A telecommunications anti-fraud method based on machine learning

Through the telecommunications anti-fraud method based on machine learning, the multi-dimensional characteristics of telecommunications users are analyzed using LightGBM algorithm and pseudo-label technology, and high-accuracy early warning for telecommunications fraud users is achieved, solving the problem of the inability to early warning and misinterception in the existing technology. The problem has improved the ability to prevent telecommunications fraud.

CN114513791BActive Publication Date: 2024-11-29ZHEJIANG HONGCHENG COMP SYST
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202210034458.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-13
Publication Date
2024-11-29
Estimated Expiration
2042-01-13

AI Technical Summary

Technical Problem

The existing technology is difficult to effectively warn and prevent telecommunications fraud. Traditional anti-fraud methods mainly rely on manual experience and rule judgments, cannot be warning in advance, and are prone to intercept legal users by mistake.

Method used

Using a telecommunications anti-fraud method based on machine learning, we can analyze historical notification numbers and notification content, segment potential telecommunications fraud users, collect multi-dimensional feature data, train models using LightGBM algorithm, and combine pseudo-label and SMOTE algorithm to achieve high accuracy User risk forecast and tiered warning.

Benefits of technology

It realizes accurate identification and early warning of high-risk users, improves prediction effect, and can be early warning when no fraud occurs, reduces the impact on legal users, and improves the ability to prevent telecommunications fraud.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114513791B_ABST
    Figure CN114513791B_ABST
Patent Text Reader

Abstract

The present invention relates to a telecommunications anti-fraud method based on machine learning. The present invention aims to solve the problem of timely interception of telecommunications fraud, with the goal of providing early warning and preventing telecommunications fraud in advance. The present invention uses the LightGBM algorithm in machine learning to train and predict positive and negative samples, and divides common fraud scenarios into four categories, and performs modeling training according to different categories to improve the accuracy of prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer big data analysis and machine learning technology, and in particular to a telecommunications anti-fraud method based on machine learning. Background Art

[0002] In recent years, telecom fraud cases have become increasingly common. Criminals use scam text messages and phone calls to target victims in real time, creating a highly organized pattern of fraud. The Ministry of Industry and Information Technology (MIIT) mandates the establishment of early warning monitoring and big data analysis mechanisms to prevent the sending of commercial text messages or calls without user consent or request. This is crucial to prevent and combat telecom fraud, safeguard the legitimate rights and interests of the public, and ensure national security and social stability. Beyond strengthening anti-fraud awareness and improving fraud detection and prevention capabilities, rapidly curbing telecom fraud at its source is also crucial. Telecom operators face significant technical challenges in establishing mechanisms to predict fraudulent user behavior and build fraud warning and analysis capabilities through various technical means to prevent and combat telecom fraud. How to prevent telecom fraud, promptly block the calls of criminals, and predict the fraud risk level of operator SIM cards is also crucial; traditional anti-fraud methods mainly use experience and manual rules to intercept low-package and new network users while generating high call volume, but cannot provide early warning and prevention; existing anti-telecom fraud models are mostly based on manual experience or rule judgment to intercept fraudulent users who have already made calls, which makes it difficult to provide early warning and is prone to generalizations. Fraudulent numbers that are difficult to cover by rules will be generated over time. Summary of the Invention

[0003] In order to overcome the above-mentioned shortcomings, the present invention aims to provide a telecommunications anti-fraud method based on machine learning. The present invention can accurately and effectively identify high-risk fraud-related users, and carry out early warning, verification and shutdown of fraud-related users in a layered and graded manner.

[0004] The present invention achieves the above-mentioned purpose through the following technical solutions: a telecommunications anti-fraud method based on machine learning, comprising the following steps:

[0005] (1) Analyze and sort out all the historical notification numbers and notification contents, sort out different scenarios, segment potential telecommunications fraud users, deeply explore user groups, and characterize the characteristics of target users;

[0006] (2) Data collection: collect features of various dimensions for existing users. Starting from the six dimensions of user basic attributes, call behavior data, package network information, consumption and usage characteristics, terminal tag information, and derived variables, the data is subdivided into 178 features, and a wide table of active behavior of the number card is constructed to collect user data.

