Big data real-time processing system

Abstract

The application discloses a big data real-time processing system, and a running method thereof comprises the following steps: step one, collecting multi-source fraud data through a data collection module; step two, extracting multi-features from the multi-source data and creating a feature matrix; step three, constructing a fraud detection model through a long short-term memory network; step four, detecting fraud behaviors by analyzing user calls; and step five, realizing real-time fraud detection and model optimization through user feedback, wherein the fraud data collection and preprocessing module is used for obtaining fraud case data from multiple sources, and performing cleaning and text preprocessing; and the user call analysis and real-time feedback module is used for analyzing user call behaviors, real-time monitoring of fraud situations, real-time fraud detection through user-authorized audio recording permission, and providing instant feedback; and the application has the characteristics of more accurate fraud detection and better real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a big data real-time processing system. BACKGROUND

[0002] The current fraud detection technology mainly focuses on the analysis of call numbers and home locations, and uses the methods of establishing rules and statistical models to identify potential fraudulent behaviors. This traditional method has significant limitations in dealing with fraud activities involving private numbers.

[0003] In the field of communication, fraudsters usually use private numbers for fraudulent activities, and by frequently changing numbers and avoiding traditional number analysis methods, the traditional fraud detection method is difficult to effectively intercept such behaviors. The use of private numbers increases the anonymity of fraudsters, making the traditional fraud detection method helpless. In addition, the traditional rules and statistical models have poor adaptability to emerging fraud methods, and often require manual updating of the rule library, resulting in lag in fraud detection. Therefore, the traditional technology is relatively powerless in the face of constantly changing and innovative fraud methods. Therefore, it is necessary to design a big data real-time processing system for fraud detection that is more accurate and has better real-time performance. SUMMARY

[0004] The present application aims to provide a big data real-time processing system to solve the problems raised in the background art.

[0005] In order to solve the above technical problems, the present application provides the following technical solution: a big data real-time processing system, the operation method of the system comprising the following steps:

[0006] Step one: multi-source fraud data collection through a data collection module;

[0007] Step two: multi-feature extraction of multi-source data and creation of a feature matrix;

[0008] Step three: construction of a fraud detection model through a long short-term memory network;

[0009] Step four: fraud behavior detection through analysis of user calls;

[0010] Step five: real-time fraud detection and model optimization through user feedback.

[0011] According to the above technical solution, the step of collecting multi-source fraud data through a data collection module comprises:

[0012] The system uses a data collection module and network crawler technology to automatically search, crawl and extract relevant fraud cases on the network, including but not limited to fraud phone recordings, fraud analysis reports, fraud case databases, and can also be integrated with various data sources through API interface. Subsequently, the collected data is cleaned, first by comparing key fields such as case ID and timestamp to remove duplicate data, and for data fields containing text information, text data preprocessing is performed to reduce text noise and improve text data quality through word segmentation, stop word removal, and stem extraction techniques.

[0013] According to the above technical solution, the step of multi-feature extraction and creation of a feature matrix from multi-source data includes:

[0014] Selecting a speech recognition engine for text conversion and building an emotion vocabulary for sentiment analysis;

[0015] Using the PyDub speech processing library for tone and speed analysis.

[0016] According to the above technical solution, the step of selecting a speech recognition engine for text conversion and building an emotion vocabulary for sentiment analysis includes:

[0017] Selecting a commercial speech recognition engine or open source engine to convert the collected fraud audio data into text, then performing sentiment analysis on the converted text. First, build an emotion vocabulary table that contains words and phrases related to different emotions. Then, split the text information into individual words. For each word, use the built emotion vocabulary table to match and identify which words belong to positive, negative, or neutral emotions. Finally, assign an emotion score to each word match result and obtain the overall emotional features. In addition, when performing sentiment recognition, speaker recognition using speech signal processing is required, which identifies the fraudster and the victim.

