Message pushing resource intelligent scheduling method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence application and computer information processing, can be applied to the field of financial technology / digital medical treatment, and discloses a message pushing resource intelligent scheduling method and device, an electronic device and a storage medium. The method comprises the following steps: acquiring message characteristic data of a to-be-pushed message, inputting a message value prediction model, and outputting message value parameters; acquiring channel state data of a pushing channel, inputting a channel load prediction model, and outputting predicted load parameters of the pushing channel; acquiring target receiving end running state data, inputting a processing state calculation model, and outputting available processing parameters; inputting the above parameters into a scheduling decision model, and outputting a target pushing strategy comprising a target pushing channel, a target pushing time and target resource occupation parameters; and according to the target resource occupation parameters, the pushing resources are allocated, and the to-be-pushed message is pushed through the target pushing channel at the target pushing time. The method reduces message delivery delay and improves pushing resource utilization.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence applications and computer information processing technology, and can be applied to the fields of fintech / digital healthcare, particularly to a method, device, electronic device and storage medium for intelligent scheduling of message push resources. Background Technology

[0002] With the development of internet services, push notification systems are widely used in fintech, digital healthcare, e-commerce, and online customer service to send messages such as transaction alerts, business notifications, risk warnings, and service requests to user terminals or business processing terminals. Traditional push notification systems typically need to receive messages from multiple business systems simultaneously and transmit them through different push channels such as web pages, mobile applications, and SMS. Because different messages vary in their generation time, business urgency, data volume, and timeliness requirements, the transmission capacity, queue status, and processing latency of each push channel may also differ. Therefore, push notification systems need to sort, cache, and send a large number of messages with limited channel resources. Traditional push notification systems usually determine the message sending order based on pre-set message type, priority tags, or message generation time, and allocate push resources according to a fixed channel capacity or preset ratio. However, message traffic in actual business operations exhibits significant fluctuations and suddenness. Fixed message sorting rules are insufficient to reflect the current business urgency of messages in a timely manner, and fixed resource allocation methods are difficult to adapt to changes in the real-time processing capabilities of different push channels. For example, in the fintech sector, when transaction alerts, risk warnings, marketing notifications, and customer inquiries are generated in a concentrated manner, the large number of messages with low timeliness requirements may consume queue and channel resources, increasing the waiting time for risk warnings or abnormal transaction messages. In the digital healthcare sector, when appointment notifications, examination results, health reminders, and emergency response messages enter the push system simultaneously, some time-sensitive messages may also be delayed due to queue congestion. Furthermore, different push channels may simultaneously experience resource idleness and resource shortages; some channels may experience long processing delays due to message backlog, while the available resources of other channels may not be fully utilized. When message traffic, network status, or business processing capacity changes, traditional message push systems struggle to adjust message sending order and channel resource configuration in a timely manner, easily leading to problems such as high-timeliness messages being blocked by low-timeliness messages, persistent congestion in local channels, and low utilization of push resources. Summary of the Invention

[0003] The main technical problem addressed by the embodiments of this application is the delay in the delivery of pushed messages and the low utilization rate of push resources in traditional message push technology.

[0004] To address the aforementioned technical problems, the first technical solution adopted in this application is: providing an intelligent scheduling method for message push resources, comprising: acquiring message feature data of a message to be pushed; inputting the message feature data into a pre-trained message value prediction model and outputting message value parameters of the message to be pushed; acquiring channel status data of multiple push channels, inputting the channel status data into a pre-trained channel load prediction model and outputting predicted load parameters of each push channel; acquiring the operating status data of the target receiving end corresponding to the message to be pushed, inputting the operating status data into a pre-trained processing status calculation model and outputting available processing parameters of the target receiving end; inputting the message value parameters, the predicted load parameters of each push channel, and the available processing parameters into a pre-trained scheduling decision model and outputting a target push strategy corresponding to the message to be pushed, wherein the target push strategy includes a target push channel, a target push timing, and a target resource occupancy parameter; allocating push resources corresponding to the target push channel according to the target resource occupancy parameter, and pushing the message to be pushed through the target push channel at the target push timing.

[0005] To address the aforementioned technical problems, the second technical solution adopted in this application is: providing an intelligent scheduling device for message push resources, comprising: a message feature acquisition module for acquiring message feature data of a message to be pushed; a message value prediction module for inputting the message feature data into a pre-trained message value prediction model and outputting message value parameters of the message to be pushed; a channel load prediction module for acquiring channel status data of multiple push channels, inputting the channel status data into a pre-trained channel load prediction model, and outputting predicted load parameters of each push channel; and a processing status calculation module for acquiring the operating status of the target receiving end corresponding to the message to be pushed. The system comprises: a state data processing module, a push strategy generation module, and a push policy generation module. The push strategy generation module takes the message value parameters, the predicted load parameters of each push channel, and the available processing parameters as input to a pre-trained scheduling decision model and outputs a target push strategy corresponding to the message to be pushed. The target push strategy includes a target push channel, a target push timing, and target resource occupancy parameters. The push allocation and execution module allocates push resources corresponding to the target push channel according to the target resource occupancy parameters and pushes the message to be pushed through the target push channel at the target push timing.

[0006] To solve the above-mentioned technical problems, the third technical solution adopted in the embodiments of this application is: to provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the message push resource intelligent scheduling method as described above.

[0007] To solve the above-mentioned technical problems, the fourth technical solution adopted in the embodiments of this application is: to provide a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by an electronic device, the electronic device executes the message push resource intelligent scheduling method as described above.

[0008] Unlike related technologies, this application obtains message value parameters by analyzing the content characteristics and historical interaction patterns of the message to be pushed; it obtains predicted load parameters for each push channel by analyzing the operational status changes of multiple push channels; and it obtains available processing parameters by analyzing the operational status of the target receiver. Then, it jointly processes the message value parameters, predicted load parameters, and available processing parameters to generate a target push strategy that includes target push channels, target push timing, and target resource usage parameters. Push resources are then allocated and message pushes are executed according to the target push strategy. Therefore, this application enables the push strategy to match message value, channel load, and receiver processing status, reducing message delivery delays and improving push resource utilization. Attached Figure Description

[0009] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0010] Figure 1 This is a schematic diagram of the operating environment of the intelligent scheduling method for message push resources provided in the embodiments of this application.

[0011] Figure 2 This is a schematic diagram of the execution flow of the intelligent scheduling method for message push resources provided in the embodiments of this application.

[0012] Figure 3 This is a schematic diagram of the execution flow for generating predicted load parameters in the intelligent scheduling method for message push resources provided in this application embodiment.

[0013] Figure 4 This is a schematic diagram of the execution flow of obtaining the target push strategy in the intelligent scheduling method for message push resources provided in the embodiments of this application.

[0014] Figure 5 This is a schematic diagram of the system structure of the intelligent scheduling device for message push resources provided in the embodiments of this application.

[0015] Figure 6 This is a schematic diagram of the hardware structure of the electronic device for implementing the intelligent scheduling method for push message resources provided in this application embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Software tools, components, or servers not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0017] It should be noted that, unless otherwise specified, the various features in the embodiments of this application can be combined with each other, all of which are within the protection scope of this application. Furthermore, although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device schematic diagram or the order in the flowchart.

[0018] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0019] To facilitate understanding of this embodiment, a detailed description of the intelligent scheduling method for message push resources disclosed in this application embodiment will be provided first. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a schematic diagram of the operating environment of the intelligent scheduling method for message push resources provided in the embodiments of this application, such as... Figure 1 As shown, the execution subject of the intelligent scheduling method for message push resources provided in this application embodiment is generally an electronic device with a certain computing power, such as a computer device. In some possible implementations, this intelligent scheduling method for message push resources can be implemented by the processor calling computer-readable instructions stored in the memory. Figure 1The computer equipment mentioned can be a server. A server can be a standalone server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. This can be understood as... Figure 1 The number of computer devices shown is merely illustrative and can be expanded in any number according to actual needs.

[0020] Please continue reading. Figure 2 , Figure 2 This is a schematic diagram of the execution flow of the intelligent scheduling method for message push resources provided in the embodiments of this application, as shown below. Figure 2 As shown, it includes the following steps: S1. Obtain message feature data of the message to be pushed.

[0021] Specifically, message feature data refers to a structured set of features used to characterize message content attributes, business relationships, and historical interaction patterns. The formation process of message feature data can include text normalization, field mapping, context association, time series processing, and feature encoding, converting message content from different sources and in different formats into a numerical representation that can be uniformly processed by subsequent models. This processing method aims to eliminate differences in text length, field format, and business semantics between original messages, while retaining semantic and interactive information relevant to message value judgment. In fintech scenarios, the text content, business type, generation time, and historical processing records of account change notifications, transaction risk alerts, repayment reminders, and service notifications can be encoded to form message feature data reflecting the urgency and business relevance of the message. In digital healthcare scenarios, the notification content, business category, sending time, and historical viewing records of appointment reminders, examination result notifications, medication time reminders, and health follow-up messages can be encoded to form message feature data for subsequent push processing. These features are only used for message scheduling and are not used to evaluate patient life value or make medical diagnoses. The above processing can improve the consistency and computability of message data representation, providing an accurate and complete input basis for subsequent message value prediction.