[0007] (3) Perform data cleaning and preprocessing, and perform missing value detection and outlier inspection based on user data;

[0008] (4) Data sampling is performed by adding pseudo-labels, that is, the number cards that have been predicted to be involved in fraud and have not been restored are added as positive samples to the reported number cards, so that the ratio of positive and negative samples is controlled at 1:15;

[0009] (5) Based on machine learning algorithm training model, predict the probability that a user is a telecom fraud user, and use cross-validation and grid search to find the optimal parameters of the model;

[0010] (6) Iterate and optimize the model to achieve better prediction results.

[0011] Preferably, the notification number analysis in step (1) is specifically as follows: based on the verification and analysis of the notification number issued by the public security and 12321, the differences between normal users and fraudulent users are analyzed from the user's network access status, package consumption, historical active behavior, Internet behavior, call indicators and location information, and the characteristics of the target user are portrayed.

[0012] Preferably, the data cleaning and preprocessing of step (3) are as follows: the user age, monthly billing information, and user stickiness in the data are filled with the median value, the user billing, intra-provincial and extra-provincial traffic usage, the number of voice calls, and the duration of voice calls are filled with zero values, and abnormal values ​​are corrected with the average value; errors in the values, types, dates, etc. that appear in the data are corrected; missing values, such as missing age, are filled with the median value, and monthly consumption amount is filled with zero values; in the preprocessing, some categorical variables, such as the user's terminal model and the user's package name, are processed with one-hot encoding to convert the original Chinese categories into digital information that can be processed by computers.

[0013] Preferably, the data sampling in step (4) is performed using the SMOTE algorithm. The SMOTE algorithm uses a linear interpolation method to calculate two recently reported number samples, generate new artificial positive samples, and add them to the overall positive samples to expand the number of positive samples; that is, the formula is used:

[0014] X new =X i +(X ij -X i )×γ

[0015] X ij is a sample X i The j neighboring samples of the fraudulent sample i are the j neighboring fraudulent samples of the fraudulent sample i; γ is a random number between 0 and 1; the newly generated fraudulent sample is represented by Xi and the linear expression of its neighboring sample j.

[0016] Preferably, when using pseudo labels, the model is first trained based on the labeled positive and negative sample data, and then the unlabeled data is predicted; the pseudo labels extract a part of the training set as a new verification set, add the predicted data to the new training set, and then train a new model to predict subsequent unlabeled data; in the actual anti-fraud application scenario, there is the following type of user data: the user's number card is predicted to be shut down by the anti-fraud model, and has not been restored through any online or offline restoration channels for a long time, then such users will be identified as users with fraud risks, and they will be added to the positive samples as pseudo label data to expand the positive samples; in this step, the pseudo label is the fraud-related and non-fraud-related label data predicted by the algorithm model, and the notification number forms a new training set to train the optimal model to predict whether the user has fraud risks.

[0017] Preferably, the machine learning algorithm in step (5) adopts the LightGBM algorithm, and the obtained data is divided into a training set and a test set, with the ratio of the training set to the test set being 3:7; and a 5-fold cross validation method and a grid search are used to screen out the optimal solution of the model, and the best evaluation index has been achieved; in terms of model evaluation, the F1 value and AUC value are mainly used to evaluate the effect of the model, wherein the F1 value of the model retest set is 0.871824, and the AUC value is 0.995512.

[0018] Preferably, the LightGBM algorithm is implemented using the Histogram algorithm, GOSS unilateral gradient sampling and EFB proprietary feature bundling algorithm;

[0019] The histogram algorithm discretizes the continuous value of the feature into k discrete features and constructs a bin with a width of k. The k bins are set to 256, the total number of sample data is 20 million, and the feature has a total of 178 dimensions. The complexity of XGBoost is reduced from O = 20000000 * 178 to O = 256 * 178.

