Intelligent task processing method and equipment based on LSTM (Long Short Term Memory) and medium

Abstract

The invention discloses an intelligent task processing method and device based on LSTM and a medium, and the method comprises the steps: collecting historical task data and real-time state data of an approver, and carrying out the preprocessing of the historical task data and the real-time state data; extracting a task feature vector of a to-be-allocated task and a state feature vector of an approver from the preprocessed data, and splicing the task feature vector and the state feature vector into an input vector; inputting the input vector into a double-layer LSTM prediction model, and outputting a prediction processing duration and a prediction passing probability of each approver on the to-be-allocated task; the double-layer LSTM prediction model is trained by adopting an improved loss function, and the improved loss function comprises a time sequence gradient penalty term and a key time step weight coefficient; and generating a task allocation scheme according to the predicted processing duration, the predicted passing probability and a preset task priority rule, and allocating the to-be-allocated task to the target approver according to the task allocation scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of task processing, and in particular to a task intelligent processing method based on LSTM, a device and a medium. BACKGROUND

[0002] With the in-depth development of enterprise digital transformation, centralized business processing platforms such as financial shared centers have become a standard configuration for large organizations, and the efficiency of their task allocation directly affects the operating costs and response speed of enterprises. Traditional task allocation methods mainly rely on static rules and simple load balancing strategies, which cannot adapt to complex and variable business scenarios and real-time changes in approver status. Static allocation mechanisms lack the ability to learn from historical behavior patterns of approvers, making it difficult to dynamically adjust allocation strategies based on task characteristics, resulting in problems such as high-priority task processing delays and uneven resource utilization.

[0003] Although existing machine learning-based task allocation methods have improved allocation efficiency to some extent, they have obvious shortcomings in handling time-dependent features. Traditional algorithms such as random forests and support vector machines cannot effectively capture the impact of task backlog on subsequent allocation and the dynamic rules of approver processing efficiency over time. In particular, during business peak periods, these allocation decisions often deviate from the optimal solution, leading to increased task backlog. SUMMARY

[0004] The embodiments of the present application provide a task intelligent processing method based on LSTM, a device and a medium to solve the above technical problems.

[0005] In one aspect, the embodiments of the present application provide a task intelligent processing method based on LSTM, comprising: Collecting historical task data and real-time state data of approvers and preprocessing the historical task data and the real-time state data; Extracting a task feature vector of a to-be-allocated task and a state feature vector of an approver from the preprocessed data, and concatenating the task feature vector and the state feature vector into an input vector; Inputting the input vector into a double-layer LSTM prediction model to output a predicted processing duration and a predicted pass probability of each approver for the to-be-allocated task; wherein the double-layer LSTM prediction model is trained using an improved loss function, and the improved loss function includes a time gradient penalty term and a key time step weight coefficient; Generating a task allocation scheme according to the predicted processing duration, the predicted pass probability and a preset task priority rule, and allocating the to-be-allocated task to a target approver according to the task allocation scheme.

[0006] In an implementation form of the present application, a task feature vector of the to-be-assigned task and a state feature vector of the approver are extracted from the preprocessed data, specifically comprising: The amount of the to-be-assigned task is normalized, the urgency of the task is quantified, and the type of the task is encoded; From the preprocessed historical task data and real-time state data, a time-out handling flag, an urgent document flag and a task type priority feature are derived; According to the preprocessed historical task data, the load rate of the approver is calculated as a load feature, and the cosine similarity between the professional field of the approver and the type of the task is calculated as a professional matching feature; wherein the load rate is the ratio of the number of current tasks to be processed to the maximum load of tasks; The historical behavior of the approver is aggregated and statistically analyzed to generate the average handling time, the approval pass rate and the task type processing proportion in the historical preset period.

[0007] In an implementation form of the present application, the improved loss function is represented by the following formula:

[0008]

[0009] wherein, represents the improved loss function, represents the Huber loss function, represents the time gradient penalty term, represents the key time step weight coefficient, represents the weight coefficient, represents the core loss term, represents the true value, represents the predicted value, represents the hyperparameter.

