Fraud prevention and control processing method and system applied to privacy number

Abstract

The invention belongs to the technical field of communication services, and provides a fraud prevention and control processing method and system applied to privacy numbers. The method comprises: according to a screening condition, performing risk behavior screening for multiple times on bill data, including: performing feature selection on number behavior features to obtain a corresponding first feature vector as an input of a first-layer model; taking the output of the first-layer model as the input of a second-layer model to obtain a judgment result; content capturing is carried out on ticket data obtained through screening according to the judgment result so as to check whether violation occurs or not through quality inspection; determining violation ticket data according to quality inspection audit, updating the user score, and updating the customer rating at the same time; and managing and controlling the privacy number to be evaluated according to the updated user score and customer level. According to the method, the clients are comprehensively evaluated from the aspects of risk evaluation, listed clients, income contribution, code number coverage, reporting scenes, call completing rates and the like, and differentiated management and control are carried out on the clients of each level.

Description

Technical Field

[0001] The present invention relates to the field of communication service technology, and in particular to a fraud prevention and control processing method and system applied to a privacy number. Background Art

[0002] Privacy communication products are communication products that protect the privacy of both parties in communication. They solve the problems of privacy protection in interpersonal communication, data protection in the interaction between people and platforms and applications, and enterprise business management. While meeting the rigid needs of users such as "preventing harassment, privacy protection, enterprise data security, and service quality control", privacy communication services also need to protect the privacy of both parties through virtual codes. Existing privacy communication products are used in multiple industries such as online car-hailing, food delivery, and intermediaries.

[0003] However, since January 2024, privacy communication products have frequently been involved in fraud, and complaints about fraud involving middlemen such as food delivery, freight, taxis, and recruitment platforms have increased rapidly. The regulatory authorities have become increasingly strict in supervising operators' privacy communication products, which has had a huge impact on business expansion. Currently, privacy communication products only conduct quality inspections on call content by manually listening to recordings. Manual quality inspections are inefficient and the quality inspection ratio is seriously insufficient. It is impossible to effectively inspect and identify the tens of millions of call recordings of privacy communication products every day.

[0004] The following problems exist in the existing technology: the risk identification method of the privacy communication product platform is mainly manual quality inspection, which is inefficient and leads to insufficient risk identification capabilities; the privacy communication product platform currently manages customer risks in a scattered manner, and has not established an effective customer risk assessment system, and has not carried out refined and differentiated management and control of customers based on customer risk assessment; the risk management and disposal methods are single, and the privacy communication product platform has a single means of management and disposal for customers, which cannot effectively perceive, identify, discover and deal with risks. The links are scattered and a full-link, systematic dynamic risk control capability has not been formed.

[0005] Therefore, it is necessary to provide a fraud prevention and control method and system for privacy numbers to solve the above problems. Summary of the invention

[0006] The present invention aims to provide a fraud prevention and control processing method and system applied to privacy numbers, so as to solve the technical problems in the prior art such as low efficiency of manual quality inspection of privacy communication products, insufficient risk identification ability, failure to establish an effective customer risk assessment system, inability to perform refined and differentiated management and control based on customer risk assessment, single means of management and disposal of customers by the privacy communication product platform, inability to effectively perceive, identify, discover and deal with risks, scattered links, and failure to form a full-link, systematic dynamic risk control capability. The technical problems to be solved by the present invention are achieved through the following technical solutions.

[0007] The first aspect of the present invention proposes a fraud prevention and control processing method applied to privacy numbers, comprising: performing multiple risk behavior screening on call sheet data according to screening conditions, specifically including: constructing various feature tables of fraud-related number behaviors based on the original call sheets of the privacy numbers, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first layer model; using the output of the first layer model as the input of the second layer model to obtain a judgment result; performing content capture on the call sheet data screened according to the judgment result to conduct quality inspection and review to see if there are any violations; based on the call sheet data determined to be in violation of the quality inspection and review, updating the user score and simultaneously updating the customer rating; and managing the privacy numbers to be evaluated based on the updated user scores and customer levels.

[0008] The second aspect of the present invention proposes a fraud prevention and control processing system applied to privacy numbers, which adopts the fraud prevention and control processing method applied to privacy numbers described in the first aspect of the present invention. The fraud prevention and control processing system includes: a screening processing module, which performs multiple risk behavior screening on call sheet data according to screening conditions, specifically including: constructing various feature tables of fraud-related number behaviors based on the original call sheets of the privacy numbers, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first layer model; using the output of the first layer model as the input of the second layer model to obtain a judgment result; a content quality inspection module, which performs content capture on the call sheet data screened according to the judgment result to perform quality inspection and review whether it is in violation of regulations; an update processing module, which determines the call sheet data that violates regulations based on the quality inspection and review, updates the user score, and updates the customer rating at the same time; a management and control processing module, which manages the privacy numbers to be evaluated according to the updated user scores and customer levels.

