Resource pushing method and device, equipment, storage medium and computer program product

Abstract

The application discloses a resource pushing method and device, equipment, a storage medium and a computer program product. The method comprises the following steps: calling a first interaction prediction network to predict an interaction probability of a to-be-pushed object performing a target interaction on a to-be-pushed resource; determining a matching degree between the to-be-pushed object and the to-be-pushed resource based on the interaction probability; if the matching degree meets a pushing rule, pushing the to-be-pushed resource to the to-be-pushed object; the first interaction prediction network is trained based on a second interaction prediction network with a larger network scale, a first sample resource and a historical short-time interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and a historical long-time interaction sequence related to the first sample resource, and the historical long-time interaction sequence comprises the historical short-time interaction sequence; the first interaction prediction network and the second interaction prediction network are both used for target interaction prediction; and the pushing accuracy and the pushing timeliness during resource pushing can be considered simultaneously.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a resource push method, apparatus, device, storage medium, and computer program product. Background Technology

[0002] In the field of resource push, it is common practice to predict the probability that a user will perform a target interaction with a push resource, and then determine whether to push the resource to that user based on the predicted probability. For example, it is possible to predict the probability that a user will click on a push resource (i.e., perform click-through rate estimation), and then determine whether to push the resource to that user based on the predicted probability (i.e., the estimated click-through rate). Therefore, the accuracy of the prediction when predicting the probability that a user will perform a target interaction with a push resource directly affects the accuracy of the resource push, and the prediction rate when predicting the probability that a user will perform a target interaction with a push resource directly affects the timeliness of the resource push. However, existing methods for predicting the probability that a user will perform a target interaction with a push resource cannot simultaneously meet the requirements of both prediction accuracy and prediction rate, thus making it impossible to simultaneously achieve both push accuracy and push timeliness when pushing resources. Summary of the Invention

[0003] This application provides a resource push method, apparatus, device, storage medium, and computer program product that can simultaneously ensure the accuracy and timeliness of resource push.

[0004] On the one hand, embodiments of this application provide a resource push method, including:

[0005] In response to a push request to push resources to a target object, the first interaction prediction network is invoked to predict the interaction probability of the target object performing the target interaction on the push resource;

[0006] The matching degree between the object to be pushed and the resource to be pushed is determined based on the interaction probability.

[0007] If the matching degree between the object to be pushed and the resource to be pushed meets the push rules, then the resource to be pushed is pushed to the object to be pushed.

[0008] The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network.

[0009] On one hand, embodiments of this application provide a resource push device, including:

[0010] The processing unit is configured to, in response to a push request to push resources to a target object, invoke a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource.

[0011] The processing unit is further configured to determine the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability;

[0012] An output unit is used to push the resource to be pushed to the object if the matching degree between the object to be pushed and the resource to be pushed meets the push rules.

[0013] The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network.

[0014] On one hand, embodiments of this application provide a resource push device, characterized in that the resource push device includes an input interface and an output interface, and further includes:

[0015] A processor, adapted to implement one or more instructions; and,

[0016] A computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded by the processor and executed by the above-described resource push method.

[0017] On one hand, this application provides a computer storage medium, characterized in that the computer storage medium stores computer program instructions, which, when executed by a processor, are used to execute the above-mentioned resource push method.

[0018] On one hand, embodiments of this application provide a computer program product or a computer program, the computer program product including a computer program stored in a computer storage medium; the processor of the resource push device reads the computer program from the computer storage medium, the processor executes the computer program, causing the resource push device to execute the above-described resource push method.

[0019] In this embodiment, in response to a push request to push resources to a target object, a first interaction prediction network is invoked to predict the interaction probability of the target object performing a target interaction on the target resource; and the matching degree between the target object and the target resource is determined based on the interaction probability; if the matching degree between the target object and the target resource meets the push rules, the target resource is pushed to the target object; wherein, the first interaction prediction network is trained based on a second interaction prediction network, a first sample resource, and historical short-term interaction sequences related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and historical long-term interaction sequences related to the first sample resource; the historical long-term interaction sequences include historical short-term interaction sequences; both the first and second interaction prediction networks are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network. Because long-term historical interaction sequences carry richer historical interaction features than short-term historical interaction sequences, the second interaction prediction network trained based on the long-term historical interaction sequences related to the first sample resource can predict the probability of an object performing a target interaction with the resource based on these rich historical interaction features, and the prediction accuracy is high. However, due to the large amount of data in the long-term historical interaction sequences, the second interaction prediction network, which can process such large amounts of data, has a large network size. Generally speaking, large-scale models take a long time to process large amounts of data, which does not meet the timeliness requirements in the resource push domain. Therefore, a smaller-scale first interaction prediction network can be trained using short-term historical interaction sequences, which have less data than long-term historical interaction sequences, to ensure the first interaction prediction network's performance. The lightweight nature of the first interaction prediction network improves the prediction rate of the probability that an object will perform a target interaction with a resource, thus ensuring the timeliness of resource push. Furthermore, by training the first interaction prediction network on the second interaction prediction network, the first interaction prediction network can effectively learn the processing capabilities of the second interaction prediction network. Therefore, when the first interaction prediction network is trained using historical short-term interaction sequences with a small amount of data, the prediction accuracy of the probability of an object performing a target interaction with a resource can be guaranteed. Based on the above description, it can be seen that when the first interaction prediction network is used to predict the interaction probability of an object to be pushed performing a target interaction with a resource to be pushed, both prediction accuracy and prediction rate can be taken into account, thereby ensuring both the accuracy and timeliness of resource push. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the structure of a resource push system provided in an embodiment of this application;

[0022] Figure 2 This is a flowchart illustrating a training method for a second interactive prediction network provided in an embodiment of this application;

[0023] Figure 3 This is an exemplary structural diagram of a second interactive prediction network provided in an embodiment of this application;

[0024] Figure 4 This is a flowchart illustrating a training method for a first interactive prediction network provided in an embodiment of this application;

[0025] Figure 5 This is a schematic diagram illustrating the training of a first interactive prediction model based on a second prediction result, a third prediction result, and a second interactive label, according to an embodiment of this application.

[0026] Figure 6 This is a flowchart illustrating a resource push method provided in an embodiment of this application;

[0027] Figure 7 This is a schematic diagram of the structure of a resource push device provided in an embodiment of this application;

[0028] Figure 8 This is a schematic diagram of the structure of a resource push device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0030] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0031] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision (CV), speech processing, natural language processing, and machine learning (ML) / deep learning (DL).

[0032] This application provides a resource push scheme based on machine learning technology in the field of artificial intelligence. In response to a push request to push resources to a target object, it invokes a first interaction prediction network to predict the interaction probability of the target object performing a target interaction with the target resource; and determines the matching degree between the target object and the target resource based on the interaction probability. If the matching degree between the target object and the target resource conforms to the push rules, the target resource is pushed to the target object. The first interaction prediction network is trained based on a second interaction prediction network, a first sample resource, and historical short-term interaction sequences related to the first sample resource. The second interaction prediction network is trained based on the first sample resource and historical long-term interaction sequences related to the first sample resource. The historical long-term interaction sequences include historical short-term interaction sequences. Both the first and second interaction prediction networks are used for target interaction prediction, and the network size of the first interaction prediction network is smaller than that of the second interaction prediction network. Furthermore, both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction. Optionally, the network structures of the first interaction prediction network and the second interaction prediction network can be the same or different. This application embodiment does not limit the specific network structures of the first interaction prediction network and the second interaction prediction network. As long as the network can achieve target interaction prediction, it is within the protection scope of this application embodiment. For ease of explanation, the following embodiments of this application will be described with the first interaction prediction network and the second interaction prediction network having the same network structure.

[0033] The aforementioned resource push solution can be executed through a resource push device, which can be a terminal device, including but not limited to smartphones, tablets, laptops, desktop computers, smart voice interaction devices, smart home appliances, in-vehicle terminals, and smart wearable devices; it can also be a server, such as a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. This resource push solution can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0034] Based on the above resource push scheme, this application provides a resource push system, see [link to relevant documentation]. Figure 1 This is a schematic diagram of the structure of a resource push system provided in an embodiment of this application. Figure 1 The resource push system shown may include a resource push device 101 and a terminal device 102. The resource push device 101 may be a server, such as a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal device 102 may include, but is not limited to, any one or more of the following: smartphones, tablets, laptops, desktop computers, smart voice interaction devices, smart home appliances, in-vehicle terminals, and smart wearable devices. The resource push device 101 and the terminal device 102 may communicate directly or indirectly via wired or wireless communication, which is not limited herein.

[0035] In one embodiment, a target application may run on the terminal device 102. This target application can be any legitimate application that provides resource push services, such as a social application or a news application. The target application can provide resource push services to the target user corresponding to the terminal device 102. The resources pushed by the target application are resources related to the services provided by the target application. For example, if the target application is a news application, it can push current events news, current events commentary articles, etc., to the target user. The resource push device 101 is a server corresponding to the target application, providing service support for the target application, specifically providing resource push service support. For example, if the target application is a news application, the resource push device 101 can be a server providing service support for that news application. The target user can be any user of the services provided by the target application.

