Resource overdue prediction method and device, graph network restart method and device
By building and updating resource overdue prediction map networks and using memory mapping features to process user behavior data, the problem of low prediction accuracy caused by slow data updates in the existing technology is solved, and a higher precision resource overdue prediction is achieved, ensuring that Stability and accuracy of resource management.
Patent Information
- Application Number
- CN202410897083.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-05
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-05
AI Technical Summary
In the case of failure recovery or slow data update, the existing technology cannot achieve high-precision resource overdue prediction, which makes it difficult for the server to effectively manage resource allocation projects. Especially when the failure occurs or data is sparse, the prediction accuracy is low and the accuracy cannot be accurately predicted. User resource overdue behavior.
By extracting historical data of resource allocation projects and user behavior data, building memory mapping data, and using memory units in resource overdue prediction map network to process these data, obtain memory mapping features, update resource overdue prediction map network, and learn the project recovery process The intrinsic association of lost information or sparse user data, thereby improving the accuracy of resource overdue prediction.
It realizes more accurate resource overdue prediction when the data update cycle is slow or failure recovery, improves prediction accuracy, avoids the impact of project failure or data sparse problems in the graph network, and ensures the stability and accuracy of resource management.
Smart Images

Figure CN118428565B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of this specification relate to the field of machine learning technology, and in particular to a resource overdue prediction method and device, and a graph network restart method and device. Background Art
[0002] With the development of computer technology, more and more service projects have begun to use neural network models for online prediction. By using models trained for the current service scenario, the prediction process can be completed quickly and efficiently, so that the service project can perform downstream business processing according to the prediction results. In the resource allocation scenario, users can apply for / lease resources in the resource pool provided by the server according to their needs, and the use of resources is time-limited. When the application / lease time is reached, the user needs to actively return the resource so that the server can provide users with resources with a more complete life cycle. However, since the user's behavior of using resources is relatively random, there may be overdue problems. Therefore, overdue prediction for user resource use is very important for the server. In the prior art, it is common to use neural network models to predict user overdue behavior. However, in the service project scenario, the server is more concerned about the prediction accuracy, and pays less attention to fault problems and slow data updates. This results in the inability to perform higher-precision overdue predictions through the model when a fault occurs or the data is updated slowly. Therefore, an effective solution is urgently needed to solve the above problems. Summary of the invention
[0003] In view of this, the embodiments of this specification provide a resource overdue prediction method. One or more embodiments of this specification also relate to a resource overdue prediction device, a graph network restart method, a graph network restart device, a computing device, a computer-readable storage medium and a computer program product to solve the technical defects existing in the prior art.
[0004] According to a first aspect of an embodiment of this specification, a resource overdue prediction method is provided, comprising:
[0005] In response to the overdue prediction request, extract historical data of the resource allocation project and determine user behavior data of the target user;
[0006] Constructing memory mapping data according to the historical data and the user behavior data, and processing the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain memory mapping features;
[0007] Using the memory mapping feature, the resource overdue prediction graph network is updated to a target resource overdue prediction graph network;
[0008] The user behavior data is input into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
[0009] According to a second aspect of an embodiment of this specification, another resource overdue prediction method is provided, which is applied to a cloud-side device, including:
[0010] The receiving end-side device submits an overdue prediction request for a target user participating in a resource allocation project;
[0011] In response to the overdue prediction request, extracting historical data of the resource allocation project and determining user behavior data of the target user;
[0012] Constructing memory mapping data according to the historical data and the user behavior data, and processing the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain memory mapping features;
[0013] Using the memory mapping feature, the resource overdue prediction graph network is updated to a target resource overdue prediction graph network;
[0014] Inputting the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user;
[0015] The resource overdue prediction information is fed back to the terminal side device for performing resource allocation processing on the target user.
[0016] According to a third aspect of an embodiment of this specification, a graph network restart method is provided, including:
[0017] In response to a restart instruction submitted to a preset to-be-restarted graph network of a target project, reading the to-be-updated data and the updated data of the target project;
[0018] Querying the updated data for key data according to the data to be updated, and constructing memory mapping data according to the key data;
[0019] Determining neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and constructing a memory replacement feature corresponding to the neighbor node information;
[0020] The memory replacement feature is used to restart the graph network to be restarted, and a target graph network matching the current project state of the target project is determined according to the restart result.
[0021] According to a fourth aspect of the embodiments of this specification, a resource overdue prediction device is provided, including:
[0022] an extraction module configured to extract historical data of resource allocation items and user behavior data of determined target users in response to the overdue prediction request;
[0023] a processing module configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain a memory mapping feature;
[0024] An updating module, configured to update the resource overdue prediction graph network to a target resource overdue prediction graph network using the memory mapping feature;
[0025] The prediction module is configured to input the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
[0026] According to a fifth aspect of an embodiment of this specification, another resource overdue prediction device is provided, which is applied to a cloud-side device, including:
[0027] A request receiving module, configured to receive an overdue prediction request submitted by a terminal device for a target user participating in a resource allocation project;
[0028] a data extraction module configured to extract historical data of the resource allocation project and determine user behavior data of the target user in response to the overdue prediction request;
[0029] a data processing module configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain a memory mapping feature;
[0030] An updating network module is configured to update the resource overdue prediction graph network to a target resource overdue prediction graph network by using the memory mapping feature;
[0031] A prediction information module is configured to input the user behavior data into the target resource overdue prediction graph network for processing, and obtain resource overdue prediction information corresponding to the target user;
[0032] The information sending module is configured to feed back the resource overdue prediction information to the terminal side device for performing resource allocation processing on the target user.
[0033] According to a sixth aspect of an embodiment of this specification, a network restart device is provided, including:
[0034] A data reading module is configured to read the to-be-updated data and the updated data of the target project in response to a restart instruction submitted by the to-be-restarted graph network preset for the target project;
[0035] a data query module, configured to query key data in the updated data according to the data to be updated, and construct memory mapping data according to the key data;
[0036] A feature construction module is configured to determine neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and construct a memory replacement feature corresponding to the neighbor node information;
[0037] The restart network module is configured to restart the graph network to be restarted by using the memory replacement feature, and determine a target graph network that matches the current project state of the target project according to the restart result.
[0038] According to a seventh aspect of an embodiment of this specification, there is provided a computing device, including:
[0039] Memory and processor;
[0040] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the above-mentioned resource overdue prediction method are implemented.
[0041] According to an eighth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the above-mentioned resource overdue prediction method are implemented.
[0042] According to a ninth aspect of the embodiments of this specification, there is provided a computer program product, comprising a computer program or instructions, which implement the steps of the above-mentioned resource overdue prediction method when executed by a processor.
[0043] The resource overdue prediction method provided in this embodiment can more accurately complete the user's overdue prediction in combination with the resource overdue prediction graph network that has not been iteratively updated in order to complete the user's overdue prediction more accurately when the data update cycle is slow or the fault is recovered. It can first respond to the overdue prediction request, extract the historical data of the resource allocation project and determine the user behavior data of the target user; on this basis, it can first construct memory mapping data based on the historical data and user behavior data, and map the internal connection of the data missing due to the service project failure or the slow data iteration cycle to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 is a schematic diagram of a resource overdue prediction method provided by an embodiment of this specification;
[0045] Figure 2 is a flowchart of a first resource overdue prediction method provided by an embodiment of this specification;
[0046] Figure 3 is a flow chart of a second resource overdue prediction method provided by an embodiment of this specification;
[0047] Figure 4 is a flowchart of a third resource overdue prediction method provided by an embodiment of this specification;
[0048] Figure 5 is a flow chart of a network restart method provided by an embodiment of this specification;
[0049] Figure 6 It is a structural schematic diagram of a resource overdue prediction device provided by an embodiment of this specification;
[0050] Figure 7 is a schematic diagram of the structure of another resource overdue prediction device provided by an embodiment of this specification;
[0051] Figure 8 It is a structural schematic diagram of a network restart device provided by an embodiment of this specification;
[0052] Fig. 9 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION
[0053] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.
[0054] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.
[0055] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0056] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.
[0057] In one or more embodiments of this specification, a large model refers to a deep learning model with large-scale model parameters, which usually contains hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than 10 trillion model parameters. A large model can also be called a foundation model / foundation model. The large model is pre-trained with large-scale unlabeled corpus to produce a pre-trained model with more than 100 million parameters. This model can adapt to a wide range of downstream tasks, and the model has good generalization ability, such as a large-scale language model (LLM), a multi-modal pre-training model, etc.
[0058] When the big model is used in practice, only a small number of samples are needed to fine-tune the pre-trained model and it can be applied to different tasks. The big model can be widely used in natural language processing (NLP), computer vision and other fields. Specifically, it can be applied to computer vision tasks such as visual question answering (VQA), image caption (IC), image generation, as well as natural language processing tasks such as text-based sentiment classification, text summary generation, and machine translation. The main application scenarios of the big model include digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc.
[0059] First, the terms involved in one or more embodiments of this specification are explained.
[0060] SM-TGNN: (Single Memory-based Temporal Graph Neural Network), a single memory temporal graph neural network, which uses a single memory module to store and update historical information in the graph. This memory module can be an RNN (such as LSTM or GRU) or other types of memory networks to capture dynamic changes in nodes or edges.
[0061] DM-TGNN: (Dual Memory-based Temporal Graph Neural Network), dual memory temporal graph neural network, this network is different from SM-TGNN, DM-TGNN uses two memory modules. These two memory modules may be used to capture long-term and short-term historical dependencies, or to store different types of graph information (such as structural information and attribute information).
[0062] Neural ODE: (Neural Ordinary Differential Equations) is a deep learning method based on ordinary differential equations (ODEs). It combines traditional ODE numerical solution technology with neural network models. By using ODE to model the evolution of data, NeuralODE can automatically learn the dynamic characteristics of data to perform tasks such as prediction, interpolation, and generation.
[0063] APE: (Absolute Position Encoding) usually refers to absolute position encoding. In fields such as machine learning and natural language processing, position encoding is a technique used to input the position information of elements in a sequence into the model. Absolute position encoding is a type of position encoding that assigns a unique and fixed position vector to each position.
[0064] RNN: (Recurrent Neural Network) is a special neural network structure used to process sequence data. RNN is called "recurrent" because it performs the same task for each element in the sequence, and the output depends on the previous calculation. This structure of RNN enables it to capture dynamic information in the sequence, such as time dependency or order relationship.
[0065] MHSA: (Multi-Head Self-Attention) is a key technology in the field of natural language processing and is widely used in various tasks such as machine translation, text generation, and sentiment analysis. The core idea of the multi-head self-attention mechanism is to divide a linear transformation into multiple heads, each of which performs a self-attention operation, and concatenate the outputs of all heads as the final representation.
[0066] MLP: (Multi-Layer Perceptron) is a feedforward artificial neural network model that consists of an input layer, one or more hidden layers, and an output layer. MLP can learn and approximate complex nonlinear functions through multiple layers of nonlinear transformations, thereby achieving classification or regression prediction of input data.
[0067] HA-TGA: (History-Augmented Temporal Graph Attention) is a model or method that combines historical information and graph attention mechanism to process and analyze dynamic graphs or temporal graph data. In this mechanism, "history enhancement" means that while considering the current graph structure, the historical information of past graph states or nodes / edges is also introduced to more comprehensively capture the dynamics and temporal dependencies in the graph.
[0068] TARO: (time-aligning RNN-ODE) is a deep learning model that combines recurrent neural networks (RNNs) and ordinary differential equations (ODEs), and also incorporates a time alignment mechanism.
