Online task processing model training, task processing method and device

Abstract

This disclosure relates to an online task processing model training, task processing method, and apparatus. The online task processing model training method includes: acquiring a real-time operation sequence and a target graph structure of a first sample object, wherein the target graph structure is a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are interaction records generated by the first sample object performing preset interaction operations on preset multimedia resources before the last update of the target graph structure; and training the task processing sub-model and the graph representation sub-model in the online task processing model in parallel based on the target graph structure and the real-time operation sequence to obtain a target online task processing model. Utilizing embodiments of this disclosure improves model training efficiency and model processing performance, and eliminates the need for feature table creation during online task processing model training, effectively reducing storage space consumption and significantly improving system performance.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to an online task processing model training, task processing method and apparatus. Background Technology

[0002] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields. Graph structures, as data structures containing rich information, are often used as training data for deep learning models in scenarios such as click-through rate (CTR) ranking and conversion ranking. In related technologies, the training process of graph-based deep learning models often adopts a two-stage approach: first, the vector representations of graph nodes in the graph structure are trained offline and stored in a table; then, they are integrated into deep learning models for CTR ranking and conversion ranking for secondary training. However, the two-stage training method in the aforementioned technologies requires offline graph learning and image vector feature storage before downstream model training. This process is time-consuming, and the training data is often not updated in a timely manner, resulting in poor timeliness, low training efficiency, and poor processing performance of the trained model. Furthermore, feature storage also incurs significant storage resource overhead, affecting system performance. Summary of the Invention

[0003] This disclosure provides an online task processing model training, task processing method, and apparatus to at least address the technical problems in related technologies, such as poor timeliness, low training efficiency, poor processing performance of trained models, and significant storage resource overhead due to feature table mapping, which affects system performance. The technical solution of this disclosure is as follows:

[0004] According to a first aspect of the present disclosure, an online task processing model training method is provided, comprising:

[0005] The real-time operation sequence and target graph structure of the first sample object are obtained. The real-time operation sequence is a sequence of resource identification information of multimedia resources in which the first sample object performs preset interactive operations within the current time period. The current time period is the time period from the last update of the target graph structure to the current time. The target graph structure is a graph structure constructed based on the historical interaction records of the first sample object. The historical interaction records are the interaction records generated by the first sample object performing preset interactive operations on preset multimedia resources before the last update of the target graph structure.

[0006] Based on the target graph structure and the real-time operation sequence, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model.

[0007] In an optional embodiment, the target graph structure is a graph structure with the first object identifier information of the first sample object, the first resource identifier information of the historical operation resource corresponding to the first sample object, and the first entity identifier information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges; the node where the first resource identifier information is located carries the resource attribute information corresponding to the first resource identifier information; the historical operation resource is a multimedia resource in which the first sample object has performed the preset interaction operation.

[0008] The parallel training of the task processing sub-model and the graph representation sub-model in the online task processing model based on the target graph structure and the real-time operation sequence, to obtain the target online task processing model, includes:

[0009] From the target graph structure, the resource attribute information and graph sampling sequence are obtained; the graph sampling sequence includes a first identifier sequence and a second identifier sequence, wherein the first identifier sequence is a sequence composed of identifier information on a first branch in the target graph structure; the first branch includes a branch whose starting node is the node where the object identifier information of any object in the first sample object is located; the second identifier sequence is a sequence composed of identifier information on a second branch in the target graph structure; the second branch includes a branch whose starting node is the node where the resource identifier information of any resource in the historical operation resources is located;

[0010] Obtain the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources;

[0011] Based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model.

[0012] In an optional embodiment, the online task processing model further includes a first decoding network; the parallel training of the task processing sub-model and the graph representation sub-model in the online task processing model based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence to obtain the target online task processing model includes:

[0013] The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the task processing sub-model for task processing to obtain the first prediction task result;

[0014] The graph sampling sequence and the resource attribute information are input into the graph representation sub-model for graph representation learning to obtain the target graph features;

[0015] The target image features are decoded to obtain the result of the second prediction task;

[0016] Based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

[0017] In an optional embodiment, the task processing sub-model includes a task feature learning network and a second decoding network; the step of inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task processing sub-model for task processing to obtain the first predicted task result includes:

[0018] The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the task feature learning network for feature learning to obtain the target task features;

[0019] The target task features are input into the second decoding network for decoding processing to obtain the first prediction task result.

[0020] In an optional embodiment, the task feature learning network includes: a feature layer normalization network, a feature partitioning network, a first feature weighting processing network, and a first encoding network; the step of inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task feature learning network for feature learning to obtain target task features includes:

[0021] The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the feature layer normalization network for layer normalization processing to obtain standardized features;

[0022] The standardized features are input into the feature partitioning network for feature partitioning to obtain target sparse features and target dense features.

[0023] The target sparse features are input into the first feature weighting processing network for weighting processing to obtain weighted features. The weighted features are features that are weighted based on the contribution degree of each feature in the target sparse features.

[0024] The weighted features and the target dense features are input into the first encoding network for task encoding processing to obtain the target task features.

[0025] In an optional embodiment, the graph representation sub-model includes: a second feature weighting processing network and a second encoding network; the step of inputting the graph sampling sequence and the resource attribute information into the graph representation sub-model for graph representation learning to obtain target graph features includes:

[0026] The graph sampling sequence and the resource attribute information are input into the second feature weighting processing network for weighting processing to obtain weighted graph features. The weighted graph features are graph features after weighting processing based on the degree of contribution of each attribute information in the resource attribute information to the representation of the first sample object and the historical operation resources.

[0027] The weighted graph features are input into the second encoding network for graph encoding processing to obtain the target graph features.

[0028] In an optional embodiment, the step of training the task processing sub-model and the graph representation sub-model in parallel based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result to obtain the target online task processing model includes:

[0029] Based on the results of the first prediction task and the results of the first preset task, the first loss information is determined;

[0030] Based on the results of the second prediction task and the second preset task, determine the second loss information;

[0031] Based on the first loss information and the second loss information, the third loss information is determined;

[0032] Based on the third loss information, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

[0033] In an optional embodiment, the target graph structure is a graph structure with the first object identifier information of the first sample object, the first resource identifier information of the historical operation resource corresponding to the first sample object, and the first entity identifier information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges; the method further includes:

[0034] When the preset update time is reached, the second object identifier information of the second sample object that has performed the preset operation in the current update cycle, the second resource identifier information of the operation resource corresponding to the second sample object in the current update cycle, and the second entity identifier information of the associated entity corresponding to the operation resource are obtained.

[0035] The target graph structure is updated based on the second object identification information, the second resource identification information, and the second entity identification information;

[0036] The preset update time is the current update time determined based on the preset update frequency corresponding to the target graph structure, and the current update cycle is the time period between the last update time of the target graph structure and the current update time.

[0037] According to a second aspect of the present disclosure, a task processing method is provided, comprising:

[0038] In response to a task processing request for a target object, the system obtains the target object's real-time operation sequence and target graph structure. The target real-time operation sequence is a sequence of resource identification information of multimedia resources in which the target object performs preset interactive operations within the current time period. The current time period is the time period corresponding to the last update of the target graph structure up to the current time. The target graph structure is a graph structure constructed based on the historical interaction records of the first sample object. The historical interaction records are the interaction records generated by the first sample object for the preset interactive operations performed on the preset multimedia resources before the last update of the target graph structure. The first sample object includes the target object.

