Takeaway number identification method, device, electronic device and storage medium

Abstract

The present invention relates to a method, device, electronic device and storage medium for take-out number identification. The method for take-out number identification comprises the following steps: S1, inputting take-out number blacklist and whitelist data and signaling call sheet data, cleaning the data, and obtaining the original data set required by the model through data association fusion; S2, adopting the SMOTE TomeK algorithm to perform comprehensive sampling on the original data set to form a model sample data set; S3, dividing the model sample data set into a training set and a validation set, and then dividing a retention set on the training set, and respectively adopting an XGboost model, an NGboost model and a Catboost model to train the model on the remaining sample training set to form a first-layer basic classifier model; S4, respectively inputting the retention set on the training set and all the validation sets into the first-layer base classifier model to respectively output a training set prediction value set and a validation set prediction value set; S5, merging to form a new validation set prediction value set; S6, constructing a second-layer model based on the LightGBM algorithm to form an XNCLBoost model.

Description

Technical Field

[0001] The present invention relates to network communication technology, and in particular to a method, device, electronic device and storage medium for identifying a takeout number. Background Art

[0002] With the continuous development of the new generation of 5G communication technology, communication users are experiencing the convenience brought by communication technology in life and work; at the same time, harassing calls bring troubles to people, not only disrupting the normal life and work order of communication users, but also greatly endangering the calls of some normal numbers. For example, takeaway calls are identified as harassing calls, which brings harm to the vital interests of the people. How to identify takeaway numbers from the existing network, so as to provide users with accurate identification and prompts of incoming call numbers, has become a technical issue of common concern to communication operators. However, because the problem of takeaway number identification is too complicated, it has not yet been completely solved.

[0003] In real life, both takeaway numbers and fraud numbers have the characteristics of high call frequency, large call volume, and short call time. At the same time, takeaway numbers are strongly correlated with meal time periods.

[0004] The prior art does not involve a solution for efficient identification of takeout numbers. Patent application title: A method for identifying express numbers based on call behavior (CN109274834A), which is based on constructing a blacklist and whitelist call record table and express feature identification rules, and then obtaining the threshold of each communication indicator to identify the express number. This method does not fully consider the content of communication indicator information, and the blacklist and whitelist call record table has certain limitations. Patent application title: A method, device and computer storage medium for identifying express numbers (CN110519466A), in which the XGB model is used to train communication indicator information data for express number identification. Because the XGB model has many hyperparameters, it is easy to fall into a local optimal solution. When processing ultra-high dimensional feature variables, the performance will be reduced.

[0005] Therefore, effectively identifying takeout numbers from signaling call record data with high accuracy and high stability has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a highly accurate and stable automated intelligent takeout number identification method, which can provide correct identification and display of incoming call numbers for the majority of mobile users, so as to improve the connection rate and communication efficiency of takeout numbers and create a good communication network environment.

[0007] In order to solve the above technical problems, according to one aspect of the present invention, a method for identifying takeout numbers is provided, and the method comprises the following steps: S1, inputting takeout number blacklist and whitelist data and signaling call list data, and cleaning the data by data ETL engineering, wherein the data ETL engineering comprises data extraction (Extract), conversion (Transform), and loading (Load), and then obtaining the original data set required by the model through data association fusion; S2, using the SMOTE TomeK (comprehensive sampling) algorithm to perform comprehensive sampling on the original data set to form a model sample data set; S3, dividing the model sample data set into a training set and a validation set according to the training samples in the ratio of x:y, wherein x∈[0,1], y∈[0,1], x+y=1, and then dividing the training set into a% holdout set, and using the XGboost (eXtreme gradient boosting, extreme gradient boosting tree) model, the NGboost (Natural gradient boosting, natural gradient boosting tree) model, and the Catboost (Categorical gradient boosting) model on the 1-a% sample training set. boosting, category gradient boosting tree) model to train the model to form the first-layer basic classifier model; S4, respectively input the a% retention set and the entire validation set on the training set into the XGboost model, NGboost model, and Catboost model of the first-layer basic classifier model, and the XGboost model, NGboost model, and Catboost model respectively output the training set prediction value set and the validation set prediction value set; S5, merge the training set prediction value set output by the first-layer basic classifier model to form a new training set feature value set, and merge the validation set prediction value set output by the first-layer basic classifier model to form a new validation set prediction value set; S6, based on LightGBM (Light GradientBoosting Machine, gradient boosting machine) algorithm to build the second-layer model, input the new training set feature values ​​and the original test set into the LightGBM model, and output a model that meets the pre-set model accuracy after training, thereby forming an XNCLBoost model (the abbreviation of the above four algorithms XGboost, NGboost, Catboost, and LightGBM). Among them, the XNCLBoost model is based on the fusion of Blending strategies. Among them, when the LightGBM model has more hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. Among them, the Bayesian optimization algorithm is improved based on the Gaussian process, and the covariance function in the form of a multi-strategy combination is used to simultaneously capture the smoothness and amplitude of the objective function.