[0069] KD: (Knowledge Distillation) Knowledge distillation is a model compression and transfer learning technology commonly used in deep learning. Its main idea is to let one model (usually a large model or teacher model) guide the training process of another model (usually a small model or student model), so that the small model can learn the knowledge of the large model and keep the performance as close to that of the large model as possible.
[0070] TGN (Temporal Graph Networks) is a general and effective framework for deep learning of dynamic graphs represented by time event sequences. It can not only capture non-Euclidean structures to generate embeddings, but also utilize the timing information in dynamic graphs. TGN is mainly composed of an encoder and a decoder. The encoder is responsible for encoding each node of the dynamic network into a vector, while the decoder calculates the predicted attribute value of the encoded vector according to the specific training problem. Finally, the loss function is designed according to the value obtained by the decoder, and the entire network is optimized by optimizing the loss function.
[0071] JODIE: (Joint Dynamic User-Item Embeddings) is a single-memory temporal graph model specifically designed for modeling user-item interactions in social network scenarios. JODIE is designed to capture dynamic embeddings of users and items, which are updated over time to reflect the interaction between users and items.
[0072] TIGE: (Temporal Interaction Graph Embedding) is a dual-memory temporal graph model used for user behavior analysis in digital financial scenarios.
[0073] In this specification, a resource overdue prediction method is provided. This specification also relates to a resource overdue prediction device, a graph network restart method, a graph network restart device, a computing device, a computer-readable storage medium and a computer program product, which are described in detail one by one in the following embodiments.
[0074] See also Figure 1 As shown in the schematic diagram, the resource overdue prediction method provided by this embodiment can be combined with the resource overdue prediction graph network that has not been iteratively updated to more accurately complete the user's overdue prediction when the data update cycle is slow or the fault is recovered. It can first respond to the overdue prediction request, extract the historical data of the resource allocation project and determine the user behavior data of the target user; on this basis, it can first construct memory mapping data based on the historical data and user behavior data, and map the internal connection of the data missing due to the service project failure or the slow data iteration cycle to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems.
[0075] See also Figure 2 , Figure 2 A flowchart of a first resource overdue prediction method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.
[0076] Step S202 , in response to the overdue prediction request, extracting historical data of the resource allocation project and determining user behavior data of the target user.
[0077] The resource overdue prediction method provided in this embodiment can be applied to any form of resource allocation scenario. For example, the resources can be computing resources, storage resources, etc. provided by the cloud computing server; they can also be items provided by the item rental platform, such as computers, mobile phones, and game consoles; they can also be funds provided by financial service providers, etc. Regardless of the type of resource, after providing available resources to users for use, there will be a resource overdue problem, and the overdue problem can easily lead to the normal maintenance of the service project. Therefore, overdue prediction for users is more important. The resource overdue prediction method provided in this embodiment can make it easier for the server to maintain the resources in the resource allocation project and make decisions for the target users by making accurate resource overdue predictions for the target users, thereby avoiding losses that affect project operations.
[0078] This embodiment uses a resource allocation project as an example of a loan project to illustrate a resource overdue prediction method. The resource overdue prediction process in other scenarios can refer to the same or corresponding description in this embodiment, and this embodiment will not be elaborated in detail here.
[0079] Specifically, the overdue prediction request specifically refers to a request triggered when a resource overdue prediction is made for a target user participating in a resource allocation project. The triggering of the request can be triggered by setting rules through the server to which the resource allocation project belongs, such as triggering at a specified time point, or detecting abnormal behavior of the user. This embodiment does not make any restrictions here. Accordingly, the resource allocation project specifically refers to a project in which the target user participates. The project can provide resources to the user for use, and the resources provided to the user need to be returned by the user on time, thus forming a leasing form. The resources can be computing resources, funds, items, etc. Accordingly, the historical data specifically refers to the attribute data of each user in the resource allocation project and the overall data of the resource usage details of each user. By analyzing the historical data, the inherent relationship in the data can be mined. Correspondingly, the target user is the user who needs to make a resource overdue prediction at the current moment, and the user behavior data is the data related to the target user. By analyzing the data, it can be determined whether the user has overdue behavior.
[0080] Based on this, in order to be able to more accurately complete the user's overdue prediction in combination with the resource overdue prediction graph network that has not been iteratively updated when the data update cycle is slow or the fault is recovered, the historical data of the resource allocation project and the user behavior data of the target user can be extracted in response to the overdue prediction request; on this basis, the memory mapping data can be first constructed according to the historical data and the user behavior data, and the internal connection of the data missing due to the service project failure or the slow data iteration cycle can be mapped to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of the sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems.
[0081] Step S204, extracting the data to be updated and the updated data from the historical data, searching the updated data for key data based on the data to be updated, and constructing memory mapping data based on the key data and the user behavior data.
[0082] In practical applications, with the continuous expansion of resource allocation projects, the existing resource overdue prediction model can no longer meet the resource overdue prediction needs for large-scale heterogeneous data processing and complex nonlinear relationship mining. The application of memory graph model provides a new idea for solving the problem. By combining the data representation ability of graph structure and the temporal memory function of memory network, the complex relationship between different users and their risk evolution pattern over time can be effectively captured. In addition, overdue prediction requires the online model performance to be highly stable over a long period of time. Since the memory graph strictly updates the memory continuously, it can ensure that the latest state is used to reason about the online data. However, when the online link fails or the online model can be replaced, it often causes the graph model to stagnate in updating the memory. Before online recovery, a lot of extra time is required to pre-update the model memory on massive data offline. This means that even if the service project is restored, the graph model built based on the memory graph cannot use the latest data to complete the new resource overdue prediction.
[0083] In view of this, the resource overdue prediction method provided in this embodiment can restart the memory graph network when the resource allocation project is restored, so as to quickly recover the resource overdue prediction, which can greatly reduce the update time overhead and memory resource usage, and flexibly use the model for online new data reasoning under time complexity. At the same time, the graph model has the ability to perceive the timing of neighbor nodes, thereby avoiding the impact on the resource allocation project and improving the accuracy of resource overdue prediction for users.
[0084] Specifically, the data to be updated specifically refers to the knowledge data missing in the resource overdue prediction graph network associated with the resource allocation project due to the link failure of the resource allocation project in the historical data; correspondingly, the updated data specifically refers to the knowledge data not missing in the resource overdue prediction graph network in the historical data; correspondingly, the key data specifically refers to the data associated with the entity matching the data to be updated in the updated data, that is, the data associated with the entity is missing due to the link failure of the project. By querying the key data, it is used to build the memory mapping data in combination with the user behavior data in the subsequent process, so that the data to be updated can be mapped to the updated data on the memory mapping data, so that the resource overdue prediction graph network can use the latest knowledge data to complete a new round of predictions. Correspondingly, the memory mapping data specifically refers to the data structure obtained after mapping the data to be updated to the key data, which is convenient for subsequent graph network updates.
[0085] Based on this, when a link failure in a resource allocation project causes the resource overdue prediction graph network to be missing data to be updated, in order to enable the resource overdue prediction graph network to add the missing data to be updated to the graph network, thereby completing the resource overdue prediction for the target user, the data to be updated and the updated data can be first extracted from the historical data, and the key data can be queried in the updated data based on the data to be updated to determine the relevant data of the entity associated with the data to be updated, and then the memory mapping data can be constructed based on the key data and user behavior data to facilitate subsequent use.
[0086] Step S206: Process the memory mapping data through the memory unit in the resource overdue prediction graph network to obtain memory mapping features.
[0087] Specifically, after the memory mapping is digitized as mentioned above, further, in order to be able to restart the resource overdue prediction graph network and apply it to the current prediction needs, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping features, and then the resource overdue prediction graph network is updated based on the features, so that the resource overdue prediction graph network can learn the missing knowledge, and then predict the resource overdue behavior of the target user after updating it to the latest state, so as to obtain more accurate resource overdue prediction information.
[0088] The resource overdue prediction graph network specifically refers to a graph network used to predict the resource overdue situation of users in a resource allocation project. Correspondingly, the memory mapping feature specifically refers to a vector expression obtained by converting the memory mapping data through a memory unit.
[0089] Furthermore, when the memory mapping data is processed into memory mapping features using the memory units in the resource overdue prediction graph network, this process needs to fully combine the memory unit characteristics to achieve feature mapping, so as to ensure that the obtained memory mapping features can fully represent the data representation corresponding to the data to be updated. In this embodiment, the specific implementation method is as follows: Step S2062 to Step S2068:
[0090] Step S2062, determining the neighbor node information associated with the memory mapping data in the resource allocation graph data of the resource allocation project, and obtaining the left memory feature corresponding to the first time node by aggregating the neighbor node information.
[0091] Specifically, the resource allocation graph data specifically refers to the graph data structure corresponding to the data associated with the resource allocation project. Correspondingly, the neighbor node information is the attribute information of the neighbor node associated with the memory mapping data in the resource allocation graph data. Correspondingly, the left memory feature specifically refers to the feature expression of the corresponding first time node obtained after aggregating the neighbor node information.
[0092] In specific implementation, the calculation of the left memory feature can be achieved as follows:
[0093] By parsing the neighbor node information, the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are obtained; the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are spliced; the splicing result is input into the aggregation unit in the resource overdue prediction graph network for processing to obtain the left memory feature corresponding to the first time node.
[0094] Based on this, when the resource overdue prediction graph network of the resource allocation project is in a period of stagnation in memory update, it will accumulate offline data to be updated, establish a historical data event record table, and determine the new data to be predicted corresponding to the target user. On this basis, the key data can be searched in the historical record table first, and then the graph attention aggregation can be performed through the neighbor node information retrieved by the online user to be predicted, so as to generate the left memory initial representation at the t"-time for restarting the calculation.
[0095] In this process, TARO can be used to restart the memory graph. First, the HA-TGA layer is used to aggregate the initial information of the target node of t"-, and the initial attribute V i As the main node information, the neighbor attribute V is used for the information of the neighbor node N(i) k , predefined absolute position encoding APE, historical interaction behavior characteristics of the master node and neighbor nodes e ik (t k ), and t"- t k The concatenation of the historical time interval codes Φ (.) is used as the neighbor node information, and then input into the HA-TGA layer for aggregation calculation. The calculation process is as follows formula (1-1):
[0096] (1-1)
[0097] in, Obtained by the following formula (1-2):
[0098] (1-2)
[0099] That is to say, by parsing the neighbor node information, the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information can be obtained; thereafter, by splicing the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information; and inputting the splicing result into the aggregation unit in the resource overdue prediction graph network for processing, the left memory feature corresponding to the first time node can be obtained.
[0100] Step S2064: Calculate the right memory replacement feature corresponding to the second time node and the left memory replacement feature corresponding to the third time node based on the left memory feature.
[0101] Specifically, after obtaining the left memory feature, in order to ensure that the final representation is obtained later, the restart memory calculation process can be decoupled, and t "+ right memory replacement feature and t'- left memory replacement feature are calculated respectively. In the calculation process, the decoupled calculation is actually completed through different models. In this embodiment, the specific implementation method is as follows:
[0102] The left memory feature is input into an event occurrence prediction model for processing to obtain a right memory replacement feature corresponding to a second time node, wherein the event occurrence prediction model is used to predict feature changes from the first time node to the second time node; the right memory replacement feature is input into a feature transfer model for processing to obtain a left memory replacement feature corresponding to a third time node, wherein the feature transfer model is used to predict feature transfer from the second time node to the third time node.
