Customer complaint rate prediction method, device and computer equipment based on multi-model fusion

Abstract

The present application relates to a customer complaint rate prediction method, device, computer equipment and storage medium based on multi-model fusion. The method comprises: obtaining the business features to be predicted; inputting the business features to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtaining the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model; determining the corresponding prediction value weight according to the weight proportion value when each customer complaint rate prediction model is fused; performing weighted calculation and fusion according to each prediction weight and the corresponding daily customer complaint rate prediction value, and obtaining the target daily customer complaint rate prediction value of the prediction period. The use of this method can improve the accuracy of customer complaint rate prediction.

Description

Technical Field

[0001] The present application relates to the field of big data technology, and in particular to a method, device and computer equipment for predicting customer complaint rate based on multi-model fusion. Background Art

[0002] With the development of artificial intelligence technology, artificial intelligence technology has been applied in various fields, such as logistics. With the development of various logistics businesses in the logistics field, there are scenarios of customer complaints. Solving customer complaints and reducing customer complaint rates are urgent problems to be solved in logistics scenarios, that is, through customer complaint rate prediction, intervention measures such as resource allocation and related indicators can be combined with prediction results to reduce customer complaint rates.

[0003] However, most of the customer complaint rate prediction models in the field of artificial intelligence technology are first trained offline. The trained models are packaged and applied to actual production, which cannot guarantee the timeliness of the prediction results. The customer complaint rate prediction models currently used online are relatively simple, resulting in low accuracy in customer complaint rate prediction. Summary of the invention

[0004] Based on this, it is necessary to provide a customer complaint rate prediction method, device, computer equipment and storage medium based on multi-model fusion that can improve the accuracy of customer complaint rate prediction in response to the above technical problems.

[0005] A customer complaint rate prediction method based on multi-model fusion, the method comprising:

[0006] Obtaining business characteristics to be predicted;

[0007] Input the service characteristics to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model;

[0008] Determine the corresponding prediction value weight according to the weight ratio and error weight value when each customer complaint rate prediction model is fused;

[0009] A weighted calculation and fusion is performed based on each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the prediction period.

[0010] In one embodiment, a method for training a customer complaint rate prediction model includes:

[0011] Determine the data set used to train the customer complaint rate prediction model;

[0012] Based on the data set, determine the optimal training window length, the number of most important business features, and hyperparameter combinations for each customer complaint rate prediction model;

[0013] Construct a training dataset for model training;

[0014] Based on the optimal training window length, the number of the most important business features and the hyperparameter combination, each customer complaint rate prediction model is trained through the training data set to obtain a trained customer complaint rate prediction model.

[0015] In one embodiment, determining a data set for training a customer complaint rate prediction model includes:

[0016] Obtaining a comprehensive table of business feature sets related to customer complaint rate prediction values ​​and corresponding customer complaint rates;

[0017] Determine, according to the historical values ​​of each service feature and the historical values ​​of the customer complaint rate, a first correlation coefficient value between the service features and a second correlation coefficient value between each service feature and the customer complaint rate;

[0018] Determining target business features related to the customer complaint rate from the integrated table according to each first correlation coefficient value and each second correlation coefficient value, and obtaining an updated integrated table; and

[0019] The update integration table is copied and merged according to the number of copies to obtain a data set.

[0020] In one embodiment, the method further comprises:

[0021] Obtain the importance of each target business feature in the update integration table;

[0022] The target business features are sorted in descending order of importance to obtain a corresponding importance sorting table.

[0023] In one embodiment, determining the optimal training window length of each customer complaint rate prediction model includes:

[0024] Get the preset training window length set;

[0025] The time series model is used to predict the predicted value of each target business feature in the first period to obtain the first backtest set;

[0026] Determine, according to the first backtest set, a first backtest error value of each customer complaint rate prediction model in the first period;

[0027] An optimal training window length is determined from the prediction training window set based on the first backtest error value.

[0028] In one embodiment, determining the number of most important business features of each customer complaint rate prediction model includes:

[0029] Obtain a set of preset important business features;

[0030] According to the preset important business feature number and importance ranking table in the preset important business feature number set, determining the preset important business feature number important business features from the first backtest set in descending order of importance, to obtain a second backtest set;

[0031] Determine, according to each second backtest set, a second backtest error value of each customer complaint rate prediction model in the first period;

[0032] The number of most important business features is determined from the prediction training window set according to each second backtest error value.

[0033] In one embodiment, determining a hyperparameter combination of each customer complaint rate prediction model includes:

[0034] Determine candidate hyperparameters for each customer complaint rate prediction model;

[0035] Determine the minimum backtest error value of each customer complaint rate prediction model in the first period according to the third backtest set determined by the optimal training window length, the number of most important business features, and the number of most important business features in the first backtest set;

[0036] Based on the minimum backtest error value, the optimal parameter combination of each customer complaint rate prediction model is determined.

[0037] In one embodiment, constructing a training data set for model training includes:

[0038] Determine the target business features and date features from the data set based on the optimal training window length and the number of most important business features;

[0039] The data of target business characteristics and date characteristics are used as training data sets.

[0040] A customer complaint rate prediction device based on multi-model fusion, the device comprising:

[0041] An acquisition module, used to acquire the business characteristics to be predicted;

[0042] A prediction module, used to input the service features to be predicted into at least two trained customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model;

[0043] A determination module is used to determine the corresponding prediction value weight according to the weight ratio value and error weight value when each customer complaint rate prediction model is fused;

[0044] The fusion module is used to perform weighted calculation and fusion according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the prediction period.

[0045] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0046] Obtaining business characteristics to be predicted;

[0047] Input the service characteristics to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model;

[0048] Determine the corresponding prediction value weight according to the weight ratio and error weight value when each customer complaint rate prediction model is fused;

[0049] A weighted calculation and fusion is performed based on each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the prediction period.

[0050] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0051] Obtaining business characteristics to be predicted;

[0052] Input the service characteristics to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model;

[0053] Determine the corresponding prediction value weight according to the weight ratio and error weight value when each customer complaint rate prediction model is fused;

[0054] A weighted calculation and fusion is performed based on each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the prediction period.