[0008] According to an embodiment of the present invention, in the above a% reserved set, a may be 30. It may also be other values ​​between 1-100, depending on actual needs and is not limited thereto.

[0009] According to an embodiment of the present invention, the method for takeout number identification may further include the following steps: S7, the signaling call record data to be tested is input into the XNCLBoost model, and after prediction by the model, the prediction result is output, and objectionable data is eliminated based on the prediction result.

[0010] Furthermore, the method for takeout number identification may also include the following steps: S8, applying the output model prediction data to the caller name card and code number recognition business scenario to identify and mark the caller number for the user; S9, collecting data and complaint data feedback from the takeout marking platform in the business scenario application, and updating the takeout number blacklist and whitelist data with the takeout platform feedback data and complaint data to form a closed loop of the modeling process.

[0011] According to an embodiment of the present invention, the input takeout blacklist and whitelist data may be the existing number identification data of the platform, and may be optimized and updated by the predicted data calculated by the model and the data verified after the platform application.

[0012] According to an embodiment of the present invention, the input signaling bill data may be the call bill data to be tested (number encrypted), the call bill is based on the call record data set of the corresponding number, and the call record data collection period may be the past month, the past two months, or the past three months. The specific data collection period is determined according to actual needs and is not limited thereto.

[0013] According to an embodiment of the present invention, call record data includes call record data of a calling number and call record data of a called number, and the call record data has characteristic variables, which may include: a thermal value of a call area of ​​a calling number, which is used to record the total number of cell / building locations in a call signaling when the number initiates an active call; an outgoing call frequency of a calling number, which is used to record the total number of calls in a call signaling when the number initiates an active call; a connected call frequency of a calling number, which is used to record the total number of successful connections among the number of active calls in a call signaling when the number initiates an active call; an average ringing duration of a calling number, which is used to record the total number of ringing durations of calls in a call signaling when the number initiates an active call; an average call duration of a calling number, which is used to record the total number of normal call durations of calls in a call signaling when the number initiates an active call; a thermal value of a call area of ​​a called number, which is used to record the total number of cell / building locations in a call signaling when the number accepts a passive call; an incoming call frequency of a called number, which is used to record the total number of outgoing calls in a call signaling when the number accepts a passive call. Specific characteristic variables are not limited to this, for example, they can be: calling contact dispersion (the deduplicated sum of called numbers in call signaling when the number initiates an active call), daily active hours of the calling number (based on call bill signaling data, counting the number of natural hours that the number initiates active calls per day), weekly active days of the calling number (based on call bill signaling data, counting the number of natural days that the number initiates active calls per week), monthly active days of the calling number (based on call bill signaling data, counting the number of days that the number initiates active calls each month), monthly voice call fee consumption of the calling number (based on call bill signaling data, The monthly voice call charges of the statistical number), the monthly traffic expenditure of the calling number (based on the call bill signaling data, the monthly traffic expenditure of the statistical number), etc.; it can also be: the frequency of called number connection (when the number accepts passive calls, the sum of the number of successful connection times in the call signaling), the ringing duration of the called number (when the number accepts passive calls, the sum of the ringing duration in the call signaling), the call duration of the called number (when the number accepts passive calls, the sum of the normal call duration in the call signaling), the dispersion of called contacts (when the number accepts passive calls, the de-duplicated sum of the calling party numbers in the call signaling), etc.