[0103] Based on this, when calculating the right memory replacement feature corresponding to the second time node and the left memory replacement feature corresponding to the third time node, the original representation of t″- Finally, in order to generate single / dual alternative representations of memory cells, a set of RNN-ODE models can be used to decouple the process of memory state change, use RNN to model the event from t"- to t"+, and generate alternative features. ; Neural ODE can also be used to model the transition process between two observations from t"+ to t'- to generate alternative representations ; Among them, the replacement feature and alternative representations It can be calculated by the following formulas (1-3) (1-4) (1-5):
[0104] (1-3)
[0105] (1-4)
[0106] (1-5)
[0107] It should be noted that the above calculation process is effective for the RNN input X at t″ t" Use the information between the main node i and the interactive node m of t', including the initial attribute V m , interactive behavior characteristics im (t"), and the concatenation of the historical interval time code Φ(.) of t'- t".
[0108] That is to say, the calculation of the right memory replacement feature and the left memory replacement feature can input the left memory feature into the event occurrence prediction model for processing to obtain the right memory replacement feature corresponding to the second time node, wherein the event occurrence prediction model is used to predict the feature changes from the first time node to the second time node, and thereafter the right memory replacement feature can be input into the feature transfer model for processing to obtain the left memory replacement feature corresponding to the third time node, wherein the feature transfer model is used to predict the feature transfer from the second time node to the third time node.
[0109] Step S2066: Time-align the right memory replacement feature and the left memory replacement feature with the memory graph corresponding to the memory unit in the resource overdue prediction graph network, and update the memory unit according to the time alignment result.
[0110] Step S2068, processing the memory mapping data through the updated memory unit in the resource overdue prediction graph network to obtain memory mapping features.
[0111] Specifically, before performing the memory mapping feature calculation, in order to enable the memory units in the resource overdue prediction graph network to realize the mapping calculation from data to features, and the mapping calculation can reflect the internal association relationship of the data, different memory units need to be updated in different ways. In this embodiment, the specific implementation method is as follows:
[0112] When the memory unit in the resource overdue prediction graph network is a single memory unit, time-aligning the left memory replacement feature with the memory graph corresponding to the single memory unit, and updating the single memory unit according to the alignment result;
[0113] When the memory unit in the resource overdue prediction graph network is a dual memory unit, the right memory replacement feature and the left memory replacement feature are time-aligned with the memory graph corresponding to the dual memory unit, and the dual memory unit is updated according to the alignment result.
[0114] On this basis, different memory mapping features will be obtained through different memory units. In this embodiment, the specific implementation method is as follows:
[0115] In the case where the memory unit is a single memory unit, the memory mapping data is processed by the updated single memory unit in the resource overdue prediction graph network to obtain a single memory mapping feature as the memory mapping feature;
[0116] In the case where the memory unit is a dual memory unit, the memory mapping data is processed by the updated dual memory unit in the resource overdue prediction graph network to obtain a dual memory mapping feature as the memory mapping feature.
[0117] Based on this, after obtaining the alternative representation, the alternative representation can be used to form a time alignment with the memory graph, and the memory unit can be reloaded to migrate the knowledge from the memory graph main model to the restart network. On this basis, the memory-based timing graph can be used to calculate the representation of the main node at time t, and the neighbors can be aggregated using the temporal attention mechanism shared with the parameters before the restart to generate the final representation of the new data node, that is, the memory map feature.
[0118] It can be understood that when the temporal graph model is restarted using the left / right substitution feature, the substitution representation can be used for the graph model of a single memory unit. Restart the memory unit. For the graph model with two memory units, you can use alternative features and alternative representations Restart the right and left memory units respectively to complete the memory reloading of the resource overdue prediction graph network, and use the backbone network of the time series graph model to generate the target node representation , that is, the memory mapping feature. Among them, the single memory timing graph network can be expressed as the following formula (1-6):
[0119] (1-6)
[0120] Among them, the dual memory timing graph network can be expressed as the following formula (1-7):
[0121] (1-7)
[0122] It should be noted that a variety of methods can be used to memorize the timing graph, for example, a single memory graph network can use a TGN model, a dual memory graph network can use a TIGE model, etc. This embodiment does not make any limitation here.
[0123] That is to say, on the one hand, when the memory unit in the resource overdue prediction graph network is a single memory unit, the left memory replacement feature can be time-aligned with the memory graph corresponding to the single memory unit, and the single memory unit can be updated according to the alignment result; thereafter, the memory mapping data can be processed through the updated single memory unit in the resource overdue prediction graph network to obtain a single memory mapping feature as the memory mapping feature.
[0124] On the other hand, when the memory unit in the resource overdue prediction graph network is a dual memory unit, the right memory replacement feature and the left memory replacement feature can be time-aligned with the memory graph corresponding to the dual memory unit, and the dual memory unit can be updated according to the alignment result. Thereafter, the memory mapping data can be processed by the updated dual memory unit in the resource overdue prediction graph network to obtain the dual memory mapping feature as the memory mapping feature.
[0125] Step S208: using the memory mapping feature to update the resource overdue prediction graph network to a target resource overdue prediction graph network.
[0126] Specifically, after obtaining the memory mapping features corresponding to the memory mapping data, in order to enable the resource overdue prediction graph network to combine the missing knowledge to complete the resource overdue behavior prediction of the target user, the resource overdue prediction graph network can be updated to the target resource overdue prediction graph network using the obtained memory mapping features for subsequent use. The target resource overdue prediction graph network specifically refers to the resource overdue prediction graph network that has learned the actual knowledge data in the link recovery process.
[0127] Furthermore, when the graph network is updated using the memory mapping feature, in order to ensure a more comprehensive update and facilitate subsequent use, the multi-dimensional information of the memory unit can be combined. In this embodiment, the specific implementation method is as follows:
[0128] The neighbor node features, second position coding information, second interaction behavior features, and second time interval coding information associated with the memory mapping features are extracted in the memory unit; the neighbor node features, the second position coding information, the second interaction behavior features, and the second time interval coding information are spliced to obtain neighbor node prior features; the neighbor node prior features and the memory mapping features are aggregated, and the resource overdue prediction graph network is updated to the target resource overdue prediction graph network according to the aggregation result.
[0129] Based on this, the memory mapping feature is obtained through the memory graph model Afterwards, the historically enhanced time-series graph attention can be used to aggregate the neighbor node information of the primary node, calculate the final fusion representation of the primary node, and use it for overdue prediction of new data for online rapid recovery.
[0130] In this process, since the graph attention may contain multiple layers, the result of TGNN calculation can be expressed as , which represents the main node information of the l-1 layer participating in the attention calculation. Participating neighbor nodes The aggregated vivid information is represented as , including the neighbor node representation obtained from the current memory unit , Absolute Position Encoding APE, Interaction Behavior Characteristics between Primary Node i and Historical Neighbor Node k , and in tt k The concatenation of the historical time interval codes Φ(.). Here, the time window is set to take the most recent Neighbor nodes participate in the graph attention calculation, making the model historically perceptible and controlling the size of the temporal receptive field. Finally, the information representation of all neighbor nodes participating in neighbor aggregation is obtained. , which is achieved by (1-8) and (1-9) as follows:
[0131] (1-8)
[0132] (1-9)
[0133] After obtaining the neighbor prior information, a standard MHSA and multiple linear layers can be used to parse the master node information. and neighbor node information Perform aggregation to obtain the final enhanced representation of the master node for new data overdue prediction , thereby completing the online rapid recovery, so that in the application stage, resource overdue prediction can be carried out for target users. The calculation of can be calculated by the following formulas (1-10), (1-11) and (1-12):
[0134] (1-10)
[0135] (1-11)
[0136] (1-12)
[0137] It should be noted that the query Q in MHSA h Features from the master node with zero time code Φ(0) , key K c Value V c From neighbor node prior , the parameters of the HA-TGA layer used before TARO and the HA-TGA layer used after TGNN are shared.
[0138] In addition, after the graph network is restarted, in order to enable the graph network to directly predict resource overdue behavior for target users during the application phase, the graph network needs to be optimized. In this embodiment, the specific implementation method is as follows:
[0139] The resource overdue prediction graph network is updated by using the memory mapping feature to obtain an initial resource overdue prediction graph network; a network optimization loss function and a knowledge transfer loss function corresponding to the initial resource overdue prediction graph network are determined; and parameters of the initial resource overdue prediction graph network are adjusted based on the network optimization loss function and the knowledge transfer loss function until a target resource overdue prediction graph network that meets the training stop condition is obtained.
[0140] Based on this, in the knowledge transfer optimization stage of the restarted graph network, in the process of training the restart module TARO and the memory graph main model, in order to transfer the main model's modeling ability of memory representation to TARO, so as to accelerate the calculation during reasoning, the knowledge distillation method can be used to minimize the distance between the real memory of the memory graph model and the alternative memory generated by the restart module TARO, using square The norm constrains the generation of memory and is added as an additional optimization target to the learning of the main target overdue prediction task to complete knowledge transfer. The overdue prediction adopts the optimization target of binary cross loss, which is as follows (1-13):
[0141] (1-13)
[0142] The optimization objectives of knowledge transfer for single memory graph and dual memory graph are as follows (1-14) and (1-15) respectively:
[0143] (1-14)
[0144] (1-15)
[0145] The final optimization goal of the task is as follows (1-16):
[0146] (1-16)
[0147] Among them, α and β are hyperparameters that control the loss size of different objectives.
[0148] In addition, when generating alternative representations for a single memory graph model, HA-TGA can also be used to directly generate alternative representations on t' It is used to restart the graph, so that the same effect can be achieved without going through RNN-ODE. Second, when modeling the impact of events using RNN, you can also use LSTM networks to consider the impact of long and short cycles, or use MLP networks for streamlined modeling. Third, when transferring knowledge, in addition to minimizing the distance between alternative memories and real memories, you can also calculate all memory values involved in an update (a single update of a single-memory graph model involves two memories, and a dual-memory graph model designs four memories), and minimize the distance constraints between all involved memories and the real memory of the main model. Fourth, in addition to the square, the distance constraint is minimized. Norm form, or norm, mean square error, etc.
[0149] Step S210: input the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
[0150] Specifically, after obtaining the resource overdue prediction graph network that is updated in time due to link failure, and after the above processing to obtain the restarted target resource overdue prediction graph network, the user behavior data can be input into the target resource overdue prediction graph network for processing, so that after recovery, overdue behavior prediction can be performed directly for the target user, thereby obtaining the resource overdue prediction information corresponding to the target user for downstream use.
[0151] In summary, in order to solve the long-standing problem of long-term offline update and time-consuming online graph models in resource allocation projects, the flexibility of the model applied to resource overdue prediction is limited, and it is difficult to switch between different data test targets. To address this defect, the resource overdue prediction method provided in this embodiment can perform knowledge distillation on the memory representation capability of the graph model and migrate it to the restart network, thereby achieving online rapid recovery of overdue prediction services with minimal data under time complexity, and being able to flexibly reason and switch between new data. In addition, the graph model has the ability to perceive the timing of neighboring nodes during aggregation, enhancing the aggregation representation of the main node.
[0152] See also Figure 3 , Figure 3 A flow chart of a second resource overdue prediction method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.
[0153] Step S302 , in response to the overdue prediction request, extracting historical data of the resource allocation project and determining user behavior data of the target user.
[0154] Step S304: extracting sparse user data associated with the target user from the historical data, and constructing memory mapping data based on the sparse user data and the user behavior data.
[0155] The second resource overdue prediction method provided in this embodiment, the contents not described in detail can refer to the same or corresponding descriptions in the above embodiments, and this embodiment will not be repeated here.