[0010] In an implementation form of the present application, according to the predicted processing time, the predicted pass probability and a preset task priority rule, a task allocation scheme is generated, specifically comprising: According to the business peak period state in the real-time state data, the weight coefficient in the improved loss function is dynamically adjusted; In the case where the urgency of the task exceeds a preset urgency threshold, the weight coefficient of the predicted processing time in the improved loss function is increased, and in the case where the amount of the task exceeds a preset amount threshold, the corresponding weight coefficient of the predicted pass probability in the improved loss function is increased.

[0011] In an implementation form of the present application, it further comprises: The model training data is updated in a rolling window manner at a preset time interval, and task data before the preset time interval is deleted when new completed task data is added to the window; A deviation between the predicted processing duration and the actual processing duration is calculated, and an online learning mechanism is triggered to adjust model parameters of the double-layer LSTM prediction model when the deviation exceeds a preset dynamic deviation threshold.

[0012] In an implementation manner of the application, historical task data and real-time state data of an approver are collected, specifically including: A historical task database in a financial sharing system is periodically crawled to extract original fields; wherein the original fields include invoice code, amount value, and task submission timestamp; A state update event of an approver terminal is listened to in real time to capture a task start processing timestamp and a task completion timestamp.

[0013] In an implementation manner of the application, the historical task data and the real-time state data are preprocessed, specifically including: Records with a key field missing rate exceeding a preset missing threshold in the historical task data are deleted, and non-key field missing values in the historical task data are filled by default values; Reimbursement tasks exceeding a preset amount threshold and tasks with a submission time earlier than a business occurrence date are identified, and tasks with an invoice type and a reimbursement category that do not match are identified to delete the identified abnormal data; An invoice code, an amount, and a date are combined to detect whether there is a repeated reimbursement task, and if so, the repeated reimbursement task is deleted; A business processing duration is calculated according to a business processing completion time and a business start processing time in the historical task data to extract business features and mark a business peak period feature.

[0014] In an implementation manner of the application, the application further includes: A task submission quantity in a current time window is monitored, and a deviation degree between the task submission quantity and a historical average task quantity in the same period is calculated; When the deviation degree is detected to exceed a preset positive fluctuation threshold, a switching threshold of an improved loss function is increased, and when the deviation degree is detected to drop to a preset negative fluctuation threshold, the switching threshold of the improved loss function is decreased.

[0015] On the other hand, the application also provides a task intelligent processing device based on LSTM, the device comprising: At least one processor; and, a memory communicatively coupled to the at least one processor; 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 so that the at least one processor can execute the above-mentioned LSTM-based task intelligent processing method.

[0016] On the other hand, an embodiment of the present application further provides a non-volatile computer storage medium storing computer-executable instructions. When the computer-executable instructions are executed, an LSTM-based task intelligent processing method as described above is implemented.

[0017] The present invention provides a method, device, and medium for intelligent task processing based on LSTM, which have at least the following beneficial effects: By processing the concatenated vector of task features and approver status through a two-layer LSTM prediction model, it explicitly captures temporal dependencies such as task queue backlogs and fluctuations in approval efficiency, solving the allocation bias problem caused by traditional machine learning models ignoring the correlation of the time dimension. Through the unique gating mechanism of LSTM, the system can learn the dynamic patterns of approver processing efficiency changes over time, as well as the impact of task backlogs on subsequent allocation decisions. It can more accurately predict the task processing performance of approvers in different working states, thereby making more reasonable allocation decisions. The loss function integrates the core advantages of Huber loss and innovatively adds a temporal gradient penalty term and a key time step weight coefficient, effectively addressing the limitations of traditional mean square error loss in time series prediction. It not only ensures the model's robustness to outliers, but also enhances the stability of long-term predictions, significantly improving the accuracy of predicted processing time and pass probability. Dynamically generate allocation plans based on prediction results and priority rules, replacing static rule configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A flowchart of an LSTM-based task intelligent processing method provided in an embodiment of the present application; Figure 2 A schematic diagram of the internal structure of an LSTM-based task intelligent processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0020] The following describes in detail the technical solutions provided by various embodiments of the present application in conjunction with the accompanying drawings.