[0055] The above-mentioned customer complaint rate prediction method, device, computer equipment and storage medium based on multi-model fusion, by respectively inputting the business characteristics to be predicted into different pre-trained customer complaint rate prediction models, obtains the daily customer complaint rate prediction values ​​predicted by different customer complaint rate prediction models, and by determining the fusion weights of each model, and performing weighted calculation fusion on the daily customer complaint rate prediction values ​​of each model according to the fusion weights and error weights of each model, further improves the disadvantage of missing feature values ​​of the prediction set in the customer complaint rate prediction model, and improves the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is an application environment diagram of a customer complaint rate prediction method based on multi-model fusion in one embodiment;

[0057] Figure 2 1 is a flow chart of a method for predicting customer complaint rate based on multi-model fusion in one embodiment;

[0058] Figure 3 A flowchart of a method for determining a data set for training a customer complaint rate prediction model in one embodiment;

[0059] Figure 4 A schematic diagram of a flow chart of a method for determining an optimal training window length for a customer complaint rate prediction model in one embodiment;

[0060] Figure 5 A flowchart of a method for determining the number of most important business features of a customer complaint rate prediction model in one embodiment;

[0061] Figure 6 A schematic diagram of a flow chart of a method for determining a hyperparameter combination of a customer complaint rate prediction model in one embodiment;

[0062] Figure 7 A flowchart of a method for training a customer complaint rate prediction model using multi-model fusion in one embodiment;

[0063] Figure 8 is a flow chart of a customer complaint rate prediction method based on multi-model fusion in another embodiment;

[0064] Fig. 9 1 is a flow chart of a method for predicting customer complaint rate based on multi-model fusion in one embodiment;

[0065] Fig.10 is a structural block diagram of a customer complaint rate prediction device based on multi-model fusion in one embodiment;

[0066] Fig.11 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0068] The customer complaint rate prediction method based on multi-model fusion provided in this application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The terminal 102 obtains the business features to be predicted from the server 104; the business features to be predicted are respectively input into at least two trained different customer complaint rate prediction models to obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model; the corresponding prediction value weight is determined according to the weight proportion value and error weight value when each customer complaint rate prediction model is fused; weighted calculation and fusion are performed according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value of the prediction period. Among them, the terminal 102 can be but is not limited to various personal computers, laptops, smart phones, tablet computers and portable wearable devices, and the server 104 can be implemented with an independent server or a server cluster composed of multiple servers.

[0069] In one embodiment, Figure 2 As shown in the figure, a customer complaint rate prediction method based on multi-model fusion is provided. Figure 1 The terminal in is taken as an example for explanation. In this embodiment, the fusion of Prophet model, XGBoost model and ANN model is taken as an example for explanation. The method includes the following steps:

[0070] Step 202: Obtain the characteristics of the service to be predicted.

[0071] Among them, business characteristics are the volume characteristics, resource characteristics and indicator characteristics related to the customer complaint rate; volume characteristics include the number of parcels received and delivered, the number of unloaded and uninvoiced items, the total number of vehicles, the timeliness achievement rate, etc. The data of each business characteristic and the customer complaint rate are stored in the form of a table; the storage method can be achieved by building a dynamic partition table, solidifying the table into a dynamic partition table, and performing incremental updates on a daily basis and overwriting and refreshing the customer complaint rate value of the previous date; each row of the table represents the values ​​of all business characteristics and the corresponding customer complaint rate value in a certain area on a certain day. As shown in Table 1:

[0072] Table 1

[0073]

[0074] The business features to be predicted are determined based on the correlation between each business feature and the customer complaint rate, that is, the business features to be predicted are determined from an importance ranking table based on the number of the most important business features determined and the importance of the business features.

[0075] Specifically, the terminal obtains a predetermined prediction set from the server, and obtains the to-be-predicted service features of the most important number of service features from the prediction set in descending order of importance of the service features.

[0076] Step 204: input the service features to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model.

[0077] Among them, each customer complaint rate prediction model is based on determining the optimal training window length, the number of the most important business features and the combination of hyperparameters, and the same training data set is obtained from different partitions of the database to perform online training on each customer complaint rate prediction model, and the training process of each customer complaint rate prediction model is the same. By performing online training on each customer complaint rate prediction model, it is possible to adjust the training according to the actual situation (for example, the accuracy of the customer complaint rate value, the real-time nature of the training data) to improve the accuracy of the prediction of each model. The customer complaint rate prediction model can be, but is not limited to, a Prophet model, an XGBoost model and an ANN model. The date features in the feature engineering of different customer complaint prediction models are different. The date features in the Prophet model are parameters for setting the holiday list, and the date features in the XGBoost and ANN models are features of date characteristics, such as year, month, date number, week number, whether it is a weekend, whether it is a holiday, etc. In the XGBoost and ANN models, for category variables such as date number and week number, the weight of evidence (WOE) conversion is performed to obtain the importance of each date feature.

[0078] The forecast time period is determined based on the forecast demand. For example, the preset time period is from T to T+7, where T is the forecast start date.

[0079] Specifically, the business features to be predicted are respectively input into the trained Prophet model, XGBoost model and ANN model to obtain the daily customer complaint rate value from T to T+7.

[0080] Step 206: Determine the corresponding prediction value weight according to the weight ratio value when each customer complaint rate prediction model is fused.

[0081] Among them, the weight ratio value refers to the value of each model when the model is integrated according to the backtest error value of each customer complaint rate prediction, that is, the error weight. For example, the customer complaint rate prediction models are Prophet model, XGBoost model and ANN model, and the error weight of the Prophet model is determined to be 1 / 3, the error weight of the XGBoost model is 1 / 3, and the error weight of the ANN model is 1 / 3.

