A machine learning based business allocation method and related apparatus

Abstract

The application provides a kind of based on machine learning's business allocation method and related device, the present application relates to computer technology field, method includes: obtaining multi-source historical data, and the multi-source historical data is extracted to feature, and initial feature data is obtained;Through random forest model and Lasso regression model, the initial feature data is screened to feature, and target feature data is obtained;Based on the target feature data and the target feature data corresponding traffic data, training set and verification set are constructed, and the training set is based on preset prediction model training, and the preset prediction model is verified after training through the verification set, and target prediction model is obtained;Multi-source real-time data is obtained, and the real-time feature corresponding to the target feature data in multi-source real-time data is obtained;The real-time feature is input into the target prediction model, and prediction result is obtained, and according to the prediction result, business demand prediction is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of business allocation, and in particular to a business allocation method based on machine learning and related devices. BACKGROUND

[0002] In the prior art, mainly based on machine learning data analysis, by collecting and analyzing business data information, to assist the installation personnel to prepare for the on-site service. However, these schemes still have some deficiencies in practical application. Specifically, the existing implementation scheme mostly adopts traditional data analysis method, such as statistical analysis and regression analysis, etc. These methods can predict business demand to a certain extent, but the accuracy and real-time performance still need to be improved. In addition, these schemes often have high weight on the terminal analysis required by the existing task, and rarely can analyze the non-scheduled business or customer fault replacement demand that may occur. For the current high mobility of installation personnel, the production work still depends on personal experience to a large extent, so the guidance provided by the prior art is not high, which leads to the problem that installation personnel still face the problem of less taking and back and forth in the on-site service process. At the same time, the average allocation method is used for business allocation, which does not consider the individual differences of business personnel, so that the allocation result is unreasonable. SUMMARY

[0003] In order to solve the problems in the prior art, the present application provides a business allocation method based on machine learning to solve the problem of inaccurate business demand prediction in the prior art, which leads to insufficient preparation of business personnel and unreasonable business allocation.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is as follows.

[0005] In a first aspect, the present application provides a business allocation method based on machine learning, comprising the following steps:

[0006] Obtain multi-source historical data, and perform feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical business data, historical marketing activity data, historical weather information and historical guidance information;

[0007] Perform feature screening on the initial feature data through a random forest model and a Lasso regression model to obtain target feature data, and record the type information of the target feature data;

[0008] construct a training set and a verification set based on the target feature data and the traffic volume data corresponding to the target feature data, train a preset prediction model based on the training set, verify the trained preset prediction model through the verification set, and obtain a target prediction model, the preset prediction model including a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average sub-model, and a long short-term memory network;

[0009] According to the multi-source historical data, multi-source real-time data, and the type information, a prediction result of business demand is obtained through the target prediction model, and business allocation is completed according to the prediction result.

[0010] Optionally, the step of extracting features from the multi-source historical data to obtain initial feature data includes:

[0011] When the multi-source historical data is historical guidance information, guidance text data corresponding to the historical guidance information is converted into structured data through natural language conversion technology.

[0012] The structured data is extracted for features by using TF-IDF and / or word embedding technology to obtain initial feature data.

[0013] Optionally, the step of obtaining a prediction result of business demand through the target prediction model according to the multi-source historical data, multi-source real-time data, and the type information includes:

[0014] Multi-source real-time data is obtained, and the multi-source real-time data and the multi-source historical data are integrated to obtain to-be-predicted data, real-time features of the to-be-predicted data are extracted according to the type information, and the multi-source real-time data at least includes real-time business data, real-time marketing activity data, real-time weather information, and real-time guidance information.

[0015] The real-time features are input into the target prediction model to obtain a prediction result, and business demand prediction is completed according to the prediction result.

[0016] Optionally, the step of obtaining a prediction result of business demand through the target prediction model according to the multi-source historical data, multi-source real-time data, and the type information includes:

[0017] The real-time features are input into the gradient boosting decision tree of the target prediction model to obtain a feature combination.

[0018] The real-time features are input into the deep neural network of the target prediction model to obtain deep feature data.

[0019] The feature combination and the deep feature data are fused to obtain fused feature data.

[0020] inputting the fusion feature data into an autoregressive integrated moving average sub-model of the target prediction model to obtain first time sequence features;

[0021] inputting the fusion feature data into a long short-term memory network of the target prediction model to obtain second time sequence features;

[0022] merging the feature combination, the deep feature data, the first time sequence features and the second time sequence features by a weighted average method to obtain a prediction result.

[0023] Optionally, the step of performing feature screening on the initial feature data by a random forest model and a Lasso regression model to obtain target feature data comprises:

[0024] inputting the initial feature data into the random forest model to obtain an importance score of each initial feature data;

[0025] performing feature screening on the initial feature data according to the importance score and a preset importance threshold to obtain first feature data;

[0026] inputting the initial feature data into the Lasso regression model to obtain a coefficient of each initial feature data, and taking initial feature data with a non-zero coefficient as second feature data;

[0027] obtaining target feature data based on the first feature data and the second feature data.

