Complaint work order prediction method optimized based on traditional festivals

Abstract

The invention discloses a complaint work order prediction method for optimization based on traditional festivals, and the method comprises the steps: collecting complaint work order data, carrying out the preprocessing of the data, carrying out the feature extraction of the preprocessed complaint work order data, and carrying out the analysis of the feature data through an ARIMA model. Time influence variables are determined based on the time from the traditional festival, then different ARIMA models are established according to different time influence variables, and a prediction result is obtained; and constructing a deep neural network model, inputting a prediction result and feature data of the ARIMA model to obtain a prediction result, carrying out weighted fusion on the prediction result of the ARIMA model and the prediction result of the deep neural network model, comparing actual complaint work order data with the prediction result, and evaluating the prediction performance of the model. According to the invention, weighted fusion is carried out based on the prediction result of the ARIMA model and the prediction result of the deep neural network model, the final complaint work order prediction result is obtained, and the prediction accuracy of complaint work order prediction in traditional festivals is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a complaint work order prediction method optimized based on traditional festivals. Background Art

[0002] Complaint ticket processing is an important part of corporate customer service. Accurately predicting the number of complaint tickets is of great significance for enterprises to reasonably allocate resources and improve customer satisfaction. Traditional time series prediction methods include ARIMA models, but generally speaking, since the ARIMA model is essentially a linear model, the autoregression (AR) and moving average (MA) parts in the model are linear and cannot capture the nonlinear relationship in the data. The ARIMA model can only process stable time series data, such as periodic seasonal data, and complaint tickets are not only affected by seasons, but also by holidays. In particular, traditional festivals such as the Spring Festival and the Mid-Autumn Festival are not on a fixed cycle, and the inflow of customer service tickets during these traditional holidays is significantly reduced compared to ordinary days, resulting in low accuracy in the prediction of complaint tickets during traditional holidays. Summary of the invention

[0003] The purpose of the present invention is to solve the problem in the prior art that the inflow of customer service tickets during traditional holidays is significantly reduced compared with ordinary days, resulting in low accuracy in complaint ticket prediction during traditional holidays, and a complaint ticket prediction method optimized based on traditional holidays is proposed.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A complaint work order prediction method optimized based on traditional festivals includes the following steps:

[0006] S1, collect complaint ticket data and preprocess the data, including cleaning, missing value filling and outlier processing;

[0007] S2, extracting features from the preprocessed complaint work order data, mainly including two-dimensional feature extraction from bidirectional data, classification feature extraction from text description data, and creating new data based on the original data to extract features from the original data;

[0008] S3, using ARIMA model to analyze feature data;

[0009] S4, based on the time away from traditional festivals, determine the time influencing variables, and then establish different ARIMA models according to different time influencing variables to obtain the prediction results;

[0010] S5, building a deep neural network model, inputting the prediction results and feature data of the ARIMA model to obtain the prediction results;

[0011] S6, weighted fusion of the prediction results of the ARIMA model in S4 and the prediction results of the deep neural network model in S5 to obtain the final complaint ticket prediction results;

[0012] S7, use the actual complaint work order data to compare with the predicted results, evaluate the prediction performance of the model, and adjust the model parameters according to the comparison results.

[0013] Preferably, in S1 pretreatment,

[0014] The cleaning includes directly discarding data with a high frequency of missing in the complaint work order data;

[0015] The missing value filling includes filling missing values ​​for data with low missing frequency in the complaint work order data, wherein the missing data of customer star rating, cumulative number of complaints, voice package name, and acceptance time are filled with their average values;

[0016] The abnormal value processing includes checking the complaint work order data and modifying the abnormal value if an abnormal value is found in the complaint work order data.

[0017] Preferably, in S2, the two-dimensional data includes whether there is a previous order, whether it is an upgraded work order, and whether there is a cooperative inquiry order, and the data is changed to 1 if it is, and changed to 0 if it is not;

[0018] The text description data is classified into description ranges including good, normal, poor and very poor.