[0022] S2. Input the message feature data into the pre-trained message value prediction model and output the message value parameters of the message to be pushed.

[0023] The message value prediction model comprehensively processes the content characteristics and corresponding historical interaction patterns of the message to be pushed, and calculates the message value parameters of the message to be pushed based on the business attributes reflected in the message content and the viewing, confirmation, and processing status of historical messages. The message value parameters characterize the likelihood that the message to be pushed will be viewed and processed in a timely manner under the current message content and historical interaction conditions, and serve as the input basis for subsequent selection of push channels, determination of push timing, and allocation of push resources.

[0024] In one specific implementation, the semantic encoding branch can employ a semantic network based on bidirectional context encoding, the temporal encoding branch can employ a long short-term memory network, and the feature fusion layer can use an attention weight calculation method. Message value parameters. V It can be calculated using the following formula: in, Ms This represents the semantic evaluation parameters obtained based on the semantic features of the message. Mt This represents the time-series evaluation parameters obtained based on interaction behavior characteristics. Indicates dynamic fusion weights, The value range is from 0 to 1, for example, it can be from 0.3 to 0.7.

[0025] As an optional implementation, the above message value prediction model includes a semantic coding branch, a temporal coding branch, a feature fusion layer, and a fully connected mapping layer. The processing of step S2 may also specifically include the following steps S21 to S24.

[0026] S21. Perform contextual semantic encoding on the message to be pushed through the semantic encoding branch to obtain the message semantic features.

[0027] The semantic encoding branch converts characters, words, or sentences in the message to be pushed into semantic vectors with fixed dimensions, and establishes the association between words by combining contextual information from different positions in the message. The semantic encoding branch can sequentially include an embedding layer, a context encoding layer, and a semantic output layer. The embedding layer maps the message content, after word segmentation and serialization, into initial feature vectors. The context encoding layer calculates the contextual association weights between the initial feature vectors. The semantic output layer aggregates the initial feature vectors from different positions according to the contextual association weights to form the message's semantic features. Contextual semantic encoding is used to avoid judging message content based solely on a single keyword, enabling the same words to form corresponding semantic representations in different sentence contexts, and providing a unified semantic input for subsequent message value calculation.

[0028] S22. According to the interaction time sequence, recursively calculate the historical interaction sequence corresponding to the message to be pushed through the time-series coding branch to obtain the interaction behavior characteristics.

[0029] The temporal coding branch processes the historical interaction sequence corresponding to the message to be pushed according to the order of occurrence of historical interaction events. Each interaction event in the historical interaction sequence can first undergo feature mapping to be converted into a corresponding interaction feature vector, and then the interaction feature vector is input into the temporal coding branch in the order of interaction time. The temporal coding branch can include a recursive calculation unit and a state update unit. The recursive calculation unit is used to combine the current interaction feature vector with the hidden state of the previous time step, and the state update unit is used to control the retention ratio of historical information and the writing ratio of current information according to the result of the combination operation, gradually forming a hidden state sequence representing the interaction change process. Generating interaction behavior features based on the hidden state sequence can preserve the temporal dependency between consecutive interaction events and avoid treating interaction events that occur at different times as independent data.

[0030] S23. Through the feature fusion layer, calculate the dynamic fusion weights corresponding to the message semantic features and interaction behavior features based on the historical message processing results.

[0031] The feature fusion layer adjusts the participation of message semantic features and interaction behavior features in message value calculation based on historical message processing results. Historical message processing results can be used to establish training samples with corresponding message semantic features and interaction behavior features. By numerically labeling the historical message processing results, supervisory information is formed for calculating fusion weights. The feature fusion layer can linearly map message semantic features and interaction behavior features separately, obtaining semantic mapping features and behavior mapping features with consistent dimensions. These features are then concatenated and normalized to output semantic weights and behavior weights. The semantic and behavior weights can be constrained within a preset numerical range using a normalization function, ensuring they satisfy a preset weight constraint relationship. This results in dynamic fusion weights that change with the input features based on historical message processing patterns. For example, the sum of the semantic and behavior weights can be set to 1 to maintain a uniform numerical scale for the fusion results of different messages to be pushed.

[0032] S24. The message semantic features and interaction behavior features are weighted and fused according to the dynamic fusion weight to obtain message feature data, and the message value parameters are calculated based on the message feature data.

[0033] The feature fusion layer multiplies message semantic features by their corresponding semantic weights and interaction behavior features by their corresponding behavior weights. It then concatenates or sums the weighted message semantic features and interaction behavior features to form message feature data, which serves as an intermediate processing result within the message value prediction model. This message feature data can then be input into the fully connected mapping layer of the message value prediction model. A linear transformation is performed using a preset weight matrix and preset bias parameters. The result is then mapped to a preset value parameter range using an activation function to obtain the message value parameters. These message value parameters represent the push value of the message under the current message content and historical interaction conditions, and are not used to evaluate the message recipient's identity, health status, economic capacity, or life value.

[0034] The semantic encoding branch and the temporal encoding branch are set up in parallel. The outputs of the semantic encoding branch and the temporal encoding branch are respectively connected to the input of the feature fusion layer, and the output of the feature fusion layer is connected to the input of the fully connected mapping layer. The semantic encoding branch is used to extract the contextual semantic information corresponding to the message to be pushed from the message feature data input in step S2. The temporal encoding branch is used to extract the temporal behavior information corresponding to the historical interaction sequence from the message feature data input in step S2. The feature fusion layer is used to dynamically weight and fuse the contextual semantic information and the temporal behavior information. The fully connected mapping layer is used to convert the fusion result into message value parameters. The message feature data input in step S2 may include the basic message features obtained after format unification, word segmentation, serialization and feature mapping of the message to be pushed, as well as the historical interaction sequence arranged according to the interaction time. Each historical interaction event in the historical interaction sequence is associated with the corresponding historical message processing result. The message feature data formed in step S24 belongs to the deep message feature data generated internally by the message value prediction model. The deep message feature data is in a different data processing stage from the message feature data input in step S2 and is used as the input of the fully connected mapping layer.

[0035] As an optional implementation, the training samples for the message value prediction model are constructed based on historical push messages, corresponding historical interaction sequences, and historical message processing results. Data anonymization, abnormal data cleanup, missing value imputation, and time alignment are performed on historical push messages and historical interaction sequences. Message value tags are generated based on message viewing status, message viewing latency, and business operation completion status. Message value tags are only used to characterize the viewing and processing status of historical push messages; the recipient's identity, economic capacity, health status, disease severity, or life value are not used as conditions for tag generation.

[0036] Specifically, historical push messages are input into the initial semantic encoding branch to obtain training message semantic features, and historical interaction sequences are input into the initial temporal encoding branch to obtain training interaction behavior features. The initial feature fusion layer calculates the fusion weights corresponding to the training message semantic features and training interaction behavior features, and then performs weighted fusion according to these weights to obtain training message feature data. The training message feature data is input into the initial fully connected mapping layer to obtain predicted message value parameters. The value prediction loss is calculated based on the difference between the predicted message value parameters and the message value label, and the model parameters of the semantic encoding branch, temporal encoding branch, feature fusion layer, and fully connected mapping layer are updated based on the value prediction loss until the value prediction loss meets the preset convergence condition, resulting in a pre-trained message value prediction model. As an example, the training batch size can be set to 32 to 64, the learning rate can be set to 0.0001 to 0.001, and the maximum number of training epochs can be set to 50 to 200.

[0037] As an example, in fintech scenarios, the input data for a message value prediction model can include anonymized account change notifications, transaction confirmation notifications, or risk warning messages, as well as historical viewing times, confirmation durations, and processing results for similar messages. A semantic encoding branch extracts contextual semantic features from the message content, while a temporal encoding branch extracts behavioral change features from historical interactions. A feature fusion layer dynamically weights and fuses these two types of features, and a fully connected mapping layer outputs message value parameters within a preset numerical range for subsequent push channel, push timing, and push resource scheduling. In digital healthcare scenarios, the input data for a message value prediction model can include anonymized test result notifications, follow-up visit reminders, appointment notifications, or medication reminders, as well as historical viewing times, confirmation times, and subsequent processing records for similar notifications. The model outputs message value parameters following the same feature extraction, dynamic fusion, and mapping process. In fintech scenarios, message value parameters are not used to evaluate the recipient's creditworthiness, economic capacity, or personal value, nor are they used to make credit or transaction approval decisions. In digital healthcare scenarios, message value parameters are not used to evaluate a patient's life value, determine treatment priorities, or replace medical personnel in making disease diagnosis or treatment decisions.