[0020] The one-sided gradient sampling method is to retain samples with larger gradients and select random sampling on samples with smaller gradients. The one-sided gradient sampling method is to sort the data in descending order according to the size of the gradient value, select the top a data, and then randomly sample b data from the remaining data, multiply the data with small gradients by a constant multiplier: (1-a) / b, and then calculate the information gain.

[0021] Preferably, in step (5), the LightGBM package in Python, i.e., lightgbm.LGBMClassifier(), is used, and its adjustment parameters include three categories: core parameters, learning control parameters, and indicator class parameters;

[0022] The core parameters include: boosting, where the GBDT algorithm is used; objective objective function: binary classification; n_estimators: the number of weak classifiers is set to 1000; learning_rate learning rate is set to 0.01; num_leaves the number of leaf nodes on a tree is set to 40, max_depth the maximum depth of the tree is set to 7, and seed seed trees are set to 4500; learning control parameters: min_child_samples the minimum number of samples is set to 46; min_child_weight the minimum Heather matrix on the leaf node and the default setting is 0.01; subsample_freq: the training sample sampling rate is set to 2; subsample subsample is set to 2; colsample_bytree column sampling parameter is selected as 0.8; reg_alpha: L1 regularization parameter; reg_lambda: L2 regularization parameter are selected as 0 and 5 respectively;

[0023] Indicator function: The metric optional indicator is set to binary_logloss; the verbose redundancy parameter is set to 1;

[0024] After the model is trained, the F1 value and AUC (area under the ROC curve) are used to further evaluate the model.

[0025] The various indicators printed out on the test set of this model are as follows:

[0026] Model Report AUC Score:0.995512

[0027] Recall: 0.925329

[0028] Accuracy: 0.824169

[0029] F1:0.871824.

[0030] Preferably, in the step (6), in the process of actual deployment and application of the anti-fraud model, first, according to the source of the data, data is collected and aggregated from different platforms, and user characteristics are generated from the user's basic data: customer information, package status, card activation channel, consumption characteristics, behavioral data: call behavior, traffic Internet behavior, call roaming location, network data: DPI data, user location and terminal data, interface data: signaling data, trial call data, notification data, platform reporting data, and anti-fraud behavior wide table is established; the model is trained by a machine learning model algorithm and it is predicted whether the user in the anti-fraud behavior wide table has fraudulent behavior; the fraud scenarios are divided into four categories: loan financial fraud, impersonation of customer service, GOIP fraud and silent card fraud, different model training predictions are generated from the four scenarios, and the predicted user's fraud probability is divided into three risk levels of high, medium and low corresponding to three anti-fraud processing processes, and an early warning verification message is sent to the user. If the verification is passed, it means that the user's real-name authentication is passed. If the verification fails, the number will be shut down.

[0031] The beneficial effects of the present invention are: (1) compared with the traditional rule division method, the present invention can more accurately and quickly predict users involved in fraud, including some number cards that have not yet committed fraud; (2) the present invention adds pseudo labels to make the model classification more accurate; (3) the present invention divides the model into four major scenarios according to the reasons why the reported number is reported, and models training and predicts data separately, which can provide hierarchical and graded early warning management for users involved in fraud. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flow chart of the method of the present invention;

[0033] Figure 2 This is a schematic diagram of the specific LightGBM parameter settings of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto:

[0035] Example: Figure 1 As shown in FIG, a telecommunications anti-fraud method based on machine learning is as follows:

[0036] The first step is to analyze the basic attributes of users reported by 12321 (and the public security bureau) in the past. The users' network access, package consumption, historical active behavior, online behavior, call indicators, and location information are used to analyze the fraudulent users and the legitimate users.