[0009] The third aspect of the present invention provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the fraud prevention and control processing method applied to a privacy number as described in the first aspect of the present invention.

[0010] The fourth aspect of the present invention provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the fraud prevention and control method applied to a privacy number as described in the first aspect of the present invention.

[0011] The embodiments of the present invention include the following advantages:

[0012] Compared with the prior art, the present invention performs multiple risk behavior screening on call data according to screening conditions, and specifically creates different number behavior feature sample libraries according to different fraud-related scenarios; performs feature selection on number behavior features to obtain the corresponding first feature vector as the input of the first layer model; uses the output of the first layer model as the input of the second layer model to obtain more accurate judgment results; performs content capture on the call data screened according to the judgment results to conduct quality inspection to verify whether there are any violations, so as to update the user score and update the customer rating at the same time; manages the privacy numbers to be evaluated based on the updated user score and customer level, and can comprehensively evaluate and grade customers from aspects such as risk assessment, list-based customers, revenue contribution, number coverage, reporting scenarios, and connection rate, and establish a comprehensive customer risk assessment system, thereby improving the ability to perceive risky behaviors and verify call content.

[0013] In addition, a risk identification model is built based on business scenarios, and big data is used to screen suspected risky users and problematic call records based on the rich call attribute information in the call record data. The user's call content is checked through ASR recognition + banned word matching to determine whether there is any illegal content, so that risky users can be accurately identified.

[0014] In addition, by refining risk control and management methods, refined control can be achieved, which can effectively and specifically control users of different risk levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of an example of a fraud prevention and control method for a private number according to the present invention;

[0016] Figure 2 It is a partial step schematic diagram of an implementation method of the fraud prevention and control processing method applied to a privacy number of the present invention;

[0017] Figure 3 It is a structural block diagram of an example of a fraud prevention and control processing system applied to a privacy number of the present invention;

[0018] Figure 4 is a schematic structural diagram of an electronic device embodiment according to the present invention;

[0019] Figure 5 It is a schematic diagram of the structure of a computer-readable medium embodiment according to the present invention. DETAILED DESCRIPTION

[0020] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0021] In view of the above problems, the present invention proposes a fraud prevention and control method and system for privacy numbers. The method comprehensively evaluates customers from the aspects of risk assessment, list-based customers, income contribution, number coverage, reporting scenarios, connection rate, etc., and performs differentiated management and control for customers at each level. The method mainly optimizes and combines four major process links, including risk behavior screening, content capture and verification, risk level assessment, and intervention of control measures. The method can effectively solve the problems of insufficient risk identification and discovery capabilities of existing privacy communication platforms, single risk control and disposal methods, and scattered risk perception, identification, discovery and disposal links. It can improve the ability to perceive risk behaviors and verify call content, establish a customer risk assessment system, enrich risk control and disposal methods, and link the risk behavior screening, content capture and verification, risk level assessment, and access to control measures of privacy communications to comprehensively process privacy numbers, realizing the full process of privacy communications from risk identification to risk disposal.

[0022] It should be noted that the method of the present invention has a wide range of applications, and is particularly suitable for business scenarios mainly including food delivery, online car-hailing, express delivery, financial services and other fields. Fraud-related scenarios mainly include false orders or fraud-related orders, ticket refunds and changes, false loans, impersonating customer service and other types.

[0023] Example 1

[0024] Refer to the following Figure 1 and Figure 2 , the contents of the present invention will be described in detail.

[0025] Figure 1 It is a step flow chart of an example of the fraud prevention and control processing method applied to a privacy number of the present invention.

[0026] like Figure 1 As shown, in step S101, the call data is subjected to multiple risk behavior screenings according to the screening conditions, specifically including: constructing various feature tables of fraudulent number behaviors based on the original call records of the privacy number, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first layer model; and using the output of the first layer model as the input of the second layer model to obtain the final judgment result.

[0027] Specifically, based on the caller-called relationship, number activity, behavior event flow, and regional distribution, the number behavior features are extracted to extract the following features: calling number features, number activity features, number social network, continuous abnormal number call events, continuous unfamiliar number calls, call intervals, call duration, called number number, unfamiliar call ratio, calling location, and called number location.