[0019] Preferably, in S3, when selecting the order of the ARIMA model, the order is determined using the autocorrelation function and the partial autocorrelation function graphs, the autocorrelation function shows the overall correlation, and the partial autocorrelation function shows the correlation between specific lag periods. By observing the autocorrelation and partial autocorrelation graphs, the order of the model is determined to establish the ARIMA model.

[0020] Preferably, in S5, the deep neural network model includes an input layer, several hidden layers and an output layer, the activation function is relu, the number of iterations is N times, the input layer receives the prediction results and feature data of the ARIMA model, the hidden layer discovers and captures the nonlinear relationship in the data, and the output layer outputs the final prediction results.

[0021] Preferably, S6 specifically includes:

[0022] Calculate the error sum of the ARIMA model based on the predicted values ​​of the ARIMA model;

[0023] Calculate the sum of squared errors of the deep neural network model using the predicted values ​​of the deep neural network model;

[0024] Through the calculation formula of the weight coefficient, the weight coefficients of the ARIMA model and the deep neural network model in the combined prediction model are calculated respectively, and the final complaint work order prediction result can be obtained.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] The present invention can determine the time influencing variables based on the time away from the traditional festivals, so as to establish different ARIMA models according to different time influencing variables, thereby obtaining prediction results, thereby improving the prediction accuracy of complaint work orders during traditional festivals; and after optimizing the ARIMA model and considering the interaction between the work order and the seasonal factors, the prediction measurement can also consider the impact of unconventional events, etc., and the prediction results of the ARIMA model can be input into the deep neural network to obtain the prediction results of the neural network, and the final prediction results can be obtained after weighted fusion, thereby further improving the prediction accuracy of the complaint work order; and then it is possible to adjust resource allocation, personnel arrangement and customer service strategy in advance according to the accurate prediction results, thereby improving customer satisfaction, reducing the complaint processing time, and optimizing the overall operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of a complaint work order prediction method based on traditional festival optimization proposed by the present invention;

[0028] Figure 2 A schematic diagram of a prediction model of a complaint work order prediction method optimized based on traditional festivals proposed by the present invention;

[0029] Figure 3 This is a schematic diagram of an ARIMA model for a complaint ticket prediction method optimized based on traditional festivals proposed in the present invention. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0031] Reference Figure 1-3 ,A complaint ticket prediction method based on traditional festival optimization, including,

[0032] S1 collects complaint ticket data and preprocesses the data, including cleaning, missing value filling, and outlier processing.

[0033] Among them, the cleaning includes directly discarding data with a high frequency of missing in the complaint work order data, such as the accumulated amount of arrears, the total available general traffic in the package, the general traffic used in the package, the Internet traffic and actual speed of the associated secondary card in the current month, etc.

[0034] Missing value filling includes filling missing values ​​for data with low missing frequency in complaint ticket data. For missing data such as customer star rating, cumulative number of complaints, voice package name or acceptance time, their average value is used for filling.

[0035] Outlier processing includes checking the complaint work order data and modifying the outlier value if an outlier is found in the data of a complaint work order.

[0036] S2, extract features from the pre-processed complaint work order data. This mainly includes two-dimensional feature extraction of bidirectional data, classification feature extraction of text description data, and creation of new data based on original data to achieve feature extraction of original data.

[0037] Among them, for two-dimensional data, such as whether there is a previous order, whether it is an upgraded work order, and whether there is a cooperative investigation order, the data is changed to 1 if it is yes, and to 0 if it is no. For text description data, the text description data is classified into description ranges. For example, the mood of complaining customers can be reduced to good, normal, poor, and very poor.

[0038] In more detail, since the season needs to be considered, a new variable season is created for the complaint ticket data according to the date corresponding to the complaint ticket data, a new variable product category is created according to the voice package name and the broadband package name, and a new variable customer satisfaction is created according to the return visit results and customer emotions. Other extracted features are not described in detail, as shown in the table.

[0039]

[0040]

[0041] S3, using the ARIMA model to analyze the characteristic data, and obtain the model orders p, d, and q, which are used to establish the ARIMA model.

[0042] Among them, when selecting the order of the ARIMA model, the autocorrelation function and partial autocorrelation function graph are often used to help determine the order. The autocorrelation function shows the overall correlation, and the partial autocorrelation function shows the correlation between specific lag periods. By observing the autocorrelation and partial autocorrelation graphs, the order of the model (p=1, d=2, q=0) is determined to establish the ARIMA model.