[0018] According to the above technical solution, the step of using the PyDub speech processing library for tone and speed analysis includes:

[0019] The prosody and speech rate analysis of the voice data is as follows: first, the voice data is loaded by using the PyDub voice processing library, then the voice signal is framed and divided into short-time frames, the fundamental frequency analysis of each frame is needed in the prosody analysis, the periodicity of the frequency spectrum is estimated by using the Fourier transform method, and then the fundamental frequency of the voice signal, i.e. the prosody, is obtained, at the same time, the prosody intensity analysis is performed, the amplitude or amplitude spectrum of the fundamental frequency is calculated to obtain the intensity information of the prosody, which provides insight into the emotional state of the speaker, on the other hand, the frame rate of the voice signal, i.e. the speech rate, is calculated by calculating the time interval between frames, the specific implementation method is as follows: the position of the frame in the entire voice signal, i.e. the time stamp, is calculated by the sampling rate and the position of the frame, then the time difference between the frames is obtained by subtracting the time stamps of adjacent frames, and finally the frame rate, i.e. the number of voice frames generated by the speaker in one second, is calculated by taking the reciprocal of the time difference, the system also calculates the average value of the frame time interval: and the variance:

[0020] The speech rate change trend of the speaker in different time periods can be obtained, in addition, the timing feature and feature vocabulary extraction need to be performed.

[0021] According to the technical scheme, the step of constructing the fraud detection model by selecting the long short-term memory network comprises:

[0022] The LSTM model parameters are set and the binary cross-entropy is used as the loss function and the Adam optimizer is used for model parameter adjustment;

[0023] The early stopping method is introduced to prevent overfitting.

[0024] According to the technical scheme, the step of setting the LSTM model parameters and using the binary cross-entropy as the loss function and the Adam optimizer for model parameter adjustment comprises:

[0025] The design of the model aims to capture the speech features and emotional information in the audio segment to provide support for accurate identification of fraud cases, specifically, a model containing two layers of LSTM units is constructed, each layer contains 50 hidden units, in the feature input stage, the timing information of each audio segment is prepared and arranged into a 3D tensor with the shape of time step, feature number, and the time step of each audio segment needs to be set to 50, in addition, in the optimization of model parameters, the binary cross-entropy is selected as the loss function and the Adam optimizer is used, the initial learning rate is set to 0.001 and the learning rate decay is used, the decay rate is set to 0.9, in the training process, the system divides the data into 80% training set and 20% validation set, the batch size used in each training is 32, and the total training iteration number is 50.

[0026] According to the technical scheme, the step of detecting fraud behavior by analyzing the user conversation comprises:

[0027] The system will also use communication pattern recognition technology to analyze the frequency, duration, and location of the conversation. By recognizing these communication patterns, the system can identify abnormal patterns that deviate from the user's normal behavior patterns, and then infer possible fraud scenarios. When the system identifies that the user may be subject to fraud through the sensitivity detection mechanism, it will trigger a prompt for the user to authorize the recording permission. This prompt will be presented to the user in real-time notification. When the system prompts the user, the user needs to authorize the recording. After the user completes the authorization, the system will transmit the user's recording data to the fraud detection module. Finally, encryption technology and security protocols are used to ensure that the user's recording data is fully protected during transmission and storage, preventing information leakage and misuse.

[0028] According to the technical scheme, the step of implementing real-time fraud detection and model optimization through user feedback comprises:

[0029] Real-time processing of audio data and application of previously trained fraud detection algorithms;

[0030] User provides multi-dimensional feedback, including confirming fraud behavior and reporting false positives;

[0031] Adopting machine learning model updating strategy to adjust weights and feature weights;

[0032] Monitoring performance indicators such as accuracy and recall rate;

[0033] Using confusion matrix to evaluate the classification effect of the model on fraud cases.

[0034] According to the technical scheme, the system comprises:

[0035] Fraud data collection and preprocessing module, for obtaining fraud case data from multiple sources, cleaning and text preprocessing to prepare data for subsequent fraud detection model training;

[0036] Fraud detection model construction module, for selecting and constructing fraud detection model, model training and parameter optimization to improve the accuracy of fraud cases;

[0037] User conversation analysis and real-time feedback module, for analyzing user conversation behavior, real-time monitoring of possible fraud scenarios, real-time fraud detection through user authorized recording permission, and providing immediate feedback, including risk prompt and user learning process.

