CN118428615A - Interactive system for scheduling and dispatching tax personnel - Google Patents

Abstract

The present disclosure provides an interactive system for tax manager to schedule tax staff to effectively perform daily tax activities. Due to the availability of historical data, the system consists of three modules: the system comprises a scheduling module, a scheduling module and a user interface. The scheduling module enables tax manager to schedule enough tax staff in advance by predicting future application counts based on an unstable historical time series using a time series prediction algorithm. The scheduling module can be matched with the application and tax staff in real time, so that the purposes of balancing workload, reducing waiting time and reducing total processing time are achieved. The user interface provides intuitive interactions and visualizations that allow a non-technical tax manager to control algorithms and check scheduling and scheduling results.

Description

Interactive system for scheduling and dispatching tax personnel

Technical Field

The present disclosure relates to an interactive system, and in particular, to an interactive system for staff scheduling and dispatching of a tax authority.

Background technique

Tax collection and payment are important public affairs. With the development of information systems in recent years, taxpayers can submit tax applications in a variety of ways, such as by visiting a tax service center or uploading documents at home via a smartphone or laptop. These applications are assigned to and processed by tax officials. There are two characteristics of Chinese taxation. First, due to the large tax base, there are a large number of applications every day. Second, various tax types require different relevant knowledge to handle.

These characteristics pose two major challenges to the tax authorities in providing efficient services. First, when many taxpayers request services through various channels at the same time, tax officers may only take on some applications, resulting in long queues or even postponing them to another day. Second, because tax types are complex and diverse, tax officers have different levels of expertise and proficiency. An application may be assigned to a staff member who needs more capabilities or knowledge to handle it, resulting in longer processing time and lower service quality. Recruiting more tax officers can better serve taxpayers, but too many tax officers may waste human resources and increase training and management costs.

Therefore, staff scheduling and dispatching are necessary. Tax authorities need to properly arrange employees' working hours, as well as match applications with employees, to better perform daily tax activities. However, traditional manual or rule-based scheduling and dispatching methods cannot cope with the increasingly complex tax schemes due to their slow response, inaccuracy, and low intelligence. Fortunately, nowadays each application can be tracked and stored in a database as historical data. The availability of historical data (such as daily application counts and processing times) makes it possible to develop a system for managers to use data to support efficient and flexible scheduling and dispatching of employees. However, it is unclear how to design and utilize such a system to achieve data-assisted scheduling and dispatching in tax authorities.

Time series forecasting has been applied in various fields, such as production planning, inventory analysis, and air pollution. Time series forecasting algorithms can be roughly classified into two categories: statistical methods and deep learning methods.

Commonly used statistical time series forecasting models include autoregressive (AR) models, moving average (MA) models, autoregressive moving average (ARMA) models, and autoregressive integrated moving average (ARIMA) models. These models are based on the assumption that the observations at the current moment are related to the observations at the previous moment, so past trends and patterns can be used to predict future trends and patterns. These algorithms are easy to use, and their forecast results are easy to interpret. However, they also have many limitations, such as they depend on the choice of parameters; they are more suitable for modeling linear series than nonlinear series; they require historical data to be fixed or fixed after difference (that is, the statistical characteristics of the series do not change over time).

Recently, researchers have attempted to design and apply deep learning methods to solve the problem of time series prediction. For example, Lai et al. proposed LSTNet, which combines convolutional neural networks (CNNs) and RNNs to extract short-term and long-term information from multivariate time series. In addition, researchers are also exploring how to apply transformers to time series prediction. However, time series in real life are usually unstable. Therefore, some recent works have proposed deep learning models for unstable time series data. AdaRNN is one of the most advanced models and provides a new perspective on time series modeling based on domain generalization in transfer learning. The time series covariance shift problem is proposed, and a two-step adaptive RNN model is constructed. The present disclosure adopts AdaRNN to predict future tax application counts because the historical application data of the tax authorities are unstable due to the impact of Covid-19 and the switch between busy tax seasons and ordinary seasons.

Resource scheduling algorithms are designed to coordinate resource demanders and service providers. In the case of taxation, we need to match applications with tax personnel based on real-time conditions.