[0028] Table 1,Table 1 includes feature grouping, feature name, and description.

[0029]

[0030] Table 1 is an example table of feature extraction.

[0031] Based on the original call records of private numbers, we build various feature tables of fraudulent number behaviors, and create different number behavior feature sample libraries according to different fraudulent scenarios. Among them, the types of fraudulent scenarios include fake orders, ticket refunds and changes, fake loans, and impersonating customer service.

[0032] The following takes "ticket refund and change" as an example to introduce the selection process of the first eigenvector:

[0033] Query the fraud-related numbers of "ticket refund and change" issued by relevant authorities, obtain the original call records of these numbers and some normal numbers in the past three months, obtain the feature data of Table 1 of the numbers based on the call record data, and select the categorical features (such as similar relevant departments, similar customer service, silent in the previous 14 days, Internet features, etc.) through the Pearson chi-square test. The calculation formula is as follows:

[0034]

[0035] Among them, χ 2 is the statistical value of the chi-square test; r is the number of variables of the first categorical variable X (such as Internet characteristics), and c is the number of variables of the second categorical variable Y (such as whether it is fraudulent); i,j is the observed value when the first categorical variable X is i and the second categorical variable Y is j, E i,j It is the expected value when the categorical variable X is i and the categorical variable Y is j.

[0036] The variance selection method is used to select numerical features (such as the number of daily calls, average call duration, number of recent calls, number of active days, sudden increase in traffic in the past three days, number of called numbers on the day, and proportion of long-distance calls on the day). The variance calculation formula is as follows:

[0037]

[0038] Among them, S 2 is the variance of the first categorical variable X (e.g., number of calls per day), x i is the observed value of the first categorical variable X (e.g., number of calls per day), is the expected value of the variable X (eg, the number of calls per day), n represents the number of variables of the first categorical variable X (eg, the number of calls per day), and n is a positive integer.

[0039] The basic features are constructed into new features through the feature construction method. For example, the two basic features of "earliest call time" and "latest call time" are constructed into the new feature of "abnormal work and rest time", and the three basic features of "continuous calls with abnormal numbers", "continuous calls with unfamiliar numbers" and "time interval between calls" are constructed into the new feature of "abnormal call behavior".

[0040] The feature vector selected through the above feature selection and feature construction is the first feature vector.

[0041] The above data is input into the first layer model as the first feature vector. The first layer model uses multiple basic models such as random forest, logistic regression, xgboost, etc. to identify and judge the same data. The output results of multiple models are as follows:

[0042]

[0043]

[0044] The output of the first layer model is used as the input of the second layer model. The second layer model uses the Voting algorithm of the Ensemble algorithm to make a final judgment result on the judgment results of multiple basic models in the first layer. The judgment result includes whether it is a number, whether it involves fraud, and the fraud scenario.

[0045] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0046] Next, in step S102, the call bill data obtained by filtering out the judgment result is captured for content inspection to check whether there is any violation.

[0047] Specifically, the following operations are performed on the filtered call record data to obtain the call record text: large file splitting, dual-channel splitting, file uploading, file parsing, and file word segmentation filtering.

[0048] Performing multi-type content recognition on the call slip text to obtain a quality inspection result, wherein the multi-type content recognition includes prohibited content recognition, text summary recognition, prohibited scene recognition, business scene recognition, and service attitude recognition.

[0049] For audit and quality inspection, the call data is recorded data, and the fraud model is used to determine the fraudulent or risky call data in the call data. Specifically, the filtered call data is transmitted to the corresponding program for content capture and review through Kafka. Content capture verification includes recording quality inspection program verification and manual review.

[0050] In a specific implementation method, for audit and quality inspection, large file audio splitting, audio dual-channel splitting, file uploading, audio to text conversion, call record text segmentation, and risk identification are performed.

[0051] In the large file audio splitting, listen to and receive the messages sent by the recording screening process through Kafka. Download the corresponding audio file according to the information in the message. Use the large file splitting tool to evenly split each large audio file into multiple small audio files of about 5 minutes in length. Pass all the split audio files to the audio dual-channel separation program for processing.

[0052] In audio dual-channel splitting, receive the split audio file set. Process these audio files in parallel and perform dual-channel splitting operations to extract independent audio channel data. Upload the separated audio files to the Minio cloud storage service.