[0038] Compared with the prior art, the present application has the beneficial effects that: firstly, the present application acquires fraud case data from multiple sources through a data acquisition module, including fraud phone recordings and fraud analysis reports, then performs multi-feature extraction on the multi-source data to create a feature matrix, next, a fraud detection model is constructed by selecting a long short-term memory network (LSTM) to train and optimize the extracted features, then, the user's call is analyzed through behavior analysis and communication mode recognition algorithms to establish a user normal behavior model and monitor abnormal patterns in real time, when the system detects that the user may be subject to fraud, a prompt to authorize the user to record the call is triggered, after the user's authorization, fraud behavior detection is performed, finally, real-time fraud detection and model optimization are realized through user feedback, users can provide different types of feedback, and the system uses machine learning models to update strategies to continuously improve, ensuring that the model more accurately identifies new fraud methods, and the method can effectively detect fraudulent calls and has higher real-time performance. BRIEF DESCRIPTION OF DRAWINGS

[0039] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the principles of the application, and do not constitute a limitation of the application. In the drawings:

[0040] Figure 1 A flowchart of a big data real-time processing method provided for the first embodiment of the present application;

[0041] Figure 2 A module composition schematic diagram of a big data real-time processing system provided for the second embodiment of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] Embodiment one: Figure 1 A flowchart of a big data real-time processing method provided for the first embodiment of the present application, the present embodiment can be applied to the scene of fraud information identification, and the method can be executed by a big data real-time processing system provided by the present embodiment, as shown in Figure 1 The method specifically includes the following steps:

[0044] Step one: multi-source fraud data acquisition through a data acquisition module;

[0045] In the embodiments of the present application, a data collection module is established in the system, which obtains fraud case data from multiple sources, including but not limited to fraud phone recordings, fraud analysis reports;

[0046] For example, the system uses the data collection module and uses web crawler technology to automatically search, crawl and extract relevant fraud cases on the network, including but not limited to fraud phone recordings, fraud analysis reports, fraud case databases, etc. It can also be integrated with various data sources through API interface. Then, the collected data is cleaned. First, data deduplication is achieved by comparing key fields such as case ID and timestamp. For data fields containing text information, text data preprocessing is performed to reduce text noise and improve the quality of text data through techniques such as word segmentation, stop word removal, and stem extraction.

[0047] Step two: multi-feature extraction is performed on the multi-source data, and a feature matrix is created;

[0048] In the embodiments of the present application, a commercial speech recognition engine or an open source engine is selected to convert the collected fraud audio data into text. Then, sentiment analysis is performed on the converted text. First, a sentiment lexicon is constructed, which contains words and phrases related to different emotions. Then, the text information is segmented into individual words. For each word, the built sentiment lexicon is used to match and identify which words belong to positive, negative or neutral emotions. Finally, a sentiment score is assigned to each word matching result, and the overall sentiment feature is obtained. In addition, when performing sentiment recognition, speaker recognition using speech signal processing is required, i.e. identifying the fraudster and the victim;

[0049] For example, pitch and speech rate analysis is performed on the speech data. First, the PyDub speech processing library is used to load the recording data. Then, the speech signal is framed and segmented into short frames. In pitch analysis, each frame needs to be analyzed for fundamental frequency. Fourier transform method is used to estimate the periodicity of the frequency spectrum, and then the fundamental frequency of the speech signal, i.e. the pitch, is obtained. At the same time, pitch intensity analysis is performed by calculating the amplitude or amplitude spectrum of the fundamental frequency to obtain the intensity information of the pitch, providing insight into the speaker's emotional state. On the other hand, by calculating the time interval between frames, the frame rate of the speech signal, i.e. the speech rate, is calculated. The specific implementation method is as follows: the position of the frame in the entire speech signal is calculated by sampling rate and frame position, i.e. the timestamp. Then, the time difference between adjacent frames is obtained by subtracting the timestamps of adjacent frames. Finally, the frame rate is calculated by taking the inverse of the time difference, i.e. the number of speech frames produced by the speaker in one second. The system also calculates the average value of the frame time interval: and variance: The speaker's speech rate change trend in different time periods can be obtained, and timing features and feature words need to be extracted. After feature extraction, a feature matrix is created, and the features extracted from the tone, speech rate, and emotion analysis are arranged into a table to ensure that each feature occupies a column. The features are standardized or normalized to ensure that they have similar scales. Next, the feature data is converted into a numerical matrix to ensure data consistency. Through this step, the feature information of the fraud data can be extracted to better train the model.