[0014] According to a second aspect of the present invention, a device for identifying takeout numbers is provided, comprising: a model sample generation unit, the model sample generation unit is used to process the input takeout number blacklist and whitelist data and signaling call list data, generate and output a model sample data set, wherein the data processing method includes data ETL engineering cleaning data and SMOTETomeK algorithm for comprehensive sampling of the original data set; a first-layer model, having an XGboost model, an NGboost model, and a Catboost model, the first-layer model divides the model sample data set into a training set and a validation set according to the ratio of x:y, wherein x∈[0,1], y∈[0,1], x+y=1, and then divides 30% of the retention set on the training set, and respectively uses the XGboost model, the NGboost model, and the Catboost model to train the model on 70% of the sample training set to form a first-layer basic classifier model, and then respectively inputs the 30% retention set and the entire validation set on the training set into the XGboost of the first-layer basic classifier model. In the model, NGboost model, and Catboost model, the XGboost model, NGboost model, and Catboost model respectively output the training set prediction value set and the validation set prediction value set, and then merge the training set prediction value set output by the first-layer basic classifier model to form a new training set feature value set, and merge the validation set prediction value set output by the first-layer basic classifier model to form a new validation set prediction value set; the second-layer model, the second-layer model is built based on the LightGBM algorithm, and the second-layer model inputs the new training set feature value and the original test set into the LightGBM model, and after training, outputs a model that meets the pre-set model accuracy, thereby forming an XNCLBoost model. Among them, when the LightGBM model has more hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. Among them, the Bayesian optimization algorithm is improved based on the Gaussian process, and the covariance function in the form of a multi-strategy combination is adopted to simultaneously capture the smoothness and amplitude of the objective function.

[0015] According to a third aspect of the present invention, there is provided an electronic device, comprising: a memory, a processor, and a takeaway number identification program stored in the memory and executable on the processor, wherein the takeaway number identification program implements the steps of the above-mentioned takeaway number identification method when executed by the processor.

[0016] According to a fourth aspect of the present invention, there is provided a computer storage medium, wherein a takeout number recognition program is stored on the computer storage medium, and when the takeout number recognition program is executed by a processor, the steps of the above-mentioned takeout number recognition method are implemented.

[0017] Compared with the prior art, the technical solution provided by the embodiments of the present invention can at least achieve the following beneficial effects:

[0018] 1. In the sampling of sample data of the XNCLBoost model, the SMOTE Tomek algorithm is used to comprehensively sample the original data. The oversampling method is first used, and then the Tomek algorithm is used to delete the points in a stalemate or very close to each other after the sample is expanded, which effectively solves the problem of data imbalance.

[0019] 2. In the XNCLBoost model framework, tree models such as the XGboost model, NGboost model and Catboost model are used as the first-level basic classifier models, while retaining the advantages of each model, effectively overcoming the problems of slow convergence of model parameters and model overfitting.

[0020] 3. In the XNCLBoost model framework, the LightGBM model is used as the second-layer meta-classifier to achieve the fusion of multiple tree models. This model effectively reduces the data dimension, removes redundant data, has good nonlinear mapping capabilities, and has high classification prediction accuracy.

[0021] 4. In the LightGBM model parameter optimization, the covariance function in the Gaussian process is used to simultaneously characterize the smoothness and amplitude of the objective function, forming an improved Gaussian Bayesian optimization algorithm.

[0022] 5. The technology of the present invention constructs a two-layer training model framework (i.e., XNCLBoost model framework) based on the Blending ensemble learning strategy, wherein the first layer model uses the XGboost model, NGboost model, and Catboost model with strong classification performance as basic classifiers, and the second layer model uses the LightGBM model with strong nonlinear performance as a meta-classifier to achieve the fusion of multiple models. The fused XNCLBoost model has better generalization ability, higher prediction accuracy, and lower memory consumption. The number recognition method based on the Blending strategy fusion model is simple and easy, with less feature information leakage and a low probability of overfitting of the model.