[0038] Through the above steps S21 to S24, the contextual semantic information of the message to be pushed and the temporal behavior information in the historical interaction sequence can be combined to form a fusion feature that dynamically changes with the message content and interaction rules. This reduces the bias caused by judging the value of the message based on a single semantic feature or a single historical behavior feature, improves the accuracy and stability of the message value parameters, and provides a reliable data foundation for the joint scheduling of subsequent push channels, push timing and push resources.

[0039] S3. Obtain channel status data for multiple push channels, input the channel status data into the pre-trained channel load prediction model, and output the predicted load parameters for each push channel.

[0040] The channel load prediction model analyzes message queuing, processing latency, and resource usage changes across multiple push channels over consecutive time periods, and predicts the load status of each push channel in subsequent preset time periods. The predicted load parameters characterize the load level and trend of the corresponding push channel, serving as input for selecting target push channels and determining resource usage parameters.

[0041] In one specific implementation, the gated recurrent coding layer can employ a gated recurrent unit network, and the self-attention coding layer can employ a coding network based on a self-attention mechanism. For any acquisition time... t Channel status data can be represented as: in, Qt This indicates the message queue occupancy parameter corresponding to the data collection time t. Tt This represents the message processing latency parameter. Bt This indicates the parameters for transmission resource usage. Arranging multiple channel status data according to acquisition time can form a channel status sequence for the input gated cyclic coding layer.

[0042] As an optional implementation, the above channel load prediction model includes a gated cyclic coding layer, a self-attention coding layer, and a load prediction layer. Please continue reading. Figure 3 , Figure 3 This is a schematic diagram of the execution flow for generating predicted load parameters in the intelligent scheduling method for message push resources provided in this application embodiment, as shown below. Figure 3 As shown, the processing procedure of step S3 above may also include steps S31 to S34.

[0043] S31. Sort the channel status data of each push channel according to the collection time to obtain the corresponding channel status sequence.

[0044] Channel status data is used to characterize the resource usage and message processing status of push channels at different collection times. Queue usage parameters, message processing latency parameters, and transmission resource usage parameters for each push channel can be continuously collected according to a preset collection cycle. Channel status data belonging to the same push channel are then sorted in ascending order based on the collection timestamp. For channel status data with discontinuous collection times, interpolation can be performed based on the variation patterns between adjacent collection times. For abnormal data that significantly exceeds the preset value range, statistical values ​​from adjacent times can be used for replacement. After sorting and preprocessing, channel status data from multiple consecutive collection times are combined into a channel status sequence, enabling the channel load prediction model to process the historical operational changes of each push channel in a unified chronological order.

[0045] S32. The channel state sequence is recursively processed by a gated cyclic coding layer to obtain the local load change characteristics between consecutive time points.

[0046] The gated cyclic coding layer is used to extract short-term load change relationships between adjacent acquisition moments. The state data of each moment in the channel state sequence can be sequentially input into the gated cyclic coding layer. The gated cyclic coding layer combines the current channel state data with the hidden state of the previous moment and calculates information retention parameters and information update parameters based on the combination result. The information retention parameter controls the retention ratio of load information from the previous moment, and the information update parameter controls the writing ratio of channel state information at the current moment. The gated cyclic coding layer updates the hidden state at the current moment according to the above parameters, gradually forming a hidden state sequence corresponding to the channel state sequence. Each hidden state in the hidden state sequence is used to characterize queue growth, processing latency changes, and transmission resource usage changes between adjacent moments, thus forming local load change characteristics between consecutive moments.

[0047] S33. Calculate the correlation weights between different times in the local load change features through the self-attention encoding layer, and aggregate the local load change features according to the correlation weights to obtain the load correlation features.

[0048] The self-attention encoding layer is used to establish load correlations between non-adjacent acquisition times. This layer linearly maps each local load change feature to corresponding query features, key features, and value features, and calculates the correlation weights between different times based on the correlation between the query and key features. These correlation weights, after normalization, are used to weighted aggregate the value features at each time point, ensuring that historical times with a high correlation to the current load change trend have a greater feature contribution. Through this process, the impact of periodic load changes, phased queue backlogs, and historical load peaks on the current prediction results can be encoded into the load correlation features, avoiding prediction of subsequent channel loads based solely on local changes at adjacent times.

[0049] S34. The load prediction layer performs mapping calculations on the load correlation features to obtain the predicted load peak and load change rate within the preset prediction period, and generates predicted load parameters based on the predicted load peak and load change rate.

[0050] The load prediction layer transforms load-related characteristics into predicted load parameters that can participate in subsequent scheduling decisions. The load prediction layer may include a peak mapping unit, a rate of change mapping unit, and a parameter combination unit. The peak mapping unit performs a linear transformation on the load-related characteristics using a preset weight matrix and preset bias parameters to obtain the predicted load peak value within a preset prediction period. The rate of change mapping unit performs another set of linear transformations on the load-related characteristics to obtain the predicted rate of change of channel load over time. The parameter combination unit normalizes the predicted load peak value and the rate of change of load, and generates predicted load parameters according to preset combination rules. The predicted load parameters represent the load level and load change trend of the corresponding push channel within the preset prediction period, providing input for subsequent selection of target push channels and calculation of target resource occupancy parameters.

[0051] The output of the gated cyclic coding layer is connected to the input of the self-attention coding layer, and the output of the self-attention coding layer is connected to the input of the load prediction layer. The gated cyclic coding layer extracts local load change features between consecutive time points from the channel state sequence. The self-attention coding layer aggregates load correlation information between different time points. The load prediction layer converts the aggregated load correlation features into predicted load peaks and load change rates. The channel state data input in step S3 is the basic state data formed by each push channel at multiple historical acquisition times. The local load change features formed in step S32 and the load correlation features formed in step S33 are intermediate processing data within the channel load prediction model. The predicted load parameters output in step S34 are prediction result data used for subsequent scheduling decisions.

[0052] As an optional implementation, the training samples for the channel load prediction model can be constructed based on historical channel state sequences and the actual channel load results within the corresponding prediction period. The historical channel state data undergoes anomaly cleaning, missing value imputation, time alignment, and numerical normalization. A continuous historical channel state sequence is then extracted according to a preset historical time window. Based on the actual load data within the preset prediction period following the historical channel state sequence, the actual load peak and actual load change rate are calculated and used as training labels. These training labels only reflect the message processing capacity and resource consumption of the push channel and do not include evaluations of the message recipient's identity, economic capacity, health status, or life value.

[0053] Specifically, the historical channel state sequence is input into the initial gated recurrent encoding layer to obtain training local load change features. These features are then input into the initial self-attention encoding layer to obtain training load association features. Finally, these features are input into the initial load prediction layer to obtain the predicted load peak and the predicted load change rate. The peak prediction loss is calculated based on the difference between the predicted and actual load peaks, and the rate prediction loss is calculated based on the difference between the predicted and actual load change rates. These two losses are then weighted and summed to obtain the load prediction loss. The model parameters in the gated recurrent encoding layer, self-attention encoding layer, and load prediction layer are updated based on the load prediction loss until the load prediction loss meets a preset convergence condition, resulting in a pre-trained channel load prediction model. As an example, the historical time window can be set to 30 minutes to 24 hours, the preset prediction period can be set to 5 minutes to 2 hours, the training batch size can be set to 32 to 128, the learning rate can be set to 0.0001 to 0.001, and the maximum number of training epochs can be set to 50 to 200.

[0054] As an example, in a fintech scenario, push channels can include application message channels, SMS channels, and in-site message channels. The input data for the channel load prediction model can include the message queuing status, processing latency status, and transmission resource occupancy status of each push channel at continuous acquisition times. A gated cyclic coding layer extracts local load change features between consecutive moments, a self-attention coding layer extracts the correlation features between peak transaction periods and the current channel status, and the load prediction layer outputs the predicted peak load and load change rate of each push channel in subsequent prediction periods, for use in channel selection and resource allocation for account change notifications, transaction confirmation notifications, and risk warning messages. In a digital healthcare scenario, push channels can include medical application message channels, SMS channels, and in-hospital terminal message channels. The channel load prediction model performs the same recursive encoding, correlation aggregation, and load mapping processing on the channel status sequences corresponding to appointment notifications, examination result notifications, follow-up visit reminders, and medication reminders, outputting the predicted load parameters for each push channel. In a digital healthcare scenario, the predicted load parameters are only used for technical resource scheduling of push channels and are not used to determine patient treatment priorities or replace medical staff in allocating medical resources.

[0055] Through the above steps S31 to S34, the short-term load change relationship and cross-time load correlation relationship in the channel state sequence can be extracted simultaneously. Based on the load correlation characteristics, the load peak and load change trend in subsequent periods can be predicted, thereby improving the accuracy and stability of push channel load prediction and providing a reliable prediction basis for target push channel selection and dynamic allocation of push resources.