[0036] In one embodiment, the target object can use the services provided by the target application through its terminal device 102, such as clicking to view resources pushed by the target application. The resource push device 101, in response to a push request to push resources to the target object, can invoke a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the resource to be pushed; and determine the matching degree between the target object and the resource to be pushed based on the interaction probability; if the matching degree between the target object and the resource to be pushed conforms to the push rules, then the resource to be pushed is pushed to the target object so that the target object can click to view the resource to be pushed through its terminal device 102. Optionally, the push request can be triggered based on the target object. For example, if the target object performs a resource search operation in the target application on its terminal device 102, a push request can be generated based on the resource search operation, and the push request can be sent to the resource push device 101 based on the terminal device 102. After receiving the push request, the resource push device 101 can respond to the push request and push resources to the target object. The push request can also be triggered by business personnel. For example, if business personnel need to promote advertising, the resource push device 101 can respond to the push request triggered by the business personnel to push advertising resources to the target object. Based on the interaction probability of the target object performing the target interaction on the advertising resource, which is predicted by calling the first interaction prediction network, it determines whether to push the advertising resource to the target object. For example, in the scenario of intelligent transportation, the terminal device 102 can be an in-vehicle terminal, and the target application can be an application used to push traffic information. If the resource to be pushed is traffic information, the resource push device 101 can determine whether to push the traffic information to the in-vehicle terminal based on the predicted interaction probability of the target object corresponding to the in-vehicle terminal performing the target interaction on the traffic information, so that the target object can understand the traffic information and improve the driving experience. Optionally, since the network scale of the first interaction prediction network is small and the amount of historical short-term interaction sequence data is small, this resource push scheme can also be executed by the terminal device 102 with general computing power.

[0037] In one embodiment, the target interaction can be specifically configured based on specific needs. When the target interaction is configured differently, a first interaction prediction network is used to predict the probability of the object performing different interactions with the resource, and a second interaction prediction network is used to predict the probability of the object performing different interactions with the resource. For example, if the target interaction is set to a click interaction, both the first and second interaction prediction networks are click-through rate prediction networks, used to predict the probability of the object performing a click interaction with the resource, that is, to predict the estimated click-through rate of the object clicking the resource. The first interaction prediction network can be called to predict the estimated click-through rate of the object to be pushed clicking the resource to be pushed; and the matching degree between the object to be pushed and the resource to be pushed is determined based on the estimated click-through rate; if the matching degree between the object to be pushed and the resource to be pushed meets the push rules, the resource to be pushed is pushed to the object to be pushed. For example, if the target interaction is set as a conversion interaction, then both the first and second interaction prediction networks are conversion rate prediction networks, used to predict the probability of an object performing a conversion interaction after clicking on a resource, i.e., to predict the estimated conversion rate of an object clicking on a resource and then performing a conversion operation. The first interaction prediction network can be called to predict the estimated conversion rate of the object to be pushed to perform a conversion operation on the resource to be pushed; and the matching degree between the object to be pushed and the resource to be pushed can be determined based on the estimated conversion rate; if the matching degree between the object to be pushed and the resource to be pushed meets the push rules, then the resource to be pushed is pushed to the object to be pushed; wherein, the conversion operation can be specifically set in different application scenarios, such as including activation, registration, payment and other operations performed after the object clicks on the resource.

[0038] In one embodiment, the target interaction can be one or more of a variety of interactions, which can be set according to specific needs. For example, if the need instructs the first and second interaction prediction networks to be click-through rate prediction networks, then the various interactions can be determined as "exposure interaction" and "click interaction," and the target interaction is the "click interaction." Here, "exposure interaction" refers to pushing a resource to an object, but the object does not click on it. Alternatively, the various interactions can be determined as "exposure interaction," "click without conversion," and "click with conversion," and the target interaction is either "click without conversion" or "click with conversion." Furthermore, if the need instructs the first and second interaction prediction networks to be conversion rate prediction networks, then the various interactions can be determined as "exposure interaction," "click without conversion," and "click with conversion," and the target interaction is "click with conversion."

[0039] In one embodiment, the first sample resource is a resource used to train the first interaction prediction network and the second interaction prediction network. The first sample resource is a resource pushed to the target object within a preset time period. The target object can be any user using the services provided by the target application. The preset time period can be set according to specific needs, such as 2 months. For example, if the preset time period is from t1 to t2, where t1 is earlier than t2, the resources pushed to the target object from morning to night during the t1 to t2 time period include: resource A, resource B, resource C, and resource D. Then, a first sample resource can be any one of the above resources A, B, C, and D, and a first sample resource corresponds to a historical long-term interaction sequence and a historical short-term interaction sequence.

[0040] Furthermore, the historical long-term interaction sequence related to the first sample resource may include N second sample resources and N first interaction tags. The N second sample resources are pushed to the target object within a first historical time period starting from the interaction between the target object and the first sample resource. The N first interaction tags correspond one-to-one with the N second sample resources, and the N first interaction tags are used to indicate the interaction between the target object and the N second sample resources, where N is a positive integer. The first preset time period can be set according to specific needs. For example, the first preset time period can be set to 1 year, then the N second sample resources are pushed to the target object within 1 year starting from the interaction between the target object and the first sample resource. Further, if the interaction between the target object and the first sample resource is an "exposure interaction," then the first historical time period starting from the interaction between the target object and the first sample resource can be defined as the first historical time period starting from when the first sample resource is pushed to the target object.

[0041] For example, if the first and second interaction prediction networks are click-through rate (CTR) prediction models, and the various interactions are "exposure interaction," "click without conversion," and "click with conversion," with the target interaction being either "click without conversion" or "click with conversion," then "exposure interaction," "click without conversion," and "click with conversion" are labeled "0," "1," and "2," respectively. If the object performs the target interaction, it is labeled "1," and if the object does not perform the target interaction, it is labeled "0." That is, when the interaction between the object and a resource is the target interaction, it can be labeled "1," and when the interaction between the object and a resource is not... During target interaction, the label "0" can be used to identify it. If the first sample resource is resource B pushed to the target object during the time period t1 to t2, and if the target object interacts with resource B at time t3, then the first historical time period from the point where the target object interacts with the first sample resource can be within one year from t3, that is, within one year before t3; if the resources pushed to the target object from morning to evening within one year before t3 are: resource E, resource F, resource G, resource H, and resource A, then the historical long-term interaction sequence related to the first sample resource (resource B) can include 5 second sample resources, namely: resource E, resource F, resource G, resource H, and resource A. F, Resource G, Resource H, Resource A; if the target object's interactions with Resources E, F, G, H, and A are respectively: "Exposure Interaction", "Click Without Conversion", "Click Converted", "Click Without Conversion", and "Exposure Interaction", then the historical long-term interaction sequence related to the first sample resource (Resource B) can include 5 second sample resources, namely: Resource E, Resource F, Resource G, H, and A, and 5 first interaction tags, namely: 0, 1, 2, 1, 0. Optionally, if the first interaction prediction network and the second interaction prediction network are conversion rate prediction models, then N second samples... Resources can be those pushed to the target object within the first historical time period after the target object interacts with the first sample resource, and which the target object clicks on. In other words, N second sample resources can be those pushed to the target object within the first historical time period after the target object interacts with the first sample resource, and which the target object performs a "click not converted" or "click converted" action on. For example, the historical long-term interaction sequence related to the first sample resource (resource B) can include three second sample resources: resource F, resource G, and resource H, and three first interaction tags: 1, 2, and 1.

[0042] In one embodiment, the historical long-term interaction sequence related to the first sample resource includes a historical short-term interaction sequence related to the first sample resource; the historical short-term interaction sequence related to the first sample resource may include M third sample resources and M third interaction tags; the M third sample resources are pushed to the target object within a second historical time period starting from the interaction between the target object and the first sample resource, and there is a one-to-one correspondence between the M third interaction tags and the M third sample resources, where M is a positive integer less than or equal to N, the first historical time period is longer than the second historical time period, and the M third sample resources belong to the N second sample resources. The second preset time period can be set according to specific needs. For example, the second preset time period can be set to 1 month, then the M third sample resources are pushed to the target object within 1 month starting from the interaction between the target object and the first sample resource. For example, if the first and second interaction prediction networks are click-through rate prediction networks, and the first sample resource is resource B, the historical long-term interaction sequence related to the first sample resource (resource B) can include five second sample resources: resource E, resource F, resource G, resource H, and resource A, and five first interaction tags: 0, 1, 2, 1, and 0. If the resources pushed to the target object from morning to night during the second historical time period after the target object interacts with the first sample resource are resource G, resource H, and resource A, then the historical short-term interaction sequence related to the first sample resource (resource B) can include three third sample resources: resource G, resource H, and resource A, and three third interaction tags: 2, 1, and 0. If the first and second interaction prediction networks are conversion rate prediction models, then the historical short-term interaction sequence related to the first sample resource (resource B) can include two third sample resources: resource G and resource H, and two third interaction tags: 2 and 1.

[0043] It should be noted that in the specific implementation of this application, user-related data is involved, such as historical long-term interaction sequences and historical short-term interaction sequences related to the first sample resource. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with local laws, regulations and standards.

[0044] Based on the aforementioned resource push scheme and system, this application provides a training method for a second interaction prediction network, and describes the corresponding process of training the second interaction prediction network based on a first sample resource and related historical long-term interaction sequences. See also... Figure 2 This is a flowchart illustrating a training method for a second interactive prediction network provided in an embodiment of this application. Figure 2The training method for the second interactive prediction network shown can be executed by a resource push device or by any electronic device capable of training the second interactive prediction network. This application embodiment describes the method using a resource push device. Figure 2 The training method for the second interactive prediction network shown may include the following steps:

[0045] S201, obtain the push-related data corresponding to the first sample resource, the historical long-term interaction sequence related to the first sample resource, and the second interaction tag of the target object performing the target interaction on the first sample resource.

