game item 2, game item 3, …, game item J-1 and game item J) to obtain the model prediction loss. Among them, the cross-entropy loss can be used to minimize the difference between the model prediction and the real behavior as the training objective of the initial pre-trained generation model. Among them, the calculation formula of the cross-entropy loss between the model prediction and the real behavior can be shown as the following formula (1).

[0075]

[0076] Among them, Loss in formula (1) is the cross-entropy loss, J is the number of game items, item′ i is the i-th predicted game item, item i is the i-th game item (i.e., the i-th actually interacted game item of the first sample object), and cross_entropy(item′ i , item i ) is the cross-entropy loss between the i-th predicted game item and the i-th game item.

[0077] The computer device can train the initial pre-trained generation model according to the model prediction loss of the initial pre-trained generation model until the initial pre-trained generation model meets the convergence condition, and obtain the pre-trained generation model. Among them, in the application process of the pre-trained generation model, the pre-trained generation model can output a list of game items that the business object may interact with in the next time period. This list is partially ordered, that is, the higher the prediction probability of the game item at the front, the higher the prediction probability indicates that the business object is more likely to be interested in the game item.

[0078] S103. Through the service recognition model, according to the second resource feature of the historical media resource in the preference media resource list, extract the key resource feature from the first resource feature to reflect the preference of the business object for the target media resource.

[0079] Specifically, the computer device can extract, through the service recognition model, from the first resource feature the key resource feature that reflects the preference of the business object for the target media resource according to the second resource feature of the historical media resource in the preference media resource list. It can be understood that the computer device extracts the resource feature that the business object is interested in from the first resource feature according to the second resource feature preferred by the business object. In this way, it is convenient to subsequently determine the resource interaction label of the business object for the target media resource.

[0080] Optionally, the specific manner in which the computer device extracts, through the service recognition model, from the first resource feature the key resource feature that reflects the preference of the business object for the target media resource according to the second resource feature of the historical media resource in the preference media resource list may include: obtaining, through the key feature extraction layer of the service recognition model, the first resource feature vector that reflects the first resource feature and the second resource feature vector that reflects the second resource feature. Generating the resource key vector and resource value vector corresponding to the second resource feature vector according to the second resource feature vector and the bias vector corresponding to the second resource feature. Extracting, according to the resource key vector and resource value vector, from the first resource feature vector the key resource feature that reflects the preference of the business object for the target media resource.

[0081] Specifically, the computer device can obtain a first resource feature vector for reflecting the first resource feature and a second resource feature vector for reflecting the second resource feature through the key feature extraction layer of the service recognition model. Among them, the key feature extraction layer can refer to a cross-attention layer, which is a network layer based on the cross-attention mechanism. The cross-attention mechanism is a variant of the neural network attention mechanism, mainly used to process the correlation information between multiple input sequences. In the cross-attention mechanism, in addition to calculating the self-attention within the input sequence, the attention between different input sequences is also calculated. In this application, the cross-attention layer can not only calculate the self-attention within the first resource feature vector and the second resource feature vector respectively, but also calculate the attention between the first resource feature vector and the second resource feature vector. In this way, the information flow and correlation between the first resource feature vector and the second resource feature vector can be improved, the similarity between the first resource feature vector and the second resource feature vector can be accurately obtained, and the performance of the model can be improved.

[0082] Among them, the computer device can vectorize the second resource feature (i.e., embedding vectorization) through the key feature extraction layer of the service recognition model to obtain a second resource feature vector for reflecting the second resource feature. Embedding vectorization is used to convert a sparse vector into a dense vector, which is convenient for the processing of the key feature extraction layer. Similarly, the computer device can convert the first feature vector through a random initialization vector through the key feature extraction layer of the service recognition model to obtain a first resource feature vector for reflecting the first resource feature.

[0083] Furthermore, the second resource feature is the resource feature corresponding to the historical media resource in the list of preferred media resources of the service object. Since the list of preferred media resources is partially ordered (i.e., arranged according to the prediction probability, the higher the ranking of the historical media resource, the greater the possibility of being preferred by the service object), the computer device can obtain the position vector corresponding to the second resource feature as the bias vector (i.e., additional bias) corresponding to the second resource feature, so that it can have a better perception of the partial order information of the list of preferred media resources (such as knowing which historical media resources at which ranking positions are more likely to be preferred by the service object). The computer device can fuse the second resource feature vector and the bias vector corresponding to the second resource feature to obtain the resource key vector and resource value vector corresponding to the second resource feature vector.

[0084] For example, the computer device may use the vector sum of the second resource feature vector and the bias vector corresponding to the second resource feature as the resource key vector (i.e., the Key (K) vector in the attention mechanism) and the resource value vector (i.e., the Value (V) vector in the attention mechanism). That is, K = V = X + P, where K is the resource key vector, V is the resource value vector, X is the second resource feature vector, and P is the bias vector of the second resource feature. Among them, X ∈ R (K×d) , P ∈ R (K×d) , K is a token in the feature vector (such as each character in the name of the media resource), and d is the dimension of the feature vector. The computer device may perform cross-attention processing on the first resource feature and the second resource feature according to the resource key vector and the resource value vector, and extract, from the first resource feature vector, the key resource feature for reflecting the preference of the service object for the target media resource.

[0085] Optionally, the specific manner in which the computer device extracts, from the first resource feature vector, the key resource feature for reflecting the preference of the service object for the target media resource may include: generating a similarity vector according to the first resource feature vector and the resource key vector; the similarity vector is used to reflect the similarity between the second resource feature and the first resource feature. Performing a multiplication process on the similarity vector, the resource value vector, and the weight vector corresponding to the resource value vector to obtain the key resource feature for reflecting the preference of the service object for the target media resource.

[0086] Specifically, the computer device may perform similarity extraction on the first resource feature vector and the resource key vector to obtain a similarity vector, which is used to reflect the similarity between the second resource feature and the first resource feature. Further, the computer device may perform a multiplication process on the similarity vector, the resource value vector, and the weight vector corresponding to the resource value vector to obtain a cross-attention feature, and the cross-attention feature may be the effective resource feature in the first resource feature. Among them, the weight vector corresponding to the resource value vector may be the weight parameter corresponding to the value vector (i.e., V) in the key feature extraction layer (i.e., the model parameter in the service recognition model). Performing a multiplication process on the resource value vector and the weight vector corresponding to the resource value vector can highlight the key features in the resource value vector to obtain the processed resource value vector. Specifically, the formula for calculating the cross-attention feature may refer to the following formula (2).

[0087]

[0088] Among them, H in formula (2) I is the cross-attention feature, I is the first resource feature, A is the similarity vector, V is the resource value vector, is the weight vector corresponding to the resource value vector, is the processed resource value vector, H I ∈R (1 ×d) , where d is the vector dimension of the first resource feature vector or the resource value vector.

[0089] Furthermore, the computer device can determine the key resource features for reflecting the preference of the business object for the target media resource according to the number of layers of the key feature extraction layer and the cross-attention features. Specifically, if the number of layers of the key feature extraction layer is one layer, the cross-attention feature is determined as the key resource feature for reflecting the preference of the business object for the target media resource. If the number of layers of the key feature extraction layer is multiple layers, the key resource feature for reflecting the preference of the business object for the target media resource is determined according to the cross-attention features output by multiple key feature extraction layers. For example, the computer device can fuse the cross-attention features output by multiple key feature extraction layers to obtain the key resource feature for reflecting the preference of the business object for the target media resource.