[0156] In practical applications, in the field of resource allocation risk control, graph network technology plays a very important role in the platform's identification of malicious fraud users and reducing the risk of overdue resource allocation for users. Among them, the accuracy of resource overdue prediction directly determines the scale of bad resources and is an important part of credit risk management. Due to the diversity of user-platform interaction behaviors, and the behaviors have strong dynamic and time-series characteristics, it is generally appropriate to use continuous-time dynamic graphs to model users. Existing methods are mainly based on memory-based and non-memory-based time-series graph neural networks. However, due to the fact that the historical interaction behaviors of a large number of users on the platform are very sparse, the features of users who have not interacted for a long time cannot be updated, resulting in the existing methods being plagued by the problem of obsolete node representations for a long time. The credit risk of users during this period may change dramatically due to personal reasons but cannot be captured; or the user is a long-term fraud user, and the fraud risk during the incubation period accumulates rapidly and cannot be perceived. These will greatly affect the effect of overdue prediction and ultimately cause direct losses to the economy.
[0157] In view of this, the resource overdue prediction method provided in this embodiment can supplement the data sparsity situation and realize the ability to learn efficient old representations. It can effectively solve the difficulties such as sparse user behavior and low activity representation that are common on the current platform, reduce the resource overdue rate, and prevent platform resource losses.
[0158] Specifically, sparse user data refers to that the user corresponding data in historical data is relatively sparse, or the user corresponding data has not been updated for a long time.
[0159] Based on this, in order to make accurate resource overdue predictions for target users when user data is sparse, sparse user data associated with the target users can be extracted from historical data. Afterwards, memory mapping data can be constructed based on the sparse user data and user behavior data for subsequent use.
[0160] Step S306: Process the memory mapping data through the memory unit in the resource overdue prediction graph network to obtain memory mapping features.
[0161] Step S308: using the memory mapping feature to update the resource overdue prediction graph network to a target resource overdue prediction graph network.
[0162] Furthermore, when the memory mapping data is processed into memory mapping features through the memory unit in the resource overdue prediction graph network, considering that the memory unit is a dual memory unit, it can be implemented through the following steps S3062 to S3068:
[0163] Step S3062, extracting left memory state features, right memory state features, and cache edge features from the memory units in the resource overdue prediction graph network.
[0164] In the specific implementation, when the memory unit is a dual memory unit, in order to ensure that the resource overdue prediction graph network can have higher prediction accuracy and can also complete the prediction for sparse data, when constructing the state feature, it can be implemented in the following way:
[0165] A left memory state feature group and a right memory state feature group are extracted from the memory unit in the resource overdue prediction graph network, wherein the left memory state feature group records the memory state features before the event is updated, and the right memory state feature group records the memory state features after the event is updated; an initial left memory state feature corresponding to the sixth time node is selected from the left memory state feature group as the left memory state feature; and an initial right memory state feature corresponding to the seventh time node is selected from the right memory state feature group as the right memory state feature.
[0166] Based on this, in this stage, the left and right representations of the initial memory unit can be input, the resource data CTDG can be input into the dual memory unit, the left and right memory states s(t"+) and s(t'-) can be loaded from the dual memory unit on the current batch, and the cached edge features e(t') can be loaded from the native message queue for subsequent use.
[0167] Specifically, the dual memory unit uses a set of left memory M (-) and right memory M (+) to represent the node at each timestamp, which respectively represent the representation before the event update and the representation after the event update. According to the correct training and update order of the model, at the beginning of the prediction phase at each time t, M (+) stores the state s (t "+) of the node representation just after t " and at t "+, and (-) stores the state s (t "-) of the node representation just before t "+ and at t ". The goal is to predict the node representation h (t-) just before the current time t. Among them, t "and t 'are the time points when the target node interacted in the last and last time, respectively.
[0168] Based on this, when calculating the memory state features, the left memory state feature group and the right memory state feature group can be extracted from the memory unit in the resource overdue prediction graph network, wherein the left memory state feature group records the memory state features before the event is updated, and the right memory state feature group records the memory state features after the event is updated; the initial left memory state feature corresponding to the sixth time node is selected in the left memory state feature group as the left memory state feature; at the same time, the initial right memory state feature corresponding to the seventh time node can be selected in the right memory state feature group as the right memory state feature for subsequent use.
[0169] Step S3064: Generate a message feature corresponding to the fourth time node according to the left memory state feature, the right memory state feature and the cache edge feature, and obtain a target right memory feature by jump updating the message feature.
[0170] Specifically, when the message feature corresponding to the fourth time node is generated according to the left memory state feature, the right memory state feature and the cache edge feature, in order to accurately map the feature to the message, a splicing mapping method can be used. In this embodiment, the specific implementation method is as follows:
[0171] Calculate the time interval between the sixth time node and the seventh time node, and determine the third time interval encoding information corresponding to the time interval; splice the left memory state feature, the right memory state feature, the cache edge feature and the third time interval encoding information, and perform message mapping on the splicing result to obtain the message feature corresponding to the fourth time node.
[0172] Based on this, when performing message mapping, at the beginning of the current batch, the state of the target node can be obtained from the dual memory unit, that is, s(t'-) is loaded from M(-), and s(t"+) is loaded from M(+). In addition, the edge feature e(t') on t' is obtained from the native message queue, and the time encoding vector Φ(.) is calculated for the time interval t'- t". Assume that i and j are two nodes of the event to be predicted at the target time t, m and n are the nodes that interact at a time t', and the message g is generated at time t'. i (t')g j (t'), where g i (t') and g j (t') can be realized by the following formulas (2-1) and (2-2):
[0173] (2-1)
[0174] (2-2)
[0175] It should be noted that Concat is a feature concatenation operation. In order to reduce computational overhead in the native message queue, only the information of the neighbor nodes m and n that most recently interacted with i and j is retained.
[0176] That is to say, the time interval between the sixth time node and the seventh time node can be calculated, and the third time interval coding information corresponding to the time interval can be determined at the same time; at this time, the left memory state feature, the right memory state feature, the cache edge feature and the third time interval coding information are spliced, and message mapping is performed on the splicing result to obtain the message feature corresponding to the fourth time node.
[0177] Furthermore, after obtaining the message features, a jump update may be performed for the message features. In this embodiment, the specific implementation is as follows:
[0178] The target state feature is determined according to the message feature, and the initial right memory feature is calculated based on the target state feature; the initial right memory feature is jump-updated to obtain the target right memory feature.
[0179] That is to say, after obtaining the left and right memory states s(t"+), s(t'-) and edge features e(t'), we can generate the time interval encoding vector Φ(t'- t"), use the existing messages for message mapping to generate the t' time message g(t'), and then use the MHSA network for jump update to generate a new right memory representation .
[0180] Based on this, when performing a jump update, MHSA modeling can be used to update the memory of the interactive event from state s i (t") to h i The updating process of (t"+) can be realized by the following formulas (2-3) and (2-4):
[0181] (2-3)
[0182] (2-4)
[0183] Among them, s i (t'-) as the query Q in MHSA s The main semantics of is concatenated with the time encoding vector Φ(0) as the query, key (K g ) and value (V g ) comes from the message g i (t').
[0184] Step S3066, calculating the target left memory feature corresponding to the fifth time node based on the target right memory feature.
[0185] Step S3068: Process the target right memory feature and the target left memory feature through the memory unit in the resource overdue prediction graph network to obtain a memory mapping feature.
[0186] Specifically, when calculating the target left memory feature, we can use the continuous memory NDP to model the state transition process and use the neural ODE network to generate an approximate estimate of the left memory representation at time t. , using link prediction labels as enhanced auxiliary representations , and generate the time interval encoding vector ; On this basis, the current and global memory prior semantic representation encoding is constructed for all edges to be predicted. Then, RNN and another neural ODE network are used to model the posterior distribution and prior distribution of state transition respectively. After that, a hidden variable can be sampled from the constructed prior distribution and injected into the new right memory The new left memory of the target node can be calculated Finally, the TGA network is used to aggregate the historical neighbor node information and fuse it with the target node to generate the final representation. .
[0187] Further, when calculating the target left memory feature corresponding to the fifth time node based on the target right memory feature, it can be implemented by a continuous memory prediction model. In this embodiment, the specific implementation method is as follows:
[0188] The target right memory feature is input into the continuous memory prediction model; the target right memory feature is processed by the semantic coding unit in the continuous memory prediction model to obtain the memory coding feature; the memory coding feature is processed by the state transfer construction unit in the continuous memory prediction model to obtain the memory distribution feature; the memory distribution feature is processed by the memory generation unit in the continuous memory prediction model to obtain the target left memory feature corresponding to the fifth time node.
[0189] Based on this, when calculating the target left memory feature, the subtle evolution trajectory of the user state in the sparse information environment during the event-free period is actually regarded as a dynamically changing random process M t , the hidden variable is Z t , using continuous memory NDP (CMNDP) to characterize the distribution of potential memory transfer from t'+ to t-, and predict the continuous dynamic changes of node representation, which can be achieved by the following formula (2-5):
[0190] (2-5)
[0191] Among them, H t To utilize the prior semantic information of memory, the goal is to generate target node features CMNDP consists of three parts: a semantic encoder, a potential distribution construction for continuous state transitions, and a new memory generator.
[0192] That is to say, the target right memory feature can be first input into the continuous memory prediction model; at this time, the target right memory feature can be processed by the semantic coding unit in the continuous memory prediction model to obtain the memory coding feature; further, the memory coding feature can be processed by the state transfer construction unit in the continuous memory prediction model to obtain the memory distribution feature; finally, the memory distribution feature can be processed by the memory generation unit in the continuous memory prediction model to obtain the target left memory feature corresponding to the fifth time node.
[0193] When processing the target right memory feature through the semantic encoder in the continuous memory prediction model, it can be achieved as follows:
[0194] The target right memory feature is input into the semantic coding unit in the continuous memory prediction model; the node similarity feature, the memory original prior feature and the enhanced auxiliary feature corresponding to the target right memory feature are constructed through the semantic coding unit; the initial memory coding feature is constructed based on the node similarity feature, the memory original prior feature and the enhanced auxiliary feature, and the initial memory coding feature is encoded to obtain the memory coding feature.
[0195] Based on this, the prior semantics used by the semantic encoder can include the , this process can be done using a neural ODE network The generated node approximation representation at t- , the calculation process can be achieved by the following formula (2-6):
[0196] (2-6)
[0197] At the same time, we can also encode the time vector Φ(.) in the event-free interval t- t', and concatenate the features to obtain the memory original prior of node i, which can be achieved by the following formula (2-7):
[0198] (2-7)
[0199] It should be noted that when using MLP to form a memory prior for the event edge between the current i and j, the label of the link prediction task can be used as an enhanced auxiliary representation , the formation of memory prior can be achieved through the following formula (2-8):
[0200] (2-8)
[0201] The memory prior information of all edges to be predicted at time t is pooled and averaged to obtain the memory prior semantics that are finally used after processing, which can be achieved by the following formula (2-9):
[0202] (2-9)
[0203] That is to say, for encoding nodes, the target right memory feature can be input into the semantic encoding unit in the continuous memory prediction model; at this time, the node similarity feature, memory original prior feature and enhanced auxiliary feature corresponding to the target right memory feature can be constructed through the semantic encoding unit; thereafter, the initial memory encoding feature can be constructed based on the node similarity feature, memory original prior feature and enhanced auxiliary feature, and then the initial memory encoding feature can be encoded to obtain the memory encoding feature.