[0039] Based on any target online task processing model obtained according to any of the online task processing model training methods provided in the first aspect, the target real-time operation sequence, and the target graph structure, task processing is performed to obtain the target task processing result.

[0040] According to a third aspect of the present disclosure, an online task processing model training apparatus is provided, comprising:

[0041] The first information acquisition module is configured to acquire a real-time operation sequence and a target graph structure of a first sample object. The real-time operation sequence is a sequence of resource identification information of multimedia resources in which the first sample object performs a preset interactive operation within the current time period. The current time period is the time period from the last update of the target graph structure to the current time. The target graph structure is a graph structure constructed based on the historical interaction records of the first sample object. The historical interaction records are the interaction records generated by the first sample object performing the preset interactive operation on the preset multimedia resources before the last update of the target graph structure.

[0042] The parallel training module is configured to perform parallel training on the task processing sub-model and the graph representation sub-model in the online task processing model based on the target graph structure and the real-time operation sequence, so as to obtain the target online task processing model.

[0043] In an optional embodiment, the target graph structure is a graph structure with the first object identifier information of the first sample object, the first resource identifier information of the historical operation resource corresponding to the first sample object, and the first entity identifier information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges; the node where the first resource identifier information is located carries the resource attribute information corresponding to the first resource identifier information; the historical operation resource is a multimedia resource in which the first sample object has performed the preset interaction operation.

[0044] The parallel training module includes:

[0045] The first information acquisition unit is configured to acquire the resource attribute information and graph sampling sequence from the target graph structure; the graph sampling sequence includes a first identifier sequence and a second identifier sequence, wherein the first identifier sequence is a sequence of identifier information on a first branch in the target graph structure; the first branch includes a branch whose starting node is the node where the object identifier information of any object in the first sample object is located; the second identifier sequence is a sequence of identifier information on a second branch in the target graph structure; the second branch includes a branch whose starting node is the node where the resource identifier information of any resource in the historical operation resources is located.

[0046] The second information acquisition unit is configured to acquire the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources.

[0047] The parallel training unit is configured to perform parallel training on the task processing sub-model and the graph representation sub-model in the online task processing model based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence, so as to obtain the target online task processing model.

[0048] In an optional embodiment, the online task processing model further includes a first decoding network; the parallel training unit includes:

[0049] The first task processing unit is configured to perform task processing by inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task processing sub-model to obtain the first prediction task result;

[0050] The graph representation learning unit is configured to perform graph representation learning by inputting the graph sampling sequence and the resource attribute information into the graph representation sub-model to obtain target graph features;

[0051] The first decoding processing unit is configured to perform decoding processing on the target map features to obtain the result of the second prediction task;

[0052] The parallel training subunit is configured to perform parallel training on the task processing submodel and the graph representation submodel based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result, to obtain the target online task processing model.

[0053] In an optional embodiment, the task processing sub-model includes a task feature learning network and a second decoding network; the first task processing unit includes:

[0054] The feature learning unit is configured to perform feature learning by inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task feature learning network to obtain target task features;

[0055] The second decoding processing unit is configured to input the target task features into the second decoding network for decoding processing to obtain the first prediction task result.

[0056] In an optional embodiment, the task feature learning network includes: a feature layer normalization network, a feature partitioning network, a first feature weighting processing network, and a first encoding network; the feature learning unit includes:

[0057] The standardization processing unit is configured to perform layer standardization processing on the feature layer standardization network by inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence to obtain standardized features;

[0058] The feature segmentation processing unit is configured to perform feature segmentation processing by inputting the standardized features into the feature segmentation network to obtain target sparse features and target dense features;

[0059] The first weighted processing unit is configured to perform weighted processing on the first feature weighting processing network by inputting the target sparse features into the first feature weighting processing network to obtain weighted features. The weighted features are features that are weighted based on the contribution degree of each feature in the target sparse features.

[0060] The task encoding processing unit is configured to input the weighted features and the target dense features into the first encoding network for task encoding processing to obtain the target task features.

[0061] In an optional embodiment, the graph representation sub-model includes: a second feature weighting processing network and a second encoding network; the graph representation learning unit includes:

[0062] The second weighted processing unit is configured to perform weighted processing by inputting the graph sampling sequence and the resource attribute information into the second feature weighted processing network to obtain weighted graph features. The weighted graph features are graph features after weighting the representation contribution of each attribute information in the resource attribute information to the first sample object and the historical operation resource.

[0063] The graph encoding processing unit is configured to input the weighted graph features into the second encoding network for graph encoding processing to obtain the target graph features.

[0064] In an optional embodiment, the parallel training subunit includes:

[0065] The first loss information determination unit is configured to determine first loss information based on the result of the first prediction task and the result of the first preset task.

[0066] The second loss information determination unit is configured to determine the second loss information based on the result of the second prediction task and the result of the second preset task.

[0067] The target loss information determination unit is configured to determine third loss information based on the first loss information and the second loss information;

[0068] The model training unit is configured to perform parallel training on the task processing sub-model and the graph representation sub-model based on the third loss information to obtain the target online task processing model.

[0069] In an optional embodiment, the target graph structure is a graph structure with nodes consisting of the first object identifier information of the first sample object, the first resource identifier information of the historical operation resource corresponding to the first sample object, and the first entity identifier information of the associated entity corresponding to the historical operation resource, and edges consisting of the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity; the device further includes:

[0070] The identification information acquisition module is configured to, upon reaching a preset update time, acquire the second object identification information of the second sample object that has performed a preset operation within the current update cycle, the second resource identification information of the operation resource corresponding to the second sample object within the current update cycle, and the second entity identification information of the associated entity corresponding to the operation resource;

[0071] The target graph structure update module is configured to update the target graph structure based on the second object identifier information, the second resource identifier information, and the second entity identifier information;

[0072] The preset update time is the current update time determined based on the preset update frequency corresponding to the target graph structure, and the current update cycle is the time period between the last update time of the target graph structure and the current update time.

[0073] According to a fourth aspect of the present disclosure, a task processing apparatus is provided, comprising:

[0074] The second information acquisition module is configured to respond to a task processing request for a target object by acquiring the target object's real-time operation sequence and target graph structure. The target real-time operation sequence is a sequence of resource identification information of multimedia resources in which the target object performs preset interactive operations within the current time period. The current time period is the time period corresponding to the last update of the target graph structure to the current time. The target graph structure is a graph structure constructed based on the historical interaction records of the first sample object. The historical interaction records are the interaction records generated by the first sample object for the preset interactive operations performed on the preset multimedia resources before the last update of the target graph structure. The first sample object includes the target object.

[0075] The second task processing module is configured to perform task processing based on any target online task processing model obtained according to any of the online task processing model training methods provided in the first aspect, the target real-time operation sequence, and the target graph structure, and obtain the target task processing result.

[0076] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the method as described in any one of the first or second aspects above.