Scheduling problems are widely modeled as multi-objective optimization problems. One of the challenges in solving multi-objective optimization problems is that the objectives are usually in conflict with each other, so it is difficult to optimize each objective to the optimal solution at the same time. Existing solutions are divided into two general categories: converting the multiple objectives into a single objective and generating a set of Pareto solutions. The first type of solution can only produce a single solution because they solve a single-objective optimization problem. Decision makers may prefer solutions in the second category to obtain a set of Pareto solutions and then make their own compromises. In one objective, the Pareto optimal solution at least dominates the other solutions. The target users of the proposed system (i.e., tax managers) want to get multiple solutions and make their own decisions. Therefore, the present disclosure selects a genetic algorithm as a starting point, which can effectively return a set of Pareto solutions. There are many variations of genetic methods, such as MOGA, NPGA, WBGA, NSGA, and NSGA-II. These genetic algorithms differ from each other in terms of fitness, allocation process, elitism, or diversification methods. The present disclosure applies the idea of genetic algorithms to solve resource scheduling problems in tax scenarios.

This paper characterizes the domain problem of employee scheduling and dispatching in tax scenarios as time series prediction and resource scheduling problems, and proposes an employee scheduling and dispatching system to meet the needs of tax authorities.

Summary of the invention

The present disclosure is an interactive system for tax managers to schedule and dispatch tax personnel. Given a large number of applications and complex processing procedures, how to effectively schedule and dispatch employees to provide good services to taxpayers is now receiving more attention from tax authorities. The availability of historical application data enables tax managers to schedule and dispatch employees with data support, but it is not clear how to properly utilize historical data. Therefore, the present disclosure provides an interactive system based on the needs of tax managers. The system can schedule employees based on predicted application counts using a time series prediction algorithm, schedule employees in real time using a genetic algorithm, and display the results using intuitive and interactive visualization. More specifically, the system first uses the most advanced deep learning model AdaRNN to predict future tax application counts and inform managers of the number of applications they will receive for each tax type in a given period. Then, based on the results and the historical processing time of each employee processing each tax type of application, the system provides the number of employees required for each day in the next week. Secondly, the system embeds a customized real-time resource scheduling algorithm based on the genetic algorithm NSGA-II. The algorithm continuously scans pending applications and optimally matches applications and employees with the goal of the most balanced workload, the least waiting time, and the least total processing time. Finally, the system provides a user interface to further facilitate managers in scheduling and dispatching tax employees by uploading historical data, training new models, and viewing the results with line graphs and Ghent charts.

On the one hand, the present disclosure provides an interactive system for scheduling and dispatching tax personnel, including: a scheduling module, which is configured to obtain historical data and use a time series prediction algorithm based on the historical data to predict future application counts to output scheduling results; a scheduling module, which is configured to receive input information and use a resource scheduling algorithm based on NSGA-II to match employees in the scheduling results with applications to be processed in real time to generate scheduling results; and a user interface, which displays the scheduling results and the scheduling results.

In an embodiment, the time series prediction algorithm includes one of a SARIMA algorithm and an AdaRNN deep learning algorithm.

In an embodiment, the time series prediction algorithm comprises an AdaRNN deep learning algorithm.

In an embodiment, the historical data includes: tax type, average processing time for each employee to process applications of the tax type, and historical daily application counts.

In an embodiment, the historical daily application count is a non-stationary time series that arranges applications in chronological order.

In an embodiment, the input information includes real-time employee availability, currently pending applications in chronological order, and each employee's proficiency in the tax type.

In an embodiment, the scheduling result includes a plurality of scheduling results that respectively achieve at least one of the following goals: minimizing the difference between employees' working hours; minimizing the total processing time; and minimizing the total waiting time of taxpayers.

In an embodiment, in the NSGA-II-based resource scheduling algorithm, the following processes are performed: encoding, population initialization, crossover, mutation, fitness calculation, non-dominated sorting, crowding distance calculation, selection, generating a new population, checking termination conditions, and returning multiple solutions.

In an embodiment, the user interface comprises:

A data panel that allows users to upload the historical data;

Algorithm selection and setting panel, which allows users to select one of the SARIMA algorithm and AdaRNN deep learning algorithm and set the corresponding parameters;

A visualization panel that displays historical daily application counts and predicted future application counts in the form of a line graph;

The shift setting panel allows users to set parameters related to shift scheduling;

The shift scheduling result panel displays the shift scheduling result in the form of a line graph.