[0021] Figure 1 A flowchart of an LSTM-based task intelligent processing method provided in an embodiment of the present application.

[0022] The analysis method involved in the embodiments of the present application can be implemented by a terminal device or a server, and the present application does not impose any special restrictions on this. For ease of understanding and description, the following embodiments are described in detail using a server as an example.

[0023] It should be noted that the server can be a single device or a system composed of multiple devices, that is, a distributed server, and this application does not make any specific restrictions on this.

[0024] like Figure 1 As shown, the embodiment of the present application provides an LSTM-based task intelligent processing method, including: Step 101: Collect historical task data and real-time status data of the approver, and pre-process the historical task data and real-time status data.

[0025] In this embodiment, the data acquisition module is responsible for acquiring the raw data required for the operation of the system. It is understandable that high-quality data collection is the basis for ensuring accurate prediction and reasonable allocation of the system. Exemplarily, the historical task data collection adopts a periodic crawling mechanism, and the system automatically accesses the database of the financial sharing system at preset time intervals. Specifically, this design not only ensures the timeliness of the data, but also avoids the pressure on the business system caused by frequent queries. It should be noted that the original fields collected have been carefully selected and contain key information such as invoice code, amount value and task submission timestamp. The combination of these fields can fully describe the basic characteristics of the task.

[0026] The collection of real-time status data adopts an event-driven mechanism. Understandably, unlike the periodic collection of historical data, real-time status monitoring requires immediate response to the changes in the operations of the approver. Illustratively, the system accurately captures the task start timestamp and task completion timestamp by listening to the status update events of the approver terminal. Specifically, this design can accurately reflect the current workload and processing progress of the approver, providing real-time basis for dynamic allocation. It should be noted that the collection accuracy of the timestamp reaches the millisecond level, ensuring sufficient accuracy of the subsequent calculation of processing time.

[0027] In this embodiment, the data preprocessing link plays an important role in data quality assurance. Understandably, various quality problems inevitably exist in the original data, which must be strictly cleaned before being used by the model. Illustratively, the system first performs integrity check on the historical task data, and adopts record deletion strategy for the case of high missing rate of key fields. Specifically, although this processing method will lose part of the data, it ensures the reliability of the remaining data. It should be noted that for missing values of non-key fields, the system will select appropriate filling strategies according to the field characteristics, such as using default values, mode or inferring through business rules.

[0028] Outlier detection is one of the core links of data preprocessing. Understandably, abnormal data will seriously interfere with model training and must be identified and processed. Illustratively, the system establishes a multi-level anomaly detection mechanism. For amount values, the system will identify records that are obviously out of the reasonable range; for time data, the system will check logical errors such as task submission time earlier than business occurrence date; in terms of classification data, the system will verify the matching of invoice type and reimbursement category. Specifically, these anomaly detection rules are based on in-depth business understanding and can effectively identify various data problems. It should be noted that all identified abnormal data will be marked and deleted, ensuring the purity of the training data.

[0029] The detection of duplicate data adopts the strategy of combining key fields. Understandably, in actual business, the same reimbursement may generate multiple records due to system problems. Illustratively, the system combines invoice code, amount and date to detect duplicates, which has sufficient discrimination. Specifically, when duplicate records are detected, the system retains the latest one and deletes the rest. It should be noted that this processing not only solves the problem of duplicate data, but also ensures the timeliness of the information.

[0030] It can be understood that the original data needs to be transformed into more meaningful features to be effectively utilized by the model. For example, the system calculates the exact business handling duration according to the handling completion time and the handling start time, which directly reflects the work efficiency of the approver. Specifically, the system also analyzes the distribution characteristics of the business handling time, automatically identifies and marks the business peak period. It should be noted that these derived features provide important timing information for subsequent model prediction.