[0077] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided, wherein when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the method described in any one of the first or second aspects of the present disclosure.

[0078] According to a seventh aspect of the present disclosure, a computer program product including instructions is provided that, when run on a computer, causes the computer to perform the method as described in any one of the first or second aspects above.

[0079] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects:

[0080] During online real-time task model training, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel by combining real-time operation sequences and target graph structures. This can compress model time delays into the graph structure update stage. By combining graph structures to effectively represent objects and multimedia resources, the training efficiency of the model is greatly improved, thereby effectively improving the processing effect and timeliness of the trained model, and significantly improving the accuracy of multimedia resource push tasks. Furthermore, the online task processing model training process does not require the table creation of features corresponding to the target graph structure, effectively reducing storage space consumption and greatly improving system performance.

[0081] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0082] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0083] Figure 1 This is a schematic diagram illustrating an application environment according to an exemplary embodiment;

[0084] Figure 2 This is a flowchart illustrating an online task processing model training method according to an exemplary embodiment;

[0085] Figure 3 This is a partial schematic diagram of a target graph structure provided according to an exemplary embodiment;

[0086] Figure 4 This is a flowchart illustrating an exemplary embodiment of a method for parallel training of a task processing sub-model and a graph representation sub-model in an online task processing model based on a target graph structure and a real-time operation sequence, to obtain a target online task processing model.

[0087] Figure 5 This is a flowchart illustrating an exemplary embodiment of a method for parallel training of a task processing sub-model and a graph representation sub-model in an online task processing model based on real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences, to obtain a target online task processing model.

[0088] Figure 6 This is a flowchart illustrating an exemplary embodiment of a method for inputting real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences into a task feature learning network to learn features and obtain target task features.

[0089] Figure 7 This is a schematic diagram illustrating the training process of an online task processing model according to an exemplary embodiment;

[0090] Figure 8 This is a flowchart illustrating a task processing method according to an exemplary embodiment;

[0091] Figure 9 This is a block diagram of an online task processing model training device according to an exemplary embodiment.

[0092] Figure 10 This is a block diagram of a task processing apparatus according to an exemplary embodiment;

[0093] Figure 11 This is a block diagram illustrating an electronic device for online task processing model training or task processing according to an exemplary embodiment;

[0094] Figure 12 This is a block diagram illustrating another electronic device for online task processing model training or task processing according to an exemplary embodiment. Detailed Implementation

[0095] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0096] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0097] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0098] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating an application environment according to an exemplary embodiment, which may include a terminal 100 and a server 200.

[0099] In an optional embodiment, terminal 100 can be used to provide business services to any user, such as multimedia resource push services. Specifically, terminal 100 can be, but is not limited to, electronic devices such as smartphones, desktop computers, tablets, laptops, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, and smart wearable devices, or software running on the aforementioned electronic devices, such as applications. Optionally, the operating system running on the electronic device can be, but is not limited to, Android, iOS, Linux, Windows, etc.

[0100] In an optional embodiment, server 200 can provide background services to terminal 100. Specifically, server 200 can combine real-time operation sequences provided by multiple terminals 100 and periodically / real-time updated target graph structures to train an online task processing model, perform task processing using the online task processing model, and provide business services to users on the terminal 100 side based on the task processing. Optionally, server 200 can be equipped with a graph engine that can perform graph retrieval, i.e., extract corresponding node data from the target graph structure. Server 200 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0101] In addition, it should be noted that, Figure 1 The example shown is merely one application environment provided by this disclosure. In practical applications, other application environments may also be included, such as more terminals.

[0102] In the embodiments described in this specification, the terminal 100 and the server 200 can be directly or indirectly connected through wired or wireless communication, and this disclosure does not impose any restrictions.

[0103] Figure 2 This is a flowchart illustrating an online task processing model training method according to an exemplary embodiment, such as... Figure 2 As shown, this method can be applied to electronic devices such as servers, and specifically, it may include the following steps:

[0104] In step S201, the real-time operation sequence and target graph structure of the first sample object are obtained.

[0105] In a specific embodiment, the target graph structure can be a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are the interaction records generated by the first sample object and the preset interactive operations performed on the preset multimedia resources before the last update of the target graph structure; specifically, the preset interactive operations can include interactive operations such as clicking, liking, commenting, and conversion (e.g., purchasing goods, downloading applications, etc.); specifically, each time an interactive operation is performed, a historical interaction record can be generated, and each historical interaction record can include interactive operation identification information (used to identify the preset interactive operation performed), object identification information and object attribute information of the object performing the preset interactive operation, and resource identification information and resource attribute information of the multimedia resource on which the preset interactive operation was performed.

[0106] In a specific embodiment, the first sample object can be a user account within a resource platform; the object identification information can be identification information that distinguishes different objects, such as a user account identifier; the object attribute information can be attribute representations used to describe the object, such as gender, age, etc.; the preset multimedia resource can be any multimedia resource within the resource platform; the aforementioned resource platform can be a platform that provides multimedia resource push services to users, such as a short video platform, a shopping platform, etc. Specifically, the multimedia resource can be dynamic media resources such as videos, or static media resources such as images and text. The resource identification information can be identification information used to distinguish different multimedia resources, and the resource attribute information can be attribute representations used to describe the multimedia resource, such as category, title, and feature representations (feature vectors) of images in the video (multimedia resource).

[0107] In an optional embodiment, the target graph structure described above can also be a graph structure with the first object identification information of the first sample object and the first resource identification information of the historical operation resource corresponding to the first sample object as nodes, and the interaction operations between the first sample object and the historical operation resource as edges. Optionally, to improve the accuracy of the graph structure in representing objects and multimedia resources, the entity identification information of the associated entities corresponding to the historical operation resource can be added to the graph structure. Accordingly, the target graph structure described above can be a graph structure with the first object identification information of the first sample object, the first resource identification information of the historical operation resource corresponding to the first sample object, and the first entity identification information of the associated entities corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, as well as the association relationship between the historical operation resource and the associated entity, as edges. The historical operation resource corresponding to the first sample object can be a multimedia resource for which the first sample object has performed a preset interaction operation. The associated entity corresponding to the historical operation resource can be an entity (an objectively existing and distinguishable thing, such as a refrigerator or a mobile phone) that the historical operation resource includes. For example, if the historical operation resource is a video, the associated entity can be an entity introduced in the video. Specifically, entity recognition processing can be performed on historical operation resources to determine the associated entities corresponding to the historical operation resources. Accordingly, the relationship between historical operation resources and associated entities can be a containment relationship between historical operation resources and associated entities.

[0108] In a specific embodiment, the first sample object includes multiple objects, and any corresponding historical operation resource can correspond to one or more multimedia resources; optionally, when an object performs multiple interactive operations on a certain historical operation resource, multiple edges can be connected between the object and the historical operation resource (one edge corresponds to one executed interactive operation); optionally, different interactive operations can correspond to different weights, and correspondingly, the edges between the object and the historical operation resource can be weighted edges. Specifically, the weights corresponding to different operations can be set according to actual application requirements. For example, the weight corresponding to the click interactive operation can be 0.5, and the weight corresponding to the conversion interactive operation can be 1.