[0053] Next, upload the audio files to Minio and obtain the audio files separated by two channels in batches. Upload these files to Minio and configure the storage bucket with public read and write permissions. Build a download path set based on the uploaded file name and send a message to the audio-to-text service through Kafka. Call the ASR capability to convert audio to text, monitor the messages in the Kafka queue in real time, and obtain the information of the audio files to be processed. Call the internal ASR system (whisper) in batches to convert the audio files into text content, that is, call bill text or audio text. For the converted call bill text, call a professional text segmentation program for processing. Word segmentation is a basic and key technology in natural language processing, which can divide continuous text strings into meaningful word units. After the word segmentation is completed, the word segmentation results are further filtered and optimized, and those contents in the call bill text that are not of substantial help for subsequent analysis are removed, such as modal particles (such as "ah", "oh"), punctuation marks, special symbols, etc. Through this processing, a more pure word set that focuses more on core information is obtained. Subsequently, a special risk control recognition is called for in-depth judgment and analysis.

[0054] When the risk control system receives the results from the text segmentation, it starts the analysis process through the real-time decision interface. Use the fraud-related sensitive and banned word library to compare and match the segmentation results. Evaluate whether the call content contains fraud-related sensitive or banned words. If there are fraud-related sensitive or banned words, use NLP natural language processing technology to understand the call content based on industry semantics, and determine whether it is fraud-related and the fraud-related scenario based on the specific meaning and context of the call.

[0055] Semantic understanding in NLP natural language processing technology aims to enable computers to understand and interpret the deep meaning of human language. It goes beyond simple lexical analysis and syntactic analysis and goes deep into the meaning level of words, phrases, sentences and even the entire text. The key technologies of semantic understanding mainly include word vectors, syntactic analysis, semantic role labeling, named entity recognition, semantic similarity, etc.

[0056] Word embedding is a technique that converts words into points or vectors in a vector space that can capture the semantic relationship between words. Similar words are close to each other in the vector space, thus achieving a quantitative representation of semantics; Syntactic analysis is the process of converting the structure of a sentence into a computer-understandable form. It helps determine the relationship and role between different words in a sentence, such as subject, object, predicate, etc., providing a basis for subsequent semantic understanding; Semantic role labeling is a technique that labels the semantic roles of different words in a sentence. It focuses on the relationship between verbs and their arguments (such as agents, patients, etc.) to help understand the semantic structure in a sentence; Named Entity Recognition is a text analysis technique that aims to identify entities with specific meanings in a text (such as names of people, places, names of organizations, etc.), which is crucial to understanding the theme and content of the text; Semantic similarity is a technique that measures the degree of semantic similarity between two sentences or words. It helps determine the relevance and difference between different elements in a text.

[0057] In addition, in the entire recording quality inspection process, the call record serial number is used as a unique identifier to connect various links such as file processing, file parsing, text segmentation, risk control identification, ASR conversion and model application to track the process and provide a query function based on the call record serial number, allowing users or system administrators to quickly view and track the processing results and status of any process step.

[0058] In addition, manual review, as a supplementary means to the recording quality inspection procedure, requires the display of the audio converted by the previous recording quality inspection procedure, and further manual verification and correction of the quality inspection results.

[0059] During the manual quality inspection process, for example, click the recording icon, send a background request, query the recording address corresponding to the call record in the call record table, download the recording to the local, return it to the front end in a streaming manner, click the player and automatically play the recording for manual quality inspection.

[0060] By clicking the recording text icon, you can query the corresponding translation result of the call record in the quality inspection result table, return the corresponding text content to the front end, and pop up the recording text page.

[0061] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0062] Next, in step S103, the user score is updated based on the call record data that is determined to be in violation of the quality inspection, and the customer rating is updated at the same time.

[0063] Specifically, the user rating is updated based on the call record data that is determined to be in violation of the regulations based on the quality inspection review.

[0064] According to the first evaluation model, the following expression is used to calculate the comprehensive score of each privacy number:

[0065] Z=∑z i *qi

[0066] Among them, Z represents the comprehensive score of each privacy number; i represents the score of the ith evaluation feature, and the evaluation features include whether the content is in violation of the rules, whether the call scenario is in violation of the rules, batch registration, association with Internet accounts, fraud, harassment, and basic score, where i is a positive integer; qi represents the weight corresponding to the ith evaluation feature.

[0067] The user score is updated based on the comprehensive score of each privacy account calculated in real time.

[0068] Next, the second evaluation model is used to calculate the initial value of the customer level corresponding to each privacy number.