[0028] Optionally, the step of completing business allocation according to the prediction result comprises:

[0029] obtaining historical image data of a business personnel performing a business, and identifying historical action information of the business personnel and target devices corresponding to the action information in the historical image data by image recognition technology;

[0030] based on the target devices, searching for standard action information corresponding to the target devices in a preset standard action information library;

[0031] determining an action score of the historical action information by assigning a similarity between the standard action information and the historical action information;

[0032] if the action score is less than a threshold, generating alarm information according to identity information of the business personnel, time information of collecting the historical image data and identification information of the target devices, and sending the alarm information to a preset client;

[0033] If the action score is greater than or equal to the threshold value, a comprehensive score of the service personnel is calculated by a weight method according to the action score, a number of complaints of the service personnel in a preset time period and efficiency data of the service personnel, and a predicted workload of the service personnel is determined based on the comprehensive score and the prediction result.

[0034] Optionally, the step of acquiring the historical image data of the service personnel performing the service comprises:

[0035] acquiring real-time image data of the service personnel performing the service, and determining whether a preset device exists in the real-time image data by image recognition technology;

[0036] when the preset device exists in the real-time image data, calculating area ratio data of the preset device in the real-time image data;

[0037] if the area ratio data is greater than or equal to a preset ratio value, a limb image of the service personnel and a solid image of the preset device are divided in the real-time image data, and relative position information of the limb image and the solid image is recorded;

[0038] the limb image and the solid image are arranged in a preset background image according to the relative position information to obtain a composite image, and the composite image is stored to obtain the historical image data of the service personnel performing the service;

[0039] when the preset device exists in the real-time image data and / or the area ratio data is less than the preset ratio value, the real-time image data is deleted.

[0040] A service allocation device based on machine learning, characterized in that the device comprises:

[0041] a feature extraction module configured to acquire multi-source historical data, and perform feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical service data, historical marketing activity data, historical weather information and historical guidance information;

[0042] a feature screening module configured to perform feature screening on the initial feature data by a random forest model and a Lasso regression model to obtain target feature data, and record type information of the target feature data;

[0043] a model training module, configured to construct a training set and a verification set based on the target feature data and the traffic volume data corresponding to the target feature data, train a preset prediction model based on the training set, verify the trained preset prediction model through the verification set, and obtain a target prediction model, the preset prediction model including a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average submodel, and a long short-term memory network;

[0044] a business prediction and distribution module, configured to obtain a prediction result of a business demand through the target prediction model according to the multi-source historical data, the multi-source real-time data, and the type information, and complete business distribution according to the prediction result.

[0045] In a third aspect, the present application also provides an electronic device, including a memory, a processor, and a first computer program stored in the memory and executable on the processor, and the processor executes the first computer program to implement the machine learning-based business distribution method as described above.

[0046] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a second computer program, and the second computer program is executed by a processor to implement the machine learning-based business distribution method as described above.

[0047] Advantages:

[0048] The application obtains multi-source historical data, and performs feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical business data, historical marketing activity data, historical weather information and historical guidance information; the initial feature data is subjected to feature screening through a random forest model and a Lasso regression model to obtain target feature data, and type information of the target feature data is recorded; a training set and a verification set are constructed based on the target feature data and business volume data corresponding to the target feature data, a preset prediction model is trained based on the training set, the trained preset prediction model is verified through the verification set, a target prediction model is obtained, and the preset prediction model includes a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average submodel and a long short-term memory network; a prediction result of business demand is obtained through the target prediction model according to the multi-source historical data, multi-source real-time data and the type information, and business allocation is completed according to the prediction result. The historical data (previous year business data, marketing activities, guidance information, etc.) is used for trend reasoning learning, the current business required terminal can be predicted in real time according to the current business environment and the change of customer demand, the accuracy and real-time performance are improved, the preparation work efficiency of the installation personnel for on-site service is significantly improved, the problems of wrong taking and missing taking and time-consuming back and forth are reduced, and customer perception and satisfaction are further improved. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a running flowchart of a business allocation method based on machine learning in an embodiment of the application;

[0050] Figure 2 is a structural schematic diagram of a business allocation system based on machine learning in an embodiment of the application;

[0051] Figure 3 is a structural schematic diagram of an electronic device provided in an embodiment of the application;

[0052] Figure 4 is a structural schematic diagram of a computer readable storage medium provided in an embodiment of the application. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions and advantages of the application clearer, the technical solutions of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0054] The embodiments described in the application are only embodiments of part of the application, not all embodiments. All other embodiments obtained by those skilled in the art without creative labor based on the spirit of the application belong to the protection scope of the application.