[0204] Furthermore, when the state transfer construction unit is used to process the encoding features, it can be implemented in the following way:
[0205] The memory encoding feature is input into the state transfer construction unit in the continuous memory prediction model; feature distribution information corresponding to the memory encoding feature is constructed by the state transfer construction unit; and memory distribution features are calculated according to the feature distribution information.
[0206] Based on this, in the potential distribution construction stage of continuous state transition, after obtaining the memory prior H at the current running time t After that, in order to obtain a global memory prior representation and model the dynamic random process M t , we can use autoregressive approach to model M t , which can be achieved by the following formula (2-10):
[0207] (2-10)
[0208] Using RNN at all running times H t To extract global information, we use latent variables r t As a global semantic representation. And because RNN estimates the distribution of information based on each new observed data, RNN models the posterior distribution , where r t It can be calculated by the following formula (2-11):
[0209] (2-11)
[0210] Take another neural ODE Construct a dynamic distribution on the observed samples, and push the memory transfer distribution at each time point in the prediction phase, so the neural ODE models the prior distribution , each time a hidden variable is sampled from the Gaussian prior to represent the current dynamic distribution , at this time rt It can be calculated by the following formula (2-12):
[0211] (2-12)
[0212] That is to say, the memory encoding feature can be input into the state transfer construction unit in the continuous memory prediction model; at this time, the feature distribution information corresponding to the memory encoding feature is constructed through the state transfer construction unit; and the memory distribution feature can be calculated according to the feature distribution information.
[0213] Finally, in the new memory generator stage, the latent variable r of memory transfer at the current moment of modeling can be obtained by sampling t Then, the hidden variables are injected into the generated , accurately track the subtle state evolution trajectory, generate a new memory representation of the target node at time t-, and prevent the problem of stale representation in a sparse information environment during the event-free period. It can be calculated by the following formula (2-13):
[0214] (2-13)
[0215] In completion After the calculation of , the temporal attention network can be used to aggregate the information of historical neighbor nodes and generate the target node for predicting the final representation , and use an additional MLP classifier network to predict the overdue of the final representation of the user node, and then get the overdue prediction result of the user's current event. This can be achieved through the following formulas (2-14) and (2-15):
[0216] (2-14)
[0217] (2-15)
[0218] In addition, in order to be able to update the dual memory unit status in real time, the right memory state feature group can also be updated according to the target right memory feature, and the left memory state feature group can be updated according to the memory mapping feature; wherein, the updated right memory state feature group and the updated left memory state feature group are used to determine the left memory state feature and the right memory state feature in the next resource overdue prediction cycle.
[0219] Based on this, for the dual memory unit state update processing, the target node representation result generated by the jump update can be used to update the right memory unit state, and the final representation result can be used to update the left memory unit state for the next timestamp processing. The update process can be implemented by the following formula (2-16):
[0220] (2-16)
[0221] In addition, firstly, the MHSA in the jump update module can be replaced by other timing models, for example, RNN is used to model the update impact of events on nodes; secondly, different neighbor node aggregation methods can be used in the TGA network module, such as maximizing Max-pooling, average pooling, and direct addition. Thirdly, in addition to a single sampling, the sampling method of the latent variable Zt can also sample multiple items at a time and take the average, thereby increasing the stability of model training. Fourthly, other types of classifiers besides MLP can be used in the final spliced classifier, such as a multi-layer convolutional neural network. Fifthly, when encoding the global memory semantics in CMNDP, in addition to RNN, other autoregressive timing models can be used, such as LSTM networks, autoregressive GPT decoders, etc. This embodiment is not limited in any way.
[0222] Step S310: input the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
[0223] Specifically, after obtaining the resource overdue prediction graph network with low model prediction accuracy due to data sparsity, after the above processing to obtain the restarted target resource overdue prediction graph network, the user behavior data can be input into the target resource overdue prediction graph network for processing, so that the overdue behavior prediction can be directly performed for the target user after recovery, thereby obtaining the resource overdue prediction information corresponding to the target user for downstream use.
[0224] In summary, in solving the problem of users with thin information, sparse behavior, and low activity that are common in resource allocation projects, the existing time series graph model has outdated representations of these users and poor risk prediction capabilities. The method provided in this embodiment can model user behavior into a continuous time dynamic graph, and through a dual memory time series graph neural network model with anti-stale representation learning capabilities, it achieves a stronger overdue risk prediction capability for users with thin information, sparse behavior, and low activity. At the same time, a continuous memory NDP process is proposed for the potential distribution of memory state transfer to effectively model the state of user memory updates in the model, and the entire overdue prediction framework has good robustness and interpretability.
[0225] See also Figure 4 , Figure 4 A flowchart of a third resource overdue prediction method provided according to an embodiment of this specification is shown, and the method is applied to a cloud-side device, including:
[0226] Step S402, receiving an overdue prediction request submitted by a target user participating in a resource allocation project by a terminal-side device;
[0227] Step S404, in response to the overdue prediction request, extracting historical data of the resource allocation project and determining user behavior data of the target user;
[0228] Step S406, constructing memory mapping data according to the historical data and the user behavior data, and processing the memory mapping data through the memory unit in the resource overdue prediction graph network to obtain memory mapping features;
[0229] Step S408, using the memory mapping feature to update the resource overdue prediction graph network to a target resource overdue prediction graph network;
[0230] Step S410, inputting the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user;
[0231] Step S412: Feedback the resource overdue prediction information to the terminal side device for performing resource allocation processing on the target user.
[0232] Specifically, the cloud-side device specifically refers to the device that runs the method on the server side that provides the resource overdue prediction method, and correspondingly, the terminal-side device specifically refers to the device held by the user who has the resource overdue prediction demand for the user. That is to say, when the user holding the terminal-side device has the resource overdue prediction demand, the overdue prediction request can be sent to the cloud-side device through the set interface, and then the resource overdue prediction information of the specified user is generated through the resource overdue prediction service provided by the cloud-side device, and the resource overdue prediction information is fed back to the terminal-side device so that the terminal-side device can perform subsequent service processing.
[0233] Among them, the resource overdue prediction process of the cloud-side device can refer to the same or corresponding description content in the above embodiment, and this embodiment will not be elaborated in detail here.
[0234] The resource overdue prediction method provided in this embodiment can more accurately complete the user's overdue prediction in combination with the resource overdue prediction graph network that has not been iteratively updated in order to complete the user's overdue prediction more accurately when the data update cycle is slow or the fault is recovered. It can first respond to the overdue prediction request, extract the historical data of the resource allocation project and determine the user behavior data of the target user; on this basis, it can first construct memory mapping data based on the historical data and user behavior data, and map the internal connection of the data missing due to the service project failure or the slow data iteration cycle to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems.
[0235] See also Figure 5 , Figure 5 A flowchart of a method for restarting a network according to an embodiment of the present specification is shown, and the method specifically includes the following steps:
[0236] Step S502 , in response to a restart instruction submitted to a preset graph network to be restarted for a target project, read the data to be updated and the updated data of the target project.
[0237] The graph network restart method provided in this embodiment can be applied to any scenario, and the graph network to be restarted that has not been updated in time can be updated to the target graph network corresponding to the current project status of the target project, so as to support the target project to use the latest graph network for subsequent predictions, wherein the target project can be any project that uses the graph network to predict information within the project, such as cloud service projects, transaction projects, credit loan projects, etc. The graph network deployed in the project can be used to make expected predictions for users to obtain prediction information for project execution. For example, the overdue behavior of the user's use of resources can be predicted, and the user's credit rating can also be predicted. This embodiment does not make any limitations here.
[0238] Specifically, since the target project uses a graph network to forecast project demand, but when the online link fails or the online model can be replaced, the graph network often stagnates in memory updates. Before the online project is restored, a lot of extra time is required to pre-update the graph network memory offline on massive data. This results in that even if the service project is restored, the graph network built based on the memory graph cannot apply the latest data to complete new predictions. In view of this, the graph network restart method provided in this embodiment can restart the graph network during the target project recovery phase, and when restarting, the graph network can be updated to the target graph network corresponding to the current state of the target project, supporting the prediction of the latest state, so as to achieve the continued provision of stable and accurate prediction services after restart. In other words, during the graph network restart phase of the target project, the graph network to be restarted can be trained to be a target graph network that matches the current state of the project, so that it has more accurate prediction performance.
[0239] Furthermore, the graph network to be restarted specifically refers to a graph memory neural network used for demand forecasting in the target project due to problems such as interruption or failure of the target project, and the graph network has not been updated and has a certain lag. Correspondingly, the restart instruction specifically refers to an instruction triggered by a pointer to the graph network to be restarted, which is used to perform subsequent automated graph restart operations. Correspondingly, the data to be updated specifically refers to the knowledge data that is missing in the historical data corresponding to the target project and associated with the graph network to be restarted due to a link failure. Correspondingly, the updated data specifically refers to the knowledge data that is not missing in the historical data of the graph network to be restarted.
[0240] Based on this, when the target project accumulates historical data due to a link failure, and the graph network to be restarted preset in the project has not learned the accumulated historical data, the graph network to be restarted needs to be updated to update the graph network to be restarted to the target graph network corresponding to the current target project state, so that it can be supported for use in the target project. Therefore, at this time, the restart instruction for the graph network to be restarted preset for the target project can be triggered, so that the missing data to be updated and the updated data learned by the graph network can be read from the historical data of the target project according to the restart instruction, so as to restart the graph network for use later.
[0241] Step S504: searching the updated data for key data according to the data to be updated, and constructing memory mapping data according to the key data.
[0242] Specifically, key data refers to the data associated with the entity that matches the data to be updated in the updated data, that is, the data associated with the entity is missing due to the link failure of the project. By querying the key data, it is used to construct the memory mapping data later, and the data to be updated can be mapped to the updated data on the memory mapping data, so that the graph network to be restarted can use the latest knowledge data to complete a new round of predictions. Correspondingly, memory mapping data refers to the data structure obtained after mapping the data to be updated to the key data, which is convenient for the subsequent restart of the graph network.
[0243] Based on this, after obtaining the data to be updated and the updated data, since the graph network lacks the knowledge corresponding to the data to be updated, but has the knowledge of the updated data, in order to enable the graph network to be restarted to quickly learn the knowledge corresponding to the data to be updated, the key data can be first queried in the updated data according to the data to be updated. After determining that there is key data with related knowledge content, it is constructed as memory mapping data for restarting the graph network to be restarted, and the restart operation can be performed.
[0244] Step S506, determining the neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and constructing a memory replacement feature corresponding to the neighbor node information.
[0245] Step S508: restart the graph network to be restarted by using the memory replacement feature, and determine a target graph network that matches the current project state of the target project according to the restart result.
[0246] Specifically, the allocation graph data specifically refers to the graph data structure corresponding to the data associated with the target project. Correspondingly, the neighbor node information is the attribute information of the neighbor nodes associated with the memory mapping data in the allocation graph data. Correspondingly, the memory replacement feature specifically refers to the feature used to update the memory unit in the network when updating the graph network to be restarted, which has the knowledge content missing from the graph network to be restarted. Correspondingly, the target graph network is the graph network obtained after restarting the graph network to be restarted, and the target graph network matches the current project status of the target project, and can provide stable and accurate prediction services.
[0247] Based on this, after completing the construction of the memory mapping data, considering that the graph network to be restarted is a graph structure network, the neighbor node information associated with the memory mapping data can be determined in the allocation graph data of the target project first, and then the memory replacement feature corresponding to the neighbor node information can be completed in combination with the neighbor node information; on this basis, the memory replacement feature can be used to restart the graph network to be restarted, and the memory unit in the graph network to be restarted can be updated to the latest state, so that the target graph network that matches the current project state of the target project can be determined according to the restart result, so that the target project can continue to perform the corresponding prediction service through the target graph network. For example, predict the overdue behavior of users.