[0109] In an optional embodiment, the target graph structure described above can also be a graph structure with the first object identification information of the first sample object and the first resource identification information of the historical operation resources corresponding to the first sample object as nodes, and the interaction operations between the first sample object and the historical operation resources as edges.

[0110] In an optional embodiment, the target graph structure described above can be updated in real time or periodically to ensure the timeliness of the model training data, thereby improving the task processing performance of the trained online task processing model. Accordingly, the method may further include:

[0111] When the preset update time is reached, obtain the second object identifier information of the second sample object that has performed the preset operation in the current update cycle, the second resource identifier information of the operation resource corresponding to the second sample object in the current update cycle, and the second entity identifier information of the associated entity corresponding to the operation resource.

[0112] The target graph structure is updated based on the second object identifier information, the second resource identifier information, and the second entity identifier information;

[0113] In one specific embodiment, the aforementioned preset update time can be the current update time determined based on a preset update frequency corresponding to the target graph structure, and the aforementioned current update period is the time interval between the last update time of the target graph structure and the current update time. The preset update frequency can be the update frequency of the target graph structure and can be set according to actual application requirements.

[0114] In one specific embodiment, the second sample object may include one or more objects. If any object in the second sample object is included in the first sample object, during the target graph structure update process, the node corresponding to that object may not be added; otherwise, a new node corresponding to the object may be added. The decision to add a node corresponding to the resource is based on whether the operation resource corresponding to the object in the current update cycle is included in the historical operation resources corresponding to the first sample object. Then, edges can be constructed based on the interaction operations performed by the object on the corresponding operation resource.

[0115] In the above embodiments, by periodically updating the target graph structure, the timeliness of the graph structure can be effectively improved, thereby enhancing the effectiveness of subsequent graph representation.

[0116] In one specific embodiment, the aforementioned real-time operation sequence can be a real-time generated operation sequence. Specifically, this real-time operation sequence can be a sequence of resource identification information of multimedia resources in which the first sample object performs a preset interactive operation within the current time period. The current time period is the time period corresponding to the last update of the target graph structure to the current time.

[0117] In a specific embodiment, such as Figure 3 As shown, Figure 3 This is a partial schematic diagram of a target graph structure provided according to an exemplary embodiment. In this diagram, the nodes corresponding to circles are the nodes containing object identification information, the nodes corresponding to rectangles are the nodes containing resource identification information, and the nodes corresponding to squares are the nodes containing entity identification information.

[0118] In step S203, based on the target graph structure and the real-time operation sequence, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model.

[0119] In a specific embodiment, the online task processing model described above can be a task processing model trained in real time. Specifically, different task processing models can correspond to different scenarios. For example, in a scenario where multimedia resources are pushed based on the click probability of multimedia resources, the task processing model described above can be a click recognition model. In a scenario where multimedia resources are pushed based on the conversion probability of multimedia resources, the task processing model described above can be a conversion recognition model.

[0120] In practical applications, data often faces significant sparsity, leading to inaccurate representations learned by the model. To mitigate this issue, resource attribute information of historical operations (as auxiliary information) can be added to the target graph structure; correspondingly, the aforementioned first resource identifier information can have its corresponding resource attribute information attached to the node.

[0121] In an optional embodiment, such as Figure 4 As shown, the above-mentioned parallel training of the task processing sub-model and the graph representation sub-model in the online task processing model based on the target graph structure and real-time operation sequence to obtain the target online task processing model may include the following steps:

[0122] In step S2031, resource attribute information and graph sampling sequence are obtained from the target graph structure;

[0123] In step S2033, the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources are obtained;

[0124] In step S2035, based on the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence, the task processing sub-model and graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model.

[0125] In a specific embodiment, the above graph sampling sequence may include a first identifier sequence and a second identifier sequence. Specifically, the first identifier sequence may be a sequence of identifier information on the first branch in the target graph structure; the first branch includes a branch with the node where the object identifier information of any object in the first sample object is located as the starting node; optionally, taking an object with an object identifier information of A as an example, the first branch may include the branch corresponding to 301 above. Accordingly, the first identifier sequence may include object A, resource 3, and entity M.

[0126] Specifically, the second identifier sequence is a sequence of identifier information on the second branch in the target graph structure; the second branch includes a branch with the node where the resource identifier information of any resource in the historical operation resources is located as the starting node; optionally, taking a multimedia resource with a resource identifier of 1 as an example, the second branch may include the branch corresponding to 302 above, and correspondingly, the second identifier sequence may include resource 1, object B, resource 2, and entity E.

[0127] In a specific embodiment, the historical operation sequence can be a sequence of resource identifiers of multimedia resources in which the target object has performed preset interactive operations within a historical time period. Specifically, the historical time period can be set according to the actual application. The historical interaction information of the aforementioned historical operation resources can be information reflecting the interaction status of the historical operation resources within the historical time period, such as the number of clicks and comments.

[0128] In a specific embodiment, the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources can be directly obtained from the historical interaction records and from the target graph structure. Correspondingly, the node where the resource identification information is located also carries the historical interaction information, and the node where the object identification information is located carries the object attribute information.

[0129] In an optional embodiment, the above-described online task processing model may further include a first decoding network; correspondingly, as... Figure 5 As shown, the parallel training of the task processing sub-model and graph representation sub-model in the online task processing model based on real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences to obtain the target online task processing model may include the following steps:

[0130] In step S501, the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information and graph sampling sequence are input into the task processing sub-model for task processing to obtain the first prediction task result;

[0131] In a specific embodiment, the task processing sub-model can adaptively learn personalized representations (feature vectors) of the first sample object and historical operation resources based on sparse and dense features, and combine the learned representations for task processing. Specifically, dense features can be features with semantic information and correlations between different features, such as attribute features like age in object attribute information, and image feature information of images in resource attribute information; sparse features can be identifiable features, such as real-time operation sequences, information represented by identification information like gender in object attribute information, information represented by identification information like category in resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences.

[0132] In one specific embodiment, the first prediction task result can be the task processing result after the task processing sub-model performs task processing. Specifically, the task processing result differs in different scenarios. For example, in a scenario where multimedia resources are pushed based on the click probability, the task processing result can be the click probability. In another scenario where multimedia resources are pushed based on the conversion probability, the task processing result can be the conversion probability.

[0133] In an optional embodiment, the task processing sub-model includes a task feature learning network and a second decoding network. Accordingly, the above-mentioned inputting the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence into the task processing sub-model for task processing to obtain the first predicted task result includes: inputting the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence into the task feature learning network for feature learning to obtain target task features; and inputting the target task features into the second decoding network for decoding processing to obtain the first predicted task result.

[0134] In a specific embodiment, the target task features can be the features of the first sample object and historical operation resources learned based on real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences.

[0135] In one specific embodiment, the second decoding network can be used to transform the target task features into task results. Optionally, the second decoding network can be the network corresponding to the sigmoid (threshold) function.

[0136] In the above embodiments, feature learning is performed by combining real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences. This allows for the learning of personalized representations of the first sample object and historical operation resources. By combining the learned target task features with task processing, the effectiveness of task processing can be effectively guaranteed.