[0046] In one embodiment, the second interaction label is used to indicate whether the target object performs a target interaction with the first sample resource, that is, to indicate whether the interaction between the target object and the first sample resource is a target interaction; if the target object performs a target interaction with the first sample resource, the second interaction label can be "1", and if the target object does not perform a target interaction with the first sample resource, the second interaction label can be "0".

[0047] In one embodiment, the push-related data corresponding to the first sample resource may include any one or more of the following: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data when the target object interacts with the first sample resource. The resource attribute data of the first sample resource refers to data used to describe the first sample resource, such as the resource identifier, name, and resource type of the first sample resource. The object attribute data of the target object refers to data used to describe the target object, such as the user identification (UID), nickname, age, and city of residence of the target object. The environmental data when the target object interacts with the first sample resource indicates the environment in which the target object interacts with the first sample resource, such as the time, location, terminal device type, and network information type (e.g., wired network, wireless network) of the interaction. Furthermore, the resource push device obtains the push-related data corresponding to the first sample resource; the historical long-term interaction sequence and the second interaction tag related to the first sample resource may be obtained from the target object's terminal device.

[0048] S202, the second interaction prediction network is invoked to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, and to perform feature transformation processing on the push association data corresponding to the first sample resource to obtain the push association feature representation corresponding to the push association data.

[0049] In one embodiment, the second interaction prediction network is used to perform target interaction prediction; further, the second interaction prediction network can be any network that can perform target interaction prediction based on the push-related data corresponding to the first sample resource and the historical long-term interaction sequence related to the first sample resource. The embodiments of this application do not limit the specific network structure of the second interaction prediction network. As long as the network can perform target interaction prediction, it is within the protection scope of the embodiments of this application.

[0050] In one embodiment, the resource push device invokes a second interaction prediction network to perform feature transformation processing on a historical long-term interaction sequence, obtaining an interaction feature representation corresponding to each of the N second sample resources. This may include: invoking the second interaction prediction network to perform feature embedding processing on each second sample resource, obtaining an embedded feature representation corresponding to each second sample resource; and performing feature embedding processing on each of the N first interaction tags, obtaining a tag feature representation corresponding to each first interaction tag; and concatenating the embedded feature representation corresponding to each second sample resource with the tag feature representation corresponding to the corresponding first interaction tag to obtain an interaction feature representation corresponding to each second sample resource. The feature embedding processing on each second sample resource may be feature embedding processing on the resource attribute data of each second sample resource, or feature embedding processing on the resource identifier of each second sample resource. Feature embedding processing on the resource attribute data and other data of the second sample resources (i.e., embedding processing) can transform the vectorized data from high-dimensional sparse vectors to dense vectors, achieving dimensionality reduction of the vectorized data, saving processing resources, and reducing the data required by the second interaction prediction network in subsequent processing.

[0051] In one embodiment, the historical long-term interaction sequence related to the first sample resource may further include N first time tags, each corresponding one-to-one with one of the N second sample resources. The N first time tags indicate the time difference between the interaction between the target object and the N second sample resources and the interaction between the target object and the first sample resource. Taking any of the N second sample resources as an example, if the time point of interaction between the target object and the second sample resource is t_history, and the time point of interaction between the target object and the first sample resource is t, then the corresponding first time tag for the second sample resource can be t-t_history. Further, the interaction feature representation corresponding to each second sample resource can be obtained by concatenating the embedded feature representation corresponding to each second sample resource, the tag feature representation corresponding to each first interaction tag, and the time feature representation corresponding to each first time tag; wherein, the time feature representation corresponding to each first time tag is obtained by performing feature embedding processing on each of the N first time tags.

[0052] In one embodiment, if the push-related data corresponding to the first sample resource includes the resource attribute data of the first sample resource, then the push-related feature representation corresponding to the push-related data includes: a resource feature representation corresponding to the first sample resource obtained by feature embedding processing on the resource attribute data of the first sample resource; if the push-related data corresponding to the first sample resource includes the object attribute data of the target object, then the push-related feature representation corresponding to the push-related data includes: an object feature representation corresponding to the target object obtained by feature embedding processing on the object attribute data of the target object; if the push-related data corresponding to the first sample resource includes environmental data, then the push-related feature representation corresponding to the push-related data includes: an environmental feature representation corresponding to the environmental data obtained by feature embedding processing on the environmental data. The embodiments of this application will subsequently be described with reference to the push-related data corresponding to the first sample resource including: the resource attribute data of the first sample resource, the object attribute data of the target object, and the environmental data.

[0053] S203, based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, perform interaction prediction processing to obtain the first prediction result of the target object performing the target interaction on the first sample resource.

[0054] In one embodiment, when the push-related data corresponding to the first sample resource includes: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data, the resource push device calls the second interaction prediction network to perform interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push-related feature representation corresponding to the push-related data, respectively, to obtain a first prediction result of the target object performing a target interaction with the first sample resource. This may include: performing feature extraction processing on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data, respectively, to obtain a sequence feature extraction vector corresponding to the historical long-term interaction sequence, a resource feature extraction vector corresponding to the first sample resource, an object feature extraction vector corresponding to the target object, and an environmental feature extraction vector corresponding to the environmental data; performing feature interaction processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object to obtain a feature interaction sub-vector; concatenating the sequence feature extraction vector corresponding to the historical long-term interaction sequence, the environmental feature extraction vector corresponding to the environmental data, and the feature interaction sub-vector to obtain a feature interaction vector; and performing probability prediction processing on the feature interaction vector to obtain the first prediction result.

[0055] In specific implementation, when the resource push device calls the second interaction prediction network to perform feature extraction processing on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data, it can be based on different feature extraction networks in the second interaction prediction network that can achieve feature extraction functions. This application embodiment does not limit the specific selection of each feature extraction network. For example, feature extraction processing can be performed on the interaction feature representation corresponding to each second sample resource based on the encoder part of the transformer network in the neural network to obtain the interaction feature extraction vector corresponding to each second sample resource; then the interaction feature extraction vectors corresponding to each second sample resource are concatenated to obtain the sequence feature extraction vector corresponding to the historical long-term interaction sequence.

[0056] Furthermore, the resource push device invokes a second interaction prediction network to perform feature interaction processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object, obtaining a feature interaction sub-vector. This may include: performing vector fusion processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object, obtaining a vector fusion result; based on the vector fusion result, predicting the interaction probability distribution when the target object performs each of the various interactions with the first sample resource; and fusing the interaction probability distribution with the resource feature extraction vector corresponding to the first sample resource, obtaining a feature interaction sub-vector. The vector dot product operation can be used to achieve vector fusion processing between the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object, obtaining a vector fusion result. A normalization function can be used to normalize the vector fusion result, predicting the interaction probability distribution when the target object performs each of the various interactions with the first sample resource, wherein the normalization function can be an exponential normalization function (i.e., a softmax function). For example, if the various interactions are "exposure interaction," "click without conversion," and "click with conversion," then the softmax function can predict the probability that the target object will perform an "exposure interaction," the probability that it will perform a "click without conversion," and the probability that it will perform a "click with conversion" on the first sample resource. Furthermore, when the resource push device calls the second interaction prediction network to fuse the interaction probability distribution with the resource feature extraction vector corresponding to the first sample resource to obtain the feature interaction sub-vector, a preset fusion function can be used, such as the matmul function.

[0057] Furthermore, the resource push device calls the second interaction prediction network to perform probability prediction processing on the feature interaction vector. When obtaining the first prediction result, the feature interaction vector can be fully connected first, and the result after the full connection processing can be normalized through a normalization function to predict the first prediction result of the target object performing target interaction on the first sample resource. The normalization function can be the sigmoid function.

[0058] See Figure 3This is an exemplary structural diagram of a second interactive prediction network provided in an embodiment of this application. The second interactive prediction network includes a sequence element encoding module, a resource feature embedding layer, an object feature embedding layer, an environment feature embedding layer, a sequence feature extraction network (including the encoder part of a transformer network and a concat module), a resource feature extraction network, an object feature extraction network, an environment feature extraction network, a vector fusion layer (dot product layer), a function layer (including: softmax function layer, sigmoid function layer, matmul function layer), a concat module, and a fully connected layer (FC layer).

[0059] Specifically, the resource push device invokes the second interaction prediction network, which performs feature transformation processing on the historical long-term interaction sequence through the sequence element encoding module to obtain the interaction feature representation corresponding to each of the N second sample resources. Then, it performs feature embedding processing on the resource attribute data of the first sample resource, the object attribute data of the target object, and the environmental data through resource feature embedding layers, object feature embedding layers, and environmental feature embedding layers, respectively, to obtain the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data. Finally, it performs feature extraction processing on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data through sequence feature extraction networks, resource feature extraction vectors corresponding to the first sample resource, object feature extraction vectors corresponding to the target object, and environmental feature extraction vectors corresponding to the environmental data, respectively.