[0248] Furthermore, when constructing the memory replacement feature corresponding to the neighbor node information, it can be achieved by aggregating the neighbor node information. In this embodiment, the specific implementation method is as follows:
[0249] The left memory feature corresponding to the target time node is obtained by aggregating the neighbor node information; and the memory replacement feature corresponding to the associated time node is calculated according to the left memory feature.
[0250] Specifically, the left memory feature specifically refers to the feature expression of the corresponding target time node obtained after aggregating the neighbor node information. The memory replacement feature specifically refers to the right memory replacement feature corresponding to the first associated time node and the left memory replacement feature corresponding to the second associated time obtained after feature calculation based on the left memory feature.
[0251] In practical applications, the calculation of the left memory feature can be achieved in the following way: by parsing the neighbor node information, the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are obtained; the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are spliced; the splicing result is input into the aggregation unit in the network to be restarted for processing, and the left memory feature can be obtained.
[0252] In this process, TARO can be used to restart the memory graph. First, the HA-TGA layer is used to aggregate the initial information of the target node of t"-, and the initial attribute V i As the main node information, the neighbor attribute V is used for the information of the neighbor node N(i) k , predefined absolute position encoding APE, historical interaction behavior characteristics of the master node and neighbor nodes e ik (t k ), and t"- t kThe concatenation of the historical time interval codes Φ(.) is used as the neighbor node information, and then input into the HA-TGA layer for aggregation calculation. The calculation process is implemented as shown in the above formulas (1-1) and (1-2).
[0253] Furthermore, the calculation of the memory replacement feature can be realized in the following manner: the left memory feature is input into the event occurrence prediction model for processing to obtain the right memory replacement feature corresponding to the first time node, wherein the event occurrence prediction model is used to predict the feature change from the target time node to the first time node; the right memory replacement feature is input into the feature transfer model for processing to obtain the left memory replacement feature corresponding to the second time node, wherein the feature transfer model is used to predict the feature transfer from the first time node to the second time node. The right memory replacement feature and the left memory replacement feature are used as memory replacement features, which can be used for subsequent graph restart operations.
[0254] In practical applications, when calculating the right memory replacement feature corresponding to the first time node and the left memory replacement feature corresponding to the second time node, the original representation of t″- Finally, in order to generate single / dual alternative representations of memory cells, a set of RNN-ODE models can be used to decouple the process of memory state change, use RNN to model the event from t"- to t"+, and generate alternative features. ; Neural ODE can also be used to model the transition process between two observations from t"+ to t'- to generate alternative representations ; Among them, the replacement feature and alternative representations It can be calculated using the above formulas (1-3), (1-4), and (1-5).
[0255] On this basis, after the replacement features are calculated, the memory replacement features can be time-aligned with the memory graph corresponding to the memory unit in the graph network to be restarted, and the memory unit can be updated according to the time alignment result as a restart of the graph network to be restarted.
[0256] Specifically, when the memory unit in the graph network to be restarted is a single memory unit, the left memory replacement feature can be time-aligned with the memory graph corresponding to the single memory unit, and the single memory unit can be updated according to the alignment result; when the memory unit in the graph network to be restarted is a dual memory unit, the right memory replacement feature and the left memory replacement feature can be time-aligned with the memory graph corresponding to the dual memory unit, and the dual memory unit can be updated according to the alignment result. After completing the update of the memory unit, the graph network to be restarted can be updated to the target graph network corresponding to the current project state of the target project, which can be used for project prediction.
[0257] It should be noted that the contents not fully described in the graph network restart method provided in this embodiment, as well as the description contents that are the same or corresponding to the above-mentioned resource overdue prediction method embodiment, can all be referred to the description in the above-mentioned embodiment, and this embodiment will not be elaborated in detail here.
[0258] The graph network restart method provided in this embodiment can solve the problem of limited network prediction flexibility caused by the time-consuming update of the graph network over a long period of time. It can realize the knowledge distillation of the memory representation ability of the graph network and migrate it to the restart network, and realize the online rapid recovery of the project under time complexity with very small amounts of data, and can flexibly reason and switch between new data. In addition, the graph network has the ability to perceive the timing of neighbor nodes during aggregation, enhancing the aggregation representation of the main node.
[0259] Corresponding to the above method embodiment, this specification also provides a resource overdue prediction device embodiment, Figure 6 FIG. 1 is a schematic diagram showing the structure of a resource overdue prediction device provided by an embodiment of the present specification. Figure 6 As shown, the device comprises:
[0260] An extraction module 602 is configured to extract historical data of resource allocation items and determine user behavior data of target users in response to the overdue prediction request;
[0261] The processing module 604 is configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through the memory unit in the resource overdue prediction graph network to obtain memory mapping features;
[0262] An updating module 606 is configured to update the resource overdue prediction graph network to a target resource overdue prediction graph network using the memory mapping feature;
[0263] The prediction module 608 is configured to input the user behavior data into the target resource overdue prediction graph network for processing, and obtain resource overdue prediction information corresponding to the target user.
[0264] In an optional embodiment, the processing module 604 is further configured to:
[0265] Extract the data to be updated and the updated data from the historical data, query the key data in the updated data based on the data to be updated, and build memory mapping data based on the key data and the user behavior data; or extract the sparse user data associated with the target user from the historical data, and build memory mapping data based on the sparse user data and the user behavior data.
[0266] In an optional embodiment, when the memory mapping data is constructed based on the key data and the user behavior data, the processing module 604 is further configured to:
[0267] In the resource allocation graph data of the resource allocation project, the neighbor node information associated with the memory mapping data is determined, and the left memory feature corresponding to the first time node is obtained by aggregating the neighbor node information; based on the left memory feature, the right memory replacement feature corresponding to the second time node and the left memory replacement feature corresponding to the third time node are calculated; the right memory replacement feature and the left memory replacement feature are time-aligned with the memory graph corresponding to the memory unit in the resource overdue prediction graph network, and the memory unit is updated according to the time alignment result; the memory mapping data is processed by the updated memory unit in the resource overdue prediction graph network to obtain the memory mapping feature.
[0268] In an optional embodiment, the processing module 604 is further configured to:
[0269] By parsing the neighbor node information, the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are obtained; the main node attribute information, the neighbor node attribute information, the first position coding information, the first interactive behavior feature, and the first time interval coding information are spliced; the splicing result is input into the aggregation unit in the resource overdue prediction graph network for processing to obtain the left memory feature corresponding to the first time node.
[0270] In an optional embodiment, the processing module 604 is further configured to:
[0271] The left memory feature is input into an event occurrence prediction model for processing to obtain a right memory replacement feature corresponding to a second time node, wherein the event occurrence prediction model is used to predict feature changes from the first time node to the second time node; the right memory replacement feature is input into a feature transfer model for processing to obtain a left memory replacement feature corresponding to a third time node, wherein the feature transfer model is used to predict feature transfer from the second time node to the third time node.
[0272] In an optional embodiment, the processing module 604 is further configured to:
[0273] When the memory unit in the resource overdue prediction graph network is a single memory unit, the left memory replacement feature is time-aligned with the memory graph corresponding to the single memory unit, and the single memory unit is updated according to the alignment result; when the memory unit in the resource overdue prediction graph network is a dual memory unit, the right memory replacement feature and the left memory replacement feature are time-aligned with the memory graph corresponding to the dual memory unit, and the dual memory unit is updated according to the alignment result.
[0274] In an optional embodiment, the processing module 604 is further configured to:
[0275] In the case where the memory unit is a single memory unit, the memory mapping data is processed by the updated single memory unit in the resource overdue prediction graph network to obtain a single memory mapping feature as the memory mapping feature; in the case where the memory unit is a dual memory unit, the memory mapping data is processed by the updated dual memory unit in the resource overdue prediction graph network to obtain a dual memory mapping feature as the memory mapping feature.
[0276] In an optional embodiment, the updating module 606 is further configured to:
[0277] The neighbor node features, second position coding information, second interaction behavior features, and second time interval coding information associated with the memory mapping features are extracted in the memory unit; the neighbor node features, the second position coding information, the second interaction behavior features, and the second time interval coding information are spliced to obtain neighbor node prior features; the neighbor node prior features and the memory mapping features are aggregated, and the resource overdue prediction graph network is updated to the target resource overdue prediction graph network according to the aggregation result.
[0278] In an optional embodiment, the updating module 606 is further configured to:
[0279] The resource overdue prediction graph network is updated by using the memory mapping feature to obtain an initial resource overdue prediction graph network; a network optimization loss function and a knowledge transfer loss function corresponding to the initial resource overdue prediction graph network are determined; and parameters of the initial resource overdue prediction graph network are adjusted based on the network optimization loss function and the knowledge transfer loss function until a target resource overdue prediction graph network that meets the training stop condition is obtained.
[0280] In an optional embodiment, when the memory mapping data is constructed based on the sparse user data and the user behavior data, the processing module 604 is further configured to:
[0281] The left memory state feature, the right memory state feature and the cache edge feature are extracted from the memory unit in the resource overdue prediction graph network; the message feature corresponding to the fourth time node is generated according to the left memory state feature, the right memory state feature and the cache edge feature, and the target right memory feature is obtained by jump updating the message feature; the target left memory feature corresponding to the fifth time node is calculated based on the target right memory feature; the target right memory feature and the target left memory feature are processed through the memory unit in the resource overdue prediction graph network to obtain a memory mapping feature.
[0282] In an optional embodiment, the processing module 604 is further configured to:
[0283] A left memory state feature group and a right memory state feature group are extracted from the memory unit in the resource overdue prediction graph network, wherein the left memory state feature group records the memory state features before the event is updated, and the right memory state feature group records the memory state features after the event is updated; an initial left memory state feature corresponding to the sixth time node is selected from the left memory state feature group as the left memory state feature; and an initial right memory state feature corresponding to the seventh time node is selected from the right memory state feature group as the right memory state feature.
[0284] In an optional embodiment, the processing module 604 is further configured to:
[0285] Calculate the time interval between the sixth time node and the seventh time node, and determine the third time interval encoding information corresponding to the time interval; splice the left memory state feature, the right memory state feature, the cache edge feature and the third time interval encoding information, and perform message mapping on the splicing result to obtain the message feature corresponding to the fourth time node.
[0286] In an optional embodiment, the processing module 604 is further configured to:
[0287] The target state feature is determined according to the message feature, and the initial right memory feature is calculated based on the target state feature; the initial right memory feature is jump-updated to obtain the target right memory feature.
[0288] In an optional embodiment, the processing module 604 is further configured to:
[0289] The target right memory feature is input into the continuous memory prediction model; the target right memory feature is processed by the semantic coding unit in the continuous memory prediction model to obtain the memory coding feature; the memory coding feature is processed by the state transfer construction unit in the continuous memory prediction model to obtain the memory distribution feature; the memory distribution feature is processed by the memory generation unit in the continuous memory prediction model to obtain the target left memory feature corresponding to the fifth time node.
[0290] In an optional embodiment, the processing module 604 is further configured to:
[0291] The target right memory feature is input into the semantic coding unit in the continuous memory prediction model; the node similarity feature, the memory original prior feature and the enhanced auxiliary feature corresponding to the target right memory feature are constructed through the semantic coding unit; the initial memory coding feature is constructed based on the node similarity feature, the memory original prior feature and the enhanced auxiliary feature, and the initial memory coding feature is encoded to obtain the memory coding feature.