[0043] In the process of determining the p, d, and q orders of ARIMA, the root mean square error RMSE can be used as a standard and a rolling ARIMA model can be used. The rolling ARIMA model usually refers to a method of training and testing a model using a rolling window, which can help the model to be verified and evaluated in different time periods to better understand its performance. The optimal p, d, and q orders are then determined based on cross-validation. After training, the ARIMA model obtained the optimal order (p = 1, d = 2, q = 0) when the RMSE on the test set was 79.64.

[0044] S4, based on the time away from traditional festivals, determine the time influencing variables, and establish different ARIMA models according to different time influencing variables to obtain the prediction results.

[0045] Specifically, since traditional festivals such as the Spring Festival and the Mid-Autumn Festival are not on a fixed cycle, and the number of complaint tickets during these traditional festivals is significantly reduced compared to ordinary days, different models are established considering the distribution of traditional holidays at different times in different months. Figure 3 As shown in Figure 1, the model is divided into two submodules, namely the regARIMA module and the X-11 module. The regARIMA module is used to build the ARIMA model, while the X-11 module is used for seasonal adjustment to generate various segmentation models.

[0046] Among them, the regARIMA module formula is as follows:

[0047]

[0048] Among them, z t is a time series (i.e., a daily, weekly, monthly, or annual time series of complaint tickets), L is a lag operator, S is a seasonal cycle (e.g., for monthly data, S = 12, for quarterly data, S = 4, etc.), φ(L) is a non-seasonal autoregressive operator, is a seasonal autoregressive operator, θ(L) is a non-seasonal moving average operator, θ(L s ) is the seasonal moving average operator, α t It is the rate at which the value returns to its original level after time t by exponential decay.

[0049] The X-11 submodule uses the additive model, and its formula is Y t =C t +S t +l t , where C t is the trend cycle component, S t For seasonal components, l t It is an irregular component, which may refer to a time-influenced variable.

[0050] The X-11 submodule is combined with the regARIMA module, Y t Finally, it evolves into the following formula Y t =c+φ 1 Y t-1 +φ 2 Y t-2 +…+φ p Y t-p +θ 1 ε t-1 +θ 2 ε t-2 +θ q ε t-q +ε t , where c is a constant term, ε t is the error term at time point t, φ 1 to φ p is the parameter of the AR model, which is used to describe the relationship between the current value and the value at p time points in the past, and θ 1 to θ q It is the parameter of the MA sub-model in the ARIMA model, which is used to describe the relationship between the current value and the error at the past q time points.

[0051] Taking the Spring Festival as an example, 15 days before the Spring Festival (w b ), 7 days during the Spring Festival (w d ), 15 days after the Spring Festival (w a ), may be affected by the Spring Festival factor, and the time influencing variables of the three time periods are different. Three variables can be established to evaluate the impact of the Spring Festival factor. Assume that the time influencing variable is Xb before the Spring Festival, Xd during the Spring Festival, and Xa after the Spring Festival. The number of days before the Spring Festival is w b The closer to the Spring Festival, the greater the impact of the Spring Festival. b The time of day influences the variable Before the festival b -1 day time influence variable is The time influencing variable on the second day before the festival is The time influencing variable on the first day before the festival is The time-influencing variables during the Spring Festival remain unchanged, that is, the time-influencing variables for each day are The number of days affected after the Spring Festival is w a The further away from the Spring Festival, the smaller the influence of the Spring Festival. That is, the time influencing variable on the first day after the festival is After the festival a -1 day time influence variable is After the festival a The time of day influences the variable

[0052] S5, build a deep neural network model. It includes an input layer, 20 hidden layers and an output layer, the activation function is relu, and the number of iterations is 1000. The input layer receives the prediction results and feature data of the ARIMA model, the hidden layer is responsible for learning the nonlinear relationship in the data, that is, discovering and capturing the nonlinear relationship in the data, and the output layer outputs the final prediction results.