[0031] It can be understood that the entire data collection and preprocessing process forms a complete data processing chain. For example, from raw data collection to final feature generation, the quality of the data can be ensured. Specifically, the strict data processing mechanism lays a solid foundation for subsequent LSTM model training and task allocation decisions. It should be noted that the system also establishes a data quality monitoring mechanism to regularly evaluate the processing effect of each link and continuously optimize the preprocessing process.

[0032] Step 102, extracting the task feature vector of the task to be allocated and the state feature vector of the approver from the preprocessed data, and concatenating the task feature vector and the state feature vector into an input vector.

[0033] In this embodiment, the feature extraction module is responsible for transforming the preprocessed raw data into machine-understandable feature vectors. For example, the normalization of the task amount uses the max-min scaling method. Specifically, the system first analyzes the distribution range of the amount in the historical data, and then linearly maps all amounts to a fixed interval. It should be noted that this processing eliminates the bias caused by different orders of magnitude, allowing the model to more evenly learn the features of each type of amount task. It can be understood that for newly emerging amount values that exceed the historical range, the system will dynamically adjust the normalization parameters to ensure consistency in processing.

[0034] The quantification of the task urgency reflects the digital conversion of the business rules. In this embodiment, the system converts the business priority into a calculable numerical feature. For example, regular tasks are marked as a baseline value, urgent tasks are appropriately adjusted according to the urgency level, and extremely urgent tasks are assigned the highest score. It should be noted that this quantification is not a simple linear mapping, but combines the priority rules established by the business department to ensure that the score setting meets the actual business needs.

[0035] The task type encoding adopts the One-Hot technique, which is a standard method for handling categorical variables. It can be understood that the financial shared center has various types of documents, including travel expenses, office purchases, customer entertainment, and other categories. For example, the system creates an independent binary feature for each type, and only one feature is active during processing. Specifically, this encoding method avoids artificially introducing ordinal relationships for different types, ensuring that the model can treat each category equally. It should be noted that the encoding dimension will automatically expand with the emergence of new business types, maintaining the adaptability of the system.

[0036] The system extracts advanced features such as overtime handling flags, urgent document flags, and task type priorities from the basic data. For example, the overtime handling flag is determined by comparing the actual processing time with the standard time. It can be understood that these derived features encapsulate important business logic and provide more direct judgment basis for the model. Specifically, the urgent document flag not only considers the emergency flag of the task itself, but also makes a comprehensive judgment based on the relationship between the submission time and the deadline.

[0037] The extraction of approver status features adopts a multi-dimensional analysis method. It should be noted that the load rate feature dynamically reflects the current work pressure of the approver. For example, the system maintains a record of the historical maximum carrying capacity of each approver and updates this benchmark value regularly. It can be understood that this design allows the load rate calculation to adapt to the natural changes in the approver's ability. Specifically, when the approver's processing efficiency improves, the maximum carrying capacity reference value will also be adjusted accordingly, ensuring the accuracy of the load assessment.

[0038] The calculation of professional matching degree adopts the cosine similarity method. The system constructs a professional field portrait for each approver, recording the historical distribution of processing various tasks. For example, for a travel expense reimbursement expert, the proportion of travel-related tasks in the portrait will be significantly higher than other types. It can be understood that this similarity calculation can quantify the matching degree between the approver and the current task, providing a basis for professional matching allocation. Specifically, the system will periodically recalculate the professional field portrait to reflect the natural evolution of the approver's skills.

[0039] Historical behavior statistics is another important component of the status features. It should be noted that the system calculates the average processing time, approval pass rate, and task type processing proportion of the approver within a predetermined period. For example, these statistics not only consider the overall average, but also analyze the performance differences under different task types. It can be understood that this fine-grained statistical analysis can more comprehensively reflect the expertise and efficiency characteristics of the approver. Specifically, the statistical period will be optimized based on business characteristics, ensuring the timeliness of the data and the stability of the statistical results.

[0040] In this embodiment, the concatenation of feature vectors is not a simple connection operation, but a structured combination. The system determines the concatenation order according to the business relevance between features. For example, the task amount feature is arranged adjacent to the historical similar task processing efficiency feature of the approver. It can be understood that this design facilitates the model to capture the interaction between related features and improves the learning efficiency. Specifically, the system also adds an identification prefix to different types of features to facilitate feature identification during debugging and analysis.