In an embodiment, the scheduling-related parameters include default working minutes for each employee and expected forecast days.

On the other hand, the present disclosure provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor performs the following operations: obtains historical data, and uses a time series prediction algorithm based on the historical data to predict future application counts to output a scheduling result; receives input information, and uses a resource scheduling algorithm based on NSGA-II to match employees in the scheduling result with applications to be processed in real time to generate a scheduling result; and displays the scheduling result and the scheduling result to a user.

BRIEF DESCRIPTION OF THE DRAWINGS

Illustrative, non-limiting example embodiments will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings.

FIG. 1 is an architecture diagram of an interactive system according to an embodiment of the present disclosure.

FIG2 is a schematic diagram of predicting future application counts based on unstable historical time series using the AdaRNN algorithm.

FIG. 3 shows a flow chart of a customized non-dominated sorting genetic algorithm II for solving the resource scheduling problem in a tax context and matching taxpayers' applications with appropriate employees.

FIG4 shows the user interface of the scheduling module.

Detailed ways

Prior art has designed systems or algorithms for scheduling staff in clinics, call centers and general workplaces. However, none of them will have the following unique characteristics to consider for the purpose of scheduling and scheduling staff in tax offices.

First, in order to schedule tax staff in advance, tax managers need to know the future forecast application counts, so a time series forecasting algorithm is needed. Due to the impact of Covid-19 and the transition between the busy tax season and the normal season, the time series of historical applications is unstable. However, none of the existing technologies take this feature into account when scheduling staff.

Secondly, in order to schedule tax personnel, tax managers want to properly match tax employees and applications to achieve the goals of the most balanced workload, the least waiting time, and the least processing time. They also want multiple scheduling schemes to choose from to handle different situations. For example, if the day is very busy, the tax manager may sacrifice the goal of balancing the workload and assign more applications to more skilled employees in pursuit of efficiency and less taxpayer waiting time. If there are fewer taxpayers on a given day, the tax manager may be more concerned about balancing the working hours of various employees to avoid improper behavior that leads to overwork. However, the prior art mainly provides a single scheme.

Third, using algorithms to obtain scheduling and dispatching results requires a certain degree of user participation, including adjusting parameters, comparing results, and sorting multiple dispatching schemes, but tax managers do not have a technical background and cannot directly use scheduling and dispatching algorithms to operate. An intuitive interface is needed to bridge tax managers and algorithms. However, the existing technology does not provide a dedicated user interface.

The present disclosure improves the prior art and satisfies the above characteristics by utilizing the following features:

1. The present disclosure designs and develops a specialized system for tax managers to schedule and dispatch staff. The system can predict the application counts of different tax types in advance to schedule staff, can match applications with appropriate tax staff in real time to schedule staff, and provides an intuitive user interface and interaction to allow managers to obtain results. Compared with existing technologies that target call centers or other staff, the present disclosure adapts to tax situations by considering specific data types, data characteristics, and workflows. The present disclosure is able to process unstable historical data and provide prediction results for scheduling use.

2. The present disclosure adapts genetic algorithms to meet specific characteristics in tax situations, thereby matching applications with appropriate tax personnel under multiple objectives and providing tax managers with several options.

3. The present disclosure designs a user interface with data visualization capabilities, allowing non-technical tax managers to easily interact with advanced algorithms, thereby providing non-technical technicians with an integrated solution for scheduling and dispatching tax employees.

Figure 1 shows the system architecture. First, the historical data is processed to obtain useful information, such as the time series of historical daily application counts, tax types, tax employees, and their average processing time for different types of applications. Then, a scheduling module is developed to predict application counts and provide scheduling results, and a scheduling module is developed to match available employees with applications in real time during working hours. Finally, the scheduling module and scheduling module are combined with a visual user interface that displays the results with intuitive visualization.