[0055] Referring to Figure 1 As shown in the method for allocating a business based on machine learning provided by the present application, the method comprises the following steps:

[0056] S110, acquiring multi-source historical data, and performing feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical business data, historical marketing activity data, historical weather information, and historical guidance information;

[0057] In a possible implementation, the step of performing feature extraction on the multi-source historical data to obtain initial feature data comprises:

[0058] When the multi-source historical data is historical guidance information, a guidance text data corresponding to the historical guidance information is converted into structured data through a natural language conversion technology;

[0059] The structured data is subjected to feature extraction through a TF-IDF and / or word embedding technology to obtain initial feature data.

[0060] For example, data is collected and integrated from multiple sources such as a past business database, a marketing activity database, a guidance announcement platform, and a real-time business volume monitoring system. An ETL (Extract, Transform, Load) process is used to clean, format, and handle missing values of the collected data, so as to ensure the quality and consistency of the input data.

[0061] For example, for a guidance announcement (historical guidance information), a natural language processing (NLP) technology is used to extract key information such as a guidance name, an implementation time, and an influence range, and the key information is converted into structured features. Guidance keywords are extracted from the structured features through a TF-IDF, word embedding, or other technology, and a guidance influence feature (initial feature data) is constructed.

[0062] Exemplarily, key information such as guidance name, implementation time, impact range, guidance content (such as subsidy ratio, technical standard, etc.) is extracted from the guidance announcement (historical guidance information), and is converted into structured features. For example, for the guidance text “From January 1, 2024, the fiber-to-the-home project will be promoted nationwide, and the fiber access cost of each household can enjoy a 30% subsidy”, the following structured information is extracted through natural language processing (NLP) technology: guidance name: fiber-to-the-home guidance; implementation time: January 1, 2024; impact range: nationwide; guidance content: 30% fiber access cost subsidy; the influence degree of the guidance on the business demand is quantified by analyzing the change of the business volume before and after the implementation of the guidance. For example, it is found through analysis of historical data that the demand for fiber terminal equipment increases by 40% after the implementation of the guidance, and the weight of the “guidance influence degree” feature is set to 40%. These guidance-related features (such as guidance implementation time, impact range, guidance influence degree, etc.) are included in the feature engineering of the model, so that the model can consider the influence of the guidance information on the business demand, thereby improving the accuracy and timeliness of the prediction.

[0063] S120, performing feature screening on the initial feature data through a random forest model and a Lasso regression model to obtain target feature data, and recording type information of the target feature data;

[0064] In a possible implementation, the step of performing feature screening on the initial feature data through a random forest model and a Lasso regression model to obtain target feature data includes:

[0065] inputting the initial feature data into the random forest model to obtain an importance score of each initial feature data;

[0066] performing feature screening on the initial feature data according to the importance score and a preset importance threshold to obtain first feature data;

[0067] inputting the initial feature data into the Lasso regression model to obtain a coefficient of each initial feature data, and taking the initial feature data with a non-zero coefficient as second feature data;

[0068] obtaining target feature data based on the first feature data and the second feature data.

[0069] Exemplarily, random forest is an ensemble learning algorithm that improves the accuracy and stability of prediction by constructing multiple decision trees. In the random forest, the importance of each feature can be evaluated by observing the average impurity reduction (such as Gini index) of the feature in all decision trees. The higher the feature importance, the greater the contribution of the feature to the prediction ability of the model.

[0070] The specific steps of the random forest algorithm include:

[0071] Data preparation: divide the initial feature data into training set and validation set; model training: use the training set to train the random forest model. Determine the parameters of the random forest model, such as the number of trees (n_estimators), maximum depth (max_depth), etc.; feature importance calculation: after training, the random forest model will output the importance score of each feature, which can be obtained through the model's feature_importances_ attribute; feature selection: according to the feature importance score, select the features with scores higher than a certain threshold (such as the average importance score) as important features. For example, if the importance score of a feature is lower than 50% of the average value, it can be considered to be removed; validation: use the validation set to evaluate the impact of the filtered feature set on the model performance, to ensure that the filtered feature set can effectively improve the prediction accuracy and generalization ability of the model.

[0072] For example, in the random forest feature selection, the following specific data is obtained: historical traffic trend (monthly growth rate): 0.25; historical traffic trend (quarterly fluctuation rate): 0.15; marketing activity influence: 0.20; guidance change influence degree: 0.10; holiday effect: 0.30.

[0073] Assuming the average importance score is 0.18, select the features with importance scores greater than 0.18, i.e. historical traffic trend (monthly growth rate), marketing activity influence and holiday effect as important features.

[0074] For example, Lasso regression model: Lasso regression is a linear regression model that implements feature selection by adding an L1 regularization term (i.e. the sum of the absolute values of feature coefficients) to the loss function. L1 regularization will make some feature coefficients become zero, thus achieving feature sparsity. The features with non-zero coefficients are important features.