[0060] The vector fusion layer (dot product layer) performs vector fusion processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object to obtain the vector fusion result; the softmax layer predicts the interaction probability distribution of the target object to the first sample resource when performing various interactions based on the vector fusion result; the matmul function layer fuses the interaction probability distribution with the resource feature extraction vector corresponding to the first sample resource to obtain the feature interaction sub-vector.

[0061] The concat module concatenates the sequence feature extraction vector corresponding to the historical long-term interaction sequence, the environmental feature extraction vector corresponding to the environmental data, and the feature interaction sub-vector to obtain the feature interaction vector. Then, a fully connected layer is used to process the feature interaction vector. The result after the fully connected processing is then passed through a sigmoid function layer to predict the first prediction result of the target object performing the target interaction on the first sample resource.

[0062] S204, The second interaction prediction network is trained based on the difference between the second interaction label and the first prediction result.

[0063] In one embodiment, the resource push device trains a second interaction prediction network based on the difference between the second interaction label and the first prediction result. This can include: determining a target loss function corresponding to the second interaction prediction network based on the difference between the second interaction label and the first prediction result; updating the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push-related data according to the target loss function corresponding to the second interaction prediction network. Specifically, the resource push device can adjust the model parameters of the second interaction prediction network in the direction of convergence of the target loss function corresponding to the second interaction prediction network, and update the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push-related data according to the adjusted model parameters. For example, the model parameters of the second interaction prediction network, the interaction feature representation corresponding to each second sample resource, and the push association feature representation corresponding to the push-related data can be adjusted based on the backpropagation algorithm. Furthermore, the resource push device can train the second interaction prediction network offline.

[0064] In one embodiment, when multiple samples are used to train the second interaction prediction network, the target loss function corresponding to the second interaction prediction network can be the cross-entropy loss function; wherein, a sample includes push-related data corresponding to a first sample resource, a historical long-term interaction sequence related to the first sample resource, and a second interaction label of a target object performing a target interaction on the first sample resource; in this case, the target loss function corresponding to the second interaction prediction network can be specifically given by the following formula 1:

[0065]

[0066] Where I is the number of samples, and i is the independent variable of the sample; y i p is the second interactive label corresponding to the i-th sample out of I samples. i ′ represents the first prediction result obtained based on the i-th sample.

[0067] In one embodiment, to enhance the generalization ability of the second interactive prediction network and prevent it from overfitting, a regularization strategy can be used to design the target loss function corresponding to the second interactive prediction network. The specific target loss function of the second interactive prediction network can then be given by the following formula 2:

[0068]

[0069] Where θ represents the model parameters of the second interactive prediction network.

[0070] In this embodiment, the second interaction prediction network can be trained based on the push-related data corresponding to the first sample resource and the historical long-term interaction sequence related to the first sample resource. The historical long-term interaction sequence related to the first sample resource carries rich historical interaction features before the target object interacts with the first sample resource. This allows the second interaction prediction network to fully learn the historical interaction features of the target object, enabling the trained second interaction prediction network to predict the probability of the object interacting with the resource based on the rich historical interaction features of the object, and improving the accuracy of the prediction of the probability of the object interacting with the resource.

[0071] Based on the aforementioned resource push scheme, resource push system, and training method of the second interaction prediction network, this application provides a training method for a first interaction prediction network. It describes the process of training the first interaction prediction network based on the second interaction prediction network, first sample resources, and related historical short-term interaction sequences. This application embodiment assumes that the first and second interaction prediction networks have the same network structure. See also... Figure 4 This is a flowchart illustrating a training method for a first interactive prediction network provided in an embodiment of this application. Figure 4 The training method of the first interactive prediction network shown can be executed by a resource push device or by any electronic device capable of training the first interactive prediction network. In this application, the embodiment is described using a resource push device. Figure 4 The training method for the first interactive prediction network shown may include the following steps:

[0072] S401, obtain the interaction feature representation corresponding to each of the M third sample resources, and obtain the updated push association feature representation corresponding to the push association data.

[0073] The interaction feature representation corresponding to each third sample resource is determined from the updated interaction feature representation corresponding to each second sample resource. For example, if the historical long-term interaction sequence related to the first sample resource includes 5 second sample resources, namely: resource E, resource F, resource G, resource H, and resource A, the updated interaction feature representations corresponding to the second sample resources are respectively: updated interaction feature representation E, updated interaction feature representation F, updated interaction feature representation G, updated interaction feature representation H, and updated interaction feature representation A; if the historical short-term interaction sequence related to the first sample resource includes 3 third sample resources, namely: resource G, resource H, and resource A, then the interaction feature representations corresponding to the third sample resources are respectively: updated interaction feature representation G, updated interaction feature representation H, and updated interaction feature representation A.

[0074] Furthermore, the updated interaction feature representation and the updated push association feature representation corresponding to each second sample resource can be obtained through the second interaction prediction network during training or after training. If obtained through the second interaction prediction network after training, in another embodiment, the second interaction prediction network after training can also be invoked to process the push association data corresponding to the first sample resource and the historical long-term interaction sequence related to the first sample resource, predict the second prediction result of the target object performing the target interaction on the first sample resource, and use the push association feature representation corresponding to the push association data obtained through the second interaction prediction network after training as the input of the first interaction prediction network. In addition, the interaction feature representation corresponding to each third sample resource is determined from the interaction feature representation corresponding to each second sample resource obtained through the second interaction prediction network after training, and the interaction feature representation corresponding to each third sample resource is used as the input of the first interaction prediction network, so that the first interaction prediction network predicts the third prediction result of the target object performing the target interaction on the first sample resource.

[0075] S402, the trained second interaction prediction network and the first interaction prediction network are called respectively, and interaction prediction processing is performed based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data, respectively, to obtain the second prediction result and the third prediction result of the target object performing the target interaction on the first sample resource.

[0076] The resource push device calls the trained second interaction prediction network and the first interaction prediction network respectively, and performs interaction prediction processing based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data, respectively, to obtain the second prediction result and the third prediction result of the target object performing the target interaction on the first sample resource. This process is similar to the process described above where the resource push device trains the second interaction prediction network, and calls the second interaction prediction network to perform interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, respectively, to obtain the first prediction result of the target object performing the target interaction on the first sample resource. Therefore, it will not be described in detail here.

[0077] S403, the first interaction prediction network is trained based on the difference between the second prediction result and the third prediction result, as well as the difference between the second interaction label and the third prediction result.

[0078] In one embodiment, see Figure 5 This document illustrates a method for training a first interactive prediction model based on a second prediction result, a third prediction result, and a second interactive tag, as provided in an embodiment of this application. The resource push device trains the first interactive prediction network based on the differences between the second and third prediction results, and the differences between the second interactive tag and the third prediction result. This training may include: determining a first loss function corresponding to the first interactive prediction network based on the differences between the second and third prediction results; determining a second loss function corresponding to the first interactive prediction network based on the differences between the second interactive tag and the third prediction result; determining a target loss function corresponding to the first interactive prediction network based on the first and second loss functions; and adjusting the model parameters of the first interactive prediction network in the direction of convergence of the target loss function. Optionally, the model parameters of the first interactive prediction network may be adjusted based on a backpropagation algorithm.

[0079] In one embodiment, when the first interaction prediction network is trained using multiple samples, the first loss function corresponding to the first interaction prediction network can be a cross-entropy loss function or a relative entropy loss function (KL divergence) to determine the difference between the second prediction result and the third prediction result; the second loss function corresponding to the first interaction prediction network can be a cross-entropy loss function; wherein, a sample includes push-related data corresponding to a first sample resource, a historical short-term interaction sequence related to the first sample resource, and a second interaction tag of a target object performing a target interaction on the first sample resource.

[0080] The first loss function corresponding to the first interactive prediction network can be given by the following formula 3.1:

[0081]

[0082] Where I is the number of samples, and i is the independent variable of the sample; q i p represents the second prediction result corresponding to the i-th sample out of I samples. i This is the third prediction result obtained based on the i-th sample.

[0083] The second loss function corresponding to the first interactive prediction network can be given by the following formula 3.2:

[0084]

[0085] Among them, y i Let be the second interactive label corresponding to the i-th sample out of I samples.

[0086] The target loss function for the first interactive prediction network can be given by the following formula 3.3:

[0087] L=αL1+L2 (3.3)

[0088] Where α is a hyperparameter indicating the weight of the influence of the second prediction result on the first interactive prediction network; α can be set according to specific training requirements, for example, α can be set to 0.5.

[0089] Optionally, the first interactive prediction network can be trained using fewer samples than when the second interactive prediction network was trained, thus saving training time. For example, a portion of the first sample resources pushed to the target object within a preset time period can be selected as the first sample resources for training the first interactive prediction network; or, a preset portion of the first sample resources pushed to the target object within a preset time period can be used as the first sample resources for training the first interactive prediction network; wherein, the preset portion of the time period can be set according to specific needs, for example, it can be set to one week.