Scheduling Module

The scheduling module is intended to enable managers to pre-arrange enough employees whose capabilities and experience can handle future applications. Based on the requirements of tax managers, the scheduling module needs to consider 1) the tax type of application, 2) the processing time for each employee to handle each application type, and 3) the daily application count. The uncertainty of these three factors is different. The uncertainty here refers to whether the value of the factor changes over time and how much it fluctuates. Specifically, the uncertainty of the tax type is the smallest because the tax agency provides basically the same tax type or only updates them after a long period of time. The processing time is slightly uncertain because it is affected by various factors, such as whether the taxpayer or employee is more familiar with the practice process, or whether they are in a hurry. However, it is assumed that the proficiency will remain stable over a short period of time, so in an embodiment, the average value of the historical processing time is used in the scheduling module. Among these three factors, the most difficult to control is the daily application count. Compared with the other two factors, since the goal of the scheduling module is to schedule future tax employees, it is not enough to know the historical situation of the application. Therefore, it is necessary to know the application count for the next few days. Based on the above analysis, the design of the scheduling module is outlined. Figure 1 is an architecture diagram of an interactive system according to an embodiment of the present disclosure. In an embodiment, using historical data (A) and mined information (B), the interactive system supports tax managers to schedule and dispatch tax employees using a scheduling module (C), a scheduling module (D) and a user interface (E). As shown in (C) in Figure 1, the input of the module is the tax type, the average processing time and the historical daily application count. The first step is to predict the future application count. Specifically, the historical application counts are arranged in chronological order to obtain a historical time series, and then an appropriate time series prediction algorithm is used to predict future applications. Then, based on the tax type, processing time and predicted application count, the module outputs the scheduling result (i.e., the number of applications to be received for each tax type in the future days, and the corresponding number of employees with the required relevant skills). This most difficult and important step is time series prediction, which will be described in detail below.

Unstable time series forecasting

A time series is considered to be unstable when its statistical properties (e.g., mean and variance) change dynamically over time. The changes in the distribution of unstable time series data can be regular or irregular. The time series of historical application counts in tax cases is unstable due to the regular transition between busy and normal tax seasons and the irregular impact of Covid-19. Using statistical methods (e.g., AR, MA, and ARIMA) on unstable time series often produces poor forecasting performance. The reason is that these models assume that the time series has a normal distribution, but unstable time series violate this assumption.

To address this challenge, this application uses a state-of-the-art deep learning algorithm called AdaRNN to predict future application counts under the impact of tax season and Covid-19. Figure 2 shows how AdaRNN is adapted. The algorithm first divides the original data into time periods using time distribution features (TDC), and then uses time distribution matching (TDM) with a GRU network to predict future values with a given loss function. Specifically, AdaRNN defines time covariate shift as a phenomenon in which the data distribution in an unstable sequence changes dynamically over time, and adopts a transfer learning method. Time covariate shift is an extension of covariate shift, which is an important problem in machine learning. As shown in Figure 3, time covariate shift occurs when a continuous time series is divided into K segments, and the probability distribution of each segment is different. Time distribution characterization and time distribution matching are two steps to solve time covariate shift.

Time distribution characterization refers to dividing the entire time series data into K discrete segments so that the difference in probability distribution between the segments can be calculated. Assuming that if the training model can reduce large distribution differences, the training model will have higher generalization ability, the purpose of time distribution characterization is to divide the time series into the most dissimilar K segments. Therefore, it solves the optimization through a greedy strategy:

Where dist indicates the distance between the distributions of the two segments, which can be calculated in several ways, such as cosine distance and MMD (see A. Gretton, D. Sejdinovic, H. Strathmann, S. Balakrishnan, M. Pontil, K. Fukumizu, and B. K. Sriperumbudur. Optimal kernel choice for largescale two-sample tests. Advances in neural information processing systems, 25, 2012). Time distribution matching is performed after time distribution characterization. Specifically, the GRU network is used to learn the commonalities between time series segments with different distributions. It also dynamically learns the importance α of different time points based on the Boosting method. This can more accurately reduce the time correlation and obtain the optimal parameters θ of the prediction model. The process can be expressed as follows:

in is the loss function of prediction, and is the distance between the distributions of any two fragments.

Scheduling Module

The scheduling module predicts taxpayers' needs in advance and then arranges the appropriate number of tax personnel. However, it can only achieve a rough supply-demand matching and cannot solve the real-time dynamic matching. Therefore, a scheduling module is needed to dynamically match applications with appropriate employees in real time to alleviate these problems.

This problem is characterized as a resource scheduling problem, and the present disclosure aims to coordinate resource providers and demanders under some objectives. In the tax case, the providers are tax employees, and the resources are their time and expertise; the demanders are taxpayers and their applications. The goals of tax managers are 1) to minimize the difference between employees' working time to avoid overwork, 2) to minimize the total processing time by assigning employees to applications they are good at, and 3) to minimize the total waiting time of taxpayers to increase customer satisfaction.