[0075] The specific steps of the Lasso regression model include:

[0076] Data standardization: Since Lasso regression is sensitive to the scale of features, it is necessary to standardize the data before training the model, making the mean of each feature 0 and the standard deviation 1; Model training: Use the standardized data to train the Lasso regression model. Set a suitable regularization parameter λ (alpha), which controls the strength of regularization. A larger λ value will make more feature coefficients become zero; Feature coefficient calculation: After training, view the coefficients of each feature. Features with zero coefficients can be excluded, and features with non-zero coefficients can be retained as important features; Parameter optimization: Select the optimal regularization parameter λ through cross-validation (such as K-fold cross-validation). Select the λ value that makes the model perform best on the validation set; Verification: Use the validation set to evaluate the impact of the filtered feature set on the model performance, ensuring that the filtered feature set can effectively improve the prediction accuracy and generalization ability of the model.

[0077] For example, filter out key features such as historical traffic trend (such as monthly growth rate, quarterly volatility), marketing activity influence (such as the percentage of traffic growth during the activity), guidance change influence (such as the percentage of traffic change before and after guidance implementation), holiday effect (such as the increase multiple of traffic during holidays), etc. Analyze the changes in traffic before and after guidance implementation through historical data, and determine an influence threshold through analysis, for example, when the guidance influence reaches or exceeds 20%, it is considered that the guidance has a significant impact on business demand and needs to be considered in the model. For example, after the implementation of a certain fiber-to-the-home guidance, the demand for fiber terminal equipment increased by 40%, exceeding the 20% influence threshold, so the guidance feature will be given a higher weight to more accurately predict changes in business demand. Similarly, after the implementation of the rural broadband popularization guidance, the demand for broadband terminal equipment increased by 30%, also exceeding the 20% influence threshold, indicating that the guidance has a significant impact on business demand and needs to be considered in the model.

[0078] For example, after training the Lasso regression model, the following feature coefficients are obtained: historical traffic trend (monthly growth rate): 0.5; historical traffic trend (quarterly volatility): 0.0; marketing activity influence: 0.3; guidance change influence: 0.0; holiday effect: 0.4.

[0079] Retain features with non-zero coefficients, i.e. historical traffic trend (monthly growth rate), marketing activity influence, and holiday effect as important features.

[0080] By both the feature importance ranking of random forest and the coefficient sparsity of Lasso regression, the feature variables that are crucial to the prediction model can be effectively screened. The random forest method selects important features by evaluating the contribution of features to the predictive ability of the model, while the Lasso regression method achieves feature sparsity through regularization techniques to eliminate unimportant features. Combining the two methods can improve the accuracy and reliability of feature selection and provide a high-quality feature set for subsequent model training.

[0081] In S130, a training set and a validation set are constructed based on the target feature data and the traffic volume data corresponding to the target feature data, a preset prediction model is trained based on the training set, the trained preset prediction model is verified through the validation set, and a target prediction model is obtained. The preset prediction model includes a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average sub-model, and a long short-term memory network.

[0082] For example, an integrated learning method (such as gradient boosting decision tree GBDT and deep neural network DNN) is used in combination with time series analysis technology to construct the model.

[0083] Specifically, the integrated learning part includes:

[0084] Gradient Boosting Decision Tree (GBDT): GBDT is an integrated learning method based on the gradient boosting framework, which minimizes the loss function by gradually adding weak learners (usually decision trees). Each new weak learner tries to correct the residual error of the previous model, gradually improving the predictive ability of the model. GBDT can effectively capture the non-linear relationships and interaction effects between features in the data, and is suitable for processing complex business data.

[0085] Deep Neural Network (DNN): DNN is a multi-layer neural network that extracts deep features of data through multiple layers of nonlinear transformations. Each layer of neurons introduces nonlinearity through an activation function, allowing the model to learn complex patterns in the data. DNN can handle high-dimensional data and extract deep features from data, making it suitable for processing large-scale business data.

[0086] Specifically, the time series analysis part includes:

[0087] Autoregressive Integrated Moving Average (ARIMA): ARIMA is a classic linear time series model that captures the temporal patterns of time series data through autoregression (AR), differencing (I), and moving average (MA) components. Model parameters include the number of autoregressive terms (p), the differencing order (d), and the number of moving average terms (q). ARIMA can effectively capture linear trends and seasonal changes in time series data, making it suitable for handling business data with clear temporal patterns.

[0088] Long Short-Term Memory (LSTM): LSTM is a special type of recurrent neural network (RNN) that addresses the gradient vanishing problem of traditional RNNs by introducing gating mechanisms (input gate, forget gate, and output gate). LSTM can effectively capture long-term dependencies in time series data, making it suitable for handling complex business data temporal changes.

[0089] Example: Combine deep neural networks (DNN) and gradient boosting decision trees (GBDT) to build a hybrid model.

[0090] Nonlinear fitting: Use DNN to capture nonlinear relationships in data and GBDT to handle feature interactions.

[0091] Validation process: Use grid search, random search, and other methods to find the optimal model parameters.

[0092] Time series analysis: Combine ARIMA, LSTM, and other techniques to integrate current business data and monitoring data in real time.

[0093] Sliding window method: Window historical data to adapt to different time scales of business demand changes.