[0037] The second step is to collect and aggregate data. Starting from the six dimensions in the first step, we refine the user's comprehensive characteristics totaling 178, establish a broad table of active behaviors of the number card, and observe and improve the number card's early warning and monitoring capabilities.

[0038] The third step is data cleaning and preprocessing. The goal is to remove redundant data and fill in missing data. By addressing missing values, outliers, and deduplication, we can obtain high-quality positive and negative samples. For example, if a user's roaming status is empty, their home location needs to be set to their roaming location. For outliers, if a user uses an app more than the average number of times per day, the median value can be used to fill in the gaps. For example, if some phone numbers have been reported multiple times, duplicates need to be removed.

[0039] The fourth step is to adjust the positive and negative samples. In actual applications, the ratio of fraudulent users to legitimate users is approximately one in five thousand. This magnitude leads to an imbalance in the training samples used for machine learning models. This large disparity between the ratio of fraudulent users to legitimate users leads to a large disparity between positive and negative samples. This large number of negative samples directly impacts the optimization of the minimum loss function, resulting in inaccurate model training.

[0040] In the early stages of model building, a simple method is to repeatedly use positive samples and adjust the ratio of positive and negative samples. This is relatively simple and will not be discussed in detail in this article. In this invention, we try to use the SMOTE (Synthetic Minority Oversampling Technique) algorithm. The SMOTE method mainly uses linear interpolation to calculate two recently reported number samples, generate new artificial positive samples, and add them to the total positive samples to expand the number of positive samples. That is, using the formula:

[0041] X new =X i +(X ij -X i )×γ X ij is a sample X i The j neighboring samples of the fraudulent sample i are the j neighboring fraudulent samples of the fraudulent sample i. γ is a random number between 0 and 1. The newly generated fraudulent sample can be used i and the linear expression of its neighboring sample j

[0042] In the actual application of anti-fraud, only a few tagged number cards have been notified and reported. A large amount of data is normal user data that is not marked as fraud-related. It is difficult to obtain a large amount of labeled data.

[0043] Typically, when using pseudo-labeling, a model is first trained on labeled positive and negative data, and then predicted on unlabeled data. Pseudo-labeling involves extracting a portion of the training set as a new validation set, adding the previously predicted data to the new training set, and then training a new model to predict subsequent unlabeled data.

[0044] In the actual application scenario of anti-fraud, there is the following type of user data: the user's number card is predicted to be shut down by the anti-fraud model, and has not been restored through any of the provided online or offline restoration channels for a long time. These users are also identified as users with fraud risks, and they are added to the positive samples as pseudo-label data to expand the positive samples. The essence of pseudo-labels is semi-supervised learning. In the present invention, pseudo-labels are the fraud-related and non-fraud-related label data predicted by the algorithm model, and the notification number forms a new training set to train the optimal model to predict whether the user has fraud risks.

[0045] The fifth step is to select an appropriate algorithm to train the model and predict data. After adjusting the ratio of positive and negative samples in the training set, the LightGBM lightweight gradient boosting machine algorithm is used to train and predict the model.

[0046] The main technology of LightGBM is to use the Histogram algorithm, GOSS (Gradient based OneSide Sampling) based on one-side gradient sampling and EFB (Exclusive Feature Bunding) proprietary feature bundling algorithm.

[0047] The histogram algorithm discretizes the continuous value of a feature into k discrete features, and constructs a bin with a width of k. For example, when predicting the monthly traffic usage of users involved in fraud, which is a continuous value feature, it can be cut into k bins after using the histogram algorithm, so that the algorithm only needs to traverse all the data k times to obtain the optimal splitting gain of the feature. Other algorithms such as XGBoost will need to traverse all the data to obtain the optimal splitting point. Its time complexity can be greatly reduced from O(data*feature) of the XGBoost algorithm to O(k*feature). In this case, the k bins are set to 256, the total amount of sample data is 20 million, and the features have a total of 178 dimensions. The complexity of XGBoost is reduced from O=20000000*178 to O=256*178.