[0090] Optionally, the specific manner in which the computer device generates the similarity vector according to the first resource feature vector and the resource key vector may include: multiplying the first resource feature vector by the weight vector corresponding to the first resource feature vector to obtain the processed first resource feature vector. Multiplying the resource key vector by the weight vector corresponding to the resource key vector to obtain the processed resource key vector. Multiplying the processed first resource feature vector by the transpose of the processed resource key vector to obtain the similarity vector.

[0091] Specifically, the computer device can obtain the weight vector corresponding to the first resource feature vector. The weight vector corresponding to the first resource feature vector is a model parameter in the key feature extraction layer and is adjusted through model training. The computer device can multiply the first resource feature vector by the weight vector corresponding to the first resource feature vector to highlight the key features in the first resource feature vector and obtain the processed first resource feature vector. Similarly, the computer device can obtain the weight vector corresponding to the resource key vector. The weight vector corresponding to the resource key vector is also a model parameter in the key feature extraction layer and is adjusted through model training. The computer device can multiply the resource key vector by the weight vector corresponding to the resource key vector to highlight the key features in the resource key vector and obtain the processed resource key vector.

[0092] Further, the computer device can obtain the similarity between the processed first resource feature vector and the processed resource key vector. Specifically, the computer device can perform a transpose operation on the processed resource key vector to obtain the transpose of the processed resource key vector, and perform a multiplication operation on the processed first resource feature vector and the transpose of the processed resource key vector to obtain an initial similarity vector. For example, the computer device can obtain the inner product of the processed first resource feature vector and the transpose of the processed resource key vector to obtain the initial similarity vector. The computer device can also obtain the outer product of the processed first resource feature vector and the transpose of the processed resource key vector to obtain the initial similarity vector. The computer device can obtain the ratio between the initial similarity vector and the scaling factor to obtain the processed initial similarity vector, which can reduce the numerical range. The scaling factor can be the square root of the vector dimension of the first resource feature vector or the resource key vector. Further, the computer device can normalize the processed initial similarity vector through the softmax function (a normalization function) to obtain the similarity vector. Specifically, the formula for calculating the similarity vector can be seen in the following formula (3).

[0093]

[0094] Among them, A in formula (3) is the similarity vector, the softmax function is a normalization function, Q is the query vector (i.e., the first resource feature vector), is the weight vector corresponding to the first resource feature vector, is the processed first resource feature vector, K is the resource key vector, is the weight vector corresponding to the resource key vector, is the processed resource key vector, is the transpose of the processed resource key vector, is the square root of the vector dimension of the first resource feature vector or the resource key vector.

[0095] In this way, taking the first resource feature vector as the query (Q) vector (i.e., the query vector) in the cross-attention mechanism, and taking the vector sum between the second resource feature vector and the bias vector corresponding to the second resource as the key (K) vector (i.e., the key vector) and value (V) (i.e., the value vector) in the cross-attention mechanism. Performing cross-attention processing on the first resource feature vector and the second resource feature vector can accurately obtain the similarity between the first resource feature vector and the second resource feature vector, and extract the key resource features (such as the resource features preferred by the business object) in the first resource feature vector.

[0096] S104. Based on the key resource features, the first historical interaction features, and the first resource features, obtain the resource interaction label of the business object for the target media resource, and push the target media resource to the business object according to the resource interaction label.

[0097] Specifically, the computer device can identify the resource interaction label of the business object for the target media resource through the service recognition model according to the key resource features, the first historical interaction features, and the first resource features. The resource interaction label can refer to the purchase probability, click probability, usage probability, sharing probability, etc. of the business object for the target media resource. For example, taking the target media resource as a target game prop, the resource interaction label can be the prop purchase probability, prop click probability, prop usage probability, prop sharing probability, etc. of the business object for the target game prop. The computer device can determine whether the target media resource meets the push condition for pushing to the business object according to the resource interaction label, and when it is determined that the target media resource meets the push condition, push the target media resource to the business object. Among them, the first historical interaction features include sparse features and dense features. Taking the media resource as a game prop as an example, the sparse features can include the amount of money spent (purchasing game props) by the business object user in the past seven days, and the dense features can include the game props purchased by the business object in the past seven days.

[0098] It can be understood that in this application, the pre-trained generation model and the service recognition model are decoupled. The pre-trained model routinely updates the preference media resource list of the business object on a daily or weekly basis. The preference media resource list is used as an additional feature and added to the downstream service recognition model (with fewer model parameters and strong calculation real-time performance), and the cross-attention mechanism is used to extract the key resource features (i.e., effective information) existing in the first resource features. In this way, it can not only ensure that the interest description of the user by the pre-trained generation model is fully utilized, thereby improving the recommendation effect, but also ensure the real-time performance of the service recognition model.

[0099] For example, taking the target media resource as the target game prop and the resource interaction label as the prop purchase probability, the computer device can detect whether the prop purchase probability of the business object for the target game prop is greater than or equal to the purchase probability threshold. If the purchase probability of the business object for the target game prop is greater than or equal to the purchase probability threshold, it is determined that the target game prop meets the push condition. Of course, the resource interaction label can also be the prop click probability. The computer device can detect whether the prop click probability of the business object for the target game prop is greater than or equal to the click probability threshold. If the prop click probability of the business object for the target game prop is greater than or equal to the click probability threshold, it is determined that the target game prop meets the push condition. Of course, the resource interaction label can also be the prop usage probability. The computer device can detect whether the prop usage probability of the business object for the target game prop is greater than or equal to the usage probability threshold. If the prop usage probability of the business object for the target game prop is greater than or equal to the usage probability threshold, it is determined that the target game prop meets the push condition. Of course, the resource interaction label can also be multiple labels such as the prop purchase probability, prop click probability, prop usage probability, and prop sharing probability. The computer device can comprehensively determine the comprehensive probability according to multiple labels such as the prop purchase probability, prop click probability, prop usage probability, and prop sharing probability. For example, by summing up multiple labels such as the prop purchase probability, prop click probability, prop usage probability, and prop sharing probability to obtain the comprehensive probability, and detecting whether the target game prop meets the push condition according to the comprehensive probability.

[0100] Specifically, the service recognition model may include a key feature extraction layer (such as a cross-attention layer) and a service sub-model. The service sub-model is used to identify the resource interaction label of the business object for the target media resource according to the key resource feature, the first historical interaction feature, and the first resource feature. Among them, the service sub-model can be any one of the DIN model, DeepFM model, FM model, and FFM (Field-aware Factorization Machine) model.

[0101] Optionally, the service sub-model can be a DeepFM model. The DeepFM model includes two parts: the FM model part (factorization machine part) and the DNN model part (neural network part). The FM model part can be used to extract low-order features, and the DNN model part can be used to extract high-order features. Moreover, the DeepFM model has a relatively fast training speed. When the service sub-model is a DeepFM model, the specific manner in which the computer device identifies the resource interaction label of the service object for the target media resource based on the key resource features, the first historical interaction features, and the first resource features can include: performing feature embedding processing on the key resource features, the first historical interaction features, and the second resource features to obtain dense resource interaction features. Through the feature cross-combination layer in the service recognition model, performing linear and cross processing on the dense resource interaction features to obtain cross resource interaction features. Through the deep feature extraction layer in the service recognition model, performing deep feature extraction on the dense resource interaction features to obtain deep resource interaction features. Identifying the resource interaction label of the service object for the target media resource based on the dense resource interaction features and the deep resource interaction features.