[0041] It can be understood that the entire feature extraction process forms a complete feature engineering system from the original field to the basic feature, then to the advanced derived feature, and finally to the structured vector. It should be noted that the system also establishes a feature quality monitoring mechanism to regularly evaluate the contribution of each feature to the prediction accuracy and continuously optimize the feature engineering strategy. Specifically, this closed-loop optimization mechanism ensures that the feature engineering can adapt to the needs of business development and model evolution.

[0042] Step 103, input the input vector into the double-layer LSTM prediction model to output the predicted processing time and predicted pass probability of each approver for the to-be-assigned task; wherein the double-layer LSTM prediction model is trained using an improved loss function, and the improved loss function includes a time gradient penalty term and a key time step weight coefficient.

[0043] It can be understood that the architecture design of the LSTM model directly affects the prediction performance. For example, the double-layer LSTM network uses a hierarchical processing mechanism. Specifically, the first layer of LSTM units captures short-term dependency patterns, such as the influence of current workloads. It should be noted that the second layer of LSTM units learns long-term patterns, such as the characteristics of business periodicity.

[0044] Huber loss as the core loss term uses squared error when the error is small and switches to linear error when the error is large. For example, this design ensures accuracy and enhances robustness. Specifically, the time gradient penalty term constrains the prediction change amplitude of adjacent time steps. It should be noted that this design avoids dramatic fluctuations in the prediction results. It can be understood that the key time step weight coefficient highlights the prediction accuracy of important business time points.

[0045] For example, the model training uses a rolling window mechanism. Specifically, the system regularly updates the model parameters with the latest data while retaining some historical data feature representations. It should be noted that this design enables the model to adapt to business changes.

[0046] In this embodiment, a loss function more suitable for time series prediction is designed based on the Huber Loss loss function, which is achieved by adding a time gradient penalty term and a key time step weighting in the loss function, and is represented by the following formula:

[0047]

[0048] wherein, denotes the improved loss function, denotes the Huber loss function, denotes the temporal gradient penalty term, denotes the key time step weight coefficient, denotes the weight coefficient, denotes the core loss term, denotes the true value, denotes the predicted value, denotes the hyperparameter.

[0049] The improved loss function can effectively reduce the interference of abnormal values. Huber can avoid the model being biased by extreme values. And significantly alleviate the gradient explosion. In deep LSTM or multi-step prediction, Huber is more stable than pure MSE.

[0050] In step 104, a task allocation scheme is generated according to the predicted processing time length, the predicted passing probability and a preset task priority rule, and the task to be allocated is allocated to the target approver according to the task allocation scheme.

[0051] In this embodiment, the system dynamically adjusts the weight coefficient in the loss function according to the real-time business state. Specifically, when the system detects task backlog during the business peak period, it will automatically increase the weight coefficient of processing speed. It should be noted that this adjustment is not a simple linear change, but a comprehensive decision combining factors such as backlog degree, availability of approvers, etc. It can be understood that when the system identifies a task with an emergency level exceeding a preset threshold, the approver with the shortest predicted processing time length will be given priority, ensuring that critical business is processed in a timely manner.

[0052] For high-amount tasks, the system adopts a differentiated processing strategy. For example, when the task amount exceeds a certain threshold, the system will increase the weight coefficient of the predicted passing probability accordingly. Specifically, this design reduces the possibility of improper allocation of high-risk tasks, ensuring the quality of approval. It should be noted that the amount threshold is not a fixed value, but is dynamically adjusted according to business type and historical data analysis. It can be understood that the system will record the approval passing rate of tasks in different amount intervals as the basis for threshold optimization.

[0053] The system can identify characteristic patterns of business peaks, such as the end of the month or quarter. Exemplarily, during these special periods, the system will adjust the weight parameters in advance to cope with the expected increase in workload. Specifically, this forward-looking adjustment avoids the response delay of the system when the business surges. It should be noted that the system analyzes the periodicity in historical data and continuously optimizes the periodicity detection algorithm.