Due to the presence of multiple objectives, the present application models the scheduling problem as a multi-objective optimization problem. Since managers want to have multiple choices rather than a single optimization result, the present disclosure selects a genetic algorithm that can simultaneously return a set of solutions. As shown in (D) of Figure 1, the input of the algorithm includes real-time information, such as employee availability and a sequence of applications sorted by start time with tax type labels. The input also includes employee proficiency in different tax types mined from historical data. The genetic algorithm integrates these three objectives and returns multiple solutions, providing a match between applications and employees. Next, the genetic algorithm and the algorithm design of the present disclosure will be briefly introduced.

Introduction to Genetic Algorithms

Genetic algorithms are inspired by natural selection and evolution theory. In their terminology, the solution to the optimization problem is called an individual or chromosome. A chromosome consists of discrete units called genes, and a collection of chromosomes constitutes a population.

When solving a single objective optimization problem, the genetic algorithm starts from an initial population and evolves the population with three operators (including crossover, mutation and selection). In the crossover operator, the genetic algorithm relies on a fitness function (i.e., reflecting the possibility of survival) to select two parent chromosomes that show good fitness, and combines them to form offspring chromosomes. The mutation operator introduces random changes to the chromosome by changing some genes with a low probability. The chromosomes produced by crossover and mutation are added to the population. Then, in the selection operator, the genetic algorithm will select chromosomes to form a new generation of population based on fitness values. When solving a multi-objective optimization problem, the genetic algorithm will involve multiple fitness functions, each of which is used for a target. The genetic algorithm will return a group of non-dominated solutions. The solution in this group of non-dominated solutions is inferior to other non-dominated solutions in terms of the target subset, but is superior to the dominated solution (i.e., the remaining solutions in the search space) in terms of all targets.

An intuitive but time-consuming way to obtain a non-dominated solution is to compare each pair of chromosomes in all target populations. To speed up, Deb et al. proposed a variant of a genetic algorithm called non-dominated sorting genetic algorithm II (NSGA-II), which uses non-dominated sorting to determine chromosomes with different levels of non-dominated fronts and calculates crowding distances to maintain diversity. The present disclosure proposes a resource scheduling algorithm based on NSGA-II in a tax situation because it has been proven to be very effective in many fields.

Resource Scheduling Algorithm of NSGA-II

The algorithm flow chart (Figure 3) shows the design of the NSGA-II algorithm to solve the resource scheduling problem and match the taxpayer's application with the appropriate employee under the three objectives. Based on the above introduction to the general genetic algorithm and NSGA-II, the algorithm design of this application will be described in detail from the following aspects.

Gene and chromosome encoding (Figure 3 (A)). The first step is to establish a mapping mechanism between the solution space and the chromosome so that the chromosome corresponds to a unique solution. In the tax case, given a sequence of applications according to the submission time, the expected solution is the sequence of employees who will handle the applications accordingly. Establish a mapping between the solution and the chromosome shown in Figure 3 (A). Assume that there are M taxpayers and N employees. Design a one-dimensional chromosome of length M, where each gene represents an employee and its index represents the serial number of the application. For example, the value of the third gene is 2, which means that the third application is handled by the second tax employee.

Population initialization (Fig. 3(B)): Given a population size, an initial population will be randomly generated.

Crossover operator ((C) of FIG3 ) A single-point crossover is used, in which a point is randomly picked, the two parent chromosomes are cut at this point, and the genes to the right of this point are exchanged between the two parent chromosomes.

Mutation operator (Fig. 3(D)): Given a mutation rate (i.e., the probability of changing a gene attribute), the gene attribute is changed within a qualified range.

Fitness function and calculation (FIG. 3 (E)). The tax manager gives three goals: minimize the difference between employees' working hours, minimize the total processing time of employees, and minimize the total waiting time of taxpayers. For each goal, the present disclosure has a fitness function. The difference between working hours is transferred by calculating the standard deviation of each employee's working time. When the following information is known, the three fitness functions can be intuitively obtained: the sequence of employees who will process M applications (e.g., chromosomes), the sequence of submission times of the M applications, the sequence of corresponding tax types of the M applications, and the sequence of processing times of the M applications.