[0050] Step three: build a fraud detection model by selecting a long short-term memory network;

[0051] In the embodiments of the present application, a long short-term memory network (LSTM) is selected to build an effective fraud detection model.

[0052] For example, the design of the model aims to capture speech features and emotional information in the audio segment to support accurate identification of fraud cases. Specifically, a model containing two layers of LSTM units is built, each layer containing 50 hidden units. In the feature input stage, the timing information of each audio segment is prepared and arranged into a 3D tensor with a shape of time steps, feature numbers. The time steps of each audio segment also need to be set to 50. In addition, in optimizing the model parameters, binary cross-entropy is selected as the loss function, and the Adam optimizer is used with an initial learning rate of 0.001 and a learning rate decay of 0.9. During training, the system divides the data into 80% training set and 20% validation set, and the batch size used for each training is 32, with a total of 50 training iterations.

[0053] For example, to prevent overfitting, the system introduces early stopping, monitors the validation set loss, and automatically stops the training process when the validation set loss does not decrease for 5 consecutive iterations. In the model monitoring and evaluation stage, the system uses performance indicators such as accuracy and recall to monitor the system. Through tools such as TensorBoard, the system can visualize the trends of the loss function and accuracy. Finally, the confusion matrix is generated using the test set to evaluate the classification effect of the model on fraud cases. After training, the model parameters, including weights and biases, need to be saved, and a model version management mechanism is established to track and manage different versions of the model. This step ensures the effectiveness and controllability of the LSTM model in fraud case detection tasks.

[0054] Step four: analyze user calls to detect fraud;

[0055] In the embodiment of the present application, the system adopts behavior analysis algorithm, through real-time monitoring and analysis of multi-dimensional data such as user dialing mode, call duration, and frequently used contacts, a normal behavior model of the user is established, and through machine learning algorithm, the system can gradually learn the normal behavior pattern of the user, forming a benchmark understanding of normal communication behavior.

[0056] For example, the system will also adopt communication mode recognition technology to analyze the frequency of calls, call time period, and call location information. Through the recognition of these communication modes, the system can identify abnormal patterns that deviate from the user's normal behavior patterns, and then infer possible fraud scenarios. When the system identifies that the user may be subject to fraud through the sensitivity detection mechanism, it will trigger a prompt for the user to authorize the recording permission. This prompt will be presented to the user in the form of an instant notification. When the system prompts the user, the user needs to authorize the recording. After the user completes the authorization, the system will transmit the user's recording data to the fraud detection module. Finally, encryption technology and security protocols are used to ensure that the user's recording data is fully protected during transmission and storage to prevent information leakage and misuse.

[0057] Step five: Real-time fraud detection and model optimization through user feedback.

[0058] In the embodiment of the present application, the acquired audio data will be sent to the fraud detection model for real-time processing and analysis. In this stage, the system will preprocess the audio data, including speech signal conversion, feature extraction, etc. Then, the model will apply the fraud detection algorithm trained in the previous steps, including keyword recognition, tone analysis, emotion feature extraction, etc., to identify potential fraudulent behavior. After the model detection is completed, the system will provide real-time feedback based on the model output results. If the model believes that there are signs of fraudulent behavior in the call, the system will trigger the corresponding early warning mechanism, such as notifying the user immediately, recording fraud behavior data, etc.

[0059] For example, after receiving the risk prompt, the user can choose different types of feedback, such as confirming whether fraudulent behavior has been encountered, reporting false positives, or providing additional information. These feedback types provide multi-dimensional information for the system's learning process. Then, the system uses a machine learning model update strategy to continuously improve using the collected user feedback data, which includes weight updates and feature weight adjustments to ensure that the model can more accurately identify new fraud methods.