[0094] Use the LSTM model to predict future business volume for the next week, and adjust the prediction granularity using the sliding window method.

[0095] Visualization tools: Develop interactive visualization interfaces to display prediction results and carry suggestions.

[0096] Personalized recommendations: Based on prediction results and engineer experience preferences, generate customized carry suggestions.

[0097] Validation process: Evaluate the effectiveness of recommendations through user satisfaction surveys, business indicator improvements, and other methods.

[0098] Feedback mechanism: Establish an installer feedback system to collect feedback data from actual applications.

[0099] Online learning: Use online learning techniques (such as incremental learning, transfer learning) to dynamically update the model.

[0100] Verification process: Continuously monitor model performance, evaluate optimization effectiveness, and ensure the model adapts to changing business needs.

[0101] Example of ensemble learning model training:

[0102] Gradient Boosting Decision Tree (GBDT):

[0103] Parameter settings: n_estimators: number of weak learners, e.g., 100. learning_rate: learning rate, e.g., 0.1. max_depth: maximum depth of decision tree, e.g., 3.

[0104] Training process: Train GBDT model using training set data. Minimize loss function by adding weak learners one by one, each new weak learner tries to correct the residual error of the previous model. Use validation set to evaluate model performance, adjust parameters to optimize model performance.

[0105] Deep Neural Network (DNN):

[0106] Input layer: number of features.

[0107] Hidden layers: e.g., 3 layers, each with 128 neurons.

[0108] Activation function: ReLU.

[0109] Output layer: 1 neuron, predicts business volume.

[0110] Train DNN model using training set data. Choose appropriate loss function (e.g., Mean Squared Error, MSE) and optimization algorithm (e.g., Adam). Update network weights through backpropagation, use early stopping mechanism during training to prevent overfitting. Use validation set to evaluate model performance, adjust learning rate, hidden layer size, and other hyperparameters.

[0111] Autoregressive Integrated Moving Average Model (ARIMA):

[0112] Parameter selection: Use ACF (Autocorrelation Function) and PACF (Partial Autocorrelation Function) plots to determine ARIMA model parameters (p, d, q). For example, choose p=2, d=1, q=1.

[0113] Training process:

[0114] Fit ARIMA model using training set data. Estimate model parameters using maximum likelihood estimation method. Use validation set to evaluate model performance, adjust parameters to optimize model performance.

[0115] Long Short-Term Memory Network (LSTM):

[0116] Input layer: number of features in time series.

[0117] LSTM layer: e.g., 1 layer, 128 neurons.

[0118] Output layer: 1 neuron, predicting traffic volume.

[0119] Training process:

[0120] Train the LSTM model using the training set data. Choose an appropriate loss function (e.g., mean squared error, MSE) and optimization algorithm (e.g., Adam). Update network weights through backpropagation, and use early stopping to prevent overfitting during training. Evaluate model performance using the validation set, and adjust learning rate, LSTM layer size, and other hyperparameters.

[0121] GBDT is used to capture the non-linear relationship and interaction effect between features, and DNN is used to process high-dimensional data and extract deep features. Combine time series analysis (e.g., ARIMA, LSTM) to capture the time series regularity of business data.

[0122] Use cross-validation (e.g., K-fold cross-validation) to evaluate model performance and select the optimal combination of hyperparameters.

[0123] Use sliding window methods (e.g., fixed window, adaptive window) to process historical data to adapt to changes in business demand at different time scales. Use online learning techniques (e.g., incremental learning, transfer learning) to dynamically update the model.

[0124] S140, according to the multi-source historical data, multi-source real-time data and the type information, through the target prediction model, get the prediction result of business demand, and complete the business allocation according to the prediction result.

[0125] In one possible implementation, the step of inputting multi-source real-time data into the target prediction model to obtain the prediction result of business demand includes:

[0126] Obtain multi-source real-time data, integrate the multi-source real-time data and the multi-source historical data to obtain predicted data, extract real-time features of the predicted data according to the type information, and the multi-source real-time data at least includes real-time business data, real-time marketing activity data, real-time weather information and real-time guidance information;

[0127] Input the real-time features into the target prediction model to obtain the prediction result, and complete the prediction of business demand according to the prediction result.

[0128] Exemplarily, the multi-source real-time data includes, but is not limited to, current traffic data: current traffic data obtained from a real-time traffic monitoring system, including but not limited to current order quantity, traffic type distribution, processed traffic volume, etc. End-to-end monitoring data: real-time network status data obtained from a network monitoring system, including network delay, packet loss rate, bandwidth utilization, etc. These data can help determine the impact of the current network environment on the business. External data: such as weather conditions, holiday information, guidance changes, etc. These data may affect traffic volume and customer behavior.

[0129] The real-time data is cleaned and formatted to ensure data integrity and consistency.

[0130] The multi-source real-time data and the multi-source historical data are integrated to obtain the to-be-predicted data, and real-time features of the to-be-predicted data are extracted according to the type information to form a complete input feature set.