[0054] In this embodiment, both the financial approval rules and the approver performance change over time, and the model must be continuously updated to maintain accuracy. Exemplarily, the system adds newly completed task data to the training window at preset intervals while removing expired data. Specifically, the window size is set considering the balance between business change speed and data volume demand. It should be noted that the system monitors the degree of change in data distribution and dynamically adjusts the window size and update frequency.

[0055] The system continuously compares the deviation between the predicted value and the actual result. Exemplarily, when the deviation continuously exceeds the threshold, the system triggers the incremental training process. It can be understood that this design enables the model to quickly adapt to changes in approver efficiency or the introduction of new business rules. Specifically, online learning uses a small learning rate and batch size to ensure the stability of parameter updates. It should be noted that the system records the effect of each online learning for optimizing the trigger threshold and learning parameters.

[0056] The deviation detection algorithm takes into account multiple factors. Exemplarily, the system not only focuses on the absolute deviation value, but also analyzes the duration and distribution characteristics of the deviation. Specifically, short-term and isolated deviations may be random fluctuations, while long-term and systematic deviations indicate that the model needs to be adjusted. It can be understood that this comprehensive judgment avoids overreaction and improves system stability. It should be noted that the deviation detection parameters are personalized for different approvers and task types.

[0057] In this embodiment, the task volume monitoring system realizes real-time perception of business situation. The system continuously tracks the number of tasks submitted within a unit time window. Exemplarily, by comparing with historical same-period data, the system can accurately identify abnormal fluctuations in business volume. Specifically, the deviation calculation not only considers the absolute difference, but also combines business periodicity and trend factors. It should be noted that this comprehensive evaluation improves the accuracy of fluctuation detection and reduces false positives.

[0058] The dynamic adjustment of loss function parameters is a key technology for coping with business fluctuations. Understandably, when a positive fluctuation (a sudden surge in business volume) is detected, the system will appropriately increase the switching threshold of the Huber loss function. Illustratively, this adjustment makes the model pay more attention to overall trends and reduces sensitivity to individual outliers under business pressure. Specifically, the adjustment amplitude is proportional to the fluctuation degree, ensuring the appropriateness of the response. It should be noted that the system will record the effects of parameter adjustment to form empirical knowledge for subsequent optimization.

[0059] The reverse fluctuation processing mechanism ensures system stability. In this embodiment, when the business volume falls to normal levels, the system will gradually restore the original parameter settings. Illustratively, this gradual adjustment avoids performance fluctuations and ensures consistency in user experience. Specifically, the recovery speed is optimized according to the business scenario, ensuring timeliness while avoiding excessive adjustment. Understandably, the system maintains multiple sets of parameter configurations for quick switching in different business scenarios.

[0060] A close collaborative relationship is formed among the modules. Understandably, dynamic weight adjustment, online learning, and business volume response mechanisms do not operate in isolation. Illustratively, business fluctuations trigger weight adjustment, and the effects of weight adjustment affect the triggering conditions of online learning. Specifically, this integrated design enables the system to collaboratively respond to complex business scenarios from multiple dimensions. It should be noted that the system analyzes the interaction between mechanisms and continuously optimizes the collaborative strategy.

[0061] The actual effects of each allocation decision are recorded and analyzed. Illustratively, these feedback data are used to evaluate the effectiveness of each mechanism and guide parameter optimization. Specifically, the system establishes a multi-dimensional evaluation index system to comprehensively measure allocation quality. Understandably, this data-driven optimization enables the system to continuously adapt to business development and maintain long-term effectiveness. It should be noted that the optimization process considers the balance between short-term and long-term goals to avoid local optimal solutions.

[0062] Understandably, the entire task allocation and optimization system forms an intelligent decision-making loop. Illustratively, from data collection to feature extraction, from model prediction to dynamic allocation, and from feedback optimization, each link is closely connected. Specifically, this end-to-end design ensures the adaptability and reliability of the system in complex and changing business environments. It should be noted that the system also has a perfect observability design, and all decision-making processes and results are traceable and analyzable, providing a solid foundation for continuous optimization.