[0056] S4. Obtain the running status data of the target receiver corresponding to the message to be pushed, input the running status data into the pre-trained processing status calculation model, and output the available processing parameters of the target receiver.

[0057] The processing status calculation model comprehensively analyzes the target receiver's operational activity, interface focus changes, and historical message processing time over a continuous period, and calculates the target receiver's current availability for processing new push messages. The available processing parameter characterizes the target receiver's likelihood of promptly viewing and processing pending push messages in its current operating state, and serves as input for subsequently determining the target push timing and generating the target push strategy. In the customer service agent scenario, the processing status calculation model can be implemented as an agent attention model, calculating the available processing parameter A based on keyboard operation frequency, mouse operation frequency, window focus change records, and historical message processing time. The available processing parameter A can be in the range of 0 to 1; a larger value indicates a higher degree of capability for the target receiver to process new push messages.

[0058] As an optional implementation, the above-mentioned processing state calculation model includes a state segmentation layer, a temporal coding layer, and a state fusion layer. The processing procedure of step S4 may also specifically include the following steps S41 to S44.

[0059] S41. The running state data is segmented according to the sliding time window through the state segmentation layer to obtain multiple running state data segments.

[0060] The operational status data characterizes the interactive operation status and message processing status of the target receiver during continuous operation. The status segmentation layer continuously moves a sliding time window along the time axis corresponding to the operational status data, according to a preset window length and a preset sliding step size, dividing the operational status data covered by each sliding time window into an operational status data segment. Adjacent operational status data segments may have partially overlapping data intervals to preserve the continuity of the target receiver's status changes within adjacent time ranges. For operational status data with discontinuous acquisition times, time alignment can be performed according to the acquisition timestamp, and missing time positions can be filled with null values ​​or supplemented with adjacent values. Using a sliding time window for segmentation is to convert continuously changing operational states into time-series processing units of relatively consistent length, avoiding direct overall processing of continuous operational data of uncertain length.

[0061] S42. Extract operation activity features, interface focus change features, and message processing duration features from each running status data segment.

[0062] Among them, the operation activity feature characterizes the frequency of interactive operations on the target receiver within the corresponding operational status data segment, and can be calculated based on the number of interactive operations, the time interval between adjacent interactive operations, and the duration of continuous inactivity. The interface focus change feature characterizes the focus transfer between different application interfaces or function windows, and can be calculated based on the number of focus switches, the duration of focus, and the focus occupancy ratio of the message processing interface. The message processing duration feature characterizes the time taken by the target receiver to process historical push messages, and can be generated based on the time difference between the message display time, the message viewing time, and the message processing completion time. The above features can be numerically normalized to convert operational status features of different dimensions to a unified numerical range, providing consistent input for subsequent time-series coding. Operational status features are only used to characterize the technical status of the target receiver in processing push messages; they do not collect personal content unrelated to message processing, nor are they used to evaluate the message recipient's work ability, health status, or personal value.

[0063] S43. By using a time-series coding layer to perform time-series coding on the operation activity features, interface focus change features, and message processing duration features, multiple processing state change features are obtained.

[0064] The temporal coding layer is used to extract the state change relationships between consecutive running state data segments of the target receiver. It concatenates the operation activity features, interface focus change features, and message processing duration features corresponding to the same running state data segment to obtain the state feature vector at that moment. This state feature vector is then sequentially input into the temporal coding layer according to the chronological order of each running state data segment. The temporal coding layer performs a combination operation between the current moment's state feature vector and the hidden state from the previous moment, updating the current moment's hidden state based on the result. This ensures that the current hidden state retains both the previous and current running state information. The multiple hidden states output by the temporal coding layer serve as multiple processing state change features, representing the process of the target receiver changing from an active state to an idle state, from a message processing state to another operation state, or from a short-term processing state to a continuously occupied state.

[0065] S44. Calculate the state weights corresponding to the state change features of each processing state through the state fusion layer, and perform weighted aggregation of the state change features of each processing state according to the state weights to obtain the available processing parameters.

[0066] The state fusion layer adjusts the participation ratio of processing state change features in the calculation of available processing parameters based on the correlation between each processing state change feature and the current message processing capability. The state fusion layer can linearly map each processing state change feature to obtain the corresponding state mapping feature, then calculate the correlation between each state mapping feature and a preset state query vector to obtain the initial state weight corresponding to each processing state change feature. A normalization function is used to process multiple initial state weights to ensure that each state weight is within a preset numerical range and the sum of the weights satisfies preset constraints. The state fusion layer performs a weighted summation of multiple processing state change features according to the state weights to obtain aggregated state features. These aggregated state features are then input into the parameter mapping layer, where a linear transformation is performed using a weight matrix and bias parameters to obtain the available processing parameters. The available processing parameters characterize the availability of the target receiver for processing new push messages in the current operating state and are not used to evaluate the target receiver's capabilities, identity, or social value.

[0067] The output of the state segmentation layer is connected to the input of the feature extraction unit, the output of the feature extraction unit is connected to the input of the temporal coding layer, the output of the temporal coding layer is connected to the input of the state fusion layer, and the output of the state fusion layer is connected to the input of the parameter mapping layer. The state segmentation layer converts continuous running state data into multiple running state data segments. The feature extraction unit generates operation activity features, interface focus change features, and message processing duration features from each running state data segment. The temporal coding layer extracts the state change relationships between continuous running state data segments. The state fusion layer aggregates multiple processing state change features, and the parameter mapping layer converts the aggregated state features into usable processing parameters. The running state data input in step S4 belongs to the basic running data formed by the target receiver at multiple acquisition times. The processing state change features obtained in step S43 and the aggregated state features formed in step S44 belong to the intermediate processing data within the processing state calculation model. The usable processing parameters belong to the model output data used for generating subsequent push strategies.

[0068] As an optional implementation, the training samples for the processing status calculation model can be constructed based on historical operational status data and historical message processing results within the corresponding time range. The historical operational status data undergoes data anonymization, abnormal data cleanup, missing value imputation, and time alignment, and multiple historical operational status data segments are generated according to a sliding time window. Available processing tags are generated based on whether message viewing operations exist within a preset time range following each historical operational status data segment, the message viewing latency, and the message processing completion time. Available processing tags are only used to characterize the technical status of the target receiving end in processing push messages; the recipient's identity, occupation, economic capacity, health status, disease type, or life value are not used as conditions for tag generation.

[0069] Specifically, historical operational state data is input into the initial state segmentation layer to obtain multiple training operational state data segments. Training operation activity features, training interface focus change features, and training message processing duration features are extracted from each training operational state data segment and input into the initial temporal encoding layer to obtain multiple training processing state change features. The initial state fusion layer calculates the training state weights corresponding to each training processing state change feature, and performs weighted aggregation according to the training state weights to obtain training aggregated state features. The training aggregated state features are input into the initial parameter mapping layer to obtain predicted available processing parameters. The state prediction loss is calculated based on the difference between the predicted available processing parameters and the available processing labels, and the model parameters in the temporal encoding layer, state fusion layer, and parameter mapping layer are updated based on the state prediction loss until the state prediction loss meets the preset convergence condition, resulting in a pre-trained processing state calculation model. As an example, the window length of the sliding time window can be set from 1 minute to 30 minutes, the sliding step size can be set from 10 seconds to 5 minutes, the training batch size can be set from 32 to 128, the learning rate can be set from 0.0001 to 0.001, and the maximum number of training rounds can be set from 50 to 200.

[0070] As an example, in a fintech scenario, the target receiving end can be a business terminal used to process account service messages or transaction confirmation messages. The input data for the processing state calculation model can include de-identified and aggregated terminal interaction operation times, business interface focus change records, and historical message processing durations. The state segmentation layer divides continuous running state data into multiple running state data segments, the time-series coding layer extracts the operation state changes between different time periods, and the state fusion layer performs weighted aggregation according to the correlation between each processing state change feature and the message processing state, outputting available processing parameters within a preset numerical range for subsequent push timing and channel selection. In a digital healthcare scenario, the target receiving end can be a terminal used to receive appointment notifications, examination result notifications, or follow-up messages. The processing state calculation model can perform the same segmentation, time-series coding, and state fusion processing on the terminal interaction state, message interface focus state, and historical notification processing duration, outputting available processing parameters. In fintech scenarios, available processing parameters are not used to evaluate staff’s business capabilities, performance, or personal value. In digital healthcare scenarios, available processing parameters are not used to evaluate the life value of patients or medical staff, to determine treatment priorities, or to replace medical staff in making decisions on treatment or allocation of medical resources.

[0071] Through the above steps S41 to S44, the continuously changing operating status data can be converted into multiple processing status change features with time correlation. Based on the correlation between the status features at different times and the current message processing capacity, dynamic weighted aggregation is performed to improve the accuracy of the available processing parameters in representing the real-time processing status of the target receiving end, and to provide a reliable status basis for target push timing selection and push resource scheduling.