[0069] According to the call record data with violations determined by quality inspection and review, the model input for calculating the initial value of the customer grade is updated to re-determine the initial value of the customer grade, and the customer grade is updated again according to the updated user score.

[0070] In a specific implementation, user scoring is to evaluate the privacy account from aspects such as whether the content is illegal, scenario compliance, etc., and a score card model is used for comprehensive scoring. The following expression is used to represent the user score card model:

[0071] Z'=∑z i '*qi'

[0072] Among them, Z' represents the comprehensive score of each privacy number; z i ' represents the comprehensive score of the i'th evaluation feature, and the evaluation features include whether the content is in violation of the rules, whether the call scenario is in violation of the rules, batch registration, associated Internet accounts, fraud, harassment, and basic score, and i is a positive integer; qi' represents the weight corresponding to the i'th evaluation feature.

[0073] It should be noted that in this example, the comprehensive score z corresponding to each feature is i Both ' and weight qi' are generated by the user score card model during the model training process.

[0074] Take user ratings as an example. Figure 2As shown in the figure, the comprehensive score and weight generation process of each feature is as follows:

[0075] Step S201: preparing training data for training a user score card model.

[0076] Specifically, the training data includes numbers marked with the following multiple labels: number rating, whether the call content is in violation of regulations, whether the call scenario is in violation of regulations, whether it is registered in batches, whether it is associated with an Internet account, whether it involves fraud, whether it involves harassment, and whether there is a basic score.

[0077] Step S202: Generate a comprehensive score and weight corresponding to each feature during model training.

[0078] Specifically, the training data is imported into the user score card model for model training, and the comprehensive score z corresponding to each feature is calculated by the user score card model. i 'and weights qi'.

[0079] Step S203: Score and rate each feature, and verify whether each feature and weight is reasonable.

[0080] After determining the weights corresponding to each feature, the training data is calculated according to the model to generate the comprehensive score z i 'And weight qi' are used to score and rate, and verify whether the comprehensive score and weight value of each feature are reasonable.

[0081] If not, repeat step S201 and step S202 until the comprehensive scores and weight values ​​of each feature are reasonable. For the specific grouping, comprehensive scores and weight values ​​of each feature, see Table 2.

[0082]

[0083]

[0084] After obtaining the comprehensive score Z', the users are divided into low-risk users, medium-risk users and high-risk users according to the comprehensive score Z', wherein the threshold of each level is calculated by the comprehensive score and level of the training number.

[0085] According to the above training data, the comprehensive scores of high-risk numbers are all less than 80 points, the comprehensive scores of medium-risk numbers are between 90 and 100 points, and the comprehensive scores of low-risk numbers are between 110 and 120 points. Therefore, it can be concluded that users with a comprehensive score of less than 80 points are high-risk users, users with a comprehensive score of ≥80 and less than 100 points are medium-risk users, and users with a comprehensive score of ≥100 points are low-risk users.

[0086] It should be noted that in this example, the implementation method of customer rating is roughly the same as that of user rating. The difference is that the characteristics of customer rating include risk assessment, list-based customers, revenue contribution, number coverage, reporting scenario, connection rate, etc., that is, there are different characteristic data. Therefore, the description of the same parts is omitted.

[0087] Next, in step S104, the privacy number to be evaluated is managed and controlled according to the updated user score and customer level.

[0088] In a specific implementation, specific control measures are taken on the privacy number and the customer based on the updated user rating (for example, 2.7) and the customer rating.

[0089] For example, different fraud control measures are specified for users and customers, and the users and customers that need to be controlled are managed in the [Black and Gray List] menu. The [Black and Gray List] menu supports black and gray list query and disposal operations. The initial query list on this page does not display content. After querying by conditions, the list displays the corresponding content according to the filter conditions.

[0090] Specifically, the blacklist and graylist types on this page include user blacklist and customer blacklist. The user blacklist shows users with high-risk ratings, and the customer blacklist shows customers with D ratings. Measures such as number freezing or number removal are taken for blacklist users, and punitive measures such as concurrent halving and number application restriction are taken for blacklist customers. In addition, it also includes the addition of number-level call interception, call restriction, number removal and other control measures, and customer-level business warnings, concurrent restrictions, customer removal, customer blacklist and other control measures.

[0091] By refining risk control and management methods, we can achieve refined management and control, and effectively and specifically manage users at different risk levels.