[0023] 6. The present invention provides an automated intelligent identification method for takeout numbers, improves the connection rate and communication efficiency of takeout numbers, creates a good communication network environment, provides a high-accuracy and high-stability takeout number identification method, and provides correct identification and display of incoming call numbers for the majority of mobile users. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present invention, but are not intended to limit the present invention.

[0025] Figure 1 is a flow chart showing a method for identifying a takeout number according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] Unless otherwise defined, the technical or scientific terms used herein shall have the common meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "a" do not indicate a quantity limitation, but indicate the existence of at least one.

[0028] Figure 1 is a flow chart showing a method for identifying a takeout number according to an embodiment of the present invention.

[0029] like Figure 1 As shown, the method for identifying a takeaway number includes the following steps:

[0030] S1. Input the takeaway number blacklist and whitelist data and signaling call record data, and clean the data using data ETL engineering. Data ETL engineering includes data extraction, conversion, and loading, and then obtains the original data set required by the model through data association and fusion;

[0031] S2, use the SMOTE TomeK algorithm to perform comprehensive sampling on the original data set to form a model sample data set;

[0032] S3. Use train_test_split() to divide the model sample data set into a training set and a validation set according to the ratio of x:y, where x∈[0,1], y∈[0,1], x+y=1, and then divide the a% hold-out set (Hold-outset) on the training set. Use the XGboost model, NGboost model, and Catboost model to train the model on the 1-a% sample training set to form the first-layer basic classifier model.

[0033] S4, respectively inputting a% of the held-out set on the training set and all the validation sets into the XGboost model, NGboost model, and Catboost model of the first-layer basic classifier model, and the XGboost model, NGboost model, and Catboost model respectively output a set of predicted values ​​for the training set and a set of predicted values ​​for the validation set;

[0034] S5, merging the training set prediction value set output by the first-layer basic classifier model to form a new training set feature value set, and merging the validation set prediction value set output by the first-layer basic classifier model to form a new validation set prediction value set;

[0035] S6. Construct the second layer model based on the LightGBM algorithm, input the new training set feature values ​​and the original test set into the LightGBM model, and output a model that meets the pre-set model accuracy after training, thereby forming an XNCLBoost model. The XNCLBoost model is based on the Blending strategy fusion. When the LightGBM model has many hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. The Bayesian optimization algorithm is improved based on the Gaussian process, and the covariance function in the form of a multi-strategy combination is used to simultaneously capture the smoothness and amplitude of the objective function.

[0036] The technical solution of the present invention adopts the SMOTE TomeK algorithm to comprehensively sample the original data in the sampling of the sample data of the XNCLBoost model. The oversampling method is first adopted, and after the sample is expanded, the Tomek algorithm is used to delete the points in a stalemate or the points that are very close, so as to effectively solve the problem of data imbalance. In the XNCLBoost model framework, tree models such as the XGboost model, the NGboost model and the Catboost model are used as the first-layer basic classifier models, while retaining the advantages of each model, effectively overcoming the problems of slow convergence of model parameters and model overfitting. In the XNCLBoost model framework, the LightGBM model is used as the second-layer meta-classifier to realize the fusion of multiple tree models. This model effectively reduces the data dimension, removes redundant data, has good nonlinear mapping ability, and has high classification prediction accuracy. In the optimization of LightGBM model parameters, the covariance function in the Gaussian process is used to simultaneously characterize the smoothness and amplitude of the objective function to form an improved Gaussian Bayesian optimization algorithm.

[0037] According to one or some embodiments of the present invention, in the above a% holdout set, a may be 30. It may also be other values ​​between 1 and 100, depending on actual needs and not limited thereto.

[0038] According to one or some embodiments of the present invention, the takeaway number identification further includes the following steps:

[0039] S7. The signaling call record data to be tested is input into the XNCLBoost model. After prediction by the model, the prediction result is output, and objectionable data is eliminated based on the prediction result.