[0090] In this embodiment, the first interaction prediction network can be trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequences related to the first sample resource. Specifically, the first interaction prediction network can be trained based on the updated interaction feature representation corresponding to each second sample resource and the updated push association feature representation corresponding to the push association data, obtained by processing the historical long-term interaction sequences related to the first sample resource and the push association data corresponding to the first sample resource, obtained through the second interaction prediction network during training or after training. This can improve the generalization of the first interaction prediction network. Furthermore, the second prediction result of the target object performing the target interaction on the first sample resource, predicted by the second interaction prediction network, can be compared with the target object performing the target interaction on the first sample resource predicted by the first interaction prediction network. The differences between the third prediction results of the target interaction and the differences between the third prediction results and the second interaction label used to indicate whether the target object performs the target interaction with the first sample resource are used to train the first interaction prediction network. This allows the first interaction prediction network to fully reference the prediction results of the second interaction prediction network during training, effectively learn the processing capabilities of the second interaction prediction network, accelerate the training rate of the first interaction prediction network, shorten the training time of the first interaction prediction network, and improve the prediction accuracy of the first interaction prediction network in predicting the probability that the target object will perform the target interaction with the resource. Furthermore, the amount of data in historical short-term interaction sequences is less than the amount of data in historical long-term interaction sequences, and the network size of the first interaction prediction network is smaller than that of the second interaction prediction network, enabling the first interaction prediction network to improve the prediction rate of the probability that the target object will perform the target interaction with the resource.

[0091] Based on the above embodiments, this application provides a resource push method; see also Figure 6 This is a flowchart illustrating a resource push method provided in an embodiment of this application. Figure 6 The resource push method shown can be executed by the resource push device. Figure 6 The resource push method shown may include the following steps:

[0092] S601, in response to a push request to push resources to the object to be pushed, the first interaction prediction network is invoked to predict the interaction probability of the object to be pushed performing the target interaction on the resource to be pushed.

[0093] The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequences related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequences related to the first sample resource; the historical long-term interaction sequences include historical short-term interaction sequences; both the first and second interaction prediction networks are used to predict target interactions, and the network size of the first interaction prediction network is smaller than that of the second interaction prediction network.

[0094] In one embodiment, in response to a push request to push resources to a target object, the resource push device invokes a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the resource to be pushed. This may include: in response to the push request to push resources to the target object, obtaining push-related data corresponding to the resource to be pushed and a historical short-term interaction sequence related to the target object; wherein, the historical short-term interaction sequence related to the target object is pushed to the target object within a second historical time period starting from the current time point; invoking the first interaction prediction network to predict the interaction probability based on the push-related data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the target object. The current time point is the time when the resource push device responds to the push request to push resources to the target object; the push-related data corresponding to the resource to be pushed may include any one or more of the following data: resource attribute data of the resource to be pushed, object attribute data of the target object, and environmental data at the current time point. The process by which the resource push device calls the first interaction prediction network to predict the interaction probability based on the push-related data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed is similar to the process by which the second interaction prediction network is called to predict the first prediction result of the target object performing the target interaction on the first sample resource based on the push-related data corresponding to the first sample resource and the historical long-term interaction sequence related to the first sample resource. It will not be described in detail here.

[0095] S602, determine the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability.

[0096] In one embodiment, the resource push device determines the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability. This may include: determining the interaction probability as the matching degree between the object to be pushed and the resource to be pushed; or, based on the conversion rule between interaction probability and matching degree, performing matching degree conversion processing on the interaction probability to obtain the matching degree between the object to be pushed and the resource to be pushed. Optionally, the benefit generated by pushing the resource to be pushed to the object to be pushed can be determined as the matching degree between the object to be pushed and the resource to be pushed. In this case, the benefit generated by pushing the resource to be pushed to the object to be pushed can be given by the following formula 4:

[0097] eCPM=bid×pCTR×pCVR (4)

[0098] Here, eCPM represents the revenue generated by pushing the resource to the target audience, bid is the bid for the resource, pCTR is the estimated click-through rate of the target audience clicking on the resource, and pCVR is the estimated conversion rate of the target audience performing conversion operations on the resource.

[0099] S603: If the matching degree between the object to be pushed and the resource to be pushed meets the push rules, then push the resource to be pushed to the object to be pushed.

[0100] In one embodiment, the push rule can instruct that resources with a matching degree greater than a matching degree threshold be pushed to the target object; wherein, the matching degree threshold can be set according to specific needs. For example, if the matching degree between the target object and the resource is determined as the interaction probability of the target object performing a target interaction on the resource, then the resource can be pushed to the target object when the interaction probability is greater than a certain probability threshold, for example, when the interaction probability is greater than 0.8.

[0101] In one embodiment, a target resource to be pushed can be determined from a plurality of resources based on the matching degree between the object to be pushed and each of the plurality of resources to be pushed, and the target resource to be pushed can be pushed to the object. Optionally, a preset number of target resources to be pushed can be determined in descending order of the matching degree between the object to be pushed and each of the plurality of resources to be pushed. For example, three target resources to be pushed can be determined in descending order of the matching degree between the object to be pushed and each of the resources to be pushed. That is, the top three resources to be pushed when sorted by matching degree from largest to smallest can be determined as target resources to be pushed.

[0102] In this embodiment, in response to a push request to push resources to a target object, a first interaction prediction network is invoked to predict the interaction probability of the target object performing a target interaction on the target resource; and the matching degree between the target object and the target resource is determined based on the interaction probability; if the matching degree between the target object and the target resource meets the push rules, the target resource is pushed to the target object; wherein, the first interaction prediction network is trained based on a second interaction prediction network, a first sample resource, and historical short-term interaction sequences related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and historical long-term interaction sequences related to the first sample resource; the historical long-term interaction sequences include historical short-term interaction sequences; both the first and second interaction prediction networks are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network. Because long-term historical interaction sequences carry richer historical interaction features than short-term historical interaction sequences, the second interaction prediction network trained based on the long-term historical interaction sequences related to the first sample resource can predict the probability of an object performing a target interaction with the resource based on these rich historical interaction features, and the prediction accuracy is high. However, due to the large amount of data in the long-term historical interaction sequences, the second interaction prediction network, which can process such large amounts of data, has a large network size. Generally speaking, large-scale models take a long time to process large amounts of data, which does not meet the timeliness requirements in the resource push domain. Therefore, a smaller-scale first interaction prediction network can be trained using short-term historical interaction sequences, which have less data than long-term historical interaction sequences, to ensure the first interaction prediction network's performance. The lightweight nature of the first interaction prediction network improves the prediction rate of the probability that an object will perform a target interaction with a resource, thus ensuring the timeliness of resource push. Furthermore, by training the first interaction prediction network on the second interaction prediction network, the first interaction prediction network can effectively learn the processing capabilities of the second interaction prediction network. Therefore, when the first interaction prediction network is trained using historical short-term interaction sequences with a small amount of data, the prediction accuracy of the probability of an object performing a target interaction with a resource can be guaranteed. Based on the above description, it can be seen that when the first interaction prediction network is used to predict the interaction probability of an object to be pushed performing a target interaction with a resource to be pushed, both prediction accuracy and prediction rate can be taken into account, thereby ensuring both the accuracy and timeliness of resource push.

[0103] Based on the embodiments related to the above resource push method, this application provides a resource push device. See also... Figure 7 This is a schematic diagram of a resource push device provided in an embodiment of this application. The resource push device may include an acquisition unit 701, a processing unit 702, and an output unit 703. Figure 7The resource push device shown can operate the following units:

[0104] Processing unit 702 is configured to, in response to a push request to push resources to a target object, invoke a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource.

[0105] The processing unit 702 is further configured to determine the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability;

[0106] The output unit 703 is used to push the resource to be pushed to the object if the matching degree between the object to be pushed and the resource to be pushed meets the push rules.

[0107] The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network.

[0108] In one embodiment, the historical long-term interaction sequence includes N second sample resources and N first interaction tags; the N second sample resources are pushed to the target object within a first historical time period from the time the target object interacts with the first sample resources, and the N first interaction tags correspond one-to-one with the N second sample resources. The N first interaction tags are used to indicate the interaction between the target object and the N second sample resources, where N is a positive integer.

[0109] When the acquisition unit 701 and the processing unit 702 train the second interaction prediction network based on the first sample resource and the historical long-term interaction sequence related to the first sample resource, they specifically perform the following operations:

[0110] The acquisition unit 701 is used to acquire push-related data corresponding to the first sample resource, historical long-term interaction sequence related to the first sample resource, and second interaction tag of the target object performing target interaction on the first sample resource.

[0111] The processing unit 702 is used to call the second interaction prediction network to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, and to perform feature transformation processing on the push association data corresponding to the first sample resource to obtain the push association feature representation corresponding to the push association data.

[0112] The processing unit 702 is further configured to perform interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, respectively, to obtain a first prediction result of the target object performing a target interaction on the first sample resource;

[0113] The processing unit 702 is further configured to train the second interaction prediction network based on the difference between the second interaction label and the first prediction result.

[0114] In one embodiment, when the processing unit 702 invokes the second interaction prediction network to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, the following operations are specifically performed:

[0115] The second interaction prediction network is invoked to perform feature embedding processing on each of the second sample resources to obtain the embedded feature representation corresponding to each of the second sample resources; and feature embedding processing is performed on each of the N first interaction labels to obtain the label feature representation corresponding to each of the first interaction labels;

[0116] The embedded feature representation corresponding to each second sample resource is concatenated with the tag feature representation corresponding to the corresponding first interaction tag to obtain the interaction feature representation corresponding to each second sample resource.