[0053] S6, weighted fusion of the prediction results of the ARIMA model and the deep neural network model to obtain the final complaint ticket prediction results.

[0054] Specifically, the predicted values ​​of the ARIMA model are first used to calculate the sum of square errors of the ARIMA model, and the predicted values ​​of the deep neural network model are used to calculate the sum of square errors of the deep neural network model, and then the weight coefficient is calculated by a calculation formula, such as the calculation formula of the weighted least squares method. The calculation formula of the weighted least squares method is L(β)=∑(w i ×(y i -β 0 -β 1 × i1 -...-β n × in ) 2 ), where w i is the weight, x i is the independent variable, y i is the dependent variable, β is the coefficient, and the weight coefficients of the ARIMA model and the deep neural network model in the combined prediction model are calculated respectively, such as (9.52, 90.48), and the final complaint ticket prediction result can be obtained.

[0055]

[0056] S7, use the actual complaint work order data to compare with the prediction results, evaluate the prediction performance of the model, and then adjust the model parameters according to the comparison results to improve the prediction accuracy.

[0057] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A complaint ticket prediction method optimized based on traditional festivals, characterized in that: The following steps are involved: S1, collect complaint ticket data and preprocess the data, including cleaning, missing value filling and outlier processing; S2, extracting features from the preprocessed complaint work order data, mainly including two-dimensional feature extraction from bidirectional data, classification feature extraction from text description data, and creating new data based on the original data to extract features from the original data; S3, using ARIMA model to analyze feature data; S4, based on the time away from traditional festivals, determine the time influencing variables, and then establish different ARIMA models according to different time influencing variables to obtain the prediction results; S5, building a deep neural network model, inputting the prediction results and feature data of the ARIMA model to obtain the prediction results; S6, weighted fusion of the prediction results of the ARIMA model in S4 and the prediction results of the deep neural network model in S5 to obtain the final complaint ticket prediction results; S7, use the actual complaint work order data to compare with the predicted results, evaluate the prediction performance of the model, and adjust the model parameters according to the comparison results.

2. According to claim 1, a method for predicting complaint work orders based on traditional festival optimization is characterized in that: In S1 preprocessing, The cleaning includes directly discarding data with a high frequency of missing in the complaint work order data; The missing value filling includes filling missing values ​​for data with low missing frequency in the complaint work order data, wherein the missing data of customer star rating, cumulative number of complaints, voice package name, and acceptance time are filled with their average values; The abnormal value processing includes checking the complaint work order data and modifying the abnormal value if an abnormal value is found in the complaint work order data.

3. According to claim 1, a method for predicting complaint work orders based on traditional festival optimization is characterized in that: In S2, the two-dimensional data includes whether there is a previous order, whether it is an upgraded work order, and whether there is a cooperative inquiry order, and the data is changed to 1 if it is yes and 0 if it is no; The text description data is classified into description ranges including good, normal, poor and very poor.

4. According to claim 1, a method for predicting complaint work orders based on traditional festival optimization is characterized in that: In S3, when selecting the order of the ARIMA model, the order is determined using the autocorrelation function and partial autocorrelation function graphs. The autocorrelation function shows the overall correlation, and the partial autocorrelation function shows the correlation between specific lag periods. By observing the autocorrelation and partial autocorrelation graphs, the order of the model is determined to establish the ARIMA model.

5. According to claim 1, the method for predicting complaint tickets based on traditional festival optimization is characterized in that: In S5, the deep neural network model includes an input layer, several hidden layers and an output layer. The activation function is relu and the number of iterations is N. The input layer receives the prediction results and feature data of the ARIMA model, the hidden layer discovers and captures the nonlinear relationship in the data, and the output layer outputs the final prediction results.

6. According to claim 1, a method for predicting complaint work orders based on traditional festival optimization is characterized in that: S6 specifically includes: Calculate the error sum of the ARIMA model based on the predicted values ​​of the ARIMA model; Calculate the sum of squared errors of the deep neural network model using the predicted values ​​of the deep neural network model; Through the calculation formula of the weight coefficient, the weight coefficients of the ARIMA model and the deep neural network model in the combined prediction model are calculated respectively, and the final complaint work order prediction result can be obtained.

Images