[0092] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0093] In addition, the drawings are only schematic illustrations of the processes included in the method according to an exemplary embodiment of the present invention, and are not intended to be limiting. It is easy to understand that the processes shown in the drawings do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0094] Compared with the prior art, the present invention performs multiple risk behavior screening on call data according to screening conditions, and specifically creates different number behavior feature sample libraries according to different fraud-related scenarios; performs feature selection on number behavior features to obtain the corresponding first feature vector as the input of the first layer model; uses the output of the first layer model as the input of the second layer model to obtain more accurate judgment results; performs content capture on the call data screened according to the judgment results to conduct quality inspection to verify whether there are any violations, so as to update the user score and update the customer rating at the same time; manages the privacy numbers to be evaluated based on the updated user score and customer level, and can comprehensively evaluate and grade customers from aspects such as risk assessment, list-based customers, revenue contribution, number coverage, reporting scenarios, and connection rate, and establish a comprehensive customer risk assessment system, thereby improving the ability to perceive risky behaviors and verify call content.

[0095] In addition, a risk identification model is built based on business scenarios, and big data is used to screen suspected risky users and problematic call records based on the rich call attribute information in the call record data. The user's call content is checked through ASR recognition + banned word matching to determine whether there is any illegal content, so that risky users can be accurately identified.

[0096] In addition, by refining the risk control and management methods, refined control can be achieved, which can effectively and specifically control users with different risk levels.

[0097] Example 2

[0098] The following is a system embodiment of the present invention, which can be used to implement the method embodiment of the present invention. For details not disclosed in the system embodiment of the present invention, please refer to the method embodiment of the present invention.

[0099] Figure 3 It is a structural diagram of an example of a fraud prevention and control processing system applied to a privacy number according to the present invention.

[0100] Reference Figure 3 The second aspect of the present disclosure provides a fraud prevention and control processing system 300 applied to a privacy number, which adopts the fraud prevention and control processing method described in the first aspect of the present invention.

[0101] Specifically, the fraud prevention and control processing system 300 includes a screening processing module 310 , a content quality inspection module 320 , an update processing module 330 and a management and control processing module 340 .

[0102] In a specific embodiment, the screening processing module 310 performs multiple risk behavior screening on the call slip data according to the screening conditions, specifically including: constructing various feature tables of fraudulent number behaviors based on the original call slips of the privacy number, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first layer model; using the output of the first layer model as the input of the second layer model to obtain the judgment result; the content quality inspection module 320 performs content capture based on the call slip data screened according to the judgment result to conduct quality inspection and review to see if there are any violations. The update processing module 330 updates the user score and customer rating at the same time based on the call slip data that has been determined to be in violation of the quality inspection. The management and control processing module 340 manages the privacy number to be evaluated based on the updated user score and customer level.

[0103] According to the relationship between the caller and the called party, number activity, behavior event flow, and geographical distribution, the number behavior characteristics are extracted to extract the following features: calling number characteristics, number activity characteristics, continuous abnormal number call events, continuous unfamiliar number calls, call intervals, call duration, number of called numbers, unfamiliar call ratio, calling number location, and called number location.

[0104] According to an optional implementation manner, the chi-square test, variance selection, and feature construction are used to perform feature discretization processing and selection calculation to obtain the corresponding first feature vector, which is used as the input of the first layer model. The output of the first layer model is used as the input of the second layer model to obtain a judgment result, which includes whether the number is a fraudulent number and the type of fraudulent number.

[0105] Use random forest algorithm, logistic regression algorithm or xgboost algorithm to build the first layer model.

[0106] The Ensemble algorithm is used to build the second layer model.

[0107] According to an optional implementation manner, the call bill data screened according to the determination result is subjected to content capture to conduct a quality inspection to check whether there is any violation.

[0108] The following operations are performed on the screened call record data to obtain the call record text: large file splitting, dual-channel splitting, file uploading, file parsing, and file word segmentation filtering; multi-type content recognition is performed on the call record text to obtain quality inspection and audit results. The multi-type content recognition includes prohibited content recognition, text summary recognition, prohibited scenario recognition, business scenario recognition, and service attitude recognition.

[0109] According to an optional implementation manner, updating the user score according to the call record data that is determined to be in violation of the quality inspection review includes:

[0110] According to the first evaluation model, the following expression is used to calculate the comprehensive score of each privacy number:

[0111] Z=∑z i *qi

[0112] Among them, Z represents the comprehensive score of each privacy number; i It represents the score of the ith evaluation feature, and the evaluation features include whether the content is in violation of the rules, whether the call scenario is in violation of the rules, batch registration, associated Internet accounts, fraud, harassment, and basic score, where i is a positive integer; qi represents the weight corresponding to the ith evaluation feature; the user score is updated according to the comprehensive score of each privacy number calculated in real time.