[0117] In one embodiment, the push-related data corresponding to the first sample resource includes any one or more of the following data: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data when the target object interacts with the first sample resource;

[0118] If the push association data corresponding to the first sample resource includes the resource attribute data of the first sample resource, then the push association feature representation corresponding to the push association data includes: the resource feature representation corresponding to the first sample resource obtained by performing feature embedding processing on the resource attribute data of the first sample resource;

[0119] If the push-related data corresponding to the first sample resource includes the object attribute data of the target object, then the push-related feature representation corresponding to the push-related data includes: the object feature representation corresponding to the target object obtained by performing feature embedding processing on the object attribute data of the target object;

[0120] If the push-related data corresponding to the first sample resource includes the environmental data, then the push-related feature representation corresponding to the push-related data includes: the environmental feature representation corresponding to the environmental data obtained by performing feature embedding processing on the environmental data.

[0121] In one embodiment, the push-related data corresponding to the first sample resource includes: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data;

[0122] When the processing unit 702 performs interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, and obtains the first prediction result of the target object performing the target interaction on the first sample resource, it specifically performs the following operations:

[0123] Feature extraction processing is performed on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data, respectively, to obtain the sequence feature extraction vector corresponding to the historical long-term interaction sequence, the resource feature extraction vector corresponding to the first sample resource, the object feature extraction vector corresponding to the target object, and the environmental feature extraction vector corresponding to the environmental data.

[0124] The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to feature interaction processing to obtain a feature interaction sub-vector;

[0125] The sequence feature extraction vector corresponding to the historical long-term interaction sequence, the environmental feature extraction vector corresponding to the environmental data, and the feature interaction sub-vector are concatenated to obtain the feature interaction vector.

[0126] The feature interaction vector is subjected to probability prediction processing to obtain the first prediction result.

[0127] In one embodiment, the target interaction is one or more of a plurality of interactions;

[0128] When the processing unit 702 performs feature interaction processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object to obtain the feature interaction sub-vector, it specifically performs the following operations:

[0129] The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to vector fusion processing to obtain the vector fusion result.

[0130] Based on the vector fusion result, predict the interaction probability distribution of the target object when performing each of the multiple interactions on the first sample resource;

[0131] The interaction probability distribution is fused with the resource feature extraction vector corresponding to the first sample resource to obtain the feature interaction sub-vector.

[0132] In one embodiment, when the processing unit 702 trains the second interaction prediction network based on the difference between the second interaction label and the first prediction result, it specifically performs the following operations:

[0133] Based on the difference between the second interactive label and the first prediction result, the target loss function corresponding to the second interactive prediction network is determined;

[0134] The interaction feature representation corresponding to each second sample resource is updated according to the target loss function corresponding to the second interaction prediction network, and the push association feature representation corresponding to the push association data is also updated.

[0135] In one embodiment, the historical short-term interaction sequence includes M third sample resources and M third interaction tags; the M third sample resources are pushed to the target object within a second historical time period starting from the interaction between the target object and the first sample resource, the M third interaction tags correspond one-to-one with the M third sample resources, M is a positive integer less than or equal to N, the first historical time period is longer than the second historical time period, and the M third sample resources belong to the N second sample resources;

[0136] When the acquisition unit 701 and the processing unit 702 train the first interaction prediction network based on the second interaction prediction network, the first sample resource, and the historical short-time interaction sequence related to the first sample resource, they specifically perform the following operations:

[0137] The acquisition unit 701 is used to acquire the interaction feature representation corresponding to each of the M third sample resources, wherein the interaction feature representation corresponding to each third sample resource is determined from the updated interaction feature representation corresponding to each second sample resource, and to acquire the updated push association feature representation corresponding to the push association data.

[0138] The processing unit 702 is used to call the trained second interaction prediction network and the first interaction prediction network respectively, and perform interaction prediction processing based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data respectively, to obtain the second prediction result and the third prediction result of the target object performing target interaction on the first sample resource.

[0139] The processing unit 702 is further configured to train the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result.

[0140] In one embodiment, when the processing unit 702 trains the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result, it specifically performs the following operations:

[0141] Based on the difference between the second prediction result and the third prediction result, a first loss function corresponding to the first interactive prediction network is determined;

[0142] Based on the difference between the second interactive label and the third prediction result, a second loss function corresponding to the first interactive prediction network is determined.

[0143] Based on the first loss function and the second loss function, the target loss function corresponding to the first interactive prediction network is determined; and the model parameters of the first interactive prediction network are adjusted in the direction of convergence of the target loss function corresponding to the first interactive prediction network.

[0144] In one embodiment, when the processing unit 702 responds to a push request to push resources to a target object and calls the first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource, it specifically performs the following operations:

[0145] In response to a push request to push resources to the object to be pushed, the push association data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed are obtained; the historical short-term interaction sequence related to the object to be pushed is pushed to the object within the second historical time period starting from the current time point;

[0146] The first interaction prediction network is invoked to predict the interaction probability based on the push-related data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed.

[0147] In one embodiment, when the processing unit 702 determines the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability, it specifically performs the following operations:

[0148] The interaction probability is determined as the matching degree between the object to be pushed and the resource to be pushed;

[0149] or,

[0150] Based on the conversion rules between interaction probability and matching degree, the interaction probability is converted into matching degree to obtain the matching degree between the object to be pushed and the resource to be pushed.

[0151] According to one embodiment of this application, Figure 2 , Figure 4 as well as Figure 6 The resource push method shown can involve various steps that can be achieved by: Figure 7 This is performed by the various units within the resource delivery device shown. For example, Figure 2 The step S201 shown can be performed by Figure 7 The acquisition unit 701 in the resource push device shown performs this action. Figure 2 Steps S202 to S204 shown can be derived from Figure 7 The processing unit 702 in the resource push device shown executes this. For example, Figure 4 The step S401 shown can be performed by Figure 7 The acquisition unit 701 in the resource push device shown performs this action. Figure 4 Steps S402 to S403 shown can be derived from... Figure 7 The processing unit 702 in the resource push device shown executes this. For example, Figure 6 Steps S601 to S602 shown can be derived from... Figure 7 The processing unit 702 in the resource push device shown executes this. Figure 6 The step S603 shown can be performed by Figure 7 The output unit 703 in the resource push device shown is used to perform this operation.

[0152] According to another embodiment of this application, Figure 7 The resource push device shown can be composed of individual or combined units into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the resource push device based on logical function division may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0153] According to another embodiment of this application, the following can be achieved by running on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), a device capable of performing operations such as... Figure 2 , Figure 4 as well as Figure 6 The computer program (including program code) for each step involved in the corresponding method shown, to construct such... Figure 7 The resource push device shown herein, and the resource push method for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable storage medium, loaded onto the aforementioned computing device via the computer-readable storage medium, and run therein.

[0154] In this embodiment, in response to a push request to push resources to a target object, a first interaction prediction network is invoked to predict the interaction probability of the target object performing a target interaction on the target resource; and the matching degree between the target object and the target resource is determined based on the interaction probability; if the matching degree between the target object and the target resource meets the push rules, the target resource is pushed to the target object; wherein, the first interaction prediction network is trained based on a second interaction prediction network, a first sample resource, and historical short-term interaction sequences related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and historical long-term interaction sequences related to the first sample resource; the historical long-term interaction sequences include historical short-term interaction sequences; both the first and second interaction prediction networks are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network. Because long-term historical interaction sequences carry richer historical interaction features than short-term historical interaction sequences, the second interaction prediction network trained based on the long-term historical interaction sequences related to the first sample resource can predict the probability of an object performing a target interaction with the resource based on these rich historical interaction features, and the prediction accuracy is high. However, due to the large amount of data in the long-term historical interaction sequences, the second interaction prediction network, which can process such large amounts of data, has a large network size. Generally speaking, large-scale models take a long time to process large amounts of data, which does not meet the timeliness requirements in the resource push domain. Therefore, a smaller-scale first interaction prediction network can be trained using short-term historical interaction sequences, which have less data than long-term historical interaction sequences, to ensure the first interaction prediction network's performance. The lightweight nature of the first interaction prediction network improves the prediction rate of the probability that an object will perform a target interaction with a resource, thus ensuring the timeliness of resource push. Furthermore, by training the first interaction prediction network on the second interaction prediction network, the first interaction prediction network can effectively learn the processing capabilities of the second interaction prediction network. Therefore, when the first interaction prediction network is trained using historical short-term interaction sequences with a small amount of data, the prediction accuracy of the probability of an object performing a target interaction with a resource can be guaranteed. Based on the above description, it can be seen that when the first interaction prediction network is used to predict the interaction probability of an object to be pushed performing a target interaction with a resource to be pushed, both prediction accuracy and prediction rate can be taken into account, thereby ensuring both the accuracy and timeliness of resource push.

[0155] Based on the aforementioned embodiments of the resource push method and the resource push device, this application also provides a resource push device. See also... Figure 8 This is a schematic diagram of the structure of a resource push device provided in an embodiment of this application. Figure 8The resource push device shown may include at least a processor 801, an input interface 802, an output interface 803, and a computer storage medium 804. The processor 801, input interface 802, output interface 803, and computer storage medium 804 may be connected via a bus or other means.

[0156] The computer storage medium 804 can be stored in the memory of the resource push device. The computer storage medium 804 is used to store computer programs, which include program instructions. The processor 801 is used to execute the program instructions stored in the computer storage medium 804. The processor 801 (or CPU (Central Processing Unit)) is the computing and control core of the resource push device. It is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions to realize the above-mentioned resource push method flow or corresponding functions.