Non-dominated sorting (FIG. 3 (F)). Multi-objective optimization problems have a set of non-dominated solutions due to multiple objectives. The NSGA-II algorithm can sort chromosomes at different levels of non-dominated fronts. Taking two fitness values (i.e., total processing time and total waiting time) as an example, FIG. 3 (F) shows two fitness values of chromosomes and three levels of non-dominated fronts, based on which chromosomes can be sorted. The smaller the two fitness values, the more suitable the chromosomes are, and the better their corresponding solutions are. The non-dominated sorting in NSGA-II is directly used to find these non-dominated fronts and sort the chromosomes.

Crowding distance (Figure 3 (G)). Non-dominated sorting can sort chromosomes based on the rank of the non-dominated front. In order to further sort the chromosomes in the same non-dominated front, it is necessary to calculate the crowding distance of each chromosome, which reflects how the chromosome is different from other chromosomes of the same rank. The crowding distance calculation method in NSGA-II can also be directly used.

Selection operator and new colony generation (Fig. 3 (H) and Fig. 3 (I)). The purpose of selection is to select more suitable and more diverse chromosomes, add them to the next generation of the colony, and use them as the starting point for the next round of evolution. Chromosomes are selected based on non-dominated sorting and crowding distance. Specifically, chromosomes with higher non-dominated rankings will be preferentially added to the new generation until the colony reaches a predefined colony size. At the last level of the non-dominated front, the selection operator may need to select chromosomes from all chromosomes at this level. Therefore, chromosomes with larger crowding distances will be selected to maintain diversity and avoid falling into local solutions.

Final solution (Fig. 3(J) and Fig. 3(K)). After obtaining the population of the new generation, the algorithm will check the termination condition (e.g., the number of iterations and whether the objective is optimized to a certain threshold). If not, the algorithm will start another iteration. Otherwise, the current chromosome will be arranged and returned in the same way as the selection operator or based on the tax manager's preference (e.g., arranging chromosomes in a specific order of the three objectives).

user interface

Considering that tax managers do not have a technical background, the process is automated to use these algorithms as much as possible. However, these algorithms still require a certain degree of user involvement, including adjusting parameters, comparing results, and sorting multiple scheduling solutions. In addition, since the prediction algorithms are imperfect and largely dependent on the data set, humans need to be involved and asked to select the most appropriate algorithm. In order to further reduce the barriers to utilizing the algorithms, a web application has been developed that includes a backend and a frontend. The backend mainly integrates the algorithms, and the frontend is a user interface that uses data visualization to show the prediction and scheduling results (Figure 4).

The interface for the scheduling module (Figure 4) consists of six linked panels, which respectively allow the user to upload historical data sets (A), select prediction algorithms (B), observe prediction results using line graphs (C), set parameters (D), and obtain scheduling results (E and F). Specifically, the six panels include the following panels (A) to (F).

In the data panel (A), the tax manager can upload a data file containing the dates and corresponding counts of applications for different tax types. After being uploaded, the file will be sent to the backend for later use.

The algorithm selection and settings panel (B) allows tax managers to select which algorithms to use for training and prediction, and set the necessary parameters. Currently, AdaRNN and SARIMA are provided to users. After setting the parameters, tax managers can click the "Start Training" button, and the backend will train the specified machine learning model based on historical data and parameters.

In the visualization panel (C), the model training and prediction results are plotted with line graphs to facilitate non-intuitive understanding and analysis by tax managers. By observing the difference between the model-predicted data and the original historical data, tax managers can qualitatively compare the advantages and disadvantages of each algorithm.

In the Schedule Settings panel (D), the tax manager needs to set some parameters specific to the tax situation, including the number of days they want to predict and the default working minutes for each tax employee. Also, with the prediction results shown in the visualization panel, the tax manager can compare them and select the best trained model for each tax type. After clicking the "Get Schedule Results" button, the backend will use the selected, trained model to predict data for each tax type.

In the Scheduling Results Panel (E)-(F), two panels show the forecast and scheduling results. Based on the tax manager's request (R1), the tax type, application count, and the number of employees required in a certain future period are displayed.

The disclosed system has been tested with real and desensitized historical data obtained from local tax authorities. It is also used by tax managers of the tax authorities. Next, the data, experimental settings, and results are introduced.