[0137] In an optional embodiment, the task feature learning network described above may include: a feature layer normalization network, a feature partitioning network, a first feature weighting processing network, and a first encoding network; correspondingly, as... Figure 6 As shown, the above-mentioned input of real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences into the task feature learning network for feature learning to obtain target task features may include the following steps:

[0138] In step S601, the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence are input into the feature layer normalization network for layer normalization processing to obtain standardized features;

[0139] In step S603, the standardized features are input into the feature partitioning network for feature partitioning processing to obtain target sparse features and target dense features;

[0140] In step S605, the target sparse features are input into the first feature weighting processing network for weighting processing to obtain weighted features;

[0141] In step S607, the weighted features and target dense features are input into the first encoding network for task encoding processing to obtain the target task features.

[0142] In a specific embodiment, in order to eliminate the influence of differences in units and scales among real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences, and to treat each feature dimension equally, layer standardization can be performed first by combining a feature layer standardization network; then, a feature partitioning network can be combined to re-partition the standardized features into target sparse features and target dense features.

[0143] In one specific embodiment, the contribution of each feature (each bit feature) in the target sparse features to the overall feature learning is different. Accordingly, a first feature weighting processing network can be used to learn the contribution of each feature in the target sparse features; specifically, this contribution can be a network parameter in the first feature weighting processing network. During the weighting process using the first feature weighting processing network, this network parameter can be used to learn the aforementioned weighted features. Specifically, the aforementioned weighted features are features obtained by weighting each feature in the target sparse features based on its contribution.

[0144] In one specific embodiment, the first encoding network can be used to concatenate weighted features and target dense features to obtain target task features for predicting task results.

[0145] In the above embodiments, after performing layer standardization on the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence, they are divided into target sparse features and target dense features, which can better improve the model training effect. Furthermore, after weighting the target sparse features, they are input together with the target dense features into the first encoding network for task encoding processing to obtain target task features for task result prediction, which can greatly improve the accuracy of the representation of the first sample object and historical operation resources during task processing.

[0146] In step S503, the graph sampling sequence and resource attribute information are input into the graph representation sub-model for graph representation learning to obtain the target graph features;

[0147] In an optional embodiment, the graph representation sub-model includes: a second feature weighting processing network and a second encoding network; inputting the graph sampling sequence and resource attribute information into the graph representation sub-model for graph representation learning to obtain target graph features includes: inputting the graph sampling sequence and resource attribute information into the second feature weighting processing network for weighting processing to obtain weighted graph features; inputting the weighted graph features into the second encoding network for graph encoding processing to obtain target graph features.

[0148] In one specific embodiment, different resource attribute information contributes differently to the representation learning of the same object or multimedia resource. Accordingly, this contribution level can be learned by combining a second feature weighting processing network. Optionally, this contribution level can be a network parameter in the second feature weighting processing network. During the weighting process using the second feature weighting processing network, this network parameter can be incorporated to learn the aforementioned weighted graph features. Accordingly, the aforementioned weighted graph features can be graph features obtained by weighting the representation contribution of each attribute information in the resource attribute information to the first sample object and the historical operation resources.

[0149] In one specific embodiment, the second encoding network can be a network used for graph representation. Optionally, the second encoding network can be a deepwalk network, a GCN (Graph Convolutional Network), etc.

[0150] In the above embodiments, during the graph representation learning process combining graph sampling sequences and resource attribute information, the degree of contribution of each attribute information in the resource attribute information to the representation of the first sample object and historical operation resources can be integrated to improve the accuracy of graph representation, thereby improving the accuracy of subsequent graph-based task processing.

[0151] In step S505, the target map features are decoded to obtain the second prediction task result;

[0152] In a specific embodiment, decoding processing can be performed by combining the target graph features of the first decoding network object corresponding to the task. The first decoding network can be used to transform the target graph features into task results. Specifically, the task processing results (second predicted task results) are different in different scenarios. Taking a scenario where multimedia resources are pushed based on the click probability of multimedia resources as an example, the second predicted task result can be the click probability. Taking a scenario where multimedia resources are pushed based on the conversion probability of multimedia resources as an example, the second predicted task result can be the conversion probability. Optionally, the first decoding network can be the network corresponding to the sigmoid (threshold) function.

[0153] In step S507, based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

[0154] In an optional embodiment, based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model, including:

[0155] Based on the results of the first prediction task and the results of the first preset task, determine the first loss information;

[0156] Based on the results of the second prediction task and the second preset task, determine the second loss information;

[0157] Based on the first and second loss information, determine the third loss information;

[0158] Based on the third loss information, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

[0159] In a specific embodiment, the first loss information can characterize the degree of difference between the result of the first prediction task and the result of the first preset task; the second loss information can characterize the result of the second prediction task and the result of the second preset task. In determining the first and second loss information, a preset loss function, such as a cross-entropy loss function or an absolute value loss function, can be used. Optionally, in determining the first and second loss information, the interaction operation weights between the corresponding object and the corresponding historical operation resources can also be used. These interaction operation weights are multiplied by the corresponding prediction task result before calculating the first and second loss information. Optionally, the first and second loss information can be weighted and summed to obtain the third loss information; specifically, the third loss information can characterize the performance of the online task processing model.

[0160] In a specific embodiment, the parallel training of the task processing sub-model and the graph representation sub-model based on the third loss information to obtain the target online task processing model may include: updating the model parameters of the task processing sub-model and the graph representation sub-model in the online task processing model according to the third loss information; repeating the training iteration operation of inputting real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information and graph sampling sequence into the task processing sub-model for task processing based on the updated online task processing model to obtain the first predicted task result until the third loss information is determined, until a preset convergence condition is met, and taking the online task processing model corresponding to the preset convergence condition as the target online task processing model.

[0161] In an optional embodiment, satisfying the preset convergence condition can be that the number of training iterations reaches a preset number of training iterations. Optionally, satisfying the preset convergence condition can also be that the third loss information is less than a specified threshold. In the embodiments of this specification, the preset number of training iterations and the specified threshold can be preset in conjunction with the training speed and accuracy of the network in practical applications.

[0162] In the above embodiments, the loss of the online task processing model is determined by combining the loss information corresponding to the task processing sub-model and the graph representation sub-model. This can effectively ensure the effectiveness of training the online task processing model while achieving parallel processing of task processing and graph representation.

[0163] In one specific embodiment, the aforementioned target online task processing model can be the latest online task processing model. Specifically, the target online task processing model can also be trained in parallel again when new real-time operation sequences are generated, so as to continuously obtain the latest online task processing model.

[0164] In the above embodiments, during the task processing process combined with the task processing sub-model, graph representation is performed in conjunction with the graph representation sub-model, and task decoding is performed in conjunction with the target graph features. This enables parallel processing of task processing and graph representation, greatly improving the training efficiency of the task model. Furthermore, there is no need to perform graph representation table storage, effectively reducing storage space consumption and significantly improving system performance.