[0113] According to an optional implementation manner, the simultaneous updating of customer ratings includes: using a second evaluation model to calculate an initial value of the customer grade corresponding to each privacy number.

[0114] According to the call record data with violations determined by quality inspection and review, the model input for calculating the initial value of the customer grade is updated to re-determine the initial value of the customer grade, and the customer grade is updated again according to the updated user score.

[0115] Compared with the prior art, the present invention performs multiple risk behavior screening on call data according to screening conditions, and specifically creates different number behavior feature sample libraries according to different fraud-related scenarios; performs feature selection on number behavior features to obtain the corresponding first feature vector as the input of the first layer model; uses the output of the first layer model as the input of the second layer model to obtain more accurate judgment results; performs content capture on the call data screened according to the judgment results to conduct quality inspection to verify whether there are any violations, so as to update the user score and update the customer rating at the same time; manages the privacy numbers to be evaluated based on the updated user score and customer level, and can comprehensively evaluate and grade customers from aspects such as risk assessment, list-based customers, revenue contribution, number coverage, reporting scenarios, and connection rate, and establish a comprehensive customer risk assessment system, thereby improving the ability to perceive risky behaviors and verify call content.

[0116] In addition, a risk identification model is built based on business scenarios, and big data is used to screen suspected risky users and problematic call records based on the rich call attribute information in the call record data. The user's call content is checked through ASR recognition + banned word matching to determine whether there is any illegal content, so that risky users can be accurately identified.

[0117] In addition, by refining risk control and management methods, refined control can be achieved, which can effectively and specifically control users of different risk levels.

[0118] Figure 4is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

[0119] like Figure 4 As shown, the electronic device is presented in the form of a general computing device. The processor may be one or more and work in coordination. The present invention does not exclude distributed processing, that is, the processor may be dispersed in different physical devices. The electronic device of the present invention is not limited to a single entity, but may also be the sum of multiple physical devices.

[0120] The memory stores a computer executable program, which is usually a machine-readable code. The computer-readable program can be executed by the processor to enable the electronic device to perform the method of the present invention, or at least part of the steps in the method.

[0121] The memory includes a volatile memory, such as a random access memory unit (RAM) and / or a cache memory unit, and may also be a non-volatile memory, such as a read-only memory unit (ROM).

[0122] Optionally, in this embodiment, the electronic device further includes an I / O interface, which is used for the electronic device to exchange data with an external device. The I / O interface can represent one or more of several types of bus structures, including a storage unit bus or a storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus structures.

[0123] It should be understood that Figure 4 The electronic device shown is only an example of the present invention, and the electronic device of the present invention may also include elements or components not shown in the above examples. For example, some electronic devices also include display units such as display screens, and some electronic devices also include human-computer interaction elements such as buttons, keyboards, etc. As long as the electronic device can execute the computer-readable program in the memory to implement the method of the present invention or at least part of the steps of the method, it can be considered as an electronic device covered by the present invention.

[0124] Through the above description of the implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by combining software with necessary hardware. Figure 5 As shown, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes a number of commands to enable a computing device (which can be a personal computer, a server, or a network device, etc.) to execute the above method according to the embodiment of the present invention.

[0125] The software product may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0126] The computer readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with a command execution system, device, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0127] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0128] The computer-readable medium carries one or more programs (eg, computer-executable programs). When the one or more programs are executed by a device, the computer-readable medium implements the method of the present disclosure.

[0129] Those skilled in the art will appreciate that the above modules can be distributed in the device according to the description of the embodiment, or can be changed accordingly and only used in one or more devices different from the embodiment. The modules of the above embodiments can be combined into one module, or further divided into multiple sub-modules.

[0130] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several commands to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiment of the present invention.

[0131] The exemplary embodiments of the present invention are specifically shown and described above. It should be understood that the present invention is not limited to the detailed structure, configuration or implementation method described herein; on the contrary, the present invention is intended to cover various modifications and equivalent configurations included in the spirit and scope of the appended claims.

Claims

1. A fraud prevention and control method applied to a private number, characterized in that: include: According to the screening conditions, the call data is screened for risk behaviors multiple times, including: constructing various feature tables of fraudulent number behaviors based on the original call records of the privacy number, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first-layer model; using the output of the first-layer model as the input of the second-layer model to obtain the judgment result; The call bill data obtained by screening according to the judgment results is captured for quality inspection to check whether there are any violations; Update the user score and customer rating based on the call record data that has been identified as illegal by quality inspection; The privacy accounts to be evaluated are managed and controlled based on the updated user ratings and customer levels.