The raw data consists of several tables covering relatively static information (such as tax type and tax employee) and dynamic information (such as historical applications). The data fields used in this disclosure are explained as follows.

Tax types. Tax business in the local tax bureau involves nine tax types and many sub-categories, each of which requires different knowledge to handle. The nine tax types are called first-level types, and the sub-types are called second-level types. This project focuses on the first-level types and represents them as Type 1, Type 2, Type 3, ..., Type 9.

Tax employees. The data given by the local tax authorities is related to five tax employees. Each tax employee has a unique identification number and they are represented by their ID.

Application event. As mentioned above, taxpayers can submit tax application forms in various ways, and employees will process these applications. The process from application submission to application processing is regarded as an application event. This application describes an event with a five-tuple {application ID, tax employee ID, tax type, start timestamp, end timestamp}. The start timestamp and end timestamp contain date and time information. Each event is recorded in the database of the tax authority to form a time series.

Data analysis is performed to gain a better understanding of the tax situation. The raw application event data is first preprocessed by removing events containing noise, outliers, or missing values. A total of 9420 historical application events over 386 days are obtained from the sample dataset. In practice, the time series data of daily application counts is unstable, the application distribution is uneven, the workload of employees is uneven, and employees have different proficiency levels.

Prediction Algorithm Evaluation

When training the AdaRNN model on historical data, a single-step forecast is first performed, where data is taken from every 24 days to predict the next day's data. Since it is expected to predict data for multiple days in the future, the model needs to be a multi-output model. In order to predict multiple time steps, multi-step recursive forecasting is used. With AdaRNN single-step forecasting, the value predicted in each step is used as the input of the next step for iterative forecasting. The advantage of this is that it is faster and only one model needs to be trained to output multiple results.

In order to evaluate the trained AdaRNN model, three benchmarks are used for numerical comparison. The first benchmark is seasonal autoregressive integrated moving average (SARIMA), which is an extension of the classic time series prediction algorithm ARIMA with seasonal time series data. Since SARIMA has seven parameters to be determined, an automatic parameter selection method is used when SARIMA is used for historical tax data. In this disclosure, the specific values of the parameters are not directly given. However, a range is given for each parameter, a group of cross multiplication of the parameters is obtained, a model is trained based on each group of parameters in the group, and the model with the best result is selected as the final trained SARIMA model. The second benchmark is the gradient boosted regression tree (GBRT), which is a newly developed machine learning method for time series prediction. Since XGboost is an effective implementation of GBRT with high accuracy and fast training speed, it is adapted to a window-based regression model to improve its efficiency in the implementation scheme. The third benchmark is LSTNet, which is a commonly used benchmark that uses CNN and RNN to capture long-term and short-term features from time series data. According to the data set, model parameters such as learning rate and batch size are adjusted.

In the embodiment, two commonly used metrics are selected to evaluate the performance of these time series prediction models on historical data. The first is the mean absolute error (MAE), which can reflect the actual situation of the prediction value error and is calculated as follows:

The second metric is the root mean square error (RMSE), which measures the deviation between the observed value and the true value and is calculated as follows:

Table 1 shows the metrics of the four models on the historical dataset for different major tax types. The results show that AdaRNN generally has better performance.

Table 1. Comparison between AdaRNN and three baseline algorithms on datasets containing different tax types.

Scheduling Algorithm Evaluation

One day in the data set is randomly selected as an example to evaluate the effectiveness of the designed scheduling algorithm. On that day, four tax employees processed 70 applications. The working hours of the four employees were 190, 102, 49 and 239 minutes. Their total working time was about 580 minutes, and the standard deviation of their working time was 85.5. The total waiting time of the taxpayer was not recorded. The design scheduling algorithm based on the application sequence was used and multiple solutions were obtained. Table 2 compares the ten scheduling solutions (Solution 1 to Solution 10) given by the algorithm with the actual data on the selected date. It can be seen that all ten scheduling solutions allow a more balanced workload, in which the total working minutes have some increase. It can also be observed that these three goals conflict with each other in many cases and cannot be optimal at the same time. In this regard, human intervention is needed to decide which scheduling solution to use. For example, if a day is very busy, the tax manager may sacrifice the goal of balancing the workload and assign more applications to more skilled employees in pursuit of efficiency and less taxpayer waiting time. If there are fewer taxpayers on a given day, tax managers may be more concerned with balancing the work schedules of individual employees to avoid improper behavior that leads to overwork.