[0102] Specifically, the service sub-model can include a feature embedding layer, a feature cross-combination layer, a deep feature extraction layer, and a first output layer. The computer device can perform feature embedding processing on the key resource features, the first historical interaction features, and the second resource features through the feature embedding layer to obtain dense resource interaction features. The computer device can call the feature cross-combination layer, and this feature cross-combination layer can refer to the FM neural network. The computer device can perform linear and cross processing on the dense resource interaction features through the feature cross-combination layer to obtain cross resource interaction features. Further, the computer device can call the deep feature extraction layer, and this deep feature extraction layer can refer to the DNN neural network. The computer device can perform deep feature extraction on the dense resource interaction features through the deep feature extraction layer, that is, performing multi-layer linear mapping and non-linear transformation on the dense resource interaction features to obtain deep resource interaction features. The computer device can fuse the cross resource interaction features and the deep resource interaction features through the first output layer to obtain fused resource interaction features, and perform feature transformation on the fused resource interaction features through the Sigmoid function to obtain the resource interaction label of the service object for the target media resource. Among them, the Sigmoid function is often used as the threshold function of the neural network to map the variable (i.e., the fused resource interaction features) to the range between 0 and 1.

[0103] Optionally, the specific manner in which the computer device performs feature embedding processing on the key resource feature, the first historical interaction feature, and the second resource feature to obtain the dense resource interaction feature may include: respectively encoding the key resource feature, the first historical interaction feature, and the second resource feature to obtain sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature respectively. Concatenating the sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature to obtain a concatenated sparse feature, and performing feature embedding processing on the concatenated sparse feature to obtain the dense resource interaction feature.

[0104] Specifically, since general categorical features cannot be directly input into the service recognition model, the feature embedding layer may include an encoding sub-layer, a feature concatenation sub-layer, and an embedding sub-layer. The computer device can use the onehot (i.e., one-hot encoding) encoding method in the encoding sub-layer to encode the key resource feature to obtain the sparse feature corresponding to the key resource feature (i.e., the sparse 01 vector representation), and encode the first historical interaction feature to obtain the sparse feature corresponding to the first historical interaction feature, and encode the second resource feature to obtain the sparse feature corresponding to the second resource feature. At the same time, the computer device can use the feature concatenation sub-layer to concatenate the sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature to obtain a concatenated sparse feature. Further, the computer device can use the embedding sub-layer (Dense Embeddings) to perform feature embedding processing on the concatenated sparse feature to obtain the dense resource interaction feature.

[0105] Optionally, the service sub-model may be a DIN model. The DIN model is a deep learning model that captures the interest points of service objects through the attention mechanism to improve the model's expressive ability and prediction effect. The design of the DIN model takes into account the diversity of user interests and their changes over time, as well as the differences in the impact of different products on users' clicking on advertisements. When the service sub-model is a DIN model, the specific manner in which the computer device obtains the resource interaction label of the service object for the target media resource based on the key resource feature, the first historical interaction feature, and the first resource feature may include: determining the feature correlation degree between the first historical interaction feature and the first resource feature through the attention layer in the service recognition model, and performing feature weighting processing on the first historical interaction feature according to the feature correlation degree to obtain the processed first historical interaction feature. Concatenating the key resource feature, the processed first historical interaction feature, and the second resource feature through the feature concatenation layer in the service recognition model to obtain a concatenated resource feature. Performing feature classification on the concatenated resource feature through the fully connected network layer in the service recognition model to obtain the resource interaction label of the service object for the target media resource.

[0106] Specifically, the business sub-model may include an attention layer, a feature splicing layer, and a fully connected network layer. The computer device may call the attention layer, which is also called a local activation unit. It can set weights according to the feature correlation degree between the first historical interaction feature and the first resource feature, and dynamically represent the interest of the business object in the media resource. The higher the feature correlation degree, the greater the weight; of course, the lower the feature correlation degree, the smaller the weight. The computer device can determine the feature correlation degree between the first historical interaction feature and the first resource feature through the attention layer, and perform feature weighting processing on the first historical interaction feature according to the feature correlation degree to obtain the weighted first historical interaction feature. Since not all historical interaction features are useful for extracting the interest of the business object, performing feature weighting processing on the first historical interaction feature according to the feature correlation degree and setting different weights for different features can highlight the key interaction features in the first historical interaction feature. The computer device can splice the key resource feature, the processed first historical interaction feature, and the second resource feature through the feature splicing layer to obtain the spliced resource feature.

[0107] Furthermore, the computer device can perform activation processing on the spliced resource feature through the activation sub-layer in the fully connected network layer to obtain the activated spliced resource feature. For example, the activation sub-layer in the fully connected network layer may include the PReLU (an activation function) activation function, and the computer device can use the PReLU (an activation function) activation function to perform activation processing on the spliced resource feature. For example, the activation sub-layer in the fully connected network layer may include the adaptive activation function Dice, and the computer device can use the adaptive activation function Dice to perform activation processing on the spliced resource feature. Further, the computer device can perform feature transformation on the activated spliced resource feature through the Sigmoid function in the fully connected network layer to obtain the resource interaction label of the business object for the target media resource. Among them, the Sigmoid function is often used as the threshold function of the neural network to map the variable (i.e., the fused resource interaction feature) to between 0 and 1.

[0108] Optionally, the attention layer may include a feature correlation extraction sub-layer, a fully connected sub-layer, and an output sub-layer. The specific way for the computer device to obtain the feature correlation degree between the first historical interaction feature and the second resource feature through the attention layer in the business recognition model may include: extracting the correlation feature between the first historical interaction feature and the second resource feature through the feature correlation feature sub-layer to obtain the resource similarity feature. Performing feature mapping on the resource similarity feature through the fully connected sub-layer to obtain the resource mapping feature. Determining the feature correlation degree between the first historical interaction feature and the second resource feature according to the resource mapping feature through the output sub-layer.

[0109] Specifically, the computer device can multiply the first historical interaction feature and the second resource feature (such as inner product processing or outer product processing, etc.) through the feature correlation feature sub-layer in the attention layer, extract the similarity between the first historical interaction feature and the second resource feature, and obtain the resource similarity feature between the first historical interaction feature and the second resource feature. Of course, the feature difference between the first historical interaction feature and the second resource feature can also be used as the resource similarity feature between the first historical interaction feature and the second resource feature. Further, the computer device can activate the resource similarity feature through the activation sub-layer in the attention layer to obtain the resource mapping feature after activation processing. The activation sub-layer in this attention layer can include the PReLU (an activation function) activation function or the adaptive activation function Dice. The computer device can linearly process the resource mapping feature after activation processing through the linear processing layer in the attention layer to obtain the feature correlation degree between the first historical interaction feature and the second resource feature. Among them, the thread processing layer can include the Linear function, and the Linear function can map the input feature variable to an output feature variable.

[0110] Optionally, the service generation model can be a pre-trained and converged model. Specifically, the training process of the service recognition model can include: obtaining the third historical interaction feature of the second sample object for the already pushed third sample media resource, the third resource feature of the to-be-pushed fourth sample media resource, and the labeled resource interaction label of the second sample object for the fourth sample media resource. Through the pre-training generation model, according to the third historical interaction feature, predict the second predicted media resource list; the second predicted media resource list includes the third sample media resources preferred by the second sample object. Through the initial service recognition model, according to the fourth resource feature of the historical media resources in the second predicted media resource list, extract the predicted key resource features that reflect the preference of the second sample object for the fourth sample media resource from the third resource features. According to the predicted key resource features, the third historical interaction feature, and the third resource feature, identify the predicted resource interaction label of the second sample object for the fourth sample media resource. According to the predicted resource interaction label and the labeled resource interaction label, train the initial service recognition model until the initial service recognition model meets the convergence condition to obtain the service recognition model.