[0060] Embodiment two: The second embodiment of the present application provides a big data real-time processing system, Figure 2 The module composition diagram of the big data real-time processing system provided in the second embodiment of the present application is shown in Figure 2 The system includes:

[0061] Fraud data collection and preprocessing module, for obtaining fraud case data from multiple sources, cleaning and text preprocessing to prepare data for subsequent fraud detection model training;

[0062] Fraud detection model construction module, for selecting and constructing fraud detection model (such as LSTM), model training and parameter optimization to improve the accuracy of fraud cases;

[0063] User call analysis and real-time feedback module, for analyzing user call behavior, real-time monitoring of possible fraud situations, real-time fraud detection through user authorized recording permission, and providing immediate feedback, including risk prompt and user learning process;

[0064] In some embodiments of the present application, the fraud data collection and preprocessing module comprises:

[0065] Data collection module, for obtaining fraud case data from multiple sources through the built-in data collection module of the system;

[0066] Data cleaning and text preprocessing module, for cleaning the collected data, removing duplicates by comparing key fields, and preprocessing text information containing fields;

[0067] Multi-feature extraction and feature matrix creation module, for multi-feature extraction of multi-source data and arrangement of extracted features into a feature matrix;

[0068] In some embodiments of the present application, the fraud detection model construction module comprises:

[0069] LSTM model selection and construction module, for selecting long short-term memory network (LSTM) as the fraud detection model, and constructing a model containing two layers of LSTM units;

[0070] Model training and parameter optimization module, for model training using fraud case data, using optimization algorithms to adjust model parameters to improve the recognition accuracy of fraud cases;

[0071] Model application and version management module, for applying the trained LSTM model for fraud detection, saving the parameters and architecture of the model, and establishing a model version management mechanism;

[0072] In some embodiments of the present application, the user call analysis and real-time feedback module comprises:

[0073] User behavior analysis and communication mode recognition module, for using behavior analysis algorithms and communication mode recognition techniques to monitor and analyze multi-dimensional data such as user dialing mode, call duration, call frequency, etc., and establish a normal behavior model of the user;

[0074] A user authorization recording permission module is configured to trigger a prompt for user authorization of recording permission when the system detects that the user is likely to be defrauded, to ensure that the user understands the operation and purpose of the system;

[0075] A real-time fraud detection and user feedback module is configured to record the user and apply the fraud detection algorithm trained in the previous steps to achieve real-time fraud detection during a call after the user has authorized.