[0157] This application embodiment also provides a computer storage medium (Memory), which is a memory device in a resource push device used to store programs and data. It is understood that the computer storage medium here can include the built-in storage medium in the terminal, or it can include an extended storage medium supported by the terminal. The computer storage medium provides storage space, which stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 801. These instructions can be one or more computer programs (including program code). It should be noted that the computer storage medium here can be a high-speed random access memory (RAM), or it can be non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer storage medium located remotely from the aforementioned processor.

[0158] In one embodiment, the processor 801, input interface 802, and output interface 803 may load and execute one or more instructions stored in the computer storage medium to achieve the aforementioned related... Figure 2 , Figure 4 as well as Figure 6 In the resource push method embodiment, the corresponding steps of the method are as follows: In a specific implementation, one or more instructions in the computer storage medium are loaded and executed by the processor 801 and the output interface 803:

[0159] The processor 801 is configured to, in response to a push request to push resources to a target object, invoke a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource.

[0160] The processor 801 is further configured to determine the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability;

[0161] The output interface 803 is used to push the resource to be pushed to the object if the matching degree between the object to be pushed and the resource to be pushed meets the push rules.

[0162] The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network.

[0163] In one embodiment, the historical long-term interaction sequence includes N second sample resources and N first interaction tags; the N second sample resources are pushed to the target object within a first historical time period from the time the target object interacts with the first sample resources, and the N first interaction tags correspond one-to-one with the N second sample resources. The N first interaction tags are used to indicate the interaction between the target object and the N second sample resources, where N is a positive integer.

[0164] When the processor 801 and the input interface 802 train the second interaction prediction network based on the first sample resource and the historical long-term interaction sequence related to the first sample resource, the following operations are specifically performed:

[0165] The input interface 802 is used to obtain push-related data corresponding to the first sample resource, the historical long-term interaction sequence related to the first sample resource, and the second interaction tag of the target object performing target interaction on the first sample resource.

[0166] The processor 801 is used to call the second interaction prediction network to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, and to perform feature transformation processing on the push association data corresponding to the first sample resource to obtain the push association feature representation corresponding to the push association data.

[0167] The processor 801 is further configured to perform interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, respectively, to obtain a first prediction result of the target object performing a target interaction on the first sample resource;

[0168] The processor 801 is further configured to train the second interaction prediction network based on the difference between the second interaction label and the first prediction result.

[0169] In one embodiment, when the processor 801 invokes the second interaction prediction network to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, the following operations are specifically performed:

[0170] The second interaction prediction network is invoked to perform feature embedding processing on each of the second sample resources to obtain the embedded feature representation corresponding to each of the second sample resources; and feature embedding processing is performed on each of the N first interaction labels to obtain the label feature representation corresponding to each of the first interaction labels;

[0171] The embedded feature representation corresponding to each second sample resource is concatenated with the tag feature representation corresponding to the corresponding first interaction tag to obtain the interaction feature representation corresponding to each second sample resource.

[0172] In one embodiment, the push-related data corresponding to the first sample resource includes any one or more of the following data: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data when the target object interacts with the first sample resource;

[0173] If the push association data corresponding to the first sample resource includes the resource attribute data of the first sample resource, then the push association feature representation corresponding to the push association data includes: the resource feature representation corresponding to the first sample resource obtained by performing feature embedding processing on the resource attribute data of the first sample resource;

[0174] If the push-related data corresponding to the first sample resource includes the object attribute data of the target object, then the push-related feature representation corresponding to the push-related data includes: the object feature representation corresponding to the target object obtained by performing feature embedding processing on the object attribute data of the target object;

[0175] If the push-related data corresponding to the first sample resource includes the environmental data, then the push-related feature representation corresponding to the push-related data includes: the environmental feature representation corresponding to the environmental data obtained by performing feature embedding processing on the environmental data.

[0176] In one embodiment, the push-related data corresponding to the first sample resource includes: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data;

[0177] When the processor 801 performs interaction prediction processing based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, and obtains the first prediction result of the target object performing the target interaction on the first sample resource, it specifically performs the following operations:

[0178] Feature extraction processing is performed on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data, respectively, to obtain the sequence feature extraction vector corresponding to the historical long-term interaction sequence, the resource feature extraction vector corresponding to the first sample resource, the object feature extraction vector corresponding to the target object, and the environmental feature extraction vector corresponding to the environmental data.

[0179] The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to feature interaction processing to obtain a feature interaction sub-vector;

[0180] The sequence feature extraction vector corresponding to the historical long-term interaction sequence, the environmental feature extraction vector corresponding to the environmental data, and the feature interaction sub-vector are concatenated to obtain the feature interaction vector.

[0181] The feature interaction vector is subjected to probability prediction processing to obtain the first prediction result.

[0182] In one embodiment, the target interaction is one or more of a plurality of interactions;

[0183] When the processor 801 performs feature interaction processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object to obtain the feature interaction sub-vector, it specifically performs the following operations:

[0184] The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to vector fusion processing to obtain the vector fusion result.

[0185] Based on the vector fusion result, predict the interaction probability distribution of the target object when performing each of the multiple interactions on the first sample resource;

[0186] The interaction probability distribution is fused with the resource feature extraction vector corresponding to the first sample resource to obtain the feature interaction sub-vector.

[0187] In one embodiment, when the processor 801 trains the second interaction prediction network based on the difference between the second interaction label and the first prediction result, it specifically performs the following operations:

[0188] Based on the difference between the second interactive label and the first prediction result, the target loss function corresponding to the second interactive prediction network is determined;

[0189] The interaction feature representation corresponding to each second sample resource is updated according to the target loss function corresponding to the second interaction prediction network, and the push association feature representation corresponding to the push association data is also updated.

[0190] In one embodiment, the historical short-term interaction sequence includes M third sample resources and M third interaction tags; the M third sample resources are pushed to the target object within a second historical time period starting from the interaction between the target object and the first sample resource, the M third interaction tags correspond one-to-one with the M third sample resources, M is a positive integer less than or equal to N, the first historical time period is longer than the second historical time period, and the M third sample resources belong to the N second sample resources;

[0191] When the processor 801 and the input interface 802 train the first interaction prediction network based on the second interaction prediction network, the first sample resource, and the historical short-time interaction sequence related to the first sample resource, the following operations are specifically performed:

[0192] The input interface 802 is used to obtain the interaction feature representation corresponding to each of the M third sample resources, wherein the interaction feature representation corresponding to each third sample resource is determined from the updated interaction feature representation corresponding to each second sample resource, and to obtain the updated push association feature representation corresponding to the push association data.

[0193] The processor 801 is used to call the trained second interaction prediction network and the first interaction prediction network respectively, and perform interaction prediction processing based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data respectively, to obtain the second prediction result and the third prediction result of the target object performing target interaction on the first sample resource.

[0194] The processor 801 is further configured to train the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result.

[0195] In one embodiment, when the processor 801 trains the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result, it specifically performs the following operations:

[0196] Based on the difference between the second prediction result and the third prediction result, a first loss function corresponding to the first interactive prediction network is determined;

[0197] Based on the difference between the second interactive label and the third prediction result, a second loss function corresponding to the first interactive prediction network is determined.

[0198] Based on the first loss function and the second loss function, the target loss function corresponding to the first interactive prediction network is determined; and the model parameters of the first interactive prediction network are adjusted in the direction of convergence of the target loss function corresponding to the first interactive prediction network.

[0199] In one embodiment, when the processor 801 responds to a push request to push resources to a target object and calls the first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource, it specifically performs the following operations:

[0200] In response to a push request to push resources to the object to be pushed, the push association data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed are obtained; the historical short-term interaction sequence related to the object to be pushed is pushed to the object within the second historical time period starting from the current time point;

[0201] The first interaction prediction network is invoked to predict the interaction probability based on the push-related data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed.

[0202] In one embodiment, when the processor 801 determines the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability, it specifically performs the following operations:

[0203] The interaction probability is determined as the matching degree between the object to be pushed and the resource to be pushed;

[0204] or,

[0205] Based on the conversion rules between interaction probability and matching degree, the interaction probability is converted into matching degree to obtain the matching degree between the object to be pushed and the resource to be pushed.

[0206] This application provides a computer program product or computer program, the computer program product including a computer program stored in a computer storage medium; the processor of the resource push device reads the computer program from the computer storage medium, and the processor executes the computer program, causing the resource push device to perform the above-described actions. Figure 2 , Figure 4 as well as Figure 6 The method embodiment shown. The computer-readable storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0207] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A resource push method, characterized in that, include: In response to a push request to push resources to a target object, the first interaction prediction network is invoked to predict the interaction probability of the target object performing the target interaction on the push resource; The matching degree between the object to be pushed and the resource to be pushed is determined based on the interaction probability. If the matching degree between the object to be pushed and the resource to be pushed meets the push rules, then the resource to be pushed is pushed to the object to be pushed. The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network. The historical long-term interaction sequence includes N second sample resources, which are pushed to the target object within a first historical time period starting from the interaction between the target object and the first sample resource, where N is a positive integer; the historical short-term interaction sequence includes M third sample resources and M third interaction tags; the M third sample resources are pushed to the target object within a second historical time period starting from the interaction between the target object and the first sample resource, and the M third interaction tags correspond one-to-one with the M third sample resources, where M is a positive integer less than or equal to N, the first historical time period is longer than the second historical time period, and the M third sample resources belong to the N second sample resources; The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and historical short-term interaction sequences related to the first sample resource, including: The interaction feature representation corresponding to each of the M third sample resources is obtained, wherein the interaction feature representation corresponding to each third sample resource is determined from the updated interaction feature representation corresponding to each second sample resource, and the updated push association feature representation corresponding to the push association data corresponding to the first sample resource is obtained. The trained second interaction prediction network and the first interaction prediction network are invoked respectively, and interaction prediction processing is performed based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data, respectively, to obtain the second prediction result and the third prediction result of the target object performing the target interaction on the first sample resource; Based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result, the first interaction prediction network is trained. The second interaction label is used to indicate whether the target object performs the target interaction on the first sample resource.