[0111] Specifically, the second sample object may refer to an object that has generated interaction data for the pushed media resource. The second sample object may be different from or the same as the first sample object. The pushed third sample media resource may refer to the media data for which the second sample object has generated interaction data. The third historical interaction feature may refer to the historical interaction feature of the second sample object for the third sample media resource, and the third historical interaction feature may also include resource usage features, resource click features, resource purchase features, resource sharing features, etc. The to-be-pushed fourth sample media resource may refer to a media resource whose push time is later than that of the third sample media resource. The third resource feature refers to the resource feature of the fourth sample media resource, which may include one or more of the resource features such as media resource ID, media resource name, media resource type, media resource usage, media resource style, and media resource value. The labeled resource interaction label of the second sample object for the fourth sample media resource may also include purchase probability, usage probability, click probability, and sharing probability, etc., and the labeled resource interaction label may be determined according to the actual behavior of the second sample object for the fourth sample media resource.

[0112] The computer device can generate a second predicted media resource list through a pre-trained generation model according to the third historical interaction feature. The second predicted media resource list includes the third sample media resources preferred by the second sample object. The specific prediction process can refer to the content of step S102 above, and this embodiment of the present application will not elaborate here. Further, the computer device can obtain the fourth resource feature of the historical media resources in the second predicted media resource list, and through the initial service recognition model, extract the predicted key resource features for reflecting the preference of the second sample object for the fourth sample media resource from the third resource features according to the fourth resource feature of the historical media resources in the second predicted media resource list. The extraction process of the predicted key resource features can refer to the content of step S103 above, and this embodiment of the present application will not elaborate here. The computer device can identify the predicted resource interaction label of the second sample object for the fourth sample media resource through the initial service recognition model according to the predicted key resource features, the third historical interaction feature, and the third resource feature. The recognition process of the predicted resource interaction label can refer to the content of step S104 above, and this embodiment of the present application will not elaborate here.

[0113] Furthermore, the computer device can obtain the label prediction error between the predicted resource interaction label and the annotated resource interaction label. For example, when the resource interaction label is the purchase probability, the computer device can obtain the difference between the predicted purchase probability and the annotated purchase probability as the label prediction error. The computer device can determine the label prediction error as the model prediction loss of the initial business identification model, and adjust the model parameters of the initial business identification model according to the model prediction loss to obtain the adjusted initial business identification model. The computer device can detect whether the adjusted initial business identification model meets the convergence condition. The convergence condition can refer to the number of model training times reaching the target number, or the model prediction loss is less than or equal to the loss threshold. If the adjusted initial business identification model meets the convergence condition, the adjusted initial business identification model is determined as the business identification model; if the adjusted initial business identification model does not meet the convergence condition, the adjusted initial business identification model continues to be trained until the initial business identification model meets the convergence condition to obtain the business identification model.

[0114] This application uses the data accumulated over half a year in the game props recommendation scenario as the training set, and the data for the next 7 days as the test set. Verification was performed on the purchase rate estimation task. The baseline method uses the traditional statistical feature method (such as using the FM model for purchase rate prediction). The effect measurement criteria on the test set use the offline model evaluation indicator AUC (area under the curve) and the online indicator ARPU (Average Revenue Per User). For the model evaluation indicator AUC, see Table 1 below:

[0115]

[0116] Table 1

[0117] As can be seen from Table 1, compared with the baseline system, the service identification model in the embodiment of the present application has a stable improvement in AUC. In addition, it also has a stable improvement in ARPU.

[0118] like Figure 5 As shown, Figure 5 This is a schematic diagram of a prediction resource interaction tag provided by an embodiment of the present application, such as Figure 5 As shown, the media resource is a game prop in a game application, the resource interaction label can be a purchase probability, and the pre-trained generation model is an example of a GPT model. The computer device can obtain the data set D GPT =[s1,S2,s3,…,s n ],s i Contains the first historical interaction features of the i-th business object generated in the game for historical game props, such as prop usage features, prop click features, prop purchase features, etc.GPT Input the pre-trained generation model, and input the list of preferred game items y of each business object on a daily or weekly basis, etc. i = [y1, y2, … y K , and this y i = [y1, y2, … y K is partially ordered, that is, the earlier it is, the greater the probability that the user is interested in the item. Further, the computer device can use the item features of each game item in the list of preferred game items, that is, y i = [y1, y2, … y K as additional features and input them into the end-to-end lightweight business recognition model. At the same time, the computer device can obtain the data set D Rec = [d1, d2, d3, …, d m , and this d i represents the first historical interaction feature of the business object i for historical game items. The computer device can input y i = [y1, y2, … y K , the first historical interaction feature, and the item features of the target game item to be pushed into the end-to-end lightweight business recognition model. Through the cross-attention mechanism, according to the item features of the game items in the list of preferred game items y i = [y1, y2, … y K , extract the key item features (i.e., effective information) from the target game item to be pushed. Then, according to the key item features, the first historical interaction feature, and the item features of the target game item to be pushed, identify the purchase probability of the business object for the target game item That is, for the business recognition model F Rec :

[0119] In the embodiment of the present application, through a pre-trained generation model, according to the first historical interaction feature of the business object for the historical media resources, a list of preferred media resources of the historical media resources including the preferences of the business object is predicted. Since the pre-trained generation model is a large model with a large number of model parameters and a relatively deep number of model layers, it can accurately characterize the interest of the business object in the media resources, and at the same time can accurately extract the sequential information of the first historical interaction feature in time, reduce the loss of sequential information, and improve the prediction accuracy of the list of preferred media resources. Further, taking this list of preferred media resources as an additional feature and inputting it into the business recognition model, the business recognition model extracts key resource features from the first resource features to reflect the preferences of the business object for the target media resources. Then, according to the key resource features, the first historical interaction features, and the first resource features, the resource interaction label of the business object for the target media resources is identified, and the target media resources are pushed to the business object according to this resource interaction label. In this way, by the business recognition model outputting the resource interaction label in real time, it can not only ensure that the media resource interest of the business object is accurately characterized, improve the accuracy of pushing the target media resources, but also ensure the real-time nature of pushing the target media resources.

[0120] Further, please refer to Figure 6 , Figure 6 which is a schematic flowchart of a data processing method provided by an embodiment of the present application. As Figure 6 shown, this method can be executed by any terminal device in Figure 1 , or can be executed by server 10 in Figure 1 , or can also be jointly executed by the terminal device and the server in Figure 1 . The devices used to execute this data processing method in the present application can be collectively referred to as computer devices. Among them, this data processing method may include but is not limited to the following steps:

[0121] S201, Obtain the first historical interaction feature of the business object for the pushed historical media resources, and the first resource feature of the to-be-pushed target media resources.

[0122] S202, Through the pre-trained generation model, according to the first historical interaction feature, predict to obtain a list of preferred media resources.

[0123] S203, Through the business recognition model, according to the second resource feature of the historical media resources in the list of preferred media resources, extract key resource features from the first resource features to reflect the preferences of the business object for the target media resources.