[0082] Specifically, the minimum backtest error value of each customer complaint rate prediction model is obtained, and the prediction value weight of each corresponding customer complaint rate prediction model is obtained by weighted calculation according to the minimum backtest error value of each customer complaint rate prediction model and the weight ratio value during fusion. For example, the weight ratio values ​​of the Prophet model, XGBoost model, and ANN model are all 1 / 3, and the corresponding minimum backtest error values ​​are Epro, Exgb, and Eann, respectively. The prediction value weight of the Prophet model, the prediction value weight of the XGBoost model, and the prediction value weight of the ANN model can be obtained through weighted calculation, where:

[0083] The prediction value weight of the Prophet model is expressed as:

[0084]

[0085] The prediction weight of the XGBoost model is expressed as

[0086]

[0087] The prediction value weight of the ANN model is expressed as

[0088]

[0089] Step 208 , weighted calculation and fusion are performed according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain a target daily customer complaint rate prediction value for the prediction period.

[0090] Specifically, the daily customer complaint rate prediction value of each customer complaint rate prediction model is obtained, and the prediction value weight of each customer complaint rate prediction model and the daily customer complaint rate prediction value are weighted and fused to obtain the target daily customer complaint rate prediction value of the prediction period. The weighted calculation fusion process can be expressed as:

[0091] After fusion, the predicted value of daily customer complaint rate = Prophet's predicted value of daily customer complaint rate * Prophet's predicted value weight + XGBoost's predicted value of daily customer complaint rate * XGBoost's predicted value weight + ANN's predicted value of daily customer complaint rate * ANN's predicted value weight

[0092] In the above-mentioned customer complaint rate prediction method based on multi-model fusion, the business features to be predicted are respectively input into different pre-trained customer complaint rate prediction models to obtain the daily customer complaint rate prediction values ​​predicted by different customer complaint rate prediction models, and the fusion weights of each model are determined, and the daily customer complaint rate prediction values ​​of each model are weighted and fused according to the fusion weights and error weights of each model, thereby further improving the disadvantage of missing feature values ​​of the prediction set in the customer complaint rate prediction model and improving the prediction accuracy of the model.

[0093] In one embodiment, Figure 3As shown, a method for determining a data set for training a customer complaint rate prediction model is provided, and the method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0094] Step 302: Obtain an integrated table of a service feature set related to the customer complaint rate prediction value and the corresponding customer complaint rate.

[0095] The business feature set includes the number of parcels received and delivered, the number of unloaded and uninvoiced items, the total number of vehicles, the timeliness achievement rate, etc. The integration table refers to mapping the characteristic values ​​of the business feature set related to the customer complaint rate prediction value and the corresponding customer complaint rate value into the data table. By constructing a dynamic partition table, the table is solidified into a dynamic partition table to obtain a data table including business characteristics and customer complaint rate.

[0096] Step 304: Determine a first correlation coefficient value between the service features and a second correlation coefficient value between the service features and the customer complaint rate based on the historical value of each service feature and the historical value of the customer complaint rate.

[0097] The first correlation coefficient value and the second correlation coefficient value are determined by calculating the strategy data according to the Pearson formula.

[0098] Step 306: Determine the target business feature related to the customer complaint rate from the integrated table according to each first correlation coefficient value and each second correlation coefficient value, and obtain an updated integrated table.

[0099] Specifically, the first correlation coefficient value and the second correlation coefficient value are calculated by the Pearson formula; according to each first correlation coefficient value and each second correlation coefficient value, the target business feature related to the customer complaint rate is determined from the integrated table to obtain an updated integrated table, that is, the business feature pairs consisting of every two business features are traversed in turn, and the business feature pairs whose first correlation coefficient values ​​are greater than a first preset value (for example, 0.8) are determined, and the business feature with the smallest second correlation coefficient value between the business feature and the customer complaint rate in the business feature pair is deleted, so that there is no collinearity between the determined target business features, and an updated integrated table is obtained.

[0100] Step 308: copy and merge the update integration table according to the number of copies to obtain a data set.

[0101] The number of copies is determined according to the number of customer complaint rate prediction models to be merged. For example, the customer complaint rate prediction models are Prophet model, XGBoost model and ANN model. The update integration table is copied into three copies and merged to obtain a data set.

[0102] Specifically, the update integration table is copied and merged according to the number of replications to obtain a data set. A column identifier is added to each table obtained by copying the update integration table, and the column identifier has a one-to-one correspondence with the customer complaint rate prediction model to be merged; and the data set is transferred from the database Hive to the target database Kafka to facilitate the acquisition of the data set and improve the timeliness of the acquisition of the data set.

[0103] Optionally, in one embodiment, the importance of each target business feature in the update integration table is obtained; and each target business feature is sorted in order from high to low importance to obtain a corresponding importance sorting table.

[0104] Specifically, the importance of each target business feature in the updated integration table is obtained through the feature importance identification function feature_importances of the customer complaint rate prediction model (for example, XGBoost); the target business features are sorted in order from high to low importance to obtain the corresponding importance sorting table; wherein, the customer complaint rate prediction model is saved in py format, and the py file can be called daily through spark, and the model is run every day according to the latest target business feature data of the day to obtain the importance of each target business feature.

[0105] The method for determining the data set for training the customer complaint rate prediction model determines the first correlation coefficient value between the business features related to the customer complaint rate and the second correlation coefficient value between each business feature and the customer complaint rate, deletes the business features with a weak correlation with the customer complaint rate, and obtains the target business features with a strong correlation with the customer complaint rate. The data set for training each customer complaint rate prediction model is obtained based on the target business features, thereby improving the reliability of the training data.

[0106] In one embodiment, Figure 4 As shown in FIG. 1 , a method for determining the optimal training window length of a customer complaint rate prediction model is provided, and the method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0107] Step 402: Obtain a preset training window length set.

[0108] The preset training window length W set may include 360 ​​days, 450 days and 720 days.

[0109] Step 404, predicting the predicted value of each target business feature in the first period through the time series model to obtain a first backtest set.

[0110] The first period refers to a period before the forecast start date T, for example, from T-30 to T. The first backtest set is obtained by training the Prophet model by taking the data of each target business feature within n (n can be 30) days before the forecast start date T in the data set, that is, the first period (for example, from T to T-30 days) as the training set of the time series model Prophet, and then predicting the predicted value of each target business feature from T to T-30 days through the Prophet model, and determining the predicted value based on the obtained predicted value.