2. The method as described in claim 1, characterized in that, The historical long-term interaction sequence also includes N first interaction tags, each of which corresponds one-to-one with the N second sample resources. The N first interaction tags are used to indicate the interaction between the target object and the N second sample resources. The second interaction prediction network is trained based on the first sample resource and the related historical long-term interaction sequences, including: Obtain the push-related data corresponding to the first sample resource, the historical long-term interaction sequence related to the first sample resource, and the second interaction tag of the target object performing the target interaction on the first sample resource; The second interaction prediction network is invoked to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources, and to perform feature transformation processing on the push association data corresponding to the first sample resource to obtain the push association feature representation corresponding to the push association data. Interaction prediction processing is performed based on the interaction feature representation corresponding to each second sample resource and the push association feature representation corresponding to the push association data, respectively, to obtain a first prediction result of the target object performing a target interaction on the first sample resource; The second interaction prediction network is trained based on the difference between the second interaction label and the first prediction result.

3. The method as described in claim 2, characterized in that, The step of invoking the second interaction prediction network to perform feature transformation processing on the historical long-term interaction sequence to obtain the interaction feature representation corresponding to each of the N second sample resources includes: The second interaction prediction network is invoked to perform feature embedding processing on each of the second sample resources to obtain the embedded feature representation corresponding to each of the second sample resources; and feature embedding processing is performed on each of the N first interaction labels to obtain the label feature representation corresponding to each of the first interaction labels; The embedded feature representation corresponding to each second sample resource is concatenated with the tag feature representation corresponding to the corresponding first interaction tag to obtain the interaction feature representation corresponding to each second sample resource.

4. The method as described in claim 2, characterized in that, The push-related data corresponding to the first sample resource includes any one or more of the following data: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data when the target object interacts with the first sample resource; If the push association data corresponding to the first sample resource includes the resource attribute data of the first sample resource, then the push association feature representation corresponding to the push association data includes: the resource feature representation corresponding to the first sample resource obtained by performing feature embedding processing on the resource attribute data of the first sample resource; If the push-related data corresponding to the first sample resource includes the object attribute data of the target object, then the push-related feature representation corresponding to the push-related data includes: the object feature representation corresponding to the target object obtained by performing feature embedding processing on the object attribute data of the target object; If the push-related data corresponding to the first sample resource includes the environmental data, then the push-related feature representation corresponding to the push-related data includes: the environmental feature representation corresponding to the environmental data obtained by performing feature embedding processing on the environmental data.

5. The method as described in claim 4, characterized in that, The push-related data corresponding to the first sample resource includes: resource attribute data of the first sample resource, object attribute data of the target object, and environmental data; The step of performing interaction prediction processing based on the interaction feature representation corresponding to each of the second sample resources and the push association feature representation corresponding to the push association data, respectively, to obtain a first prediction result of the target object performing a target interaction on the first sample resource, includes: Feature extraction processing is performed on the interaction feature representation corresponding to each second sample resource, the resource feature representation corresponding to the first sample resource, the object feature representation corresponding to the target object, and the environmental feature representation corresponding to the environmental data, respectively, to obtain the sequence feature extraction vector corresponding to the historical long-term interaction sequence, the resource feature extraction vector corresponding to the first sample resource, the object feature extraction vector corresponding to the target object, and the environmental feature extraction vector corresponding to the environmental data. The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to feature interaction processing to obtain a feature interaction sub-vector; The sequence feature extraction vector corresponding to the historical long-term interaction sequence, the environmental feature extraction vector corresponding to the environmental data, and the feature interaction sub-vector are concatenated to obtain the feature interaction vector. The feature interaction vector is subjected to probability prediction processing to obtain the first prediction result.

6. The method as described in claim 5, characterized in that, The target interaction can be one or more of a variety of interactions; The step of performing feature interaction processing on the resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object to obtain a feature interaction sub-vector includes: The resource feature extraction vector corresponding to the first sample resource and the object feature extraction vector corresponding to the target object are subjected to vector fusion processing to obtain the vector fusion result. Based on the vector fusion result, predict the interaction probability distribution of the target object when performing each of the multiple interactions on the first sample resource; The interaction probability distribution is fused with the resource feature extraction vector corresponding to the first sample resource to obtain the feature interaction sub-vector.

7. The method as described in claim 2, characterized in that, The step of training the second interaction prediction network based on the difference between the second interaction label and the first prediction result includes: Based on the difference between the second interactive label and the first prediction result, the target loss function corresponding to the second interactive prediction network is determined; The interaction feature representation corresponding to each second sample resource is updated according to the target loss function corresponding to the second interaction prediction network, and the push association feature representation corresponding to the push association data is also updated.

8. The method as described in claim 7, characterized in that, The step of training the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result, includes: Based on the difference between the second prediction result and the third prediction result, a first loss function corresponding to the first interactive prediction network is determined; Based on the difference between the second interactive label and the third prediction result, a second loss function corresponding to the first interactive prediction network is determined. Based on the first loss function and the second loss function, the target loss function corresponding to the first interactive prediction network is determined; and the model parameters of the first interactive prediction network are adjusted in the direction of convergence of the target loss function corresponding to the first interactive prediction network.

9. The method as described in claim 1, characterized in that, The step of responding to a push request to push resources to a target object by invoking a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the resource to be pushed includes: In response to a push request to push resources to the object to be pushed, the push association data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed are obtained; the historical short-term interaction sequence related to the object to be pushed is pushed to the object within the second historical time period starting from the current time point; The first interaction prediction network is invoked to predict the interaction probability based on the push-related data corresponding to the resource to be pushed and the historical short-term interaction sequence related to the object to be pushed.

10. The method as described in claim 1, characterized in that, Determining the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability includes: The interaction probability is determined as the matching degree between the object to be pushed and the resource to be pushed; or, Based on the conversion rules between interaction probability and matching degree, the interaction probability is converted into matching degree to obtain the matching degree between the object to be pushed and the resource to be pushed.

11. A resource delivery device, characterized in that, include: The processing unit is configured to, in response to a push request to push resources to a target object, invoke a first interaction prediction network to predict the interaction probability of the target object performing a target interaction on the push resource. The processing unit is further configured to determine the matching degree between the object to be pushed and the resource to be pushed based on the interaction probability; An output unit is used to push the resource to be pushed to the object if the matching degree between the object to be pushed and the resource to be pushed meets the push rules. The first interaction prediction network is trained based on the second interaction prediction network, the first sample resource, and the historical short-term interaction sequence related to the first sample resource; the second interaction prediction network is trained based on the first sample resource and the historical long-term interaction sequence related to the first sample resource; the historical long-term interaction sequence includes the historical short-term interaction sequence; both the first interaction prediction network and the second interaction prediction network are used to perform target interaction prediction, and the network size of the first interaction prediction network is smaller than the network size of the second interaction prediction network. The historical long-term interaction sequence includes N second sample resources, which are pushed to the target object within a first historical time period starting from the interaction between the target object and the first sample resource, where N is a positive integer; the historical short-term interaction sequence includes M third sample resources and M third interaction tags; the M third sample resources are pushed to the target object within a second historical time period starting from the interaction between the target object and the first sample resource, and the M third interaction tags correspond one-to-one with the M third sample resources, where M is a positive integer less than or equal to N, the first historical time period is longer than the second historical time period, and the M third sample resources belong to the N second sample resources; The acquisition unit is used to acquire the interaction feature representation corresponding to each of the M third sample resources, wherein the interaction feature representation corresponding to each third sample resource is determined from the updated interaction feature representation corresponding to each second sample resource, and to acquire the updated push association feature representation corresponding to the push association data corresponding to the first sample resource. The processing unit is used to call the trained second interaction prediction network and the first interaction prediction network respectively, and perform interaction prediction processing based on the interaction feature representation corresponding to each third sample resource and the updated push association feature representation corresponding to the push association data respectively, to obtain the second prediction result and the third prediction result of the target object performing target interaction on the first sample resource. The processing unit is further configured to train the first interaction prediction network based on the difference between the second prediction result and the third prediction result, and the difference between the second interaction label and the third prediction result, wherein the second interaction label is used to indicate whether the target object performs the target interaction on the first sample resource.

12. A resource push device, characterized in that, The resource push device includes an input interface and an output interface, and also includes: A processor, adapted to implement one or more instructions; and, A computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded by the processor and executed as described in any one of claims 1-10.

13. A computer storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed by a processor, are used to perform the resource push method as described in any one of claims 1-10.

14. A computer program product or computer program, characterized in that, The computer program product includes a computer program, which, when executed by a processor, is used to load and execute the resource push method as described in any one of claims 1-10.

Images