[0063] The above is an embodiment of the method proposed in the present application. Based on the same inventive concept, the present embodiment also provides an LSTM-based task intelligent processing device, the structure of which is as shown in Figure 2 .

[0064] Figure 2 An internal structure schematic diagram of a task intelligent processing device based on LSTM is provided in the embodiments of the present application. As shown in the figure, the device comprises: Figure 2 at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable 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: collect historical task data and real-time state data of an approver, and pre-process the historical task data and the real-time state data; extract a task feature vector of a to-be-assigned task and a state feature vector of the approver from the pre-processed data, and splice the task feature vector and the state feature vector into an input vector; input the input vector into a double-layer LSTM prediction model to output a predicted processing time length and a predicted passing probability of each approver for the to-be-assigned task; wherein the double-layer LSTM prediction model is trained by using an improved loss function, and the improved loss function comprises a time gradient penalty term and a key time step weight coefficient; generate a task assignment scheme according to the predicted processing time length, the predicted passing probability and a preset task priority rule, and assign the to-be-assigned task to a target approver according to the task assignment scheme.

[0065] The embodiments of the present application also provide a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions can be executed to: collect historical task data and real-time state data of an approver, and pre-process the historical task data and the real-time state data; extract a task feature vector of a to-be-assigned task and a state feature vector of the approver from the pre-processed data, and splice the task feature vector and the state feature vector into an input vector; input the input vector into a double-layer LSTM prediction model to output a predicted processing time length and a predicted passing probability of each approver for the to-be-assigned task; wherein the double-layer LSTM prediction model is trained by using an improved loss function, and the improved loss function comprises a time gradient penalty term and a key time step weight coefficient; generate a task assignment scheme according to the predicted processing time length, the predicted passing probability and a preset task priority rule, and assign the to-be-assigned task to a target approver according to the task assignment scheme.

[0066] ​The various embodiments in the present application are described in a progressive manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, the device and medium embodiments are described simply because they are substantially similar to the method embodiments, and the relevant parts can be referred to the description of the method embodiments.

[0067] The device and medium provided by the embodiments of the present application are one-to-one corresponding to the method, and therefore, the device and medium also have the similar beneficial technical effects as the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the device and medium will not be described here.

[0068] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) containing computer-usable program code.

[0069] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the function specified in the flow or flows and / or block or blocks.

[0070] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the function specified in the flow or flows and / or block or blocks.

[0071] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0072] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0073] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, which can provide temporary storage for information. The memory can also include non-volatile memory, such as read-only memory (ROM) and / or flash memory, which can provide longer-term storage for information. The memory is an example of computer-readable media.

[0074] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology for storing information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition herein, computer-readable media does not include transitory media, such as modulated data signals and carriers.

[0075] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element.

[0076] ​​The above merely provides an example of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall fall into the scope of claims of the present application.