[0124] S204. Based on the key resource features, the first historical interaction features, and the first resource features, obtain the resource interaction label of the business object for the target media resource; the target media resource to be pushed is the target game item in the game application, and the resource interaction label is the purchase probability.

[0125] Specifically, for the content of steps S201 - S204 in the embodiments of the present application, reference can be made to the content of steps S101 - S104 above, and details are not repeated here. Among them, the target media resource to be pushed is the target game item in the game application, and the resource interaction label can be the purchase probability of the business object for the target game item.

[0126] S205. Compare the purchase probability of the target game item with the purchase probability threshold.

[0127] S206. If the purchase probability of the target game item is greater than or equal to the probability threshold, then push the target game item to the business object.

[0128] Specifically, the computer device can obtain the purchase probability threshold. This purchase probability threshold can be set by the management object of the game application, can also be set by the computer device according to specific requirements, or can be the purchase probability set for the business object individually. For example, if the business object rarely purchases game items (i.e., has a low purchase desire), then the purchase probability of the business object can be set low to avoid the situation where there is no game item with a purchase probability greater than or equal to the purchase probability threshold, thus avoiding the risk of not being able to push game items to the business object. The computer device can compare the purchase probability of the target game item with the purchase probability threshold. If the purchase probability of the target game item is greater than or equal to the probability threshold, it indicates that the business object has a high possibility of purchasing the target game item, and then the target game item is pushed to the business object. If the purchase probability of the target game item is less than the probability threshold, it indicates that the business object has a low possibility of purchasing the target game item, and then the target game item will not be pushed to the business object. In this way, it is possible to accurately push the target game item to the business object, improve the return rate and activity of the business object in the game application, and at the same time, avoid pushing the target game item to invalid business objects (i.e., game objects that do not need the target game item), thereby greatly saving the pushing cost of game items.

[0129] As Figure 7 shown, Figure 7 is a schematic diagram of game item pushing provided by the embodiments of the present application. As Figure 7As shown in the figure, taking the target media resource as the target game item in the game application as an example, after the computer device generates a prediction through pre-training to obtain the list of preferred game items of the business object, according to the second resource features of the historical game items in the list of preferred game items, the key resource features for reflecting the preference of the business object for the target game item are extracted from the first resource features. Then, through the business recognition model, according to the key resource features, the first historical interaction features, and the first resource features, the purchase probability of the business object for the target game item is recognized. For example, Figure 7 As shown in the figure, if the computer device determines that the purchase probabilities corresponding to the target media resources 70b, 70c, and 70d are all greater than the purchase probability threshold, the target media resources 70b, 70c, and 70d can be displayed in the push game item interface 70a of the game application mall. For example, Figure 7 As shown in the figure, the target media resource 70b is a battle vehicle, the target media resource 70c is a battle suit, and the target media resource 70d is battle pants. At the same time, the computer device can also display the purchase controls corresponding to the target media resources 70b, 70c, and 70d in the push game item interface 70a, facilitating the purchase operations for the target media resources 70b, 70c, and 70d.

[0130] Optionally, the present application can also be applied to the video data push scenario. The target media resource can be the target video data in the video application, and the resource interaction label can be the click probability of the business object for the target video data. For example, Figure 8 As shown in the figure, Figure 8 is a schematic diagram of a video data push provided by an embodiment of the present application. For example, Figure 8As shown, the terminal device 80a can be a terminal device installed with a video application, and the business object 80b can be a user in the video application installed on the terminal device 80a. The terminal device 80a can obtain the historical video interaction features (i.e., historical interaction features) of the business object for the pushed historical video data. The historical video interaction features can include one or more of the video click feature, video playback feature, video sharing feature, etc. of the business object 80b for the historical video data. The terminal device 80a can also obtain the video features of the target video data to be pushed. The video features can include one or more of the video ID, video identifier, video type, video style, etc. The terminal device 80a can send the historical video interaction features and the video features of the target video data to the server 80c. The server 80c can input the historical video interaction features into a pre-trained generation model, which is trained based on the sample video data in the video application scenario. Through the pre-trained generation model, according to the historical video interaction features, a list of preferred video data of the business object 80b is predicted.

[0131] Furthermore, the server 80c can input the video features of the target video data and the video features of the video data in the list of preferred video data into the business recognition model. Through the cross-attention layer in the business recognition model, according to the video features of the video data in the list of preferred video data, the key video features preferred by the business object 80b for the target video data are extracted from the video features of the target video data. The server 80c can input the historical video interaction features, the video features of the target video data, and the key video features into the business sub-model in the business recognition model. The business sub-model can be any one of the model structures of the DIN model, DeepFM model, FM model, FFM (Field-aware Factorization Machine) model structure. By inputting the click probability of the business object 80b for the target video data through the business sub-model, the click probability of the business object 80b for the target video data is returned to the terminal device 80a. The terminal device 80a can detect whether the click probability of the business object 80b for the target video data is greater than the click probability threshold. The click probability threshold can be set by the management object of the video application, can also be set by the computer device according to specific requirements, or can be a personalized click probability threshold set for the business object 80b.

[0132] If the click probability of the target video data is greater than or equal to the click probability threshold, the target video data is pushed to the business object 80b. Figure 8As shown, if the click probabilities corresponding to the target video data 80e and the target video data 80f are both greater than the click probability threshold, the target video data 80e and the target video data 80f can be displayed in the push video data interface 80d of the video application. In this way, accurate pushing of target videos for business objects can be achieved, improving the regression rate and activity of business objects in the video application. At the same time, pushing target video data to invalid business objects (i.e., objects that do not need the target video data) can be avoided, thus greatly saving the pushing cost of video data.

[0133] In the embodiment of the present application, through a pre-trained generation model, based on the first historical interaction features of a business object for historical media resources, a preference media resource list of historical media resources including the preferences of the business object is predicted. Since the pre-trained generation model is a large model with a large number of model parameters and a relatively deep number of model layers, it can accurately depict the interest of the business object in media resources. At the same time, it can also accurately extract the sequential information of the first historical interaction features in time, reduce the loss of sequential information, and improve the prediction accuracy of the preference media resource list. Further, the preference media resource list is used as an additional feature to input into a business recognition model, and the business recognition model extracts key resource features from the first resource features to reflect the preferences of the business object for the target media resources. Then, based on the key resource features, the first historical interaction features, and the first resource features, the resource interaction label of the business object for the target media resources is identified, and the target media resources are pushed to the business object according to the resource interaction label. In this way, by the business recognition model outputting the resource interaction label in real time, it can not only ensure the accurate depiction of the media resource interest of the business object and improve the accuracy of pushing the target media resources, but also ensure the real-time nature of pushing the target media resources. The present application can be applied to scenarios such as game prop pushing and video data pushing, facilitating the accurate pushing of target game props or target video data, etc. for business objects, improving the regression rate and activity of business objects in game applications or video applications, etc. At the same time, pushing target game props or target video data, etc. to invalid business objects (i.e., objects that do not need target game props or target video data, etc.) can be avoided, thus greatly saving the pushing cost of game props or video data, etc.

[0134] Further, please refer to Figure 9 , Figure 9 is a schematic structural diagram of a data processing device provided by an embodiment of the present application. The data processing device can be a computer program (including program code) running in a computer device. For example, the data processing device is an application software; the data processing device can be used to execute the corresponding steps in the method provided by the embodiment of the present application. As Figure 9As shown, the data processing device can be any blockchain node in the blockchain network. The data processing device may include: a first acquisition module 11, a first prediction module 12, a first extraction module 13, a push module 14, a second acquisition module 15, a second prediction module 16, a first training module 17, a third acquisition module 18, a third prediction module 19, a second extraction module 20, an identification module 21, and a second training module 22.