[0131] Exemplarily, the multi-source real-time data includes the latest published guidance announcement, but in actual application, the latest published guidance announcement often cannot directly represent the specific business guidance of the business implementation unit (for example, an operator). The specific business guidance of the business implementation unit (for example, an operator) has a lag compared to the latest published guidance announcement, that is, when the specific business guidance of the business implementation unit (for example, an operator) needs to be formulated according to the latest published guidance announcement. Therefore, the present scheme is understood as providing data support for formulating the specific business guidance after the latest published guidance announcement, and can also be understood as predicting business demand under ideal conditions (the specific business guidance and the latest published guidance announcement are published at the same time).

[0132] In one possible implementation, the step of inputting the real-time features into the target prediction model to obtain a prediction result includes:

[0133] Inputting the real-time features into a gradient boosting decision tree of the target prediction model to obtain a feature combination;

[0134] Inputting the real-time features into a deep neural network of the target prediction model to obtain deep feature data;

[0135] Fusing the feature combination and the deep feature data to obtain fused feature data;

[0136] Inputting the fused feature data into an autoregressive integrated moving average sub-model of the target prediction model to obtain a first time series feature;

[0137] inputting the fusion feature data into a long short-term memory network of the target prediction model to obtain second time sequence features;

[0138] merging the feature combination, the deep feature data, the first time sequence features and the second time sequence features by a weighted average method to obtain a prediction result.

[0139] Exemplarily, the extracted features (the real-time features) are input into a GBDT model, and the important feature combination is extracted by using the nonlinear fitting capability of the GBDT for preliminary processing of the features. Meanwhile, the extracted features (the real-time features) are input into a DNN model, and the deep feature data is obtained by using the deep feature extraction capability of the DNN to further mine the complex patterns in the data. The feature combination and the deep feature data are fused to obtain fusion feature data, and the data after the feature fusion is input into an ARIMA model, and the linear trend and seasonal variation of the time sequence are captured by using the ARIMA model. Meanwhile, the data after the feature fusion is input into an LSTM model, and the long-term dependence relationship of the time sequence is captured by using the LSTM model. The output results of the GBDT, the DNN, the ARIMA and the LSTM models are fused, and the final prediction result is generated by using a weighted average or a voting mechanism.

[0140] In a possible implementation, the step of completing business allocation according to the prediction result comprises:

[0141] obtaining historical image data of a business personnel performing a business, and identifying historical action information of the business personnel and a target device corresponding to the action information in the historical image data by using an image recognition technology;

[0142] based on the target device, searching for standard action information corresponding to the target device in a preset standard action information library;

[0143] determining an action score of the historical action information by using a similarity between the standard action information and the historical action information and a value assignment method;

[0144] if the action score is less than a threshold value, generating alarm information according to identity information of the business personnel, time information of collecting the historical image data and identification information of the target device, and sending the alarm information to a preset client;

[0145] if the action score is greater than or equal to the threshold value, calculating a comprehensive score of the business personnel by using a weight method according to the action score, a number of complaints of the business personnel in a preset time period and efficiency data of the business personnel, and determining a predicted workload of the business personnel based on the comprehensive score and the prediction result.

[0146] For example, in order to allocate the service more reasonably, the work capacity (action score) of the service personnel, the work efficiency and the service satisfaction (complaint number) of the customers to the service personnel are comprehensively considered to allocate the work to the service personnel, so that the allocation result is more reasonable.

[0147] For example, according to the predicted work load of the service personnel, the starting point of the service personnel and the location information of each user, the work path of the service personnel is planned, for example, the path is planned according to the nearest principle, that is, the user closest to the starting point is selected as the target user according to the starting point of the service personnel and the location of each user, and then the user closest to the target user is selected according to the service demand of the target user, and the path is planned until the service personnel completes the predicted work load.

[0148] For example, after the user closest to the starting point is selected as the target user according to the starting point of the service personnel and the location of each user, the time node of the service personnel to arrive at the target user is predicted based on the path between the starting point and the target user, the average travel speed of the service personnel, the travel mode and the like, and the target user is informed of the time node, if the target user accepts the time node, the service personnel performs the service for the target user; if the target user does not accept the time node, the second closest user to the starting point is selected as the target user according to the starting point of the service personnel and the location of each user.

[0149] For example, after the target user accepts the time node, the time taken by the service personnel to perform the service for the target user is predicted based on the efficiency data of the service personnel, and the location of the target user is taken as the starting point of the service personnel, and then the closest user to the starting point is selected as the new target user according to the starting point of the service personnel and the location of each user except the target user, a new path between the location of the target user and the location of the new target user is obtained, a new time node of the service personnel to arrive at the new target user is predicted, and the new target user is informed of the new time node.

[0150] For example, the actual work load of the service personnel in a preset time period is counted, the actual work load and the multi-source real-time data are used to construct an actual training set, the target prediction model is continuously trained according to the actual training set, and the continuously trained target prediction model is used in the subsequent service allocation process, so that the prediction result is more accurate.