2. The fraud prevention and control method according to claim 1 is characterized in that: include: According to the relationship between the caller and the called party, number activity, behavior event flow, and geographical distribution, the number behavior characteristics are extracted to extract the following features: calling number characteristics, number activity characteristics, continuous abnormal number call events, continuous unfamiliar number calls, call intervals, call duration, number of called numbers, unfamiliar call ratio, calling number location, and called number location.

3. The fraud prevention and control method according to claim 2 is characterized in that: Further including: Using chi-square test, variance selection, and feature construction, feature discretization processing is performed, and the corresponding first feature vector is obtained by selection calculation, and the first feature vector is used as the input of the first layer model; The output of the first layer model is used as the input of the second layer model to obtain a determination result, wherein the determination result includes whether the number is a fraudulent number and the type of the fraudulent number.

4. The fraud prevention and control method according to claim 3 is characterized in that: Further including: Use random forest algorithm, logistic regression algorithm or xgboost algorithm to build the first layer model; The Ensemble algorithm is used to build the second layer model.

5. The fraud prevention and control method according to claim 1 is characterized in that: The process of capturing the contents of the call list data obtained by screening according to the determination result to check whether there is any violation includes: The following operations are performed on the filtered call record data to obtain the call record text: large file splitting, dual-channel splitting, file uploading, file parsing, and file word segmentation filtering; Performing multi-type content recognition on the call slip text to obtain a quality inspection result, wherein the multi-type content recognition includes prohibited content recognition, text summary recognition, prohibited scene recognition, business scene recognition, and service attitude recognition.

6. The fraud prevention and control method according to claim 1 is characterized in that: The updating of the user rating according to the call record data that is determined to be in violation of the quality inspection includes: According to the first evaluation model, the following expression is used to calculate the comprehensive score of each privacy number: Z=∑z i *qi Among them, Z represents the comprehensive score of each privacy number; i represents the score of the ith evaluation feature, where the evaluation features include whether the content is in violation of the rules, whether the call scenario is in violation of the rules, batch registration, associated Internet accounts, fraud, harassment, and basic score, and i is a positive integer; qi represents the weight corresponding to the ith evaluation feature; The user score is updated based on the comprehensive score of each privacy account calculated in real time.

7. The fraud prevention and control method according to claim 1 or 6, characterized in that: The customer rating is updated at the same time, including: Using the second evaluation model, calculate the initial value of the customer level corresponding to each privacy number; According to the call record data with violations determined by quality inspection and review, the model input for calculating the initial value of the customer grade is updated to re-determine the initial value of the customer grade, and the customer grade is updated again according to the updated user score.

8. A fraud prevention and control system applied to private numbers, characterized in that: It implements the fraud prevention and control processing method applied to a privacy number according to any one of claims 1 to 7, and the fraud prevention and control processing system includes: The screening processing module performs multiple risk behavior screening on the call data according to the screening conditions, specifically including: constructing various feature tables of fraudulent number behaviors based on the original call records of the privacy number, and creating different number behavior feature sample libraries according to different fraud-related scenarios; performing feature selection on the number behavior features to obtain the corresponding first feature vector as the input of the first layer model; using the output of the first layer model as the input of the second layer model to obtain the judgment result; The content quality inspection module captures the content of the call record data filtered out according to the judgment results to check and verify whether there are any violations; Update processing module, update user scores and customer ratings based on the call record data that has been identified as illegal by quality inspection; The management and control processing module manages the privacy numbers to be evaluated based on the updated user ratings and customer levels.

9. The fraud prevention and control processing system according to claim 8 is characterized in that: Using chi-square test, variance selection, and feature construction, feature discretization processing is performed, and the corresponding first feature vector is obtained by selection calculation, and the first feature vector is used as the input of the first layer model; The output of the first layer model is used as the input of the second layer model to obtain a determination result, wherein the determination result includes whether the number is a fraudulent number and the type of the fraudulent number.

10. The fraud prevention and control processing system according to claim 8, characterized in that: Based on the caller-called relationship, number activity, behavior event flow, and regional distribution, the number behavior features are extracted to extract the following features: Caller number characteristics, number activity characteristics, consecutive abnormal number call events, consecutive unfamiliar number calls, call intervals, call duration, number of called numbers, unfamiliar call ratio, caller location, called number location.

Images