[0072] S5. Input the message value parameters, the predicted load parameters of each push channel, and the available processing parameters into the pre-trained scheduling decision model, and output the target push strategy corresponding to the message to be pushed. The target push strategy includes the target push channel, the target push timing, and the target resource usage parameters.

[0073] The scheduling decision model comprehensively analyzes message value parameters, predicted load parameters for each push channel, and available processing parameters at the target receiver. Based on the correlation between these three types of parameters, it generates a target push strategy that matches the current push conditions. The target push strategy jointly determines the push channel, push timing, and resource occupancy level corresponding to the message to be pushed, serving as the basis for subsequent push resource allocation and message sending.

[0074] In one specific implementation, the scheduling decision model can be trained using a near-end policy optimization algorithm. During training, the historical scheduling state, formed by message value parameters, predicted load parameters for each push channel, and available processing parameters, serves as the state space. The historical push actions, including push channel, push timing, and resource consumption parameters, serve as the action space. Scheduling reward parameters are generated based on message push latency and message delivery results. By limiting the action probability ratio between the current policy network and the reference policy network, the parameter update magnitude of the policy output layer is controlled.

[0075] As an optional implementation, the scheduling decision model described above includes a state encoding layer and a policy output layer. Please continue reading. Figure 4 , Figure 4 This is a schematic diagram illustrating the execution flow of the target push strategy obtained in the intelligent scheduling method for message push resources provided in this application embodiment, as shown below. Figure 4 As shown, the processing procedure of step S5 above may also include the following steps S51 to S53.

[0076] S51. The message value parameters, the predicted load parameters of each push channel, and the available processing parameters are encoded and feature fused through the state coding layer to obtain the joint scheduling state features.

[0077] The state coding layer transforms message value parameters, predicted load parameters, and available processing parameters, which differ in value range, data dimension, and representational meaning, into a unified feature space. The state coding layer includes a parameter normalization unit, a feature mapping unit, and a state fusion unit. The parameter normalization unit scales the message value parameters, predicted load parameters for each push channel, and available processing parameters to ensure they fall within a preset range. The feature mapping unit performs linear mapping on the normalized parameters using corresponding weight matrices and bias parameters to obtain message value mapping features, channel load mapping features, and available processing mapping features. The state fusion unit concatenates these mapping features and, through fully connected operations and nonlinear activation processing, obtains the joint scheduling state features.

[0078] The predicted load parameters for each push channel can be arranged according to a preset channel order and combined with the corresponding channel identifier features, enabling the joint scheduling status feature to distinguish the load status of different push channels. By encoding message value, channel load, and the available processing status of the target receiver into the same joint scheduling status feature, it is possible to avoid selecting push channels, push timing, or push resources based on a single parameter, allowing the three scheduling elements to be jointly calculated under the same state conditions.

[0079] S52. The joint scheduling state features are mapped through the policy output layer to obtain the action probabilities corresponding to multiple candidate push actions. Each candidate push action includes the candidate push channel, candidate push timing and candidate resource occupation parameters.

[0080] Candidate push actions can be generated by combining selectable push channels, preset push time intervals, and preset resource usage levels. Candidate push timing can be divided into immediate push and multiple delayed push time intervals according to preset time granularity. Candidate resource usage parameters can be divided into multiple resource usage levels based on the number of push connections, the number of processing threads, transmission bandwidth, or message concurrency. Different combinations of candidate push channels, candidate push timings, and candidate resource usage parameters correspond to a single candidate push action.

[0081] The policy output layer can include an action mapping unit and a probability normalization unit. The action mapping unit maps the joint scheduling state characteristics using a preset action weight matrix and preset action bias parameters to obtain the action score corresponding to each candidate push action. The probability normalization unit normalizes multiple action scores to obtain the action probability corresponding to each candidate push action, ensuring that each action probability is within a preset range and the sum of the action probabilities is 1. For candidate push actions that are currently unavailable or whose resource usage parameters exceed the upper limit of the allocated resources for the corresponding push channel, an action blocking flag can be set. When calculating action probabilities, the policy output layer reduces or excludes candidate push actions with action blocking flags to avoid outputting push policies that cannot be actually executed. Action probabilities are used to characterize the suitability of executing the corresponding candidate push action in the current joint scheduling state and are not used to evaluate the message recipient's identity, ability, economic status, health condition, or personal value.

[0082] S53. Select the candidate push action with the highest probability, and use the candidate push channel, candidate push timing and candidate resource usage parameters corresponding to the selected candidate push action as the target push strategy.

[0083] This process involves comparing the probabilities of multiple candidate push actions and selecting the one with the highest probability. If multiple candidate push actions have the same maximum probability, a preset action selection rule can be used to select the one with the lowest predicted load parameter, lowest resource consumption parameter, or earliest push timing. The selected candidate push action is then parsed to obtain the corresponding candidate push channel, candidate push timing, and candidate resource consumption parameter, which are then used as the target push channel, target push timing, and target resource consumption parameter, respectively. By selecting the target push channel, push timing, and target resource consumption parameter as a whole, the correspondence between the three strategy parameters is maintained, avoiding mismatches that could result from selecting the push channel, push timing, and resource consumption parameter separately.

[0084] The output of the state coding layer is connected to the input of the policy output layer. The state coding layer converts message value parameters, predicted load parameters for each push channel, and available processing parameters into joint scheduling state features. The policy output layer calculates the action probability corresponding to each candidate push action based on these joint scheduling state features. The message value parameters, predicted load parameters, and available processing parameters input in step S5 are the output results of the message value prediction model, the channel load prediction model, and the processing state calculation model, respectively. The joint scheduling state features formed in step S51 are intermediate processing data within the scheduling decision model. The action probabilities output in step S52 are the decision evaluation results for each candidate push action. The target push policy output in step S53 is the model output data used for executing message push and resource allocation.

[0085] As an optional implementation, the training samples for the scheduling decision model can be constructed based on historical joint scheduling states, corresponding historical push actions, and historical push results. Historical joint scheduling states may include historical message value parameters, historical channel load parameters, and historical available processing parameters. Historical push actions may include the actual push channel used, push timing, and resource usage parameters. Historical push results may include message delivery latency, message processing latency, actual push channel load, and actual resource usage. The above data undergoes anomaly cleaning, missing value imputation, time alignment, and numerical normalization to form a training state sequence arranged according to push time.

[0086] Specifically, the historical message value parameters, historical channel load parameters, and historical available processing parameters from the training state sequence are input into the initial state encoding layer to obtain the training joint scheduling state features. These features are then input into the initial policy output layer to obtain the current action probability corresponding to each candidate push action, and the corresponding current policy probability is read based on the actually executed historical push actions. A state value evaluation branch can also be set during the training phase. The training joint scheduling state features are input into this branch to obtain the predicted state value of the corresponding scheduling state. Scheduling reward parameters are generated based on historical push results, and advantage parameters are calculated based on the scheduling reward parameters and the predicted state value. The scheduling reward parameters can be weighted based on message delivery latency, message processing latency, push channel load changes, and resource usage. When message delivery latency or message processing latency decreases, and push channel load and resource usage are within a preset range, the scheduling reward parameters are increased. When push fails, channel load exceeds a preset load threshold, or resource usage exceeds a preset resource threshold, the scheduling reward parameters are decreased. The policy probability ratio is calculated based on the current policy probability and the reference policy probability saved before the parameter update, and this ratio is limited to a preset truncation range. The strategy loss is calculated based on the truncated strategy probability ratio and dominance parameters. The value loss is calculated based on the difference between the predicted state value and the actual cumulative reward. The strategy distribution constraint loss is calculated based on the action probability corresponding to each candidate push action. The strategy loss, value loss, and strategy distribution constraint loss are weighted to obtain the scheduling training loss. The weight parameters and bias parameters in the state encoding layer, strategy output layer, and state value evaluation branch are updated in reverse based on the scheduling training loss. Joint state encoding, action probability calculation, reward calculation, loss calculation, and parameter update are repeated according to training batches until the scheduling training loss meets the preset convergence condition, resulting in a pre-trained scheduling decision model. The state value evaluation branch is mainly used in the model training process. During model inference, the target push strategy can be generated only through the state encoding layer and strategy output layer. As an example, the training batch size can be set to 32 to 128, the learning rate can be set to 0.00001 to 0.001, the truncation range of the strategy probability ratio can be set to 0.8 to 1.2, the number of parameter updates per batch can be set to 3 to 10, and the maximum number of training rounds can be set to 50 to 300.