[0135] The first acquisition module 11 is configured to acquire a first historical interaction feature of a business object with respect to a historical media resource that has been pushed, and a first resource feature of a target media resource to be pushed.

[0136] The first prediction module 12 is configured to predict a list of preferred media resources according to the first historical interaction feature through a pre-trained generation model; the list of preferred media resources includes historical media resources preferred by the business object.

[0137] The first extraction module 13 is configured to extract, from the first resource feature, a key resource feature for reflecting the preference of the business object for the target media resource according to the second resource feature of the historical media resources in the list of preferred media resources through a business identification model.

[0138] The push module 14 is configured to obtain a resource interaction label of the business object for the target media resource according to the key resource feature, the first historical interaction feature, and the first resource feature, and push the target media resource to the business object according to the resource interaction label.

[0139] Wherein, the first extraction module 13 is configured to:

[0140] Obtain a first resource feature vector for reflecting the first resource feature and a second resource feature vector for reflecting the second resource feature through a key feature extraction layer of the business identification model;

[0141] Generate a resource key vector and a resource value vector corresponding to the second resource feature vector according to the second resource feature vector and the bias vector corresponding to the second resource feature;

[0142] Extract, from the first resource feature vector, a key resource feature for reflecting the preference of the business object for the target media resource according to the resource key vector and the resource value vector. [[ID=2,6]]

[0143] Wherein, the first extraction module 13 is further configured to:

[0144] Generate a similarity vector according to the first resource feature vector and the resource key vector; the similarity vector is used to reflect the similarity between the second resource feature and the first resource feature.

[0145] Multiply the similarity vector, the resource value vector, and the weight vector corresponding to the resource value vector to obtain the key resource feature that reflects the preference of the business object for the target media resource.

[0146] Among them, the first extraction module 13 is further configured to:

[0147] Multiply the first resource feature vector by the weight vector corresponding to the first resource feature vector to obtain the processed first resource feature vector;

[0148] Multiply the resource key vector by the weight vector corresponding to the resource key vector to obtain the processed resource key vector;

[0149] Multiply the processed first resource feature vector by the transpose of the processed resource key vector to obtain the similarity vector.

[0150] Among them, the push module 14 is configured to:

[0151] Perform feature embedding processing on the key resource feature, the first historical interaction feature, and the second resource feature to obtain the dense resource interaction feature;

[0152] Through the feature cross-combination layer in the service recognition model, perform linear and cross processing on the dense resource interaction feature to obtain the cross resource interaction feature;

[0153] Through the deep feature extraction layer in the service recognition model, perform deep feature extraction on the dense resource interaction feature to obtain the deep resource interaction feature;

[0154] According to the cross resource interaction feature and the deep resource interaction feature, obtain the resource interaction label of the business object for the target media resource.

[0155] Among them, the push module 14 is further configured to:

[0156] Encode the key resource feature, the first historical interaction feature, and the second resource feature respectively to obtain the sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature respectively;

[0157] Concatenate the sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature respectively to obtain the concatenated sparse feature;

[0158] Perform feature embedding processing on the concatenated sparse feature to obtain the dense resource interaction feature.

[0159] Among them, the push module 14 is further configured to:

[0160] Determine the feature correlation degree between the first historical interaction feature and the first resource feature through the attention layer in the service recognition model, and perform feature weighting processing on the first historical interaction feature according to the feature correlation degree to obtain the processed first historical interaction feature;

[0161] Through the feature splicing layer in the service recognition model, splice the key resource feature, the processed first historical interaction feature, and the second resource feature to obtain the spliced resource feature;

[0162] Through the fully connected network layer in the service recognition model, perform feature classification on the spliced resource feature to obtain the resource interaction label of the service object for the target media resource.

[0163] Among them, the push module 14 is further used for:

[0164] Multiply the first historical interaction feature and the second resource feature to obtain a resource similarity feature;

[0165] Perform activation processing on the resource similarity feature to obtain the activated resource mapping feature;

[0166] Perform linear processing on the activated resource mapping feature to obtain the feature correlation degree between the first historical interaction feature and the second resource feature.

[0167] Among them, the pre-trained generation model includes M transformation encoders; M is an integer greater than 1;

[0168] The first prediction module 12 is used for:

[0169] Through N self-attention layers in the first transformation encoder, perform attention feature extraction on the first historical interaction feature to obtain N-dimensional attention features; the first transformation encoder is the transformation encoder ranked first in the M transformation encoders, and N is a positive integer;

[0170] [[ID=3))Through the neural network layer in the first transformation encoder, perform deep feature extraction on the N-dimensional attention features to obtain a resource preference feature;

[0171] Determine the list of preferred media resources preferred by the service object according to the resource preference feature and the remaining transformation encoders; the remaining transformation encoders are the transformation encoders other than the first transformation encoder among the M transformation encoders.

[0172] Among them, the data processing device further includes:

[0173] The second acquisition module 15 is configured to acquire the second historical interaction feature corresponding to the first sample media resource interacted by the first sample object within the first historical time period, and the second sample media resource interacted by the first sample object within the second historical time period; the first historical time period is before the second historical time period, and the first historical time period is adjacent to the second historical time period;

[0174] The second prediction module 16 is configured to, through the initial pre-trained generation model, predict a first predicted media resource list according to the second historical interaction feature; the first predicted media resource list includes the first sample media resources preferred by the first sample object;

[0175] The first training module 17 is configured to train the initial pre-trained generation model according to the second sample media resource and the first predicted media resource list until the initial pre-trained generation model meets the convergence condition, so as to obtain the pre-trained generation model.

[0176] Wherein, the data processing device further includes:

[0177] The third acquisition module 18 is configured to acquire the third historical interaction feature of the second sample object for the third sample media resource that has been pushed, the third resource feature of the fourth sample media resource to be pushed, and the labeled resource interaction label of the second sample object for the fourth sample media resource;

[0178] The third prediction module 19 is configured to, through the pre-trained generation model, predict a second predicted media resource list according to the third historical interaction feature; the second predicted media resource list includes the third sample media resources preferred by the second sample object;

[0179] The second extraction module 20 is configured to, through the initial service recognition model, extract, from the third resource feature, the predicted key resource feature used to reflect the preference of the second sample object for the fourth sample media resource according to the fourth resource feature of the historical media resources in the second predicted media resource list;

[0180] The recognition module 21 is configured to recognize the predicted resource interaction label of the second sample object for the fourth sample media resource according to the predicted key resource feature, the third historical interaction feature, and the third resource feature;

[0181] The second training module 22 is configured to train the initial service recognition model according to the predicted resource interaction label and the labeled resource interaction label until the initial service recognition model meets the convergence condition, so as to obtain the service recognition model.

[0182] Wherein, the target media resource to be pushed is the target game item in the game application, and the resource interaction label is the purchase probability;

[0183] The push module 14 is configured to:

[0184] Compare the purchase probability of the target game item with the purchase probability threshold;

[0185] If the purchase probability of the target game item is greater than or equal to the probability threshold, then push the target game item to the business object.