[0111] Optionally, the Prophet model is used to predict the predicted value of each target business feature from T to T+7 days to obtain the prediction set data.

[0112] Step 406: Determine a first backtest error value of each customer complaint rate prediction model in the first period according to the first backtest set.

[0113] Specifically, each preset training window length W is selected from the preset training window length set in turn, and the data of all target business features in the second period are determined from the data set of each customer complaint rate prediction model according to each preset training window length W to obtain a training set, and the corresponding customer complaint rate prediction model is trained to obtain a trained customer complaint rate prediction model; the trained customer complaint rate prediction model is backtested through the first backtest set to obtain the first backtest error value corresponding to each preset training window length. The second period is determined based on the preset training window length and the first n days before the prediction start date, that is, within T-30-W to T-30 days; the calculation expression of the first backtest error value MAPE can be expressed as follows:

[0114]

[0115] Where n is the number of predicted values ​​in the first backtest set.

[0116] Step 408: determine the optimal training window length from the predicted training window length set according to the first backtest error value.

[0117] Specifically, the optimal training window length is determined by taking the predicted training window length corresponding to the minimum first backtest error value.

[0118] The above method for determining the optimal training window length of the customer complaint rate prediction model predicts the predicted value of each target business feature in the first period through a time series model to obtain a first backtest set, and determines the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T based on the first backtest set, and determines the optimal training window length by taking the predicted training window length corresponding to the minimum first backtest error value, that is, obtaining the optimal training window length of each customer complaint rate prediction model. By determining the optimal training window length, the speed of model convergence can be accelerated and the duration of model training can be shortened.

[0119] In one embodiment, Figure 5 As shown in FIG. 1 , a method for determining the number of the most important business features of the customer complaint rate prediction model is provided. The method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0120] Step 502: Obtain a preset set of important business characteristics.

[0121] The preset set of important business feature numbers may be, but is not limited to, [5, 10, 15].

[0122] Step 504, according to the preset important business feature number and importance ranking table in the preset important business feature number set, determine the preset important business feature number important business features from the first backtest set in descending order of importance to obtain a second backtest set.

[0123] Step 506: Determine the second backtest error value of each customer complaint rate prediction model in the first period according to each second backtest set.

[0124] Specifically, the number of important business features is selected from the set of important business feature numbers in turn, and according to the importance ranking table of each preset number of important business features and the target business feature, the data of the date characteristics of each preset number of important business features and each customer complaint rate prediction model in the second period are determined from the first backtest set in order of importance from high to low as the training set, and each customer complaint rate prediction model is trained to obtain the trained customer complaint rate prediction model; the trained customer complaint rate prediction model is backtested through the second backtest set to obtain the second backtest error value corresponding to each preset number of important business features in the first period.

[0125] Step 508: Determine the number of most important business features from a preset set of important business features according to each second backtest error value.

[0126] Specifically, the preset number of important business features corresponding to the minimum second backtest error value is determined as the number of most important business features.

[0127] In the above method for determining the number of the most important business features of the customer complaint rate prediction model, by determining the number of each preset important business feature and the importance ranking table in the preset important business feature number set in order from high to low importance from the first backtest set, a second backtest set is obtained, and based on the second backtest set, the second backtest error value of each customer complaint rate prediction model during the period from T-30 to T is determined, and the preset number of important business features corresponding to the smallest second backtest error value is determined as the most important number of business features, thereby shortening the duration of model training.

[0128] In one embodiment, Figure 6 As shown in FIG. 1 , a method for determining the hyperparameter combination of the customer complaint rate prediction model is provided, and the method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0129] Step 602: determine candidate hyperparameters for each customer complaint rate prediction model.

[0130] Among them, the candidate hyperparameters of different customer complaint rate prediction models are different. For example, the candidate hyperparameters of the Prophet model include:

[0131] changepoint_prior_scale:[0.03,0.05,0.07]; holidays_prior_scale:[7,10,12],

[0132] seasonality_prior_scale:[7,10,12];

[0133] Candidate hyperparameters for the XGBoost model include:

[0134] n_estimator:[500,600,700,800,900,1000,1200,1400];

[0135] max_depth:[3,4,5,6,7,8,9,10]; min_child_weight:[1,2,3,4,5,6];

[0136] gamma:[0,0.1,0.2,0.3,0.4,0.5,0.6]; subsample:[0.7,0.8,0.9,1];

[0137] colsample_bytree:[0.6,0.7,0.8,0.9];reg_alpha:[0,0.05,0.1,1,2,3];

[0138] reg_lambda:[0.05,0.1,1,2,3]; learning_rate:[0.01,0.05,0.07,0.1,0.2];

[0139] Candidate hyperparameters for the ANN model include:

[0140] batch_size:[10,20,40,60,80,100]; epochs:[10,50,100];

[0141] dropout_rate:[0.1,0.4,0.7,0.9].

[0142] Step 604, determining the minimum backtest error value of each customer complaint rate prediction model in the first period according to the third backtest set determined based on the optimal training window length, the number of most important business features and the number of most important business features in the first backtest set.

[0143] Step 606: Determine the optimal parameter combination of each customer complaint rate prediction model based on the minimum backtest error value.

[0144] Specifically, according to the optimal training window length and the number of the most important business features, the data of the most important business features during T-30-Wbest and T-30, as well as the date features, are used as the third training set to train each customer complaint rate prediction model to obtain the trained customer complaint rate prediction model; the trained customer complaint rate prediction model is backtested through the third backtest set to obtain the third backtest error value corresponding to each preset number of important business features in the first period, and each third backtest error value is used as a hyperparameter search target; according to the Bayesian optimization theory and setting the maximum number of searches, the hyperparameter combination corresponding to the minimum third backtest error value of each customer complaint rate prediction model is determined as the optimal parameter combination.