[0076] It should be noted that the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or sequence between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for real-time processing of big data, the method comprising: The method comprises the following steps: ​ Step 1: Collecting multi-source fraud data through a data collection module; Step 2: Extracting multi-features from multi-source data and creating a feature matrix; Step 3: Building a fraud detection model by selecting a long short-term memory network; Step 4: Detecting fraud behavior by analyzing user calls; Step 5: Real-time fraud detection and model optimization through user feedback; The step of collecting multi-source fraud data through a data collection module comprises: The system uses a data collection module and network crawler technology to automatically search, crawl, and extract relevant fraud cases on the network, including but not limited to fraud phone recordings, fraud analysis reports, and fraud case databases. The system can also integrate with various data sources through API interfaces. After data collection, the collected data is cleaned. First, data deduplication is achieved by comparing key fields such as case ID and timestamp. For data fields containing text information, text data preprocessing is performed to reduce text noise and improve text data quality through word segmentation, stop word removal, and stem extraction techniques; The step of extracting multi-features from multi-source data and creating a feature matrix comprises: Selecting a speech recognition engine for text conversion and building an emotion lexicon for sentiment analysis; Using the PyDub speech processing library for tone and speed analysis; The step of selecting a speech recognition engine for text conversion and building an emotion lexicon for sentiment analysis comprises: Selecting a commercial speech recognition engine or open-source engine to convert the collected fraud audio data into text. Then, sentiment analysis is performed on the converted text. First, an emotion lexicon is built, which contains words and phrases related to different emotions. Then, the text information is segmented into individual words. For each word, the built emotion lexicon is used to match and identify which words belong to positive, negative, or neutral emotions. Finally, each word is assigned a sentiment score, and the overall sentiment feature is obtained. In addition, during sentiment recognition, speaker recognition is performed using speech signal processing to identify the fraudster and the victim; The step of using the PyDub speech processing library for tone and speed analysis comprises: The pitch and speech rate analysis of the voice data is as follows: first, the voice data is loaded by using the PyDub voice processing library, then the voice signal is framed and divided into short-time frames, each frame needs to be analyzed for pitch, the periodicity of the frequency spectrum is estimated by using the Fourier transform method, and then the fundamental frequency of the voice signal, i.e. the pitch, is obtained, at the same time, the intensity of the pitch is analyzed, the amplitude or amplitude spectrum of the fundamental frequency is calculated to obtain the intensity information of the pitch, which provides insight into the emotional state of the speaker, on the other hand, the frame rate of the voice signal, i.e. the speech rate, is calculated by calculating the time interval between frames, the specific implementation method is as follows: the position of the frame in the entire voice signal, i.e. the timestamp, is calculated by using the sampling rate and the position of the frame, then the time difference between the adjacent frames is obtained by subtracting the timestamp of the adjacent frame, finally the frame rate, i.e. the number of speech frames produced by the speaker in one second, is calculated by taking the reciprocal of the time difference, the system also calculates the average value of the frame time interval: and the variance: The trend of the speech rate of the speaker in different time periods can be obtained, in addition, the timing features and feature words need to be extracted; The step of building a fraud detection model by selecting a long short-term memory network comprises: Setting the LSTM model parameters and using binary cross-entropy as the loss function and Adam optimizer for model parameter adjustment; Introducing early stopping to prevent overfitting; The step of detecting fraud behavior by analyzing user calls comprises: The system also adopts communication pattern recognition technology to analyze the frequency, time period, and location information of the call. Through the recognition of these communication patterns, the system can identify abnormal patterns that deviate from the user's normal behavior patterns and infer possible fraud scenarios. When the system identifies that the user may be subject to fraud through the sensitivity detection mechanism, it will trigger a prompt for the user to authorize the recording permission. This prompt will be presented to the user in the form of an instant notification. When the system prompts the user, the user needs to authorize the recording. After the user completes the authorization, the system will transmit the user's recording data to the fraud detection module. Finally, encryption technology and security protocols are used to ensure that the user's recording data is fully protected during transmission and storage to prevent information leakage and misuse. The step of achieving real-time fraud detection and model optimization through user feedback includes: Real-time processing of audio data using previously trained fraud detection models; User provides multi-dimensional feedback, including confirming fraudulent behavior and reporting false positives; Update strategy using machine learning model, adjust weight and feature weight; Monitor performance indicators, accuracy and recall rate; Use confusion matrix to evaluate the classification effect of the model on fraud cases.

2. The method of claim 1, wherein: The step of setting LSTM model parameters and using binary cross-entropy as loss function and Adam optimizer for model parameter adjustment includes: The design of the model aims to capture the speech features and emotional information in the audio segments to support accurate identification of fraud cases. Specifically, a model containing two layers of LSTM units is constructed, each layer containing 50 hidden units. In the feature input stage, the time sequence information of each audio segment is prepared and organized into a 3D tensor with a shape of time steps, feature numbers. Additionally, the time steps of each audio segment are set to 50. Furthermore, in optimizing the model parameters, binary cross-entropy is chosen as the loss function, and the Adam optimizer is used with an initial learning rate of 0.001 and a learning rate decay of 0.

9. During training, the data is divided into 80% training set and 20% validation set, with a batch size of 32 for each training and a total of 50 training iterations.

3. The processing system of a big data real-time processing method according to claim 2, wherein: The system includes: Fraud data collection and preprocessing module, used to obtain fraud case data from multiple sources, clean and preprocess the text, to prepare the data for subsequent fraud detection model training; Fraud detection model construction module, used to select and construct fraud detection model, model training and parameter optimization, to improve the accuracy of fraud cases; User call analysis and real-time feedback module, used to analyze user call behavior, real-time monitoring of possible fraud scenarios, real-time fraud detection through user authorized recording permission, and providing instant feedback, including risk prompt and user learning process.

Images