[0040] Furthermore, the takeaway number identification also includes the following steps:

[0041] S8, applying the output model prediction data to the caller name card and number recognition business scenario to identify and mark the caller number for the user;

[0042] S9. Collect the feedback data and complaint data from the takeaway tagging platform in the business scenario application, update the takeaway number blacklist and whitelist data with the feedback data and complaint data from the takeaway platform, and form a closed loop of the modeling process.

[0043] The technology of the present invention builds a two-layer training model framework (i.e., XNCLBoost model framework) based on the Blending ensemble learning strategy. The main idea of ​​Blending is to divide the original training set into two parts, such as 70% of the data as the training set and the remaining 30% of the data as the test set. The first layer of the XNCLBoost model uses the XGboost model, NGboost model, and Catboost model with strong classification performance as the basic classifier. The first layer is to build a hold-out set (Hold-outset) on the training sample set, use the remaining data to train each base model, and then use the base model to train the Hold-out set. ou t se t for prediction; the second-layer model uses the LightGBM model with strong nonlinear performance as the meta-classifier, directly models the prediction results to generate the second-layer model, and realizes the fusion of multiple models. The fused XNCLBoost model has better generalization ability, higher prediction accuracy and lower memory consumption. The number recognition method based on the Blending strategy fusion model is simple and easy, with less feature information leakage and low probability of overfitting of the model.

[0044] The basic idea of ​​the Blending algorithm is as follows:

[0045] Input: training set

[0046] D={(x1, y1), (x2, y2), ....., (x m ,y m )}

[0047] Hold-out:

[0048] D=D1∪D2

[0049] The first layer of machine learning classification algorithms A1, A2, ..., A n

[0050] The second layer of machine learning classification algorithm A

[0051] Initialization parameters: Hold-out parameter K, learning rate α, learning step size l, tree depth n and other machine learning hyperparameters.

[0052] step:

[0053] Step1: for i=1, 2, ...., n do

[0054] Step 2: Out_A i =A i (D1);

[0055] Step 3: end for

[0056] Step 4: D′=D2;

[0057] Step 5: for j=1, 2, ...., m do

[0058] Step 6: for i=1, 2, ...., n do

[0059] Step 7: z ji =Out_A i (x j );

[0060] Step 8: end for

[0061] Step 9:

[0062] D″=D′∪((z j1 , z j2 ,......,z jn ), y j )∪((x1, x2, ..., x n ), y j );

[0063] Step 10: end for

[0064] Step 11: Out_A′=A(D”);

[0065] Output:

[0066]

[0067] According to one or some embodiments of the present invention, the input takeout blacklist and whitelist data is the platform's existing number identification data, and can be optimized and updated by the predicted data calculated by the model and the data verified after the platform application.

[0068] According to one or some embodiments of the present invention, the input signaling bill data is the call bill data to be tested, the call bill is based on the call record data set of the corresponding number, and the call record data collection period can be the past month, the past two months, or the past three months. The specific data collection period is determined according to actual needs and is not limited thereto.

[0069] Based on the sample data set, the feature variables required by the model are extracted. After analyzing the sample data, it is found that the call information behavior of takeaway calls is strongly correlated with meal times. The meal time periods are mainly defined as breakfast time: 7:00-9:30, lunch time: 11:00-14:30, dinner time: 16:00-20:30, and midnight meal time: 21:30-24:00. The data within each meal time will have different weights when generating feature variables. The weight vectors of breakfast time, lunch time, dinner time, midnight meal time, and other time periods are:

[0070] |breakfast, lunch, dinner, midnight snack, other] = [w, 1, w2, w3, w4, w5]

[0071] Where 0≤w i ≤1, For first-tier cities, developed cities in the west (Xi'an / Lanzhou / Urumqi), developed cities in the central region, cities in the south and cities in the east, the living habits (meal times) in different regions will be different, and fine-tuning can be made by setting the offset of meal times.