Claims

1. A method for intelligent processing of tasks based on LSTM, characterized in that, The method comprises: Collecting historical task data and real-time state data of the approvers, and preprocessing the historical task data and the real-time state data; Extracting a task feature vector of the to-be-assigned task and a state feature vector of the approvers from the preprocessed data, and splicing the task feature vector and the state feature vector into an input vector; Inputting the input vector into a double-layer LSTM prediction model to output a predicted processing time length and a predicted passing probability of each approver for the to-be-assigned task; wherein the double-layer LSTM prediction model is trained by using an improved loss function, and the improved loss function comprises a time gradient penalty term and a key time step weight coefficient; Generating a task assignment scheme according to the predicted processing time length, the predicted passing probability and a preset task priority rule, and assigning the to-be-assigned task to a target approver according to the task assignment scheme. 2.The LSTM-based task intelligent processing method of claim 1, wherein, The task feature vector of the to-be-assigned task and the state feature vector of the approvers are extracted from the preprocessed data, specifically comprising: Normalizing the amount of the to-be-assigned task, quantifying the task urgency and encoding the task type; Deriving an overtime handling flag, an urgent document flag and a task type priority feature from the preprocessed historical task data and real-time state data; Calculating the load rate of the approver as a load feature according to the preprocessed historical task data, and calculating the cosine similarity between the professional field of the approver and the task type as a professional matching feature; wherein the load rate is the ratio of the current to-be-processed task quantity to the maximum carrying task quantity; Aggregating and counting the historical behavior of the approver to generate the average processing time length, the approval passing rate and the task type processing proportion in a historical preset period. 3.The LSTM-based task intelligent processing method of claim 1, wherein, The improved loss function is represented by the following formula: wherein, denotes an improved loss function, denotes a Huber loss function, denotes a temporal gradient penalty term, denotes a key time step weight coefficient, denotes a weight coefficient, denotes a core loss term, denotes a true value, denotes a predicted value, denotes a hyperparameter. 4.The LSTM-based task intelligent processing method of claim 1, wherein, Generating a task assignment scheme according to the predicted processing time length, the predicted passing probability and a preset task priority rule, specifically comprising: Dynamically adjusting the weight coefficient in the improved loss function according to the business peak period state in the real-time state data; In the case that the task urgency exceeds a preset urgency threshold, increasing the weight coefficient of the predicted processing time length in the improved loss function, and in the case that the task amount exceeds a preset amount threshold, increasing the corresponding weight coefficient of the predicted passing probability in the improved loss function. 5.The LSTM-based task intelligent processing method of claim 1, wherein, The method further comprises: Updating the model training data in a rolling window manner at a preset time interval, and deleting the task data before the preset time interval when new completed task data is added to the window; Calculating the deviation between the predicted processing time length and the actual processing time length, and triggering an online learning mechanism to adjust the model parameters of the double-layer LSTM prediction model when the deviation exceeds a preset dynamic deviation threshold. 6.The LSTM-based task intelligent processing method of claim 1, wherein, Collecting historical task data and real-time state data of the approvers, specifically comprising: Periodically crawling a historical task database in a financial sharing system to extract original fields; wherein the original fields include invoice code, amount value and task submission timestamp; Real-time monitoring of the status update event of the approval terminal, capturing the task start timestamp and task completion timestamp. 7.The LSTM-based task intelligent processing method of claim 1, wherein, The historical task data and the real-time state data are preprocessed, specifically including: Records with a key field missing rate exceeding a preset missing threshold in the historical task data are deleted, and non-key field missing values in the historical task data are filled with default values. Identify reimbursement tasks exceeding a preset amount threshold and tasks with a submission time earlier than the business occurrence date, and identify tasks with an invoice type and reimbursement category that do not match, to delete the identified abnormal data. Combine the invoice code, amount and date to detect whether there is a duplicate reimbursement task, and if so, delete the duplicate reimbursement task. According to the business completion time and the start processing time of the business in the historical task data, the business processing duration is calculated to extract the business features and mark the business peak period features. 8.The LSTM-based task intelligent processing method of claim 1, wherein, The method further includes: Monitoring the number of task submissions in the current unit time window and calculating the deviation between the number of task submissions and the historical average task amount in the same period; In the case where the deviation is detected to exceed a preset positive fluctuation threshold, the switching threshold of the improved loss function is increased, and in the case where the deviation is detected to drop to a preset negative fluctuation threshold, the switching threshold of the improved loss function is reduced. 9.A LSTM-based task intelligent processing device, characterized in that, The device includes: At least one processor; and a memory connected in communication with the at least one processor; Wherein the memory stores instructions executable 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 perform the LSTM-based task intelligent processing method of any one of claims 1-8.

10. A non-transitory computer storage medium storing computer-executable instructions, the computer-executable instructions comprising instructions for: receiving a request to access a file; determining whether the file is stored in a cache; and in response to determining that the file is stored in the cache, providing access to the file from the cache. The computer executable instructions, when executed, implement the LSTM-based task intelligent processing method of any one of claims 1-8. The computer executable instructions, when executed, implement the LSTM-based task intelligent processing method of any one of claims 1-8.

Images