[0165] In a specific embodiment, such as Figure 7 As shown, Figure 7 This is a schematic diagram illustrating the training process of an online task processing model according to an exemplary embodiment. By combining the target graph structure and the real-time operation sequence, the input real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information, and graph sampling sequence of the task processing sub-model are determined; and the input graph sampling sequence and resource attribute information of the graph representation sub-model are determined. Representation processing can be performed in both sub-models to obtain the first predicted task result and the second predicted task result. Then, the task processing sub-model and the graph representation sub-model can be trained in parallel by combining the first predicted task result, the first preset task result corresponding to the first predicted task result, the second predicted task result, and the second preset task result corresponding to the second predicted task result, to obtain the target online task processing model.

[0166] In the above embodiments, resource attribute information and graph sampling sequence are obtained from the target graph structure; and combined with the object attribute information of the first sample object, historical operation sequence, historical interaction information of historical operation resources, and real-time operation sequence, the task processing sub-model and graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model. This can improve the training efficiency of the task model while improving the timeliness of data during model training, and it does not require graph representation to be written into the table, effectively reducing storage space consumption and greatly improving system performance.

[0167] As can be seen from the technical solutions provided in the embodiments of this specification above, in the online real-time task model training process of this specification, by combining the real-time operation sequence and the target graph structure, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel. This can compress the model's timeliness delay to the graph structure update stage. Based on the effective representation of objects and multimedia resources by the graph structure, the model training efficiency is greatly improved, thereby effectively improving the processing effect and timeliness of the trained model, greatly improving the accuracy of multimedia resource push tasks. Moreover, there is no need to add the corresponding features of the target graph structure to the table during the online task processing model training process, effectively reducing storage space consumption and greatly improving system performance.

[0168] Model time delay compression to graph structure update

[0169] Figure 8 This is a flowchart illustrating a task processing method according to an exemplary embodiment, such as... Figure 8 As shown, this method can be applied to electronic devices such as servers and terminals, and specifically, it may include the following steps:

[0170] In step S801, in response to a task processing request for the target object, the target real-time operation sequence and target graph structure of the target object are obtained;

[0171] In step S803, task processing is performed based on the target online task processing model, the target real-time operation sequence, and the target graph structure to obtain the target task processing result.

[0172] In one specific embodiment, the target real-time operation sequence can be a real-time generated operation sequence corresponding to the target object. Specifically, the target real-time operation sequence can be a sequence of resource identification information of the multimedia resource for which the target object is currently performing a preset interactive operation.

[0173] In one specific embodiment, the target graph structure is a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are the interaction records generated by the preset interactive operations performed on the preset multimedia resources before the last update of the target graph structure, and the first sample object includes the target object.

[0174] In an optional embodiment, the task processing based on the target online task processing model, the target real-time operation sequence, and the target graph structure to obtain the target task processing result may include: determining the object features corresponding to the target object from the target graph features and target task features obtained during the training and convergence of the target online task processing model, by combining the target object identification information corresponding to the target object; determining the resource features corresponding to the preset multimedia resources from the target graph features and target task features obtained during the training and convergence of the target online task processing model, by combining the resource identification information of the preset multimedia resources; then, determining the multimedia resources to be recommended from the preset multimedia resources by combining the similarity between the object features and the resource features; then, combining the target real-time operation sequence, target object attribute information (object attribute information of the target object), target resource attribute information (resource attribute information of the multimedia resources to be recommended), target historical operation sequence (historical operation sequence of the target object), and target historical interaction information. The first target task result is obtained by inputting the following information into the task processing sub-model of the target online task processing model: first, the historical interaction information of the multimedia resources to be recommended; second, the target task result; and third, the target resource attribute information. The first and second target task results are then input into the graph representation sub-model of the target online task processing model for graph representation learning to obtain graph features. These graph features are then input into the first decoding network of the target online task processing model for task decoding to obtain the second target task result. Finally, the first and second target task results are weighted and summed to obtain the target task processing result. Optionally, the weights of the first and second target task results can be set according to actual application requirements.

[0175] In a specific embodiment, the detailed process of performing the task to obtain the first target task result can be found in the relevant steps described above, and will not be repeated here.

[0176] In a specific embodiment, the detailed refinement of the graph representation learning process to obtain graph features can be found in the relevant steps described above, and will not be repeated here. Optionally, taking a scenario where multimedia resources are pushed based on the click probability of multimedia resources as an example, the target task processing result represents the probability (click probability) of the target object clicking the multimedia resource to be recommended; accordingly, a preset number of multimedia resources with higher click probabilities can be selected as target multimedia resources and pushed to the target object, or multimedia resources with click probabilities greater than a preset threshold can be selected as target multimedia resources and pushed to the target object.

[0177] Optionally, a secondary task model can be trained by combining the above target graph representation to better improve the task processing effect, depending on the actual application requirements.

[0178] As can be seen from the technical solutions provided in the embodiments of this specification above, in the task processing process, by combining the target object's real-time operation sequence, target graph structure, and online task processing model, the timeliness can be greatly improved, and the task processing effect and efficiency can be greatly enhanced.

[0179] Figure 9 This is a block diagram illustrating an online task processing model training apparatus according to an exemplary embodiment. (Refer to...) Figure 9 The device includes:

[0180] The first information acquisition module 910 is configured to acquire a real-time operation sequence and a target graph structure of a first sample object. The real-time operation sequence is a sequence of resource identification information of multimedia resources in which the first sample object performs a preset interactive operation within the current time period. The current time period is the time period from the last update of the target graph structure to the current time. The target graph structure is a graph structure constructed based on the historical interaction records of the first sample object. The historical interaction records are the interaction records generated by the first sample object performing the preset interactive operation on the preset multimedia resources before the last update of the target graph structure.

[0181] The parallel training module 920 is configured to perform parallel training on the task processing sub-model and the graph representation sub-model in the online task processing model based on the target graph structure and real-time operation sequence, so as to obtain the target online task processing model.

[0182] In an optional embodiment, the target graph structure is a graph structure with the first object identifier information of the first sample object, the first resource identifier information of the historical operation resource corresponding to the first sample object, and the first entity identifier information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges; the node where the first resource identifier information is located carries the resource attribute information corresponding to the first resource identifier information; the historical operation resource is a multimedia resource in which the first sample object has performed the preset interaction operation;

[0183] Parallel training module 920 includes:

[0184] The first information acquisition unit is configured to acquire resource attribute information and a graph sampling sequence from the target graph structure. The graph sampling sequence includes a first identifier sequence and a second identifier sequence. The first identifier sequence is a sequence of identifier information on a first branch in the target graph structure. The first branch includes a branch whose starting node is the node where the object identifier information of any object in the first sample object is located. The second identifier sequence is a sequence of identifier information on a second branch in the target graph structure. The second branch includes a branch whose starting node is the node where the resource identifier information of any resource in the historical operation resources is located.

[0185] The second information acquisition unit is configured to acquire the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources.

[0186] The parallel training unit is configured to perform parallel training on the task processing sub-model and graph representation sub-model in the online task processing model based on real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences, so as to obtain the target online task processing model.