Table 2. Comparison between ten potential scheduling solutions and actual records on a selected day

In summary, the present disclosure provides a dedicated system for tax managers to schedule and dispatch employees. The system can predict the application counts of different tax types in advance based on unstable historical data to schedule employees, can match applications with appropriate tax employees in real time to dispatch employees, and provide an intuitive user interface and interaction to allow managers to obtain results.

The present disclosure improves the genetic algorithm to meet the specific characteristics in the tax case to match the application with the appropriate tax staff under multiple objectives and provide several alternatives for the tax manager.

The present disclosure provides a user interface with data visualization to allow non-technical tax managers to easily interact with advanced algorithms.

According to another aspect, a non-transitory computer-readable storage medium is disclosed, on which a computer program is stored. When executed, the computer program causes a processor to perform the functions of the above-mentioned scheduling module and dispatching module.

The computer-readable storage medium may be, for example, a disk or a memory, etc. The computer program may be stored in the computer-readable storage medium in the form of instructions encoding the computer-readable storage medium.

The foregoing is an illustration of example embodiments and should not be construed as limiting thereof. Although some example embodiments have been described, it will be readily appreciated by those skilled in the art that many modifications may be made in the example embodiments without substantially departing from the novel teachings and advantages of the example embodiments. Therefore, all such modifications are intended to be included within the scope of the example embodiments defined in the claims. Therefore, it should be understood that the foregoing is an illustration of various example embodiments and should not be construed as being limited to the specific example embodiments disclosed, and modifications to the disclosed example embodiments and some example embodiments are intended to be included within the scope of the appended claims.

Claims (11)

1. An interactive system for scheduling and dispatching tax personnel, comprising: A scheduling module, which is configured to obtain historical data, and predict future application counts using a time series prediction algorithm based on the historical data to output a scheduling result; A scheduling module configured to receive input information and use a resource scheduling algorithm based on NSGA-II to match employees in the scheduling result with applications to be processed in real time to generate a scheduling result; and A user interface displays the scheduling result and the dispatching result. 2. The interactive system according to claim 1, wherein the time series prediction algorithm comprises one of a SARIMA algorithm and an AdaRNN deep learning algorithm. 3. The interactive system according to claim 1, wherein the time series prediction algorithm comprises an AdaRNN deep learning algorithm. 4. The interactive system according to claim 2, wherein: The historical data includes: tax type, average processing time for each employee to process applications of the tax type, and historical daily application counts. 5. The interactive system according to claim 2, wherein: The historical daily application count is a non-stationary time series that arranges applications in chronological order. 6. The interactive system according to claim 2, wherein: The input information includes real-time employee availability, chronologically ordered applications currently pending, and each employee's proficiency with the tax type. 7. The interactive system according to claim 2, wherein the scheduling result comprises a plurality of scheduling results that respectively achieve at least one of the following objectives: Minimize differences between employees’ working hours; Minimize overall processing time; and Minimize the total waiting time for taxpayers. 8. The interactive system according to claim 2, wherein, in the NSGA-II-based resource scheduling algorithm, the following processing is performed: encoding, population initialization, crossover, mutation, fitness calculation, non-dominated sorting, crowding distance calculation, selection, generating new populations, checking termination conditions, and returning multiple solutions. 9. The interactive system according to claim 2, wherein the user interface comprises: A data panel that allows users to upload the historical data; Algorithm selection and setting panel, which allows users to select one of the SARIMA algorithm and AdaRNN deep learning algorithm and set the corresponding parameters; A visualization panel that displays historical daily application counts and predicted future application counts in the form of a line graph; The shift setting panel allows users to set parameters related to shift scheduling; The shift scheduling result panel displays the shift scheduling result in the form of a line graph. 10. The interactive system according to claim 2, wherein the parameters related to scheduling include default working minutes and expected forecast days for each employee. 11. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the following operations: Acquire historical data, and use a time series prediction algorithm based on the historical data to predict future application counts to output a scheduling result; Receiving input information, and using a resource scheduling algorithm based on NSGA-II to match employees in the scheduling result with applications to be processed in real time to generate a scheduling result; and The scheduling result and the dispatching result are displayed to the user.