[0292] In an optional embodiment, the processing module 604 is further configured to:
[0293] The memory encoding feature is input into the state transfer construction unit in the continuous memory prediction model; feature distribution information corresponding to the memory encoding feature is constructed by the state transfer construction unit; and memory distribution features are calculated according to the feature distribution information.
[0294] In an optional embodiment, the processing module 604 is further configured to:
[0295] The right memory state feature group is updated according to the target right memory feature, and the left memory state feature group is updated according to the memory mapping feature; wherein the updated right memory state feature group and the updated left memory state feature group are used to determine the left memory state feature and the right memory state feature in the next resource overdue prediction cycle.
[0296] The resource overdue prediction device provided in this embodiment can more accurately complete the user's overdue prediction in combination with the resource overdue prediction graph network that has not been iteratively updated when the data update cycle is slow or the fault is recovered. It can first respond to the overdue prediction request, extract the historical data of the resource allocation project and determine the user behavior data of the target user; on this basis, it can first construct memory mapping data based on the historical data and user behavior data, and map the internal connection of the data missing due to the service project failure or the slow data iteration cycle to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems.
[0297] The above is a schematic scheme of a resource overdue prediction device of this embodiment. It should be noted that the technical scheme of the resource overdue prediction device and the technical scheme of the resource overdue prediction method mentioned above belong to the same concept, and the details not described in detail in the technical scheme of the resource overdue prediction device can be found in the description of the technical scheme of the resource overdue prediction method mentioned above.
[0298] Corresponding to the above method embodiment, this specification also provides a resource overdue prediction device embodiment, Figure 7 FIG. 2 shows a schematic diagram of the structure of another resource overdue prediction device provided by an embodiment of the present specification. Figure 7 As shown, the device is applied to a cloud-side device, including:
[0299] A request receiving module 702 is configured to receive an overdue prediction request submitted by a terminal device for a target user participating in a resource allocation project;
[0300] A data extraction module 704 is configured to extract historical data of the resource allocation project and determine user behavior data of the target user in response to the overdue prediction request;
[0301] A data processing module 706 is configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain a memory mapping feature;
[0302] An updating network module 708 is configured to update the resource overdue prediction graph network to a target resource overdue prediction graph network by using the memory mapping feature;
[0303] The prediction information module 710 is configured to input the user behavior data into the target resource overdue prediction graph network for processing, and obtain the resource overdue prediction information corresponding to the target user;
[0304] The information sending module 712 is configured to feed back the resource overdue prediction information to the terminal side device for performing resource allocation processing on the target user.
[0305] The resource overdue prediction device provided in this embodiment can more accurately complete the user's overdue prediction in combination with the resource overdue prediction graph network that has not been iteratively updated when the data update cycle is slow or the fault is recovered. It can first respond to the overdue prediction request, extract the historical data of the resource allocation project and determine the user behavior data of the target user; on this basis, it can first construct memory mapping data based on the historical data and user behavior data, and map the internal connection of the data missing due to the service project failure or the slow data iteration cycle to the resource overdue prediction graph network through the memory mapping data. At this time, the memory mapping data can be processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping feature; thereafter, the resource overdue prediction graph network is updated to the target resource overdue prediction graph network using the memory mapping feature; the target resource overdue prediction graph network can learn the information lost during the project recovery process, or the internal correlation relationship of sparse user data. Finally, the user behavior data can be input into the target resource overdue prediction graph network for processing, so as to obtain the resource overdue prediction information corresponding to the target user according to the processing result. When implementing resource overdue prediction for target users, the graph network can be updated first so that the graph network can learn the knowledge under the current prediction requirements and make predictions based on this. This can effectively improve the accuracy of resource overdue predictions and prevent the graph network from being affected by project failures or data sparsity problems.
[0306] The above is a schematic scheme of a resource overdue prediction device of this embodiment. It should be noted that the technical scheme of the resource overdue prediction device and the technical scheme of the resource overdue prediction method mentioned above belong to the same concept, and the details not described in detail in the technical scheme of the resource overdue prediction device can be found in the description of the technical scheme of the resource overdue prediction method mentioned above.
[0307] Corresponding to the above method embodiment, this specification also provides a network restart device embodiment, Figure 8 FIG. 1 shows a schematic diagram of a network restart device provided by an embodiment of the present specification. Figure 8 As shown, the device comprises:
[0308] The data reading module 802 is configured to read the to-be-updated data and the updated data of the target project in response to the restart instruction submitted by the to-be-restarted graph network preset for the target project;
[0309] A data query module 804 is configured to query key data in the updated data according to the data to be updated, and construct memory mapping data according to the key data;
[0310] A feature construction module 806 is configured to determine neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and construct a memory replacement feature corresponding to the neighbor node information;
[0311] The restart network module 808 is configured to restart the graph network to be restarted by using the memory replacement feature, and determine a target graph network that matches the current project state of the target project according to the restart result.
[0312] In an optional embodiment, the construct feature module 806 is further configured to:
[0313] By aggregating the neighbor node information, a left memory feature corresponding to the target time node is obtained; and a memory replacement feature corresponding to the associated time node is calculated according to the left memory feature;
[0314] Wherein, the restart network module 808 is further configured to:
[0315] The memory replacement feature is time-aligned with the memory graph corresponding to the memory unit in the graph network to be restarted, and the memory unit is updated according to the time alignment result as a restart of the graph network to be restarted.
[0316] The graph network restart device provided in this embodiment can solve the problem of limited network prediction flexibility caused by the time-consuming update of the graph network over a long period of time. It can realize the knowledge distillation of the memory representation ability of the graph network and migrate it to the restart network, realize the online rapid recovery of the project under time complexity with very small amounts of data, and can flexibly reason and switch between new data. In addition, the graph network has the ability to perceive the timing of neighbor nodes during aggregation, enhancing the aggregation representation of the main node.
[0317] The above is a schematic scheme of a graph network restart device of this embodiment. It should be noted that the technical scheme of the graph network restart device and the technical scheme of the graph network restart method described above belong to the same concept, and the details not described in detail in the technical scheme of the graph network restart device can be referred to the description of the technical scheme of the graph network restart method described above.
[0318] Fig. 9 The structure block diagram of a computing device 900 provided according to an embodiment of the present specification is shown. The components of the computing device 900 include but are not limited to a memory 910 and a processor 920. The processor 920 is connected to the memory 910 via a bus 930, and a database 950 is used to store data.
[0319] The computing device 900 also includes an access device 940 that enables the computing device 900 to communicate via one or more networks 960. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 940 may include one or more of any type of network interface (e.g., a network interface card (NIC)) that is wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, and a near field communication (NFC).
[0320] In one embodiment of the present specification, the above components of the computing device 900 and Fig. 9 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig. 9 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0321] The computing device 900 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 900 may also be a mobile or stationary server.
[0322] The processor 920 is used to execute the following computer executable instructions, which, when executed by the processor, implement the steps of the above-mentioned resource overdue prediction method.
[0323] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method.
[0324] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned resource overdue prediction method or graph network restart method.
[0325] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method.
[0326] An embodiment of the present specification further provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned resource overdue prediction method or graph network restart method.
[0327] The above is a schematic scheme of a computer program of this embodiment. It should be noted that the technical scheme of the computer program and the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method belong to the same concept, and the details not described in detail in the technical scheme of the computer program can be referred to the description of the technical scheme of the above-mentioned resource overdue prediction method or graph network restart method.
[0328] An embodiment of the present specification also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the steps of the above-mentioned resource overdue prediction method or graph network restart method.
[0329] The above is a schematic scheme of a computer program product of this embodiment. It should be noted that the technical scheme of the computer program product and the technical scheme of the resource overdue prediction method or the graph network restart method mentioned above belong to the same concept, and the details not described in detail in the technical scheme of the computer program product can be referred to the description of the technical scheme of the resource overdue prediction method or the graph network restart method mentioned above.
[0330] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0331] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.
[0332] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.
[0333] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0334] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.
Claims
1. A resource overdue prediction method, comprising: In response to the overdue prediction request, extract historical data of the resource allocation project and determine user behavior data of the target user; Memory mapping data is constructed according to the historical data and the user behavior data, and the memory mapping data is processed through the memory unit in the resource overdue prediction graph network to obtain memory mapping features, wherein the construction of the memory mapping data includes extracting data to be updated and updated data from the historical data, querying key data in the updated data based on the data to be updated, and constructing memory mapping data based on the key data and the user behavior data, the data to be updated is the knowledge data missing from the resource overdue prediction graph network, the updated data is the knowledge data not missing from the resource overdue prediction graph network, and the key data is the data associated with the entity matching the data to be updated in the updated data; Aggregating the prior features of the neighbor nodes associated with the memory unit and the memory mapping features, and updating the resource overdue prediction graph network to a target resource overdue prediction graph network according to the aggregation result, wherein the target resource overdue prediction graph network is a graph network that has learned the missing knowledge data; The user behavior data is input into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
2. The method according to claim 1, wherein the step of constructing memory mapping data based on the historical data and the user behavior data comprises: Sparse user data associated with the target user is extracted from the historical data, and memory mapping data is constructed based on the sparse user data and the user behavior data.
3. The method according to claim 1, wherein the memory mapping data is constructed based on the key data and the user behavior data, the memory mapping data is processed by the memory unit in the resource overdue prediction graph network to obtain the memory mapping features, comprising: In the resource allocation graph data of the resource allocation project, determining the neighbor node information associated with the memory mapping data, and obtaining the left memory feature corresponding to the first time node by aggregating the neighbor node information; According to the left memory feature, calculating the right memory replacement feature corresponding to the second time node and the left memory replacement feature corresponding to the third time node; Time-aligning the right memory replacement feature and the left memory replacement feature with the memory graph corresponding to the memory unit in the resource overdue prediction graph network, and updating the memory unit according to the time alignment result; The memory mapping data is processed by the updated memory unit in the resource overdue prediction graph network to obtain memory mapping features.
4. The method according to claim 3, wherein the step of aggregating the neighbor node information to obtain the left memory feature corresponding to the first time node comprises: By parsing the neighbor node information, the main node attribute information, the neighbor node attribute information, the first position coding information, the first interaction behavior feature, and the first time interval coding information are obtained; splicing the master node attribute information, the neighbor node attribute information, the first position coding information, the first interaction behavior feature, and the first time interval coding information; The splicing result is input into the aggregation unit in the resource overdue prediction graph network for processing to obtain the left memory feature corresponding to the first time node.
5. The method according to claim 3, wherein the step of calculating the right memory replacement feature corresponding to the second time node and the left memory replacement feature corresponding to the third time node according to the left memory feature comprises: Inputting the left memory feature into an event occurrence prediction model for processing to obtain a right memory replacement feature corresponding to the second time node, wherein the event occurrence prediction model is used to predict feature changes from the first time node to the second time node; The right memory replacement feature is input into a feature transfer model for processing to obtain a left memory replacement feature corresponding to a third time node, wherein the feature transfer model is used to predict the feature transfer from the second time node to the third time node.
6. The method according to claim 3, wherein the right memory replacement feature and the left memory replacement feature are time-aligned with the memory graph corresponding to the memory unit in the resource overdue prediction graph network, and the memory unit is updated according to the time alignment result, comprising: When the memory unit in the resource overdue prediction graph network is a single memory unit, time-aligning the left memory replacement feature with the memory graph corresponding to the single memory unit, and updating the single memory unit according to the alignment result; When the memory unit in the resource overdue prediction graph network is a dual memory unit, the right memory replacement feature and the left memory replacement feature are time-aligned with the memory graph corresponding to the dual memory unit, and the dual memory unit is updated according to the alignment result.