[0072] According to one or some embodiments of the present invention, call record data includes call record data of a calling number and call record data of a called number, and the call record data has characteristic variables, and the characteristic variables include: a thermal value of a call area of ​​a calling number, which is used to record the total number of cell / building locations in a call signaling when a number initiates an active call; an outgoing call frequency of a calling number, which is used to record the total number of calls in a call signaling when a number initiates an active call; a connected call frequency of a calling number, which is used to record the total number of successful connections among the number of active calls in a call signaling when a number initiates an active call; an average ringing duration of a calling number, which is used to record the total number of ringing durations of calls in a call signaling when a number initiates an active call; an average call duration of a calling number, which is used to record the total number of normal call durations of calls in a call signaling when a number initiates an active call; a thermal value of a call area of ​​a called number, which is used to record the total number of cell / building locations in a call signaling when a number accepts a passive call; an incoming call frequency of a called number, which is used to record the total number of outgoing calls in a call signaling when a number accepts a passive call. Specific characteristic variables are not limited to this, for example, they can also be: calling number repeat call index 1, calling number repeat call index 2, calling contact dispersion (when the number initiates an active call, the total of the called numbers called in the call signaling without duplication), calling number active hours per day (based on call bill signaling data, counting the number of natural hours that the number initiates active calls per day), calling number active days per week (based on call bill signaling data, counting the number of natural days that the number initiates active calls per week), calling number active days per month (based on call bill signaling data, counting the number of days that the number initiates active calls per month), calling number monthly voice call consumption The monthly data includes: amount (based on call bill signaling data, the monthly voice charge amount of the statistical number), monthly traffic cost of the calling number (based on call bill signaling data, the monthly traffic cost amount of the statistical number), etc.; it can also be: called number connection frequency (when the number accepts passive calls, the sum of the number of successful connections in the number of passive acceptances in the call signaling), called number ringing duration (when the number accepts passive calls, the sum of the ringing duration in the call signaling), called number call duration (when the number accepts passive calls, the sum of the normal call duration in the call signaling), called contact dispersion (when the number accepts passive calls, the de-duplicated sum of the calling party numbers in the call signaling), etc.

[0073] The calling behavior of takeout phone numbers often shows clustering in the spatial dimension. The dispersion of called numbers from takeout numbers is large, while the dispersion of cities of called numbers is low.

[0074] In terms of time dimension, the calling behavior of takeout phone numbers often shows a dense distribution of outgoing calls within a short period of time, generally with 7-10 dense outgoing calls within 3 minutes.

[0075] When food delivery number A calls called number B and called number B does not answer, food delivery number A will repeatedly dial called number B within a short period of time, with an average interval of nearly 35 seconds.

[0076] When the takeaway number A calls the called number B and the called number B does not answer, the takeaway number A will send a short message to the called number B within a short period of time.

[0077] The calling number repeat call index 1 is used to indicate that when takeout number A calls called number B and called number B is not connected, takeout number A will repeatedly dial called number B within a short period of time until the call is connected. This is the total number of calls between takeout number A and called number B.

[0078] The calling number repeat call index 2 is used to indicate that when food delivery number A calls called number B and called number B does not connect, food delivery number A will repeatedly call called number B within a short period of time, until the total number of calls before food delivery number A and called number B successfully send text messages.

[0079] The present invention provides an automated intelligent identification method for takeout numbers, improves the connection rate and communication efficiency of takeout numbers, creates a good communication network environment, provides a high-accuracy and high-stability takeout number identification method, and provides correct identification and display of incoming call numbers for a large number of mobile users.