[0145] The above method for determining the hyperparameter combination of the customer complaint rate prediction model backtests the trained customer complaint rate prediction model through the third backtest set, obtains the third backtest error value corresponding to each preset number of important business features in the first period, and uses each third backtest error value as the hyperparameter search target; according to the Bayesian optimization theory and setting the maximum number of searches, the hyperparameter combination corresponding to the minimum third backtest error value of each customer complaint rate prediction model is determined as the optimal parameter combination; that is, the hyperparameter search is performed based on the Bayesian optimization theory, avoiding traversal of all hyperparameters, thereby improving the efficiency of hyperparameter traversal.

[0146] In one embodiment, Figure 7 As shown in FIG, a training method for a customer complaint rate prediction model for multi-model fusion is provided, and the method is applied to Figure 1The terminal in is used as an example to illustrate, including the following steps:

[0147] Step 702: determine a data set for training each customer complaint rate prediction model.

[0148] Specifically, by obtaining a set of business features related to the predicted value of the customer complaint rate and an integrated table of the corresponding customer complaint rate; according to the historical values ​​of each business feature and the historical values ​​of the customer complaint rate, the first correlation coefficient value between each business feature and the second correlation coefficient value between each business feature and the customer complaint rate are calculated through the Pearson formula calculation strategy data; according to each first correlation coefficient value and each second correlation coefficient value, the target business feature related to the customer complaint rate is determined from the integrated table, an updated integrated table is obtained, and copying and merging are performed to obtain a data set.

[0149] Step 704: Determine the optimal training window length, the number of most important business features, and the hyperparameter combination of each customer complaint rate prediction model based on the data set.

[0150] Specifically, a preset training window length set is obtained; the predicted value of each target business feature in the first period is predicted by a time series model to obtain a first backtest set; based on the first backtest set, the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T is determined; and the optimal training window length is determined from the prediction training window set based on the first backtest error value.

[0151] By obtaining a set of preset important business feature numbers; determining each preset number of important business features from the first backtest set in order of importance from high to low according to the number of preset important business features and an importance ranking table in the set of preset important business features, and obtaining a second backtest set; determining the second backtest error value of each customer complaint rate prediction model during the period from T-30 to T according to each second backtest error value; and determining the number of most important business features from the set of preset important business features according to each second backtest error value.

[0152] According to the candidate hyperparameters, optimal training window length, number of most important business features of each customer complaint rate prediction model, and the third backtest set determined by the number of most important business features in the first backtest set, the minimum backtest error value of each customer complaint rate prediction model during the period from T-30 to T is determined; based on the minimum backtest error value, the optimal parameter combination of each customer complaint rate prediction model is determined.

[0153] Step 706, construct a training data set for model training.

[0154] Specifically, target business features and date feature data are determined from the data set according to the optimal training window length and the number of most important business features; and the target business features and date feature data are used as the training data set.

[0155] Step 708 , based on the optimal training window length, the number of most important business features and the hyperparameter combination, each customer complaint rate prediction model is trained through the training data set to obtain a trained customer complaint rate prediction model.

[0156] Specifically, according to each customer complaint rate prediction model to be trained, a data set is obtained from the corresponding partition of the database, and the data of important business features and date features are determined from the data set according to the optimal training window length and the number of the most important business features; the data of the target business features and date features are used as the training data set, and each customer complaint rate prediction model is trained through the training data set to obtain a trained customer complaint rate prediction model; that is, the data of the most important business features and date features within the optimal training window length of the prediction start date are used as the training data set, and each customer complaint rate prediction is fitted online. The weight ratio of each model fusion is determined according to the backtest error value of each customer complaint rate prediction model, and each customer complaint rate prediction model is fused according to the weight ratio.

[0157] In the above-mentioned training method for the customer complaint rate prediction model for multi-model fusion, a data set for training each customer complaint rate prediction model is determined. By determining the optimal training window length, the number of most important business features and hyperparameter combination and training data set for each customer complaint rate prediction model, each customer complaint rate prediction model can be trained online at the same time, thereby shortening the training time and improving the training effect of the model, thereby improving the accuracy of the prediction of each model.

[0158] In another embodiment, Figure 8 As shown in the figure, a customer complaint rate prediction method based on multi-model fusion is provided. Figure 1 Taking the terminal in as an example, the method includes the following steps:

[0159] Step 802: determine a data set for training a customer complaint rate prediction model.

[0160] Specifically, the terminal obtains a business feature set related to the customer complaint rate prediction value and a corresponding integrated table of the customer complaint rate from the server; determines a first correlation coefficient value between the business features and a second correlation coefficient value between each business feature and the customer complaint rate according to historical values ​​of each business feature and historical values ​​of the customer complaint rate; traverses the business feature set in turn according to each first correlation coefficient value and each second correlation coefficient value to obtain an updated integrated table of the target business features related to the customer complaint rate in the retained integrated table; and copies and merges the updated integrated table according to the number of copies, stores each copied integrated table in a partition memory, and obtains a data set for model training; each copied integrated table has a column identifier associated with the corresponding customer complaint rate prediction model; the number of copies is determined according to the number of pre-built customer complaint rate prediction models.

[0161] Step 804: Determine the optimal training window length, the number of most important business features, and the combination of hyperparameters for each customer complaint rate prediction model based on the data set.

[0162] Specifically, sort each target business feature in order of importance from high to low to obtain a corresponding importance ranking table. Obtain the importance ranking table and obtain the corresponding data set from the partition memory according to the column identifier associated with each customer complaint rate prediction model; determine the feature parameter set (including date features) of each customer complaint rate prediction model; use the historical values ​​of each target business feature before the prediction start date T-30 as the first training set, and use the Prophet model to predict the predicted values ​​of each feature from T-30 to T as the first backtest set; use the historical values ​​of each target business feature before the prediction start date T as the second training set, and use the Prophet model to predict the predicted values ​​of each feature from T to T+7 as the prediction set. Obtain a preset training window duration W set; use the historical values ​​of all target business features during T-30-W and T-30 as the second training set; determine the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T based on the first backtest set; determine the optimal training window length from the prediction training window set based on the first backtest error value.