[0151] In a possible implementation, the step of obtaining historical image data of the service personnel performing the service comprises:

[0152] acquire real-time image data when the service personnel performs the service, and determine whether a preset device exists in the real-time image data through image recognition technology;

[0153] when the preset device exists in the real-time image data, calculate area ratio data of the preset device in the real-time image data;

[0154] if the area ratio data is greater than or equal to a preset ratio value, divide a limb image of the service personnel and an entity image of the preset device in the real-time image data, and record relative position information of the limb image and the entity image;

[0155] set the limb image and the entity image in a preset background image according to the relative position information to obtain a composite image, store the composite image, and obtain historical image data when the service personnel performs the service;

[0156] when the preset device exists in the real-time image data, and / or the area ratio data is less than the preset ratio value, delete the real-time image data.

[0157] For example, to ensure customer privacy when entering a private area for services such as home services, only the image of the device by the service personnel (the limb image of the service personnel and the entity image of the preset device) is retained, and the limb image and the entity image are set in a preset background image according to the relative position information to avoid leaking user privacy. When the preset device exists in the real-time image data, and / or the area ratio data is less than the preset ratio value, the real-time image data is deleted to avoid the risk of privacy leakage caused by retaining or uploading images unrelated to the service.

[0158] In a possible implementation, a service allocation device based on machine learning includes:

[0159] The feature extraction module 201 is configured to acquire multi-source historical data, and perform feature extraction on the multi-source historical data to obtain initial feature data. The multi-source historical data at least includes historical service data, historical marketing activity data, historical weather information, and historical guidance information.

[0160] The feature screening module 202 is configured to perform feature screening on the initial feature data through a random forest model and a Lasso regression model to obtain target feature data, and record type information of the target feature data.

[0161] The model training module 203 is configured to construct a training set and a verification set based on the target feature data and the traffic volume data corresponding to the target feature data, train a preset prediction model based on the training set, verify the trained preset prediction model through the verification set, and obtain a target prediction model. The preset prediction model includes a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average submodel, and a long short-term memory network.

[0162] The business prediction and distribution module 204 is configured to obtain a prediction result of a business demand through the target prediction model based on the multi-source historical data, the multi-source real-time data, and the type information, and complete business distribution according to the prediction result.

[0163] In a possible implementation, as shown in Figure 3 The embodiment of the present application provides a terminal device 300, which comprises a memory 310, a processor 320, and a first computer program 311 stored in the memory 310 and capable of running on the processor 320. When the processor 320 executes the first computer program 311, the steps of the method for business distribution based on machine learning are implemented.

[0164] In a possible implementation, as shown in Figure 4 The embodiment of the present application provides a computer readable storage medium 400, which stores a second computer program 411. When the second computer program 411 is executed by a processor, the steps of the method for business distribution based on machine learning are implemented.

[0165] It should be noted that the information interaction and execution process between the above apparatuses / units are based on the same concept as the method embodiments of the present application, and the specific functions and technical effects thereof can be referred to the method embodiments part, which will not be described here.

[0166] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is taken as an example for illustration, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of software functional unit. In addition, the specific names of the functional units and modules are only for the convenience of mutual distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the system can be referred to the corresponding process in the foregoing method embodiments, which will not be described here.

[0167] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the present application can implement all or part of the processes in the above-mentioned embodiment methods through a computer program to instruct related hardware to complete, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium can at least include any entity or device capable of carrying the computer program code to the photographing device / terminal equipment, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium. For example, U disk, mobile hard disk, magnetic disk or optical disk, etc. In some jurisdictions, according to legislation and patent practice, the computer readable medium can not be an electrical carrier signal and a telecommunication signal.

[0168] In the above embodiments, the description of each embodiment has its own focus, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0169] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0170] In the embodiments provided by the present application, it should be understood that the disclosed apparatus / network device and method can be implemented in other ways. For example, the above-described apparatus / network device embodiments are merely schematic, for example, the division of the modules or units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed each other can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0171] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0172] The above embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

[0173] The above disclosure is only the preferred embodiment of the present application, of course, cannot limit the scope of the right of the present application, therefore, the equivalent changes made by the claims of the present application, still belong to the scope covered by the present application.

Claims

1. A method for service allocation based on machine learning, characterized in that, The method comprises the following steps: obtaining multi-source historical data, and performing feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical business data, historical marketing activity data, historical weather information and historical guidance information; performing feature screening on the initial feature data through a random forest model and a Lasso regression model to obtain target feature data, and recording type information of the target feature data; constructing a training set and a validation set based on the target feature data and business volume data corresponding to the target feature data, training a preset prediction model based on the training set, verifying the trained preset prediction model through the validation set, and obtaining a target prediction model, wherein the preset prediction model includes a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average sub-model and a long short-term memory network; obtaining a prediction result of business demand through the target prediction model according to the multi-source historical data, multi-source real-time data and the type information, and completing business allocation according to the prediction result.