[0048] The one-sided gradient sampling method retains samples with larger gradients and randomly samples samples with smaller gradients. This is because when fitting the gradient of the loss function, smaller absolute values ​​of the gradient indicate smaller sample errors and higher accuracy, which in turn reduces the training value of the model algorithm. The one-sided gradient sampling method sorts the data in descending order by gradient value. The top-ranked data, a, is selected, and then data, b, is randomly sampled from the remaining data. The data with smaller gradients is multiplied by a constant multiplier: (1-a) / b, and the information gain is calculated. Anti-fraud sample data often contains a large number of non-fraudulent, or normal, users. These are the sample data with smaller gradients and lower training value mentioned above. In this case, reducing the sample size of these normal users and randomly sampling them can reduce the sample data volume, making the histogram algorithm more effective.

[0049] The use of mutually exclusive feature merging algorithms: In the wide anti-fraud table, there are nearly 180 dimensions of various features. Some features are relatively sparse, but these features are often mutually exclusive. For example, user preference attributes, user out-of-province data usage, user roaming time in high-risk provinces and cities, and night call usage. Most of these features are highly missing and sparse. The usual practice is to merge these sparse and mutually exclusive features to construct the feature histogram mentioned in the first step above.

[0050] The present invention uses the LightGBM package in Python, namely lightgbm.LGBM Classifier(), whose main adjustment parameters include three categories: core parameters, learning control parameters and indicator class parameters.

[0051] The core parameters include: boosting, this method uses the GBDT algorithm; objective objective function: binary classification; n_estimators: the number of weak classifiers is set to 1000; learning_rate learning rate is set to 0.01; num_leaves the number of leaf nodes on a tree is set to 40, max_depth the maximum depth of the tree is set to 7, and seed seed trees are set to 4500.

[0052] Learning control parameters: min_child_samples is the minimum number of samples set to 46; min_child_weight is the minimum Heather matrix on the leaf node and the default setting is 0.01; subsample_freq: the training sample sampling rate is set to 2; subsample is set to 2; colsample_bytree column sampling parameter is selected as 0.8; reg_alpha: L1 regularization parameter; reg_lambda: L2 regularization parameter are selected as 0 and 5 respectively.

[0053] Indicator function: The optional metric is set to binary_logloss; the verbose parameter is set to 1. The specific LightGBM parameter settings are as follows Figure 2 shown.

[0054] After the model is trained, the F1 value and AUC (area under the ROC curve) are used to further evaluate the performance of the model.

[0055] The various indicators printed out on the test set of this model are as follows:

[0056] Model Report

[0057] AUC Score:0.995512

[0058] Recall: 0.925329

[0059] Accuracy: 0.824169

[0060] F1:0.871824

[0061] The sixth step involves optimization and iteration. Anti-fraud involves a variety of scenarios, including loan and financial fraud, customer service impersonation, GoI scams, and silent card scams. The model is divided into four categories for training and prediction, each focused on achieving optimal prediction results.

[0062] The actual deployment and application of the anti-fraud model begins with data collection and aggregation from various platforms based on their source. User profiles are generated from basic user data (including customer information, package status, card activation channels, and spending patterns); behavioral data (including call activity, data usage, and roaming locations); network data (including DPI data, user location, and terminal data); and interface data (including signaling data, call attempts, notifications, and platform reporting data). This generates a broad anti-fraud behavior table.

[0063] Machine learning algorithms are used to train models and predict whether users in the anti-fraud behavior table are involved in fraudulent activity. Fraud scenarios are categorized into four main categories: loan and financial fraud, customer service impersonation, GoIP fraud, and silent card fraud. Different model training predictions are generated for each of these scenarios, and the predicted probability of fraudulent activity is categorized into three risk levels: high, medium, and low. These levels correspond to three different anti-fraud processing procedures, and users are sent an alert for verification. If verified, real-name authentication has been successfully completed. Failure to pass the verification process results in the user's number being shut down.