[0163] Obtain N sets of preset important business features; according to the optimal training window length, use the N target business features and date features corresponding to the period T-30-Wbest and T-30 as the third training set; determine the second backtest error value of each customer complaint rate prediction model during the period T-30 to T according to the second backtest set determined by the N target business features in the first backtest set; determine the number of most important business features from the prediction training window set according to each second backtest error value. According to the optimal training window length and the number of most important business features, use the data of the preset number of important business features and date features during the period T-30-Wbest and T-30 as the third training set; determine the minimum backtest error value of each customer complaint rate prediction model during the period T-30 to T according to the third backtest set determined by the Nbest target business features in the first backtest set; determine the optimal parameter combination of each customer complaint rate prediction model based on the minimum backtest error value.

[0164] Step 806, construct a training data set for model training.

[0165] Step 808 , based on the optimal training window length, the number of most important business features and the hyperparameter combination, each customer complaint rate prediction model is trained through the training data set to obtain a trained customer complaint rate prediction model.

[0166] Step 810: Obtain the service characteristics to be predicted.

[0167] Specifically, the Prophet model is used to predict the predicted value of each target business feature in the third period (T to T+7 days, T is the prediction start date) to obtain the prediction set data, and the number of most important business features is determined from the prediction set data according to the importance ranking table, and the number of most important business features is used as the business features to be predicted to obtain the data of the business features to be predicted.

[0168] Step 812: Input the service features to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model.

[0169] Step 814: determine the corresponding prediction value weight according to the weight ratio of each customer complaint rate prediction model when it is fused.

[0170] Step 816: Perform weighted calculation and fusion based on each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the prediction period.

[0171] The following is an embodiment, Fig. 9 As shown, the application scenario of the customer complaint rate prediction method based on multi-model fusion is as follows: by obtaining all business features related to the customer complaint rate, the obtained business features and the corresponding customer complaint rate values ​​are integrated into an integrated table; according to the historical values ​​of each business feature and the historical values ​​of the customer complaint rate, the first correlation coefficient value between the business features and the second correlation coefficient value between each business feature and the customer complaint rate are determined by calculating the strategy data through the Pearson formula; according to each first correlation coefficient value and each second correlation coefficient value, the target business feature related to the customer complaint rate is determined from the integrated table to obtain an updated integrated table; and the updated integrated table is copied and merged according to the number of copies to obtain a data set; and the importance of each target business feature is determined through the XGBoost model to obtain a corresponding importance ranking table.

[0172] Construct feature engineering of each customer complaint rate prediction model (such as model A, model B and model C) in turn, confirm the date features of each customer complaint rate prediction model, that is, set the holiday list parameter in the Prophet model; add date characteristics to the XGBoost and ANN models, such as year, month, date number, weekday number, whether it is a weekend, whether it is a holiday. Obtain data from the database partition corresponding to each customer complaint rate prediction model (for example, model A, model B and model C obtain data from database partition A, database partition B and database partition C in turn), use the data of each target business feature from T to T-30 days as the training set of Prophet, train the Prophet model, obtain the Prophet model for predicting the target business feature value, predict the predicted value of each target business feature from T to T-30 days through the Prophet model, as the first backtest set, predict the predicted value of each target business feature from T to T+7 days through the Prophet model, and obtain the prediction set data.

[0173] According to the first backtest set, determine the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T; determine the optimal training window length from the prediction training window set according to the first backtest error value; obtain a preset important business feature number set; according to the preset important business feature number and importance ranking table in the preset important business feature number set, determine the preset important business feature number from the first backtest set in order from high to low importance, and obtain a second backtest set; according to each second backtest set, determine the customer complaint rate prediction model The second backtest error value during the period from T-30 to T; determining the number of the most important business features from a preset set of important business features according to each second backtest error value; determining the candidate hyperparameters of each customer complaint rate prediction model; determining the minimum backtest error value of each customer complaint rate prediction model during the period from T-30 to T according to the third backtest set determined by the optimal training window length, the number of the most important business features, and the number of the most important business features in the first backtest set; determining the optimal parameter combination of each customer complaint rate prediction model based on the minimum backtest error value.

[0174] According to the optimal training window length and the number of most important business features, the target business features and the date features are determined from the data set; the target business features and the date features are used as the training data set.

[0175] The business features to be predicted are obtained from the prediction set data, and the business features to be predicted are respectively input into at least two trained different customer complaint rate prediction models to obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model; the corresponding prediction value weight is determined according to the weight ratio value when each customer complaint rate prediction model is fused; weighted calculation and fusion are performed according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value of the prediction period, and the obtained target daily customer complaint rate prediction value and the importance ranking table of the determined target business features are displayed on the display screen of the terminal.

[0176] In the above-mentioned customer complaint rate prediction method based on multi-model fusion, the optimal training window length, the number of the most important business features and the hyperparameter combination of each customer complaint rate model are determined by predetermining the data set; the training data is pre-constructed, and each customer complaint rate prediction model (Prophet, XGBoost, ANN) is trained at the same time; the business feature distribution is input into the trained model to obtain three groups of customer complaint rate prediction values, and the weight ratio of the three groups of customer complaint rate prediction values ​​is calculated; the final daily customer complaint rate prediction value is calculated according to the error weight and weight ratio of each model, that is, each customer complaint rate prediction model is trained online, the fusion weight of each model is determined, and the daily customer complaint rate prediction value of each model is weighted and calculated and fused according to the fusion weight and error weight of each model, so as to further improve the disadvantage of missing feature values ​​of the prediction set in the customer complaint rate prediction model and improve the prediction accuracy of the model.

[0177] It should be understood that although Figure 2-8 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2-8 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0178] In one embodiment, Fig.10 As shown, a customer complaint rate prediction device based on multi-model fusion is provided, including: an acquisition module 1002, a prediction module 1004, a determination module 1006 and a fusion module 1008, wherein:

[0179] The acquisition module 1002 is used to acquire the characteristics of the service to be predicted.

[0180] The prediction module 1004 is used to input the service features to be predicted into at least two trained customer complaint rate prediction models respectively, and obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model.

[0181] The determination module 1006 is used to determine the corresponding prediction value weight according to the weight proportion value and the error weight value when each customer complaint rate prediction model is fused.