2. The method of claim 1, wherein, The step of performing feature extraction on the multi-source historical data to obtain initial feature data comprises: when the multi-source historical data is historical guidance information, converting guidance text data corresponding to the historical guidance information into structured data through natural language conversion technology; performing feature extraction on the structured data through TF-IDF and / or word embedding technology to obtain initial feature data.

3. The method of claim 1, wherein, The step of obtaining a prediction result of business demand through the target prediction model according to the multi-source historical data, multi-source real-time data and the type information comprises: obtaining multi-source real-time data, integrating the multi-source real-time data and the multi-source historical data to obtain to-be-predicted data, extracting real-time features of the to-be-predicted data according to the type information, wherein the multi-source real-time data at least includes real-time business data, real-time marketing activity data, real-time weather information and real-time guidance information; inputting the real-time features into the target prediction model to obtain a prediction result, and completing business demand prediction according to the prediction result.

4. The method of claim 3, wherein, The step of obtaining a prediction result of business demand through the target prediction model according to the multi-source historical data, multi-source real-time data and the type information comprises: inputting the real-time features into the gradient boosting decision tree of the target prediction model to obtain feature combinations; inputting the real-time features into the deep neural network of the target prediction model to obtain deep feature data; fusing the feature combinations and the deep feature data to obtain fused feature data; inputting the fused feature data into the autoregressive integrated moving average sub-model of the target prediction model to obtain first time sequence features; inputting the fused feature data into the long short-term memory network of the target prediction model to obtain second time sequence features; merging the feature combinations, the deep feature data, the first time sequence features and the second time sequence features through a weighted average method to obtain a prediction result.

5. The method of claim 1, wherein, The step of performing feature screening on the initial feature data by the random forest model and the Lasso regression model to obtain target feature data comprises: inputting the initial feature data into the random forest model to obtain an importance score of each initial feature data; performing feature screening on the initial feature data according to the importance score and a preset importance threshold to obtain first feature data; inputting the initial feature data into the Lasso regression model to obtain a coefficient of each initial feature data, and taking the initial feature data with a non-zero coefficient as second feature data; obtaining target feature data based on the first feature data and the second feature data.

6. The method of claim 1, wherein, The step of completing business allocation according to the prediction result comprises: obtaining historical image data of a business personnel performing a business, and identifying historical action information of the business personnel and target devices corresponding to the action information in the historical image data through image recognition technology; based on the target devices, searching for standard action information corresponding to the target devices in a preset standard action information library; calculating the similarity between the standard action information and the historical action information, and determining the action score of the historical action information through the assignment method; if the action score is less than a threshold, generating alarm information according to the identity information of the business personnel, the time information of collecting the historical image data, and the identification information of the target devices, and sending the alarm information to a preset client; if the action score is greater than or equal to the threshold, calculating the comprehensive score of the business personnel through the weight method according to the action score, the number of complaints of the business personnel within a preset time period, and the efficiency data of the business personnel, and determining the predicted workload of the business personnel based on the comprehensive score and the prediction result.

7. The method of claim 6, wherein, The step of obtaining historical image data of a business personnel performing a business comprises: obtaining real-time image data of the business personnel performing the business, and determining whether a preset device exists in the real-time image data through image recognition technology; when the preset device exists in the real-time image data, calculating area proportion data of the preset device in the real-time image data; if the area proportion data is greater than or equal to a preset proportion value, dividing a body image of the business personnel and a entity image of the preset device in the real-time image data, and recording relative position information of the body image and the entity image; placing the body image and the entity image in a preset background image according to the relative position information to obtain a composite image, and storing the composite image to obtain historical image data of the business personnel performing the business; when the preset device exists in the real-time image data, and / or the area proportion data is less than the preset proportion value, deleting the real-time image data.

8. A machine learning based service distribution apparatus, characterized by, The device comprises: The feature extraction module is configured to obtain multi-source historical data, and perform feature extraction on the multi-source historical data to obtain initial feature data, wherein the multi-source historical data at least includes historical business data, historical marketing activity data, historical weather information, and historical guidance information. The feature screening module is configured to perform feature screening on the initial feature data by using a random forest model and a Lasso regression model to obtain target feature data, and record type information of the target feature data. The model training module is configured to construct a training set and a verification set based on the target feature data and business volume data corresponding to the target feature data, train a preset prediction model based on the training set, verify the trained preset prediction model by using the verification set, and obtain a target prediction model, wherein the preset prediction model includes a gradient boosting decision tree, a deep neural network, an autoregressive integrated moving average sub-model, and a long short-term memory network. The business prediction and distribution module is configured to obtain a prediction result of business demand by using the target prediction model based on the multi-source historical data, multi-source real-time data, and the type information, and complete business distribution according to the prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the business distribution method based on machine learning as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the business distribution method based on machine learning as claimed in any one of claims 1 to 7.

Images