[0087] As an example, in a fintech scenario, the message value parameter can be used to characterize the push value of account change notifications, transaction confirmation notifications, or risk warning messages. The predicted load parameter for each push channel can be used to characterize the load status of application message channels, SMS channels, and in-site message channels in subsequent periods. The available processing parameter can be used to characterize the availability of the target business terminal for processing new messages. The state encoding layer encodes and fuses the above parameters, and the strategy output layer calculates the action probability corresponding to different combinations of push channels, push timing, and resource occupancy levels, and outputs the target push strategy. The target push strategy is only used to determine the technical push method and resource configuration of the message; it is not used to evaluate the credit level, economic capacity, or personal value of the message recipient, nor is it used to make credit granting, transaction approval, or financial product pricing decisions. In a digital healthcare scenario, the message value parameter can be used to characterize the push value of appointment notifications, examination result notifications, follow-up visit reminders, or medication reminders. The predicted load parameter can be used to characterize the load status of medical application message channels, SMS channels, and in-hospital terminal message channels. The available processing parameter can be used to characterize the availability of the target receiver for processing new notifications. The scheduling decision model outputs the target push channel, target push timing, and target resource usage parameters based on the above parameters. The target push strategy is only used for message transmission and technical resource scheduling of medical notifications; it is not used to evaluate the value of a patient's life, determine the patient's treatment priority, or replace medical personnel in making decisions on disease diagnosis, treatment plans, or medical resource allocation.

[0088] Through the above steps S51 to S53, the target push channel, target push timing, and target resource usage parameters can be jointly determined by combining message value, channel load, and receiver processing status, thereby improving the adaptability of the push strategy, reducing message delivery delay, and improving push resource utilization.

[0089] As another optional implementation, the training process of the above-mentioned scheduling decision model may further include the following steps S501 to S506: S501. Obtain the historical scheduling status, the historical push action corresponding to the historical scheduling status, and the push result corresponding to the historical push action.

[0090] Historical scheduling status can include historical message value parameters, historical predicted load parameters for each push channel, and historical available processing parameters. Historical push actions can include the actual push channel used, push timing, and resource consumption parameters. Push results can include message push latency, whether the message was successfully delivered, whether the message was viewed, and whether the corresponding business operation was completed. Historical scheduling status, historical push actions, and push results can be correlated according to message identifier and push time. The correlated data can then undergo anomaly cleanup, missing value imputation, time alignment, and numerical normalization to form training samples for scheduling decisions. Historical scheduling status serves as training input, historical push actions are the actions being evaluated, and push results are used to calculate the scheduling reward parameters obtained after executing the historical push action.

[0091] S502. Calculate the scheduling reward parameters based on the message push latency and message delivery results in the push results.

[0092] The system can normalize message push latency to obtain a latency evaluation value, and generate a reach evaluation value based on whether the message was successfully delivered, viewed within a preset time, and whether the corresponding business operation was completed. The latency evaluation value and the reach evaluation value are weighted according to preset latency weights and preset reach weights to obtain a scheduling reward parameter. A higher scheduling reward parameter can be set when the message push latency is shorter and the message is successfully delivered or processed within a preset time. The scheduling reward parameter can be reduced when the message push fails, the push latency exceeds a preset latency threshold, or the preset reach conditions are not met. The scheduling reward parameter is only used to evaluate the execution result of the push action and is not used to evaluate the message recipient's identity, ability, economic status, health condition, or life value.

[0093] S503. Input the historical scheduling status into the value calculation layer, output the status value corresponding to the historical scheduling status, and calculate the strategy advantage parameter based on the scheduling reward parameter and the status value.

[0094] The value calculation layer can include a state feature mapping unit and a value output unit. The state feature mapping unit performs linear transformation and nonlinear activation processing on the joint scheduling state features corresponding to historical scheduling states to obtain value mapping features. The value output unit maps the value mapping features according to the value weight matrix and value bias parameters, outputting the state value corresponding to the historical scheduling state. The state value represents the expected cumulative reward that can be obtained by executing a push action according to the current scheduling strategy under the corresponding historical scheduling state. A strategy advantage parameter can be calculated based on the scheduling reward parameter corresponding to the current historical scheduling state, the state value corresponding to the next historical scheduling state, and the state value corresponding to the current historical scheduling state. A positive strategy advantage parameter indicates that the actual execution result of the historical push action is better than the expected result of the current strategy; a negative strategy advantage parameter indicates that the actual execution result of the historical push action is lower than the expected result of the current strategy.

[0095] S504. Calculate the probability of historical push actions under the current policy network and the reference policy network respectively, and calculate the probability ratio of the two action probabilities.

[0096] Both the current policy network and the reference policy network can include a state encoding layer and a policy output layer, and they have the same network structure. The output of the state encoding layer is connected to the input of the policy output layer. The current policy network is used to calculate the probability of the current action corresponding to the historical push action based on the current network parameters, and the reference policy network is used to calculate the probability of the reference action corresponding to the historical push action based on the reference network parameters saved before the parameter update. At the start of training, the network parameters of the current policy network can be copied to the reference policy network. During multiple parameter updates in a training batch, the network parameters of the reference policy network remain unchanged. The ratio of the current action probability to the reference action probability is calculated to obtain the probability ratio. When the probability ratio is greater than 1, it indicates that the current policy network has increased the output probability of the historical push action; when the probability ratio is less than 1, it indicates that the current policy network has decreased the output probability of the historical push action.

[0097] S505. Limit the probability ratio to a preset range, and calculate the strategy loss parameter based on the limited probability ratio and the strategy advantage parameter.

[0098] Specifically, the probability ratio can be compared with preset lower and upper limits. When the probability ratio is less than the preset lower limit, it is set to the preset lower limit; when it is greater than the preset upper limit, it is set to the preset upper limit; when the probability ratio is within the preset range, the original probability ratio is retained. The first product of the original probability ratio and the policy advantage parameter, and the second product of the restricted probability ratio and the policy advantage parameter are calculated respectively, and the policy loss parameter is generated based on the first and second products. By restricting the probability ratio, the parameter variation of the current policy network relative to the reference policy network can be controlled, preventing excessive changes in the action probability of candidate push actions during a single training iteration.

[0099] S506. Update the network parameters of the policy output layer according to the policy loss parameters to obtain the pre-trained scheduling decision model.

[0100] The process involves backpropagation based on the policy loss parameters to calculate the parameter gradients corresponding to the action weight matrix and action bias parameters in the policy output layer, and updating the action weight matrix and action bias parameters according to a preset learning rate. This process is repeated for each training batch, involving encoding historical scheduling states, calculating action probabilities, calculating policy loss, and updating network parameters, until the policy loss parameters meet a preset convergence condition, resulting in a pre-trained scheduling decision model.

[0101] In one implementation, a value loss parameter can be calculated based on the difference between the scheduling reward parameter and the state value, and the value weight matrix and value bias parameter in the value calculation layer can be updated based on the value loss parameter. The value calculation layer is mainly used to calculate the state value and policy advantage parameters during the model training phase, and the target push policy can be directly generated through the state encoding layer and policy output layer during the model inference phase.

[0102] The output of the state coding layer is connected to the inputs of the policy output layer in the current policy network, the policy output layer in the reference policy network, and the value calculation layer, respectively. The current policy network and the reference policy network receive the same joint scheduling state features and output the action probability corresponding to each candidate push action. The value calculation layer receives the joint scheduling state features and outputs the state value. The initial network parameters of the reference policy network are based on the network parameter settings of the current policy network before the parameter update, and are resynchronized after a preset number of parameter updates.

[0103] As an example, the training batch size of the scheduling decision model can be set to 32 to 128, the learning rate can be set to 0.00001 to 0.001, the preset ratio range can be set to 0.8 to 1.2, the number of parameter updates per training batch can be set to 3 to 10, the cumulative reward discount parameter can be set to 0.9 to 0.99, and the maximum number of training epochs can be set to 50 to 300. Training can be stopped when the change in policy loss corresponding to multiple consecutive training epochs is less than the preset loss threshold.

[0104] As an example, in a fintech scenario, historical scheduling status can include historical message value parameters corresponding to account change notifications, historical predicted load parameters corresponding to application message channels and SMS channels, and historical available processing parameters corresponding to the target business terminal. Scheduling reward parameters are calculated based on the actual push channel, actual push timing, actual resource usage parameters, and their corresponding message push latency and reach results, and the strategy output layer is updated using the above training process. In a digital healthcare scenario, a scheduling decision model can be trained based on the historical scheduling status, historical push actions, and actual push results corresponding to appointment notifications, examination result notifications, or medication reminders. The trained scheduling decision model is only used for scheduling message push methods and technical resources; it is not used to determine patient treatment priorities, nor does it replace medical personnel in making decisions regarding treatment or medical resource allocation.

[0105] Through steps S501 to S506, the scheduling reward can be calculated based on the actual execution results of historical push actions, and the single update range of the scheduling strategy can be limited, thereby improving the stability of the training process of the scheduling decision model and the adaptability of the push strategy.

[0106] S6. Allocate push resources corresponding to the target push channel according to the target resource occupancy parameters, and push the message to be pushed through the target push channel at the target push time.