[0186] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other relevant parts to achieve a predetermined goal, and can be fully or partially implemented by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the function of the module or unit. According to an embodiment of the present application, Figure 9 Each module in the data processing device shown can be separately or entirely combined into one or several units to form, or a certain one (or some) of the units can be further split into at least two smaller sub-units in terms of function, and the same operations can be achieved without affecting the realization of the technical effects of the embodiments of the present application. The above modules are divided based on logical functions. In actual applications, the function of one module can also be implemented by at least two units, or the functions of at least two modules can be implemented by one unit. In other embodiments of the present application, the data processing device can also include other units. In actual applications, these functions can also be assisted by other units and can be achieved by the cooperation of at least two units.

[0187] According to an embodiment of the present application, it can be achieved by running a computer program (including program code) that can execute the respective steps involved in the corresponding method shown in Figure 3 on a general computer device such as a computer including processing elements and storage elements such as a central processing unit (CPU), a random access storage medium (RAM), a read-only storage medium (ROM), etc., to construct a data processing device as shown in Figure 9 and to implement a data processing method of an embodiment of the present application. The above computer program can be recorded on, for example, a computer-readable recording medium, loaded into the above computer device through the computer-readable recording medium, and run therein.

[0188] In the embodiment of the present application, through a pre-trained generation model, according to the first historical interaction feature of the business object for the historical media resource, a list of preferred media resources of the historical media resource including the preferences of the business object is predicted. Since the pre-trained generation model is a large model with a large number of model parameters and a relatively deep number of model layers, it can accurately characterize the interest of the business object in the media resource, and at the same time, it can accurately extract the sequential information of the first historical interaction feature in time, reduce the loss of sequential information, and improve the prediction accuracy of the list of preferred media resources. Further, the list of preferred media resources is used as an additional feature to input into the service recognition model, and the service recognition model extracts key resource features from the first resource feature to reflect the preferences of the business object for the target media resource. Then, according to the key resource features, the first historical interaction feature, and the first resource feature, the resource interaction label of the business object for the target media resource is identified, and the target media resource is pushed to the business object according to the resource interaction label. In this way, by the service recognition model outputting the resource interaction label in real time, it can not only ensure that the interest of the business object in the media resource is accurately characterized, improve the accuracy of pushing the target media resource, but also ensure the real-time nature of pushing the target media resource. The present application can be applied to scenarios such as game prop pushing and video data pushing, which is convenient for accurately pushing the target game prop or target video data to the business object, improving the return rate and activity of the business object in game applications or video applications, etc. At the same time, it can avoid pushing the target game prop or target video data to invalid business objects (i.e., objects that do not need the target game prop or target video data, etc.), thereby greatly saving the pushing cost of game props or video data, etc.

[0189] Further, please refer to Figure 10 , Figure 10 which is a schematic diagram of a computer device provided by an embodiment of the present application. As Figure 10 shown, the computer device 3000 may be the terminal device or server corresponding to the above Figure 2 embodiment. The computer device 3000 may include: at least one processor 3001, such as a CPU, at least one network interface 3004, a user interface 3003, a memory 3005, and at least one communication bus 3002. Among them, the communication bus 3002 is used to realize the connection and communication between these components. Among them, the user interface 3003 may include a display screen (Display) and a keyboard (Keyboard). The network interface 3004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 3005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The storage 3005 may optionally also be at least one storage device located far from the aforementioned processor 3001. AsFigure 10 As shown, the memory 3005, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a computer program control application.

[0190] In Figure 10 In the computer device 3000 shown, the network interface 3004 is mainly used for the second node device to communicate with the target relay server and the target oracle server; while the user interface 3003 is mainly used to provide an interface for users to input; and the processor 3001 can be used to call the computer program control application stored in the memory 3005 to implement:

[0191] Obtain the first historical interaction feature of the business object for the pushed historical media resources, and the first resource feature of the target media resource to be pushed;

[0192] Through a pre-trained generation model, predict a list of preferred media resources according to the first historical interaction feature; the list of preferred media resources includes the historical media resources preferred by the business object;

[0193] Through a service recognition model, according to the second resource features of the historical media resources in the list of preferred media resources, extract the key resource features from the first resource features to reflect the preferences of the business object for the target media resource;

[0194] According to the key resource features, the first historical interaction feature, and the first resource feature, obtain the resource interaction label of the business object for the target media resource, and push the target media resource to the business object according to the resource interaction label.

[0195] It should be understood that the computer device 3000 described in the embodiments of the present application can also execute the description of a data processing method in the corresponding embodiments mentioned above. The computer device 3000 described in the embodiments of the present application can also execute the description of the data processing device in the corresponding embodiments mentioned above. Figure 6 corresponding embodiments respectively, which will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. Figure 9 respectively corresponding embodiments of the description of the data processing device, which will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either.

[0196] In addition, it should be pointed out here that: the embodiments of the present application also provide a computer-readable storage medium, and the computer-readable storage medium stores the computer program executed by the data processing device mentioned above, and the computer program includes program instructions. When the processor executes the program instructions, it can execute the above Figure 3 or Figure 6The description of the data processing method in the corresponding embodiment will not be repeated here. In addition, the beneficial effects of using the same method will not be described again. For the technical details not disclosed in the embodiment of the computer-readable storage medium involved in this application, please refer to the description of the method embodiment of this application. As an example, the program instructions can be deployed to be executed on a computing device, or on multiple computing devices located at one location, or on multiple computing devices distributed at multiple locations and interconnected by a communication network. The multiple computing devices distributed at multiple locations and interconnected by a communication network can form a blockchain system.

[0197] On the one hand, this application provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device can execute the description of a data processing method in the foregoing Figure 3 Or Figure 6 The description of the corresponding embodiment will not be repeated here. In addition, the beneficial effects of using the same method will not be described again.

[0198] It should be noted that when collecting and processing relevant data in this application book during actual application, the informed consent or separate consent (or having a legal basis) of the personal information subject should be obtained strictly in accordance with the requirements of relevant national laws and regulations, and subsequent data use and processing behaviors should be carried out within the scope authorized by laws and regulations and the personal information subject. For example, when this application obtains interactive data such as historical interaction data (such as resource usage data, resource click data, resource purchase data, resource sharing data, etc.) and historical interaction features (such as resource usage features, resource click features, resource purchase features, and resource sharing features, etc.) of the business object and the sample object for media data respectively, the informed consent or separate consent of the corresponding business object or sample object needs to be obtained.

[0199] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0200] The above-disclosed is only the preferred embodiment of the present invention, and of course, the scope of the rights of the present invention cannot be limited thereby. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope covered by the present invention.

Claims

1. A data processing method, characterized in that, Including: Obtain the first historical interaction feature of the business object for the pushed historical media resources and the first resource feature of the target media resource to be pushed; Through a pre-trained generation model, predict a list of preferred media resources according to the first historical interaction feature; the list of preferred media resources includes the historical media resources preferred by the business object; Through a business recognition model, according to the second resource feature of the historical media resources in the list of preferred media resources, extract from the first resource feature the key resource feature for reflecting the preference of the business object for the target media resource; According to the key resource feature, the first historical interaction feature and the first resource feature, obtain the resource interaction label of the business object for the target media resource, and push the target media resource to the business object according to the resource interaction label.

2. The method according to claim 1, wherein The step of, through a business recognition model, according to the second resource feature of the historical media resources in the list of preferred media resources, extract from the first resource feature the key resource feature for reflecting the preference of the business object for the target media resource includes: Through the key feature extraction layer of the business recognition model, obtain the first resource feature vector for reflecting the first resource feature and the second resource feature vector for reflecting the second resource feature; Generate the resource key vector and resource value vector corresponding to the second resource feature vector according to the second resource feature vector and the bias vector corresponding to the second resource feature; According to the resource key vector and the resource value vector, extract from the first resource feature vector the key resource feature for reflecting the preference of the business object for the target media resource.