[0187] In an optional embodiment, the online task processing model further includes a first decoding network; the parallel training unit includes:

[0188] The first task processing unit is configured to process the real-time operation sequence, object attribute information, resource attribute information, historical operation sequence, historical interaction information and graph sampling sequence into the task processing sub-model to obtain the first prediction task result;

[0189] The graph representation learning unit is configured to perform graph representation learning by inputting graph sampling sequences and resource attribute information into the graph representation sub-model to obtain target graph features;

[0190] The first decoding processing unit is configured to perform decoding processing on the features of the target map to obtain the result of the second prediction task;

[0191] The parallel training sub-unit is configured to perform parallel training on the task processing sub-model and the graph representation sub-model based on the results of the first prediction task, the first preset task results corresponding to the results of the first prediction task, the second prediction task results, and the second preset task results corresponding to the results of the second prediction task, to obtain the target online task processing model.

[0192] In an optional embodiment, the task processing sub-model includes a task feature learning network and a second decoding network; the first task processing unit includes:

[0193] The feature learning unit is configured to perform feature learning by inputting real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences into the task feature learning network to obtain target task features;

[0194] The second decoding processing unit is configured to input the target task features into the second decoding network for decoding processing to obtain the first prediction task result.

[0195] In an optional embodiment, the task feature learning network includes: a feature layer normalization network, a feature partitioning network, a first feature weighting processing network, and a first encoding network; the feature learning unit includes:

[0196] The standardization processing unit is configured to perform layer standardization processing on the feature layer standardization network by inputting real-time operation sequences, object attribute information, resource attribute information, historical operation sequences, historical interaction information, and graph sampling sequences to obtain standardized features;

[0197] The feature partitioning processing unit is configured to perform feature partitioning processing on the feature partitioning network by inputting standardized features into the feature partitioning network to obtain target sparse features and target dense features.

[0198] The first weighted processing unit is configured to perform weighted processing on the first feature weighting processing network by inputting the target sparse features into the target sparse features to obtain weighted features. The weighted features are features that are weighted based on the contribution of each feature in the target sparse features.

[0199] The task encoding processing unit is configured to input the weighted features and target dense features into the first encoding network for task encoding processing to obtain the target task features.

[0200] In an optional embodiment, the graph representation sub-model includes: a second feature weighting processing network and a second encoding network; the graph representation learning unit includes:

[0201] The second weighted processing unit is configured to input the graph sampling sequence and resource attribute information into the second feature weighted processing network for weighted processing to obtain weighted graph features. The weighted graph features are graph features after weighting the representation contribution of each attribute information in the resource attribute information to the first sample object and the historical operation resource.

[0202] The graph encoding processing unit is configured to input the weighted graph features into the second encoding network for graph encoding processing to obtain the target graph features.

[0203] In an optional embodiment, the parallel training subunit includes:

[0204] The first loss information determination unit is configured to determine the first loss information based on the results of the first prediction task and the results of the first preset task.

[0205] The second loss information determination unit is configured to determine the second loss information based on the results of the second prediction task and the results of the second preset task.

[0206] The target loss information determination unit is configured to determine third loss information based on the first loss information and the second loss information;

[0207] The model training unit is configured to perform parallel training of the task processing sub-model and the graph representation sub-model based on third loss information to obtain the target online task processing model.

[0208] In an optional embodiment, the target graph structure is a graph structure with the first object identification information of the first sample object, the first resource identification information of the historical operation resource corresponding to the first sample object, and the first entity identification information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operations between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges; the above device further includes:

[0209] The identification information acquisition module is configured to, upon reaching a preset update time, acquire the second object identification information of the second sample object that has performed a preset operation within the current update cycle, the second resource identification information of the operation resource corresponding to the second sample object within the current update cycle, and the second entity identification information of the associated entity corresponding to the operation resource;

[0210] The target graph structure update module is configured to update the target graph structure based on the second object identifier information, the second resource identifier information, and the second entity identifier information.

[0211] The preset update time is the current update time determined based on the preset update frequency corresponding to the target graph structure, and the current update cycle is the time period between the last update time of the target graph structure and the current update time.

[0212] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0213] Figure 10 This is a block diagram of a task processing apparatus according to an exemplary embodiment. (Refer to...) Figure 10 The device includes:

[0214] The second information acquisition module 1010 is configured to respond to a task processing request for a target object by acquiring the target object's real-time operation sequence and target graph structure; the target real-time operation sequence is a sequence of resource identification information of multimedia resources in which the target object performs preset interactive operations within the current time period; the current time period is the time period corresponding to the last update of the target graph structure to the current time; the target graph structure is a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are the interaction records generated by the first sample object for the preset interactive operations performed on the preset multimedia resources before the last update of the target graph structure, and the first sample object includes the target object;

[0215] The second task processing module 1020 is configured to perform task processing based on any target online task processing model, target real-time operation sequence and target graph structure obtained according to any online task processing model training method provided in the first aspect, and obtain the target task processing result.

[0216] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0217] Figure 11 This is a block diagram illustrating an electronic device for online task processing model training or task processing according to an exemplary embodiment. The electronic device may be a terminal, and its internal structure diagram may be as follows: Figure 11As shown, the electronic device includes a processor, memory, model interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The model interface is used to communicate with external terminals via model connection. When the computer program is executed by the processor, it implements an online task processing model training or task processing method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0218] Figure 12 This is a block diagram illustrating another electronic device for online task processing model training or task processing according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as follows: Figure 12 As shown, the electronic device includes a processor, memory, and a model interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The model interface is used to communicate with external terminals via a model connection. When the computer program is executed by the processor, it implements an online task processing model training method or a task processing method.

[0219] Those skilled in the art will understand that Figure 11 or Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the electronic device to which the present disclosure is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0220] In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the online task processing model training method or task processing method as described in the embodiments of this disclosure.

[0221] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the online task processing model training method or task processing method of the present disclosure embodiments.

[0222] In an exemplary embodiment, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the online task processing model training method or task processing method of the present disclosure embodiments.

[0223] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0224] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0225] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for training an online task processing model, characterized in that, include: The real-time operation sequence and target graph structure of the first sample object are obtained. The real-time operation sequence is a sequence of resource identification information of multimedia resources in which the first sample object performs preset interactive operations within the current time period. The current time period is the time period corresponding to the last update of the target graph structure to the current time. The target graph structure is a graph structure with the first object identification information of the first sample object, the first resource identification information of historical operation resources, and the first entity identification information of the associated entities corresponding to the historical operation resources as nodes, and the interactive operations between the first sample object and the historical operation resources, as well as the association relationship between the historical operation resources and the associated entities, as edges. The node containing the first resource identifier information carries the resource attribute information corresponding to the first resource identifier information; the historical operation resource is the multimedia resource for which the first sample object has performed the preset interactive operation. From the target graph structure, the resource attribute information and graph sampling sequence are obtained; the graph sampling sequence includes a first identifier sequence and a second identifier sequence, wherein the first identifier sequence is a sequence composed of identifier information on a first branch in the target graph structure; the first branch includes a branch whose starting node is the node where the object identifier information of any object in the first sample object is located; the second identifier sequence is a sequence composed of identifier information on a second branch in the target graph structure; the second branch includes a branch whose starting node is the node where the resource identifier information of any resource in the historical operation resources is located; Obtain the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources; Based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence, the task processing sub-model and the graph representation sub-model in the online task processing model are trained in parallel to obtain the target online task processing model.