[0107] The target resource occupancy parameter limits the channel resources required for this message push, and may include the number of connections, the number of processing threads, transmission bandwidth, message concurrency, or message queue capacity. Based on the target resource occupancy parameter, the push resources required for this message push are allocated from the available resources corresponding to the target push channel, and a correspondence is established between the message to be pushed, the target push channel, and the allocated push resources. When the target push timing is reached, the message to be pushed is written to the message queue corresponding to the target push channel, and the allocated push resources are used to perform message encapsulation, transmission, and delivery confirmation. After the message push is completed, the push resources occupied by this message push can be released, or some push resources can be reserved based on the number of subsequent messages to be pushed in the target push channel. By pre-allocating push resources according to the target resource occupancy parameter, processing wait caused by temporarily requesting resources after the target push timing arrives can be avoided, while limiting the channel resources occupied by a single message to be pushed, reducing resource contention or channel congestion when multiple messages are pushed simultaneously.

[0108] As an optional implementation, after pushing the message to be pushed through the target push channel in step S6 above, the following steps S71 to S74 may also be included.

[0109] S71. Obtain the actual processing result and actual delivery delay of the message to be pushed.

[0110] The actual processing result characterizes the delivery and processing status of the message to be pushed after the push is completed. This can include whether the message was successfully delivered, whether it was viewed, the viewing time, whether the corresponding business operation was completed, and the business operation completion time. The actual delivery delay can be calculated based on the time difference between the moment the message to be pushed begins transmission through the target push channel and the moment the target receiver confirms receipt of the message. The message delivery status and time can be obtained from the delivery confirmation information returned by the target push channel, and the message viewing status and business operation status can be obtained from the message viewing record or business processing record returned by the target receiver. Associating and storing the actual processing result and actual delivery delay with the message identifier of the message to be pushed ensures that the feedback data used for subsequent deviation calculations matches the corresponding message value parameters and the target push strategy.

[0111] S72. Calculate the value prediction deviation based on the actual processing results and message value parameters, and calculate the delay deviation based on the actual delivery delay and the target push timing.

[0112] The actual processing results can be converted into actual message value according to preset numerical conversion rules. For example, a higher actual message value can be set for messages that are successfully delivered and viewed or whose business operations are completed within a preset time range; a lower actual message value can be set for messages that are not successfully delivered, not viewed, or processed after the preset time range. The numerical difference between the actual message value and the message value parameter output in step S2 is calculated to obtain the value prediction deviation. The target push timing can be associated with a corresponding baseline delivery delay, which represents the time required to complete message delivery under the target push timing and target push channel. The difference between the actual delivery delay and the baseline delivery delay is calculated to obtain the delay deviation. A positive delay deviation indicates that the actual delivery speed is lower than expected; a negative or zero delay deviation indicates that the actual delivery speed meets or exceeds expectations. By calculating the value prediction deviation and the delay deviation respectively, a correspondence can be established between the actual message processing situation and the aforementioned message value prediction results and push strategy execution results, providing feedback for subsequent adjustments to the message value prediction model and scheduling decision model.

[0113] S73. Generate message value training samples based on the value prediction bias, and use the message value training samples to update the model parameters of the message value prediction model.

[0114] This process involves associating message feature data, actual processing results, and value prediction bias for the messages to be pushed, generating message value training samples. Training labels in these samples can be generated based on actual processing results, and the value prediction bias can be used to adjust the participation weight of corresponding training samples during model updates. Training samples with larger value prediction biases can be assigned higher participation weights, allowing the message value prediction model to prioritize learning message samples where the predicted results differ significantly from the actual processing results. The message value prediction model can be periodically updated using newly added training samples after the number of training samples reaches a preset limit, or it can be updated in batches by adding training samples to the existing training sample set. The updated message value prediction model is then used to process subsequent messages to be pushed, gradually adapting the output message value parameters to the actual viewing and processing patterns of messages.

[0115] S74. Generate scheduling decision training samples based on the time delay deviation and actual processing results, and use the scheduling decision training samples to update the model parameters of the scheduling decision model.

[0116] This involves correlating the message value parameters, predicted load parameters, available processing parameters, target push strategy, actual processing results, and latency deviation of the current message push to generate scheduling decision training samples. These training samples reflect the actual push results generated after executing the target push strategy under the corresponding joint scheduling state. When the actual delivery latency is less than or equal to the corresponding baseline delivery latency, and the message to be pushed is viewed or the corresponding business operation is completed within a preset time range, the strategy evaluation result represented by the scheduling decision training sample can be improved. When the actual delivery latency exceeds the baseline delivery latency, the message push fails, or the actual processing result does not meet the preset processing conditions, the corresponding strategy evaluation result can be reduced. The scheduling decision training samples are used to update the model parameters of the scheduling decision model, reducing the output probability of candidate push actions that produce large latency deviations or poor actual processing results, and increasing the output probability of candidate push actions that produce small delivery latency and good actual processing results. Before generating the message value training samples and scheduling decision training samples, identity information irrelevant to model updates can be deleted, and the correspondence can be verified according to message identifier, push time, and target push channel. The actual processing results are used only to evaluate the message push and processing status, and are not used to evaluate the message recipient's identity, ability, economic status, health status, or life value.

[0117] As an example, in a fintech scenario, the message to be pushed could be an account change notification. Based on target resource usage parameters, a transmission connection and message processing thread are allocated to the application's message channel, and the account change notification is sent at the target push time. After the message is sent, the system obtains whether the message was successfully delivered, whether it was viewed, the viewing duration, and the actual delivery delay. If the message value parameter is high, but the account change notification is not viewed for a long time, a large value prediction deviation is generated based on the actual processing results, and a message value training sample is generated accordingly. If the actual delivery delay exceeds the baseline delivery delay corresponding to the target push time, a scheduling decision training sample is generated based on the delay deviation and the actual processing results to adjust the push channel, push time, and resource usage parameters for subsequent similar messages. In a digital healthcare scenario, the message to be pushed could be an appointment notification, examination result notification, or medication reminder. Push resources corresponding to the medical application's message channel or SMS channel are allocated based on target resource usage parameters, and the message to be pushed is sent at the target push time. After the message is pushed, the system obtains the message delivery status, viewing status, and actual delivery delay, and generates message value training samples and scheduling decision training samples based on the feedback results. The feedback data mentioned above is only used to optimize the message push process and is not used to determine the patient's treatment priority, nor does it replace medical staff in making decisions on disease diagnosis, treatment plans, or allocation of medical resources.

[0118] Through the above steps S71 to S74, feedback training samples can be generated and relevant models updated based on the actual processing results and actual delivery delays, thereby improving the accuracy of message value prediction and push strategies, reducing message delivery delays, and improving the utilization rate of push resources.

[0119] The intelligent scheduling method for message push resources provided in this application analyzes the content characteristics and historical interaction patterns of the message to be pushed to obtain message value parameters; analyzes the historical state changes of multiple push channels to predict the load level and trend of each push channel; analyzes the operating status of the target receiver to calculate the available processing parameters of the target receiver; and then comprehensively considers the message value parameters, predicted load parameters, and available processing parameters to jointly determine the target push channel, target push timing, and target resource occupancy parameters, and allocates push resources and executes message push according to the target push strategy. This enables the push strategy to be adapted to message value, channel load, and receiver processing status, reducing channel congestion and resource contention, lowering message delivery delays, and improving push resource utilization.

[0120] Please continue reading. Figure 5 , Figure 5 This is a schematic diagram of the system structure of the intelligent scheduling device for message push resources provided in the embodiments of this application, such as... Figure 5 As shown, the intelligent scheduling device 50 for message push resources includes: a message feature acquisition module 51 for acquiring message feature data of a message to be pushed; a message value prediction module 52 for inputting the message feature data into a pre-trained message value prediction model and outputting message value parameters of the message to be pushed; a channel load prediction module 53 for acquiring channel status data of multiple push channels, inputting the channel status data into a pre-trained channel load prediction model, and outputting predicted load parameters of each push channel; and a processing status calculation module 54 for acquiring the operating status data of the target receiving end corresponding to the message to be pushed and inputting the operating status data into a pre-trained channel load prediction model. The pre-trained processing state calculation model is input, and the available processing parameters of the target receiver are output. The push strategy generation module 55 is used to input the message value parameters, the predicted load parameters of each push channel, and the available processing parameters into the pre-trained scheduling decision model, and output the target push strategy corresponding to the message to be pushed. The target push strategy includes the target push channel, the target push timing, and the target resource occupancy parameters. The push allocation and execution module 56 is used to allocate the push resources corresponding to the target push channel according to the target resource occupancy parameters, and push the message to be pushed through the target push channel at the target push timing.

[0121] It should be noted that the aforementioned intelligent scheduling device for message push resources can execute the intelligent scheduling method for message push resources provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in the embodiments of the intelligent scheduling device for message push resources can be found in the intelligent scheduling method for message push resources provided in the embodiments of this application.