3. The method according to claim 2, wherein The step of, according to the resource key vector and the resource value vector, extract from the first resource feature vector the key resource feature for reflecting the preference of the business object for the target media resource includes: Generate a similarity vector according to the first resource feature vector and the resource key vector; the similarity vector is used to reflect the similarity between the second resource feature and the first resource feature; Perform a multiplication process on the similarity vector, the resource value vector, and the weight vector corresponding to the resource value vector to obtain the key resource feature for reflecting the preference of the business object for the target media resource.

4. The method according to claim 3, wherein The step of, according to the first resource feature vector and the resource key vector, generate a similarity vector includes: Perform a multiplication process on the first resource feature vector and the weight vector corresponding to the first resource feature vector to obtain the processed first resource feature vector; Perform a multiplication process on the resource key vector and the weight vector corresponding to the resource key vector to obtain the processed resource key vector; Perform a multiplication process on the processed first resource feature vector and the transpose of the processed resource key vector to obtain the similarity vector.

5. The method according to claim 1, characterized in that, The step of, according to the key resource feature, the first historical interaction feature and the first resource feature, obtain the resource interaction label of the business object for the target media resource includes: Perform feature embedding processing on the key resource feature, the first historical interaction feature, and the second resource feature to obtain a dense resource interaction feature; Through the feature cross-combination layer in the service recognition model, perform linear and cross processing on the dense resource interaction feature to obtain a cross resource interaction feature; Through the deep feature extraction layer in the service recognition model, perform deep feature extraction on the dense resource interaction feature to obtain a deep resource interaction feature; Based on the cross resource interaction feature and the deep resource interaction feature, obtain the resource interaction label of the service object for the target media resource.

6. The method according to claim 5, wherein The performing feature embedding processing on the key resource feature, the first historical interaction feature, and the second resource feature to obtain a dense resource interaction feature includes: Encode the key resource feature, the first historical interaction feature, and the second resource feature respectively to obtain sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature respectively; Concatenate the sparse features corresponding to the key resource feature, the first historical interaction feature, and the second resource feature respectively to obtain a concatenated sparse feature; Perform feature embedding processing on the concatenated sparse feature to obtain a dense resource interaction feature.

7. The method according to claim 1, characterized in that, The obtaining the resource interaction label of the service object for the target media resource based on the key resource feature, the first historical interaction feature, and the first resource feature includes: Through the attention layer in the service recognition model, determine the feature correlation degree between the first historical interaction feature and the first resource feature, and perform feature weighting processing on the first historical interaction feature according to the feature correlation degree to obtain a processed first historical interaction feature; Through the feature concatenation layer in the service recognition model, concatenate the key resource feature, the processed first historical interaction feature, and the second resource feature to obtain a concatenated resource feature; Through the fully connected network layer in the service recognition model, perform feature classification on the concatenated resource feature to obtain the resource interaction label of the service object for the target media resource.

8. The method according to claim 7, characterized in that, The determining the feature correlation degree between the first historical interaction feature and the first resource feature through the attention layer in the service recognition model includes: Multiply the first historical interaction feature and the second resource feature to obtain a resource similarity feature; Perform activation processing on the resource similarity feature to obtain an activated resource mapping feature; Perform linear processing on the activated resource mapping feature to obtain the feature correlation degree between the first historical interaction feature and the second resource feature.

9. The method according to claim 1, wherein The pre-trained generation model includes M transformation encoders; M is an integer greater than 1; The predicting a list of preferred media resources based on the first historical interaction feature through the pre-trained generation model includes: Through N self-attention layers in the first transformation encoder, attention feature extraction is performed on the first historical interaction feature to obtain N-dimensional attention features; the first transformation encoder is the transformation encoder with the first arrangement order among the M transformation encoders, and N is a positive integer; Through the neural network layer in the first transformation encoder, deep feature extraction is performed on the N-dimensional attention features to obtain resource preference features; According to the resource preference features and the remaining transformation encoders, a preference media resource list preferred by the business object is determined; the remaining transformation encoders are the transformation encoders other than the first transformation encoder among the M transformation encoders.

10. The method according to claim 1, wherein The method further includes: Obtain the second historical interaction feature corresponding to the first sample media resource interacted by the first sample object in the first historical time period, and the second sample media resource interacted by the first sample object in the second historical time period; the first historical time period is before the second historical time period, and the first historical time period is adjacent to the second historical time period; Through the initial pre-trained generation model, according to the second historical interaction feature, a first predicted media resource list is predicted; the first predicted media resource list includes the first sample media resources preferred by the first sample object; According to the second sample media resource and the first predicted media resource list, the initial pre-trained generation model is trained until the initial pre-trained generation model meets the convergence condition to obtain the pre-trained generation model.

11. The method according to claim 1, wherein The method further includes: Obtain the third historical interaction feature of the second sample object for the third sample media resource that has been pushed, the third resource feature of the fourth sample media resource to be pushed, and the labeled resource interaction label of the second sample object for the fourth sample media resource; Through the pre-trained generation model, according to the third historical interaction feature, a second predicted media resource list is predicted; the second predicted media resource list includes the third sample media resources preferred by the second sample object; Through the initial business recognition model, according to the fourth resource feature of the historical media resources in the second predicted media resource list, the predicted key resource feature reflecting the preference of the second sample object for the fourth sample media resource is extracted from the third resource feature; According to the predicted key resource feature, the third historical interaction feature, and the third resource feature, the predicted resource interaction label of the second sample object for the fourth sample media resource is identified; According to the predicted resource interaction label and the labeled resource interaction label, the initial business recognition model is trained until the initial business recognition model meets the convergence condition to obtain the business recognition model.

12. The method according to claim 1, characterized in that, The target media resource to be pushed is the target game prop in the game application, and the resource interaction label is the purchase probability; The pushing the target media resource to the business object according to the resource interaction label includes: Comparing the purchase probability of the target game prop with the purchase probability threshold; If the purchase probability of the target game item is greater than or equal to the probability threshold, the target game item is pushed to the business object.

13. A data processing device, characterized in that, Including: A first acquisition module, configured to acquire a first historical interaction feature of the business object for the pushed historical media resource, and a first resource feature of the target media resource to be pushed; A first prediction module, configured to predict a list of preferred media resources according to the first historical interaction feature through a pre-trained generation model; the list of preferred media resources includes the historical media resources preferred by the business object; A first extraction module, configured to extract, according to the second resource feature of the historical media resource in the list of preferred media resources through a business recognition model, key resource features from the first resource feature for reflecting the preference of the business object for the target media resource; A push module, configured to obtain a resource interaction label of the business object for the target media resource according to the key resource feature, the first historical interaction feature, and the first resource feature, and push the target media resource to the business object according to the resource interaction label.

14. A computer device, characterized in that, Including: A processor and a memory; The processor is connected to the memory, wherein the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method according to any one of claims 1-12.

15. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, and the computer program is suitable for being loaded and executed by the processor so that the computer device with the processor executes the method according to any one of claims 1-12.

16. A computer program product or computer program, characterized in that, The computer program product or the computer program includes computer instructions, the computer instructions are stored in the computer-readable storage medium, and the computer instructions are suitable for being read and executed by the processor so that the computer device with the processor executes the method according to any one of claims 1-12.

Images