2. The online task processing model training method according to claim 1, characterized in that, The method of training the task processing sub-model and the graph representation sub-model in the online task processing model in parallel based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence to obtain the target online task processing model includes: The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the task processing sub-model for task processing to obtain the first prediction task result; The graph sampling sequence and the resource attribute information are input into the graph representation sub-model for graph representation learning to obtain the target graph features; The target image features are decoded to obtain the result of the second prediction task; Based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

3. The online task processing model training method according to claim 2, characterized in that, The task processing sub-model includes a task feature learning network and a second decoding network; the step of inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task processing sub-model for task processing to obtain the first predicted task result includes: The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the task feature learning network for feature learning to obtain the target task features; The target task features are input into the second decoding network for decoding processing to obtain the first prediction task result.

4. The online task processing model training method according to claim 3, characterized in that, The task feature learning network includes: a feature layer normalization network, a feature partitioning network, a first feature weighting processing network, and a first encoding network; the step of inputting the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence into the task feature learning network for feature learning to obtain target task features includes: The real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence are input into the feature layer normalization network for layer normalization processing to obtain standardized features; The standardized features are input into the feature partitioning network for feature partitioning to obtain target sparse features and target dense features. The target sparse features are input into the first feature weighting processing network for weighting processing to obtain weighted features. The weighted features are features obtained by weighting each feature in the target sparse features based on the degree of contribution of each feature in the target sparse features to learning the weighted features. The weighted features and the target dense features are input into the first encoding network for task encoding processing to obtain the target task features.

5. The online task processing model training method according to claim 2, characterized in that, The graph representation sub-model includes: a second feature weighting processing network and a second encoding network; the step of inputting the graph sampling sequence and the resource attribute information into the graph representation sub-model for graph representation learning to obtain target graph features includes: The graph sampling sequence and the resource attribute information are input into the second feature weighting processing network for weighting processing to obtain weighted graph features. The weighted graph features are graph features after weighting processing based on the degree of contribution of each attribute information in the resource attribute information to the representation of the first sample object and the historical operation resources. The weighted graph features are input into the second encoding network for graph encoding processing to obtain the target graph features.

6. The online task processing model training method according to claim 2, characterized in that, The method of parallel training of the task processing sub-model and the graph representation sub-model based on the first prediction task result, the first preset task result corresponding to the first prediction task result, the second prediction task result, and the second preset task result corresponding to the second prediction task result to obtain the target online task processing model includes: Based on the results of the first prediction task and the results of the first preset task, the first loss information is determined; Based on the results of the second prediction task and the results of the second preset task, determine the second loss information; Based on the first loss information and the second loss information, the third loss information is determined; Based on the third loss information, the task processing sub-model and the graph representation sub-model are trained in parallel to obtain the target online task processing model.

7. The online task processing model training method according to claim 1, characterized in that, The target graph structure is a graph structure with the first object identification information of the first sample object, the first resource identification information of the historical operation resource corresponding to the first sample object, and the first entity identification information of the associated entity corresponding to the historical operation resource as nodes, and the interaction operation between the first sample object and the historical operation resource, and the association relationship between the historical operation resource and the associated entity as edges. The method further includes: When the preset update time is reached, the second object identifier information of the second sample object that has performed the preset operation in the current update cycle, the second resource identifier information of the operation resource corresponding to the second sample object in the current update cycle, and the second entity identifier information of the associated entity corresponding to the operation resource are obtained. The target graph structure is updated based on the second object identification information, the second resource identification information, and the second entity identification information; The preset update time is the current update time determined based on the preset update frequency corresponding to the target graph structure, and the current update cycle is the time period between the last update time of the target graph structure and the current update time.

8. A task processing method, characterized in that, include: In response to a task processing request for a target object, obtain the target object's real-time operation sequence and target graph structure; The target real-time operation sequence is a sequence of resource identification information of multimedia resources in which the target object performs a preset interactive operation within the current time period; the current time period is the time period from the last update of the target graph structure to the current time; the target graph structure is a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are the interaction records generated by the first sample object for the preset interactive operation performed on the preset multimedia resources before the last update of the target graph structure, and the first sample object includes the target object; Based on any target online task processing model obtained by the online task processing model training method according to any one of claims 1 to 7, the target real-time operation sequence, and the target graph structure, task processing is performed to obtain the target task processing result.

9. An online task processing model training device, characterized in that, include: The first information acquisition module is configured to acquire a real-time operation sequence and a target graph structure of a first sample object. The real-time operation sequence is a sequence of resource identification information of multimedia resources in which the first sample object performs preset interactive operations within the current time period. The current time period is the time period corresponding to the last update of the target graph structure to the current time. The target graph structure is a graph structure with the first object identification information of the first sample object, the first resource identification information of historical operation resources, and the first entity identification information of the associated entities corresponding to the historical operation resources as nodes, and the interactive operations between the first sample object and the historical operation resources, as well as the association relationship between the historical operation resources and the associated entities, as edges. The node containing the first resource identifier information carries the resource attribute information corresponding to the first resource identifier information; the historical operation resource is the multimedia resource for which the first sample object has performed the preset interactive operation. A parallel training module is configured to extract resource attribute information and a graph sampling sequence from the target graph structure. The graph sampling sequence includes a first identifier sequence and a second identifier sequence. The first identifier sequence is a sequence of identifier information on a first branch in the target graph structure. The first branch includes a branch whose starting node is the node containing the object identifier information of any object in the first sample objects. The second identifier sequence is a sequence of identifier information on a second branch in the target graph structure. The second branch includes a branch whose starting node is the node containing the resource identifier information of any resource in the historical operation resources. Obtain the object attribute information of the first sample object, the historical operation sequence of the first sample object, and the historical interaction information of the historical operation resources; based on the real-time operation sequence, the object attribute information, the resource attribute information, the historical operation sequence, the historical interaction information, and the graph sampling sequence, perform parallel training on the task processing sub-model and the graph representation sub-model in the online task processing model to obtain the target online task processing model.

10. A task processing device, characterized in that, include: The second information acquisition module is configured to respond to a task processing request for a target object and acquire the target object's real-time operation sequence and target graph structure. The target real-time operation sequence is a sequence of resource identification information of multimedia resources in which the target object performs a preset interactive operation within the current time period; the current time period is the time period from the last update of the target graph structure to the current time; the target graph structure is a graph structure constructed based on the historical interaction records of the first sample object; the historical interaction records are the interaction records generated by the first sample object for the preset interactive operation performed on the preset multimedia resources before the last update of the target graph structure, and the first sample object includes the target object; The second task processing module is configured to execute task processing based on any target online task processing model obtained by the online task processing model training method according to any one of claims 1 to 7, the target real-time operation sequence, and the target graph structure, and obtain the target task processing result.

11. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the online task processing model training method as described in any one of claims 1 to 7 or the task processing method as described in claim 8.

12. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the online task processing model training method as described in any one of claims 1 to 7 or the task processing method as described in claim 8.

13. A computer program product containing instructions, characterized in that, When the instructions are executed on a computer, the computer performs the online task processing model training method as described in any one of claims 1 to 7 or the task processing method as described in claim 8.

Images