[0064] The above description is a specific embodiment of the present invention and the technical principles used. If the changes made according to the concept of the present invention do not exceed the spirit covered by the description and drawings, they should still fall within the scope of protection of the present invention.

Claims

1. A telecommunications anti-fraud method based on machine learning, characterized in that: The steps include: (1) Analyze and sort out all the historical notification numbers and notification contents, sort out different scenarios, segment potential telecommunications fraud users, deeply explore user groups, and characterize the characteristics of target users; (2) Data collection: Collect the characteristics of existing users in various dimensions, starting from the six dimensions of basic user attributes, call behavior data, package network information, consumption and usage characteristics, terminal tag information, and derived variables. The characteristics are subdivided into 178 characteristics, and a broad table of active behaviors of account cards is constructed to collect user data. (3) Perform data cleaning and preprocessing, and perform missing value detection and outlier inspection based on user data; (4) Performing data sampling by adding pseudo labels, that is, adding the number cards that have been predicted to be involved in fraud and have not been restored as positive samples to the number cards that have been reported, so that the ratio of positive samples to negative samples is controlled at 1:15; The SMOTE algorithm is used for data sampling. The SMOTE algorithm uses the linear interpolation method to calculate two recently reported number samples, generate new artificial positive samples and add them to the overall positive samples to expand the number of positive samples; that is, use the formula: X new =X i +(X ij -X i )×γ X ij is a sample X i The j neighboring samples of the fraudulent sample i are represented by X, which means the j neighboring fraudulent samples of the fraudulent sample i; γ is a random number between 0 and 1; the newly generated fraudulent sample is represented by X i and the linear expression of its neighboring sample j; When using pseudo labels, first train the model based on the labeled positive and negative sample data, and then predict the unlabeled data; the pseudo label extracts a part of the training set as a new verification set, adds the predicted data to the new training set, and then trains a new model to predict the subsequent unlabeled data; in the actual anti-fraud application scenario, there is the following type of user data: the user's card is predicted to be shut down by the anti-fraud model, and has not been restored through any of the provided online or offline restoration channels for a long time. In this case, such users will be identified as users with fraud risks, and they will be added to the positive samples as pseudo label data to expand the positive samples; in this step, the pseudo label is the fraud-related or non-fraud-related label data predicted by the algorithm model, and the notification number forms a new training set to train the optimal model to predict whether the user has a fraud risk; (5) Based on the machine learning algorithm, the model is trained to predict the probability that the user is a telecommunications fraud user, and cross-validation and grid search are used to find the optimal parameters of the model; The machine learning algorithm uses the LightGBM algorithm, and the obtained data is divided into a training set and a test set, with a ratio of 3:

7. The 5-fold cross validation method and grid search are used to screen out the optimal solution of the model, achieving the best evaluation index. In terms of model evaluation, the F1 value and AUC value are mainly used to evaluate the effect of the model. The F1 value of the model retest set is 0.871824, and the AUC value is 0.995512. The LightGBM algorithm is implemented using the Histogram algorithm, GOSS unilateral gradient sampling and EFB proprietary feature bundling algorithm; The histogram algorithm discretizes the continuous value of the feature into k discrete features and constructs a bin with a width of k. The k bins are set to 256, the total amount of sample data is 20 million, and the features have a total of 178 dimensions. The complexity of XGBoost is reduced from O = 20000000*178 to O = 256*178. The one-sided gradient sampling method is to retain samples with larger gradients and select random sampling on samples with smaller gradients. The one-sided gradient sampling method is to sort the data in descending order according to the size of the gradient value, select the top a data, and then randomly sample b data from the remaining data, multiply the small gradient data by a constant multiplier: (1-a) / b, and then calculate the information gain. Among them, the LightGBM package in python is used, namely lightgbm.LGBM Classifier(), and its adjustment parameters include three categories: core parameters, learning control parameters and indicator class parameters; The core parameters include: boosting, where the gbdt algorithm is used; objective objective function: binary classification; n_estimators: the number of weak classifiers is set to 1000; learning_rate learning rate is set to 0.01; num_leaves the number of leaf nodes on a tree is set to 40, max_depth the maximum depth of the tree is set to 7, and seed seed trees are set to 4500; Learning control parameters: min_child_samples minimum sample size is set to 46; min_child_weight minimum Heather matrix on the leaf node and default setting is 0.01; subsample_freq: training sample sampling rate is set to 2; subsample subsample is set to 2; colsample_bytree column sampling parameter is selected as 0.8; reg_alpha: L1 regularization parameter; reg_lambda: L2 regularization parameters are selected as 0 and 5 respectively; Indicator function: The optional metric is set to binary_logloss; the verbose parameter is set to 1; After the model is trained, the F1 value and AUC are used to further evaluate the model. AUC is the area under the ROC curve. The various indicators printed out on the test set of this model are as follows: Model Report AUC Score: 0.995512 Recall: 0.925329 Accuracy: 0.824169 F1:0.871824; (6) Iterate and optimize the model to achieve better prediction results; The optimization and iteration of the model in step (6) is specifically as follows: first, the anti-fraud scenarios are divided into four categories: loan financial fraud, impersonation of customer service, GOIP fraud, and silent card fraud; then the models are divided into four categories for model training and prediction respectively; and the four categories of models are continuously optimized and iterated according to the restoration of the number cards that have been shut down by the model and the complaints; In the actual deployment and application process of the anti-fraud model, data is first collected and aggregated from different platforms according to the source of the data. User characteristics are generated from the user's basic data: customer information, package status, card activation channels, consumption characteristics, behavioral data: call behavior, traffic Internet behavior, call roaming location, network data: DPI data, user location and terminal data, interface data: signaling data, trial call data, notification data, platform reporting data, and anti-fraud behavior wide table is established; the model is trained through the machine learning model algorithm and it is predicted whether the user in the anti-fraud behavior wide table has fraudulent behavior; the fraud scenarios are divided into four categories: loan financial fraud, impersonation of customer service, GOIP fraud and silent card fraud, different model training predictions are generated from the four scenarios, and the predicted user's fraud probability is divided into three risk levels of high, medium and low corresponding to three anti-fraud processing processes, and an early warning verification message is sent to the user. If the verification is passed, it means that the user's real-name authentication is passed. If the verification fails, the number will be shut down.

2. The telecommunications anti-fraud method based on machine learning according to claim 1, characterized in that: The analysis of the notification number in step (1) is specifically as follows: based on the verification and analysis of the notification number issued by the public security and 12321, the difference between normal users and fraudulent users is analyzed from the user's network access status, package consumption, historical active behavior, Internet behavior, call indicators and location information, to characterize the characteristics of the target user.

3. The telecommunications anti-fraud method based on machine learning according to claim 1, characterized in that: The data cleaning and preprocessing of step (3) are as follows: the user age, monthly billing information, and user stickiness in the data are filled with the median value, and the user billing, intra-provincial and extra-provincial traffic usage, number of voice calls, and voice call duration are filled with zero values. For abnormal values, the average value is used to correct them; the numerical value, type, and date errors in the data are corrected; for missing values, such as age, the median is used to fill them, and for monthly consumption amount, zero is filled; In the preprocessing, some categorical variables, such as the user's terminal model and the user's package name, are uniquely encoded to convert the original Chinese categories into digital information that can be processed by computers.

Citation Information

Patent Citations

  • Method for analyzing fraudulent phone number

    CN107506776A

  • Telecommunication fraud process prediction method and system based on phantom evolution

    CN111815034A