[0080] According to another aspect of the present invention, a device for identifying takeout numbers is provided, comprising: a model sample generation unit, the model sample generation unit is used to process the input takeout number blacklist and whitelist data and signaling call list data, generate and output a model sample data set, wherein the data processing method includes data ETL engineering cleaning data and SMOTETomeK algorithm for comprehensive sampling of the original data set; a first-layer model, having an XGboost model, an NGboost model, and a Catboost model, the first-layer model divides the model sample data set into a training set and a validation set according to the ratio of x:y, wherein x∈[0,1], y∈[0,1], x+y=1, and then divides 30% of the retention set on the training set, and respectively uses the XGboost model, the NGboost model, and the Catboost model to train the model on 70% of the sample training set to form a first-layer basic classifier model, and then respectively inputs the 30% retention set and the entire validation set on the training set into the XGboost model of the first-layer basic classifier model. In the first layer model, NGboost model and Catboost model, the XGboost model, NGboost model and Catboost model respectively output the training set prediction value set and the validation set prediction value set, and then merge the training set prediction value set output by the first layer basic classifier model to form a new training set feature value set, and merge the validation set prediction value set output by the first layer basic classifier model to form a new validation set prediction value set; the second layer model is built based on the LightGBM algorithm, and the second layer model inputs the new training set feature value and the original test set into the LightGBM model, and outputs a model that meets the pre-set model accuracy after training, thereby forming an XNCLBoost model. When the LightGBM model has many hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. The Bayesian optimization algorithm is improved based on the Gaussian process, and the covariance function in the form of a multi-strategy combination is used to simultaneously capture the smoothness and amplitude of the objective function.

[0081] According to another aspect of the present invention, a device for takeout number identification is provided, comprising: a memory, a processor, and a takeout number identification program stored in the memory and executable on the processor, wherein the takeout number identification program implements the steps of the above-mentioned takeout number identification method when executed by the processor.

[0082] According to the present invention, a computer storage medium is also provided.

[0083] A takeout number recognition program is stored on the computer storage medium, and when the takeout number recognition program is executed by the processor, the steps of the above-mentioned takeout number recognition method are implemented.

[0084] Among them, the method implemented when the takeout number identification program running on the processor is executed can refer to the various embodiments of the takeout number identification method of the present invention, and will not be repeated here.

[0085] The present invention also provides a computer program product.

[0086] The computer program product of the present invention includes a takeout number identification program, which, when executed by a processor, implements the steps of the takeout number identification method described above.

[0087] Among them, the method implemented when the takeout number identification program running on the processor is executed can refer to the various embodiments of the takeout number identification method of the present invention, and will not be repeated here.

[0088] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0089] The above description is merely an exemplary embodiment of the present invention and is not intended to limit the protection scope of the present invention. The protection scope of the present invention is determined by the appended claims.

Claims

1. A method for identifying a takeaway number, the method comprising the following steps: S1. Input the takeaway number blacklist and whitelist data and signaling call record data, and clean the data by using data ETL engineering. The data ETL engineering includes data extraction, conversion, and loading, and then obtains the original data set required by the model through data association and fusion; S2. Using the SMOTE TomeK algorithm to perform comprehensive sampling on the original data set to form a model sample data set; S3, dividing the model sample data set into a training set and a validation set according to the ratio of x:y, wherein: x∈[0,1], y∈[0,1], x+y=1, then divide the training set into a% holdout set, and use XGboost model, NGboost model, and Catboost model to train the model on the 1-a% sample training set to form the first-layer basic classifier model; S4, respectively inputting a% of the held-out set on the training set and all the validation sets into the XGboost model, NGboost model, and Catboost model of the first-layer basic classifier model, and the XGboost model, NGboost model, and Catboost model respectively output a set of predicted values ​​for the training set and a set of predicted values ​​for the validation set; S5, merging the training set prediction value set output by the first-layer basic classifier model to form a new training set feature value set, and merging the validation set prediction value set output by the first-layer basic classifier model to form a new validation set prediction value set; S6. Build the second layer model based on the LightGBM algorithm, input the new training set feature values ​​and the original test set into the LightGBM model, and after training, output a model that meets the pre-set model accuracy, thereby forming the XNCLBoost model. The XNCLBoost model is based on the Blending strategy. When the LightGBM model has many hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. Among them, the Bayesian optimization algorithm is improved based on Gaussian process, and the covariance function in the form of multi-strategy combination is used to capture the smoothness and amplitude of the objective function at the same time; S7, the signaling call record data to be tested is input into the XNCLBoost model, and after prediction by the model, the prediction result is output, and objectionable data is eliminated according to the prediction result; S8. Apply the output model prediction data to the caller name card and number recognition business scenario to identify and label the caller number for the user.