[0182] The fusion module 1008 is used to perform weighted calculation and fusion according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain a target daily customer complaint rate prediction value for the prediction period.

[0183] The above-mentioned customer complaint rate prediction device based on multi-model fusion obtains the daily customer complaint rate prediction values ​​predicted by different customer complaint rate prediction models by inputting the business features to be predicted into different pre-trained customer complaint rate prediction models respectively, and determines the fusion weights of each model. The daily customer complaint rate prediction values ​​of each model are weighted and fused according to the fusion weights and error weights of each model, thereby further improving the disadvantage of missing feature values ​​of the prediction set in the customer complaint rate prediction model and improving the prediction accuracy of the model.

[0184] In another embodiment, a customer complaint rate prediction device based on multi-model fusion is provided, which includes, in addition to the acquisition module 1002, the prediction module 1004, the determination module 1006 and the fusion module 1008, a construction module, a training module, a merging module and a sorting module, wherein:

[0185] The determination module 1006 is further used to determine a data set for training a customer complaint rate prediction model; based on the data set, determine the optimal training window length, the number of most important business features and the hyperparameter combination of each customer complaint rate prediction model.

[0186] The building module is used to construct a training dataset for model training.

[0187] The training module is used to train each customer complaint rate prediction model through a training data set based on the optimal training window length, the number of most important business features and the hyperparameter combination, so as to obtain a trained customer complaint rate prediction model.

[0188] The acquisition module 1002 is used to obtain an integrated table of a service feature set related to a customer complaint rate prediction value and a corresponding customer complaint rate.

[0189] The determination module 1006 is also used to determine the first correlation coefficient value between the business features and the second correlation coefficient value between each business feature and the customer complaint rate based on the historical value of each business feature and the historical value of the customer complaint rate; based on each first correlation coefficient value and each second correlation coefficient value, determine the target business feature related to the customer complaint rate from the integrated table to obtain an updated integrated table.

[0190] The merging module is used to copy and merge the update integration table according to the number of copies to obtain a data set.

[0191] The acquisition module 1002 is used to acquire the importance of each target business feature in the update integration table.

[0192] The sorting module is used to sort the target business features in order from high to low importance to obtain a corresponding importance sorting table.

[0193] The acquisition module 1002 is further used to acquire a preset training window length set.

[0194] The prediction module 1004 is also used to predict the predicted value of each target business feature in the first period through the time series model to obtain a first backtest set.

[0195] The determination module 1006 is used to determine the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T according to the first backtest set; and determine the optimal training window length from the prediction training window set according to the first backtest error value.

[0196] The acquisition module 1002 is further used to acquire a preset set of important service feature numbers.

[0197] The determination module 1006 is further configured to determine the preset number of important business features from the first backtest set in descending order of importance according to the preset number of important business features and the importance ranking table in the preset number of important business features set, to obtain a second backtest set.

[0198] The determination module 1006 is further configured to determine the second backtest error value of each customer complaint rate prediction model in the first period according to each second backtest set; and determine the number of most important business features from a preset set of important business feature numbers according to each second backtest error value.

[0199] The determination module 1006 is also used to determine candidate hyperparameters of each customer complaint rate prediction model; determine the minimum backtest error value of each customer complaint rate prediction model in the first period according to the optimal training window length, the number of most important business features, and the third backtest set determined by the number of most important business features in the first backtest set; determine the optimal parameter combination of each customer complaint rate prediction model based on the minimum backtest error value; and determine the data of target business features and date features from the data set according to the optimal training window length and the number of most important business features; and use the data of target business features and date features as the training data set.

[0200] In one embodiment, by acquiring all business features related to the customer complaint rate, the obtained business features and the corresponding customer complaint rate values ​​are integrated into an integrated table, and according to the historical values ​​of each business feature and the historical values ​​of the customer complaint rate, the first correlation coefficient value between the business features and the second correlation coefficient value between each business feature and the customer complaint rate are determined by calculating the strategy data through the Pearson formula; according to each first correlation coefficient value and each second correlation coefficient value, the target business feature related to the customer complaint rate is determined from the integrated table to obtain an updated integrated table; and the updated integrated table is copied and merged according to the number of copies to obtain a data set; and the importance of each target business feature is determined through the XGBoost model to obtain a corresponding importance ranking table.

[0201] Construct the feature engineering of each customer complaint rate prediction model in turn, confirm the date features of each customer complaint rate prediction model, that is, set the holiday list parameters in the Prophet model; add date characteristics to the XGBoost and ANN models, such as year, month, date number, weekday number, whether it is a weekend, and whether it is a holiday. By using the data of each target business feature from T to T-30 days as the training set of Prophet, train the Prophet model to obtain a Prophet model for predicting the target business feature value, and use the Prophet model to predict the predicted value of each target business feature from T to T-30 days as the first backtest set, and use the Prophet model to predict the predicted value of each target business feature from T to T+7 days to obtain the prediction set data. According to the first backtest set, determine the first backtest error value of each customer complaint rate prediction model during the period from T-30 to T; determine the optimal training window length from the prediction training window set according to the first backtest error value; obtain a preset important business feature number set; according to the preset important business feature number and importance ranking table in the preset important business feature number set, determine the preset important business feature number from the first backtest set in order from high to low importance, and obtain a second backtest set; according to each second backtest set, determine the customer complaint rate prediction model The second backtest error value during the period from T-30 to T; determining the number of the most important business features from a preset set of important business features according to each second backtest error value; determining the candidate hyperparameters of each customer complaint rate prediction model; determining the minimum backtest error value of each customer complaint rate prediction model during the period from T-30 to T according to the third backtest set determined by the optimal training window length, the number of the most important business features, and the number of the most important business features in the first backtest set; determining the optimal parameter combination of each customer complaint rate prediction model based on the minimum backtest error value.

[0202] According to the optimal training window length and the number of the most important business features, the target business features and date features are determined from the data set; the target business features and date features are used as the training data set. The business features to be predicted are obtained from the prediction set data, and the business features to be predicted are respectively input into at least two trained different customer complaint rate prediction models to obtain the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model; the corresponding prediction value weight is determined according to the weight proportion value when each customer complaint rate prediction model is fused; weighted calculation and fusion are performed according to each prediction weight and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value of the prediction period.