7. The method according to claim 6, wherein the processing of the memory mapping data by the updated memory unit in the resource overdue prediction graph network to obtain the memory mapping features comprises: In the case where the memory unit is a single memory unit, the memory mapping data is processed by the updated single memory unit in the resource overdue prediction graph network to obtain a single memory mapping feature as the memory mapping feature; In the case where the memory unit is a dual memory unit, the memory mapping data is processed by the updated dual memory unit in the resource overdue prediction graph network to obtain a dual memory mapping feature as the memory mapping feature.
8. The method according to any one of claims 2 to 7, wherein the a priori features of the neighbor nodes associated with the memory unit and the memory mapping features are aggregated, and the resource overdue prediction graph network is updated to a target resource overdue prediction graph network according to the aggregation result, comprising: Extracting the neighbor node feature, the second position encoding information, the second interaction behavior feature, and the second time interval encoding information associated with the memory mapping feature in the memory unit; The neighbor node feature, the second position coding information, the second interaction behavior feature, and the second time interval coding information are concatenated to obtain a priori features of the neighbor node; Aggregation is performed on the neighbor node prior features and the memory map features, and the resource overdue prediction graph network is updated to a target resource overdue prediction graph network according to the aggregation result.
9. The method according to any one of claims 2 to 7, further comprising: Using the memory mapping feature to update the resource overdue prediction graph network to obtain an initial resource overdue prediction graph network; Determine a network optimization loss function and a knowledge transfer loss function corresponding to the initial resource overdue prediction graph network; The parameters of the initial resource overdue prediction graph network are adjusted based on the network optimization loss function and the knowledge transfer loss function until a target resource overdue prediction graph network that meets the training stop condition is obtained.
10. The method according to claim 2, wherein the memory map data is constructed based on the sparse user data and the user behavior data, the memory map data is processed by a memory unit in a resource overdue prediction graph network to obtain a memory map feature, comprising: Extract left memory state features, right memory state features, and cache edge features from the memory cells in the resource overdue prediction graph network; Generate a message feature corresponding to a fourth time node according to the left memory state feature, the right memory state feature, and the cache edge feature, and obtain a target right memory feature by performing a jump update on the message feature; Calculate the target left memory feature corresponding to the fifth time node based on the target right memory feature; The target right memory feature and the target left memory feature are processed by a memory unit in the resource overdue prediction graph network to obtain a memory mapping feature.
11. The method according to claim 10, wherein the determining of the left memory state feature and the right memory state feature comprises: Extracting a left memory state feature group and a right memory state feature group from a memory unit in a resource overdue prediction graph network, wherein the left memory state feature group records memory state features before an event is updated, and the right memory state feature group records memory state features after an event is updated; Selecting the initial left memory state feature corresponding to the sixth time node in the left memory state feature group as the left memory state feature; The initial right memory state feature corresponding to the seventh time node is selected from the right memory state feature group as the right memory state feature.
12. The method according to claim 11, wherein generating a message feature corresponding to a fourth time node according to the left memory state feature, the right memory state feature and the cache edge feature comprises: Calculating a time interval between the sixth time node and the seventh time node, and determining third time interval coding information corresponding to the time interval; The left memory state feature, the right memory state feature, the cache edge feature and the third time interval encoding information are spliced, and message mapping is performed on the splicing result to obtain a message feature corresponding to a fourth time node.
13. The method according to claim 10, wherein the step of obtaining the target right memory feature by performing jump update on the message feature comprises: Determining a target state feature according to the message feature, and calculating an initial right memory feature based on the target state feature; A jump update is performed on the initial right memory feature to obtain a target right memory feature.
14. The method according to claim 10, wherein the step of calculating the target left memory feature corresponding to the fifth time node based on the target right memory feature comprises: Inputting the target right memory feature into a continuous memory prediction model; Processing the target right memory feature through the semantic coding unit in the continuous memory prediction model to obtain a memory coding feature; Processing the memory encoding feature through the state transfer construction unit in the continuous memory prediction model to obtain a memory distribution feature; The memory distribution feature is processed by the memory generation unit in the continuous memory prediction model to obtain the target left memory feature corresponding to the fifth time node.
15. The method according to claim 14, wherein the processing of the target right memory feature by the semantic encoding unit in the continuous memory prediction model to obtain the memory encoding feature comprises: Inputting the target right memory feature into the semantic encoding unit in the continuous memory prediction model; Constructing the node similarity feature corresponding to the target right memory feature, the memory original prior feature and the enhanced auxiliary feature through the semantic coding unit; An initial memory encoding feature is constructed based on the node similarity feature, the memory original prior feature and the enhanced auxiliary feature, and encoding processing is performed on the initial memory encoding feature to obtain a memory encoding feature.
16. The method according to claim 14, wherein the memory encoding feature is processed by the state transfer construction unit in the continuous memory prediction model to obtain the memory distribution feature, comprising: Inputting the memory encoding feature into a state transition construction unit in the continuous memory prediction model; Constructing feature distribution information corresponding to the memory encoding feature through the state transfer construction unit; Memory distribution features are calculated according to the feature distribution information.
17. The method according to claim 11, further comprising: updating the right memory state feature group according to the target right memory feature, and updating the left memory state feature group according to the memory mapping feature; The updated right memory state feature group and the updated left memory state feature group are used to determine the left memory state feature and the right memory state feature in the next resource overdue prediction cycle.
18. A resource overdue prediction method, applied to a cloud-side device, comprising: The receiving end-side device submits an overdue prediction request for a target user participating in a resource allocation project; In response to the overdue prediction request, extracting historical data of the resource allocation project and determining user behavior data of the target user; Memory mapping data is constructed according to the historical data and the user behavior data, and the memory mapping data is processed through the memory unit in the resource overdue prediction graph network to obtain memory mapping features, wherein the construction of the memory mapping data includes extracting data to be updated and updated data from the historical data, querying key data in the updated data based on the data to be updated, and constructing memory mapping data based on the key data and the user behavior data, the data to be updated is the knowledge data missing from the resource overdue prediction graph network, the updated data is the knowledge data not missing from the resource overdue prediction graph network, and the key data is the data associated with the entity matching the data to be updated in the updated data; Aggregating the prior features of the neighbor nodes associated with the memory unit and the memory mapping features, and updating the resource overdue prediction graph network to a target resource overdue prediction graph network according to the aggregation result, wherein the target resource overdue prediction graph network is a graph network that has learned the missing knowledge data; Inputting the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user; The resource overdue prediction information is fed back to the terminal side device for performing resource allocation processing on the target user.
19. A graph network restart method, comprising: In response to a restart instruction submitted to a preset to-be-restarted graph network for a target project, read the to-be-updated data, the updated data of the target project, and the user behavior data of the target user, wherein the to-be-updated data and the updated data are read from the historical data of the target project; Query key data in the updated data according to the data to be updated, and construct memory mapping data according to the key data and the user behavior data, wherein the data to be updated is the knowledge data missing from the graph network to be restarted, the updated data is the knowledge data not missing from the graph network to be restarted, and the key data is the data associated with the entity matching the data to be updated in the updated data; Determining neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and constructing a memory replacement feature corresponding to the neighbor node information; The memory replacement feature is used to restart the graph network to be restarted, and a target graph network that matches the current project state of the target project is determined according to the restart result, wherein the target graph network is used to process the user behavior data to obtain resource overdue prediction information corresponding to the target user.
20. The method according to claim 19, wherein the step of constructing a memory replacement feature corresponding to the neighbor node information comprises: By aggregating the neighbor node information, a left memory feature corresponding to the target time node is obtained; Calculate the memory replacement feature corresponding to the associated time node according to the left memory feature; Wherein, restarting the graph network to be restarted by using the memory replacement feature includes: The memory replacement feature is time-aligned with the memory graph corresponding to the memory unit in the graph network to be restarted, and the memory unit is updated according to the time alignment result as a restart of the graph network to be restarted.
21. A resource overdue prediction device, comprising: an extraction module configured to extract historical data of resource allocation items and user behavior data of determined target users in response to the overdue prediction request; a processing module configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain memory mapping features, wherein the construction of the memory mapping data includes extracting data to be updated and updated data from the historical data, querying key data in the updated data based on the data to be updated, and constructing the memory mapping data based on the key data and the user behavior data, wherein the data to be updated is knowledge data missing from the resource overdue prediction graph network, the updated data is knowledge data not missing from the resource overdue prediction graph network, and the key data is data associated with an entity in the updated data that matches the data to be updated; An updating module is configured to aggregate the prior features of the neighbor nodes associated with the memory unit and the memory mapping features, and update the resource overdue prediction graph network to a target resource overdue prediction graph network according to the aggregation result, wherein the target resource overdue prediction graph network is a graph network that has learned the missing knowledge data; The prediction module is configured to input the user behavior data into the target resource overdue prediction graph network for processing to obtain resource overdue prediction information corresponding to the target user.
22. A resource overdue prediction device, applied to a cloud-side device, comprising: A request receiving module, configured to receive an overdue prediction request submitted by a terminal device for a target user participating in a resource allocation project; a data extraction module configured to extract historical data of the resource allocation project and determine user behavior data of the target user in response to the overdue prediction request; a data processing module configured to construct memory mapping data according to the historical data and the user behavior data, and process the memory mapping data through a memory unit in a resource overdue prediction graph network to obtain memory mapping features, wherein the construction of the memory mapping data includes extracting data to be updated and updated data from the historical data, querying key data in the updated data based on the data to be updated, and constructing the memory mapping data based on the key data and the user behavior data, wherein the data to be updated is knowledge data missing from the resource overdue prediction graph network, the updated data is knowledge data not missing from the resource overdue prediction graph network, and the key data is data associated with an entity in the updated data that matches the data to be updated; An updating network module is configured to aggregate the prior features of the neighbor nodes associated with the memory unit and the memory mapping features, and update the resource overdue prediction graph network to a target resource overdue prediction graph network according to the aggregation result, wherein the target resource overdue prediction graph network is a graph network that has learned the missing knowledge data; A prediction information module is configured to input the user behavior data into the target resource overdue prediction graph network for processing, and obtain resource overdue prediction information corresponding to the target user; The information sending module is configured to feed back the resource overdue prediction information to the terminal side device for performing resource allocation processing on the target user.
23. A network restart device, comprising: A data reading module is configured to read the to-be-updated data, the updated data and the user behavior data of the target user of the target project in response to a restart instruction submitted for the to-be-restarted graph network preset for the target project, wherein the to-be-updated data and the updated data are read from the historical data of the target project; a query data module, configured to query key data in the updated data according to the data to be updated, and construct memory mapping data according to the key data and the user behavior data, wherein the data to be updated is the knowledge data missing from the graph network to be restarted, the updated data is the knowledge data not missing from the graph network to be restarted, and the key data is the data associated with the entity matching the data to be updated in the updated data; A feature construction module is configured to determine neighbor node information associated with the memory mapping data in the allocation graph data of the target project, and construct a memory replacement feature corresponding to the neighbor node information; The restart network module is configured to restart the graph network to be restarted using the memory replacement feature, and determine a target graph network that matches the current project state of the target project based on the restart result, wherein the target graph network is used to process the user behavior data to obtain resource overdue prediction information corresponding to the target user.
24. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method described in any one of claims 1 to 20 are implemented.
25. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the method according to any one of claims 1 to 20.
26. A computer program product, comprising a computer program or instructions, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 20.