2. The method of claim 1, wherein: The a% holdout set, where a=30.

3. The method according to claim 1, further comprising the steps of: S9. Collect the data and complaint data fed back by the takeaway marking platform in the business scenario application, and update the takeaway number blacklist and whitelist data with the data fed back by the takeaway platform and complaint data to form a closed loop of the modeling process.

4. The method of claim 3, wherein: The input takeout blacklist and whitelist data is the platform's existing number identification data, and is optimized and updated by the predicted data calculated by the model and the data verified after the platform application.

5. The method of claim 1, wherein: The input signaling bill data is the call bill data to be tested, the call bill is based on the call record data set of the corresponding number, and the call record data collection period is the past month, the past two months or the past three months.

6. The method of claim 5, wherein: The call record data includes call record data of a calling number and call record data of a called number, and the call record data has characteristic variables, and the characteristic variables include: The call area thermal value of the calling number is used to record the total number of cell / building locations in the call signaling when the number initiates an active call; The calling number outgoing call frequency is used to record the total number of calls in the call signaling when the number initiates an active call; The calling number connection frequency is used to record the total number of successful connections among the number of active calls in the call signaling when the number initiates an active call; The average ringing duration of the calling number is used to record the total ringing duration of the call in the call signaling when the number initiates an active call; The average call duration of the calling number is used to record the total normal call duration in the call signaling when the number initiates an active call; The called number call area thermal value is used to record the total number of cell / building locations in the call signaling when the number receives a passive call; The called number's incoming call frequency is used to record the total number of outgoing calls in the call signaling when the number accepts passive calls.

7. A device for identifying a takeaway number, comprising: A model sample generating unit, the model sample generating unit is used to process the input takeaway number blacklist and whitelist data and signaling call bill data, generate and output a model sample data set, wherein the data processing method includes cleaning the data by data ETL engineering, obtaining the original data set required by the model by data association fusion, and using the SMOTETomeK algorithm to perform comprehensive sampling on the original data set to form a model sample data set; The first layer model includes an XGboost model, an NGboost model, and a Catboost model. The first layer model divides the model sample data set into a training set and a validation set according to the ratio of x:y, where x∈[0,1], y∈[0,1], and x+y=1. Then, a 30% holdout set is divided on the training set. The XGboost model, the NGboost model, and the Catboost model are used to train the model on the 70% sample training set to form the first layer basic classifier model. Then, the training set is divided into The upper 30% reserved set and all validation sets are input into the XGboost model, NGboost model and Catboost model of the first-layer basic classifier model, and the XGboost model, NGboost model and Catboost model respectively output a training set prediction value set and a validation set prediction value set, and then the training set prediction value set output by the first-layer basic classifier model is merged to form a new training set feature value set, and the validation set prediction value set output by the first-layer basic classifier model is merged to form a new validation set prediction value set; The second layer model is built based on the LightGBM algorithm. The second layer model inputs the new training set feature values ​​and the original test set into the LightGBM model, and after training, outputs a model that meets the pre-set model accuracy, thereby forming an XNCLBoost model. Among them, when the LightGBM model has many hyperparameters, the Gaussian Bayesian algorithm is used to optimize the value range of the hyperparameters to improve the prediction accuracy and robustness of the model. Among them, the Bayesian optimization algorithm is improved based on the Gaussian process, and the covariance function in the form of a multi-strategy combination is used to simultaneously capture the smoothness and amplitude of the objective function. Among them, the signaling call record data to be tested is input into the XNCLBoost model, and after model prediction, the prediction result is output, and objectionable data is eliminated based on the prediction result; the output model prediction data is applied to the caller business card and code number recognition business scenario to identify and mark the incoming call number for the user.

8. An electronic device comprising: A memory, a processor, and a takeout number identification program stored in the memory and executable on the processor, wherein the takeout number identification program, when executed by the processor, implements the steps of the takeout number identification method according to any one of claims 1 to 6.

9. A computer storage medium, wherein: The computer storage medium stores a takeout number recognition program, which, when executed by a processor, implements the steps of the takeout number recognition method according to any one of claims 1 to 6.

Images