[0203] For the specific limitations of the customer complaint rate prediction device based on multi-model fusion, please refer to the limitations of the customer complaint rate prediction method based on multi-model fusion mentioned above, which will not be repeated here. Each module in the above-mentioned customer complaint rate prediction device based on multi-model fusion can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0204] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.11 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a customer complaint rate prediction method based on multi-model fusion is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse.

[0205] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0206] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0207] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0208] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0209] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0210] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A customer complaint rate prediction method based on multi-model fusion, characterized in that: The method comprises: Acquire the business characteristics to be predicted, wherein the business characteristics to be predicted are determined according to the correlation between each business characteristic and the customer complaint rate, the business characteristics represent the quantity characteristics, resource characteristics and index characteristics related to the customer complaint rate, and the quantity characteristics include the delivery quantity, the number of unloaded and uninvoiced items, the total number of vehicles and the timeliness achievement rate; Inputting the service features to be predicted into at least two trained different customer complaint rate prediction models respectively, and obtaining the daily customer complaint rate prediction value of the prediction period output by each customer complaint rate prediction model, wherein each customer complaint rate prediction model is obtained by online training of each customer complaint rate prediction model on the basis of determining the optimal training window length, the number of the most important service features and the combination of hyperparameters; Determine the corresponding prediction value weight according to the weight ratio value when each of the customer complaint rate prediction models is fused; A weighted calculation and fusion are performed according to the weights of each of the predicted values ​​and the corresponding daily customer complaint rate predicted value to obtain a target daily customer complaint rate predicted value for the predicted period.

2. The method according to claim 1, characterized in that: The training method of the customer complaint rate prediction model comprises: Determining a data set for training the customer complaint rate prediction model; Based on the data set, determining the optimal training window length, the number of most important business features and the hyperparameter combination of each of the customer complaint rate prediction models; Construct a training dataset for model training; Based on the optimal training window length, the number of most important business features and the hyperparameter combination, each of the customer complaint rate prediction models is trained using the training data set to obtain a trained customer complaint rate prediction model.

3. The method according to claim 2, characterized in that The step of determining a data set for training the customer complaint rate prediction model comprises: Obtaining a comprehensive table of business feature sets related to customer complaint rate prediction values ​​and corresponding customer complaint rates; Determining, according to the historical values ​​of each of the service features and the historical values ​​of the customer complaint rate, a first correlation coefficient value between the service features and a second correlation coefficient value between each of the service features and the customer complaint rate; Determining target business features related to the customer complaint rate from the integrated table according to each of the first correlation coefficient values ​​and each of the second correlation coefficient values, and obtaining an updated integrated table; and The update integration table is copied and merged according to the number of copies to obtain a data set.

4. The method according to claim 3, characterized in that The method further comprises: Obtaining the importance of each target business feature in the update integration table; The target service features are sorted in descending order of importance to obtain a corresponding importance sorting table.

5. The method according to claim 3, characterized in that: Determining the optimal training window length of each customer complaint rate prediction model includes: Get the preset training window length set; Predicting the predicted value of each target business feature in the first period by using a time series model to obtain a first backtest set; Determining, according to the first backtest set, a first backtest error value of each of the customer complaint rate prediction models in the first period; An optimal training window length is determined from the prediction training window set according to the first backtest error value.

6. The method according to claim 4, characterized in that Determining the number of most important business features of each customer complaint rate prediction model includes: Obtain a set of preset important business features; According to the number of each preset important business feature in the preset important business feature number set and the importance ranking table, determining each preset number of important business features from the first backtest set in descending order of importance to obtain a second backtest set, wherein the first backtest set is obtained by predicting the predicted value of each target business feature in the first period through a time series model; Determining, according to each of the second backtest sets, a second backtest error value of each of the customer complaint rate prediction models in the first period; The number of most important business features is determined from the preset number of important business features according to each of the second backtest error values.

7. The method according to claim 3, characterized in that Determining the hyperparameter combination of each customer complaint rate prediction model includes: Determine candidate hyperparameters for each customer complaint rate prediction model; Determine the minimum backtest error value of each of the customer complaint rate prediction models in the first period according to the optimal training window length, the number of the most important business features, and the third backtest set determined by the number of the most important business features in the first backtest set, wherein the first backtest set is obtained by predicting the predicted value of each of the target business features in the first period through a time series model; Based on the minimum backtest error value, the optimal parameter combination of each of the customer complaint rate prediction models is determined.

8. The method according to claim 2, characterized in that: The training data set for constructing the model training includes: Determine the target business feature and date feature data from the data set according to the optimal training window length and the number of the most important business features; The data of the target business characteristics and date characteristics are used as a training data set.

9. A customer complaint rate prediction device based on multi-model fusion, characterized in that: The device comprises: An acquisition module is used to acquire the business characteristics to be predicted, wherein the business characteristics to be predicted are determined according to the correlation between each business characteristic and the customer complaint rate, and the business characteristics represent the quantity characteristics, resource characteristics and index characteristics related to the customer complaint rate, and the quantity characteristics include the number of dispatched items, the number of unloaded items and uninvoiced items, the total number of vehicles and the timeliness achievement rate; A prediction module, used for inputting the service features to be predicted into at least two trained customer complaint rate prediction models respectively, and obtaining the daily customer complaint rate prediction value of the prediction period output by each of the customer complaint rate prediction models, wherein each of the customer complaint rate prediction models is obtained by online training of each customer complaint rate prediction model by obtaining the same training data set from different partitions of the database on the basis of determining the optimal training window length, the number of the most important service features and the combination of hyperparameters; A determination module, used to determine the corresponding prediction value weight according to the weight proportion value and the error weight value when each of the customer complaint rate prediction models is fused; The fusion module is used to perform weighted calculation and fusion according to the weights of each of the predicted values ​​and the corresponding daily customer complaint rate prediction value to obtain the target daily customer complaint rate prediction value for the predicted period.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images