[0122] Figure 6 This is a schematic diagram of the hardware structure of the electronic device for implementing the intelligent scheduling method for push message resources provided in this application embodiment, as shown below. Figure 6 As shown, the electronic device 600 includes: One or more processors 610 and memory 620, Figure 6 Take the 610 processor as an example.

[0123] The processor 610 and the memory 620 can be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.

[0124] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the message push resource intelligent scheduling method in the embodiments of this application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 620, thereby realizing the message push resource intelligent scheduling method in the above-described method embodiments.

[0125] The memory 620 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the message push resource intelligent scheduling device. Furthermore, the memory 620 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include memory remotely located relative to the processor 610, and these remote memories can be connected to the message push resource intelligent scheduling device via a network. Examples of such networks include, but are not limited to, the Internet, enterprise intranets, local area networks, mobile communication networks, and combinations thereof.

[0126] The one or more modules are stored in the memory 620. When executed by the one or more processors 610, they perform the message push resource intelligent scheduling method in any of the above method embodiments, for example, performing the above-described method. Figure 2 Method steps S1 to S6, Figure 3 Method steps S31 to S34, Figure 4 Steps S51 to S53 in the method are implemented. Figure 5 The functions of modules 51-56 in the document.

[0127] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0128] This application provides a non-volatile computer-readable storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 6 One of the processors 610 enables the one or more processors to execute the message push resource intelligent scheduling method in any of the above method embodiments, for example, to execute the method described above. Figure 2 Method steps S1 to S6, Figure 3 Method steps S31 to S34, Figure 4 Steps S51 to S53 in the method are implemented. Figure 5 The functions of modules 51-56 in the document.

[0129] This application provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium. The computer program includes program instructions, which, when executed by an electronic device, enable the electronic device to execute the message push resource intelligent scheduling method described in any of the above method embodiments, for example, to execute the method described above. Figure 2 Method steps S1 to S6, Figure 3 Method steps S31 to S34, Figure 4 Steps S51 to S53 in the method are implemented. Figure 5 The functions of modules 51-56 in the document.

[0130] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0131] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software and a general-purpose hardware platform, or of course, using hardware. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0132] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of this application as described above, which are not provided in detail for the sake of brevity; although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent scheduling of message push resources, characterized in that, include: Obtain message feature data of the message to be pushed; The message feature data is input into a pre-trained message value prediction model, which outputs the message value parameters of the message to be pushed. Obtain channel status data for multiple push channels, input the channel status data into a pre-trained channel load prediction model, and output the predicted load parameters for each push channel. Obtain the running status data of the target receiving end corresponding to the message to be pushed, input the running status data into the pre-trained processing status calculation model, and output the available processing parameters of the target receiving end. The message value parameters, the predicted load parameters of each push channel, and the available processing parameters are input into a pre-trained scheduling decision model, and the target push strategy corresponding to the message to be pushed is output. The target push strategy includes the target push channel, the target push timing, and the target resource consumption parameters. Based on the target resource occupancy parameters, the push resources corresponding to the target push channel are allocated, and the message to be pushed is pushed through the target push channel at the target push time.

2. The intelligent scheduling method for message push resources according to claim 1, characterized in that, The message value prediction model includes a semantic coding branch, a temporal coding branch, and a feature fusion layer; The step of inputting the message feature data into a pre-trained message value prediction model and outputting the message value parameters of the message to be pushed includes: The semantic encoding branch is used to perform contextual semantic encoding on the message to be pushed to obtain message semantic features; According to the interaction time sequence, the historical interaction sequence corresponding to the message to be pushed is recursively calculated through the time-series coding branch to obtain the interaction behavior characteristics. Through the feature fusion layer, dynamic fusion weights corresponding to the message semantic features and the interaction behavior features are calculated based on the historical message processing results. The message semantic features and the interaction behavior features are weighted and fused according to the dynamic fusion weight to obtain the message feature data, and the message value parameter is calculated based on the message feature data.

3. The intelligent scheduling method for message push resources according to claim 1, characterized in that, The channel load prediction model includes a gated cyclic coding layer, a self-attention coding layer, and a load prediction layer. The step of inputting the channel status data into a pre-trained channel load prediction model and outputting the predicted load parameters for each push channel includes: The channel status data of each push channel are sorted according to the collection time to obtain the corresponding channel status sequence; The gated cyclic coding layer is used to recursively operate on the channel state sequence to obtain the local load change characteristics between consecutive time moments; The correlation weights between different times in the local load change features are calculated through the self-attention encoding layer, and the local load change features are aggregated according to the correlation weights to obtain load correlation features; The load prediction layer maps and calculates the load-related features to obtain the predicted load peak and load change rate within a preset prediction period, and generates the predicted load parameters based on the predicted load peak and load change rate.

4. The intelligent scheduling method for message push resources according to claim 1, characterized in that, The processing state calculation model includes a state segmentation layer, a temporal coding layer, and a state fusion layer; The step of inputting the running status data into a pre-trained processing status calculation model and outputting the available processing parameters for the target receiver includes: The running state data is segmented according to a sliding time window by the state segmentation layer to obtain multiple running state data segments; Extract operation activity features, interface focus change features, and message processing duration features from each of the aforementioned running status data segments; The operation activity feature, the interface focus change feature, and the message processing duration feature are sequentially encoded by the temporal coding layer in chronological order to obtain multiple processing state change features. The state fusion layer calculates the state weights corresponding to each processing state change feature, and then aggregates each processing state change feature according to the state weights to obtain the available processing parameters.

5. The intelligent scheduling method for message push resources according to claim 1, characterized in that, The scheduling decision model includes a state encoding layer and a policy output layer; The step of inputting the message value parameters, the predicted load parameters of each of the push channels, and the available processing parameters into a pre-trained scheduling decision model, and outputting the target push strategy corresponding to the message to be pushed, includes: The message value parameters, the predicted load parameters of each push channel, and the available processing parameters are encoded and feature-fused by the state coding layer to obtain joint scheduling state features; The joint scheduling state features are mapped through the policy output layer to obtain the action probabilities corresponding to multiple candidate push actions. Each candidate push action includes a candidate push channel, a candidate push timing, and a candidate resource occupation parameter. The candidate push action with the highest probability is selected, and the candidate push channel, candidate push timing, and candidate resource usage parameters corresponding to the selected candidate push action are used as the target push strategy.

6. The intelligent scheduling method for message push resources according to claim 5, characterized in that, The training process of the scheduling decision model includes: Obtain historical scheduling status, historical push actions corresponding to the historical scheduling status, and push results corresponding to the historical push actions; Calculate scheduling reward parameters based on the message push latency and message delivery results in the push results; The historical scheduling state is input into the value calculation layer, the state value corresponding to the historical scheduling state is output, and the strategy advantage parameter is calculated based on the scheduling reward parameter and the state value. Calculate the probability of the historical push action under the current policy network and the reference policy network respectively, and calculate the probability ratio of the two action probabilities; The probability ratio is restricted to a preset range, and the strategy loss parameter is calculated based on the restricted probability ratio and the strategy advantage parameter. The network parameters of the policy output layer are updated according to the policy loss parameters to obtain the pre-trained scheduling decision model.

7. The intelligent scheduling method for message push resources according to claim 1, characterized in that, After pushing the message to be pushed through the target push channel, the method further includes: Obtain the actual processing result and actual delivery delay corresponding to the message to be pushed; The value prediction deviation is calculated based on the actual processing result and the message value parameter, and the delay deviation is calculated based on the actual delivery delay and the target push timing. Message value training samples are generated based on the value prediction deviation, and the model parameters of the message value prediction model are updated using the message value training samples. Based on the delay deviation and the actual processing result, a scheduling decision training sample is generated, and the model parameters of the scheduling decision model are updated using the scheduling decision training sample.

8. A message push resource intelligent scheduling device, characterized in that, include: The message feature acquisition module is used to acquire message feature data of the message to be pushed. The message value prediction module is used to input the message feature data into a pre-trained message value prediction model and output the message value parameters of the message to be pushed. The channel load prediction module is used to acquire channel status data of multiple push channels, input the channel status data into a pre-trained channel load prediction model, and output the predicted load parameters of each push channel. The processing status calculation module is used to obtain the running status data of the target receiving end corresponding to the message to be pushed, input the running status data into the pre-trained processing status calculation model, and output the available processing parameters of the target receiving end. The push strategy generation module is used to input the message value parameters, the predicted load parameters of each push channel and the available processing parameters into a pre-trained scheduling decision model, and output the target push strategy corresponding to the message to be pushed. The target push strategy includes the target push channel, the target push timing and the target resource consumption parameters. The push allocation and execution module is used to allocate push resources corresponding to the target push channel according to the target resource occupancy parameters, and push the message to be pushed through the target push channel at the target push time.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the message push resource intelligent scheduling method according to any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that, The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by an electronic device, cause the electronic device to perform the message push resource intelligent scheduling method according to any one of claims 1-7.

Images