An anti-fraud method, system, device and storage medium based on natural language processing technology

Through natural language processing technology, the transfer remarks information and user association characteristics are processed, combined with the neural network model, the problem of low fraud identification accuracy of existing systems in complex business scenarios is solved, and the accuracy of fraud identification and rich feature information is achieved. Spend.

CN114298715BActive Publication Date: 2024-12-03ZHEJIANG BANGSUN TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202111657952.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-30
Publication Date
2024-12-03
Estimated Expiration
2041-12-30

AI Technical Summary

Technical Problem

The existing electronic banking financial anti-fraud system is difficult to adapt to complex business scenarios and operational characteristics, resulting in low accuracy of fraud identification, especially ignoring the transfer notes information and the diversity of different business scenarios.

Method used

Using anti-fraud methods based on natural language processing technology, multi-dimensional user feature data is constructed through the combination of transfer notes and user-related feature information, and fraud identification is performed using neural network models, including word segmentation and word embedding vectors of transfer notes Extraction and long-term short-term memory network feature extraction, combining user authorization and behavioral data to make fraud predictions.

Benefits of technology

It significantly improves the accuracy of anti-fraud identification, provides richer feature information, adapts to specific scenario needs in the field of financial risk monitoring technology, fast and convenient word segmentation accuracy and higher fraud prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114298715B_ABST
    Figure CN114298715B_ABST
Patent Text Reader

Abstract

The present invention discloses an anti-fraud method, system, device and storage medium based on natural language processing technology, which obtains sample data required for processing from a financial anti-fraud transfer transaction data set and labels it; parses the transfer note information of each transaction data and processes it; parses the user-related features required for each sample, and establishes a target sample data set in combination with the transfer note information processing result and the sample labeling processing result, and trains a neural network after preprocessing; deploys the trained neural network to an online anti-fraud decision server; when making online decisions, integrates user-related feature information to make anti-fraud transfer transaction decisions, and returns anti-fraud transaction decision results. The present invention takes into account the operational feature data of auxiliary business scenarios in electronic banking business, uses auxiliary business scenarios to assist in the identification of fraudulent operation requests in target business scenarios, and greatly improves the fraud identification accuracy of the anti-fraud model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of financial risk monitoring technology, and in particular to an anti-fraud method, system, device and storage medium based on natural language processing technology. Background Art

[0002] With the rapid development of the Internet and the popularization of smart devices, people have gained greater convenience in using mobile banking and online banking to check balances and transfer transactions, but this is accompanied by security risks faced by electronic banking and the risk of financial fraud.

[0003] Relevant research shows that online fraud and cybercrime have become increasingly complex and have penetrated into different industries, causing huge economic losses to users around the world every year. Financial fraud has developed into a tightly organized black industry chain with clear professional division of labor, which has brought severe challenges to banks in developing online financial services.

[0004] At present, the existing electronic banking financial anti-fraud system prevention and control strategies in the industry generally use two implementation schemes: one is based on the expert system method. The simple expert system is difficult to adapt to the needs of the current electronic banking anti-fraud system. The second is to use machine learning and neural network algorithms to learn the risk characteristics of the entire business process and conduct offline analysis of historical data. The existing machine learning and neural network-based methods directly apply the anti-fraud model to the current scenario without modification. Although it avoids human subjectivity to a certain extent compared to the expert system, it ignores the important characteristics of the transfer note information and the complexity and diversity of the operation process and operation feature data of the electronic banking business system in different business scenarios. Directly training the anti-fraud model without distinguishing the business scenario will bring difficulties to the judgment of the anti-fraud system. Summary of the invention

[0005] To solve the above problems, the present invention proposes an anti-fraud method, system and storage medium based on natural language processing technology. The present invention introduces transfer note information and natural language processing technology, considers the operation characteristic data of auxiliary business scenarios in electronic banking business, and uses auxiliary business scenarios to assist in the identification of fraudulent operation requests in target business scenarios, so that the fraud identification accuracy of the anti-fraud model is greatly improved.

[0006] The objective of the present invention is to achieve the following technical solution: an anti-fraud method based on natural language processing technology, the method comprising the following steps:

[0007] (1) Obtain the transfer note information of each transaction data from the financial anti-fraud transfer transaction data set and process it; the details are as follows:

[0008] (1.1) Split each transfer note into individual words, add a suffix to the end of each word, and count the frequency of occurrence of each word in the transfer note;

[0009] (1.2) Sort the words by frequency, select the top m words in frequency and merge them with their neighbor words in the transfer notes into a string, and put them into the subword list;

[0010] (1.3) performing a deduplication operation on the subword word list, calculating the mutual information between the individual words of the subwords, and deleting the subwords whose mutual information is less than a threshold;

[0011] (1.4) Use the above subword word list to segment the transfer transaction notes, and use the Word2Vec method to train Chinese word embedding vectors. After feature extraction through the long short-term memory network, the Chinese word embedding vectors are used as the input of the subsequent neural network;

[0012] (2) obtaining user-related feature information required for each transaction data and the transfer note information processing result obtained from step (1) from the financial anti-fraud transfer transaction data set, and labeling each transaction data in the financial anti-fraud transfer transaction data set, establishing multi-dimensional user feature data to form a target sample data set, performing feature engineering and data preprocessing on the target sample data set and dividing it into a training set and a test set, using the training set to train a neural network, and obtaining a neural network model for subsequently predicting the possibility of fraud in real-time transfer transaction requests;

[0013] (3) When making online decisions, the transfer transaction request data initiated by the user in real time is processed according to the method of step (1) and the user's associated characteristic information to form the transfer transaction decision data, which is input into the neural network model trained in step (2) to make anti-fraud transfer transaction decisions. The output of the neural network model is the user fraud probability value. The possibility of fraud in the transfer transaction request is judged based on the user fraud probability value, and the result of the anti-fraud transaction decision is returned to the associated user of the transfer transaction request or the transaction platform where the user is located.

[0014] Furthermore, the financial anti-fraud transfer transaction dataset is obtained from a bank and is a transfer transaction dataset including transfer remark information.

[0015] Furthermore, based on the user's transaction record, each sample data is labeled, and a user label is established based on the user's historical credit status corresponding to the sample data, that is, users with abnormal transfer transaction records within a specified period are defined as high-quality users and key monitored users respectively.

[0016] Furthermore, the user-associated characteristic information includes user authorization data and user behavior data corresponding to the sample data; the user authorization data includes user age and user credit record characteristics, and the user behavior data includes characteristics of the cumulative amount of transfers made by the user in the last 12 hours.

[0017] Furthermore, before training the neural network model, the target sample data set is subjected to data preprocessing, including data standardization and missing value filling; the missing value filling processing includes separately counting the missing rates of all variables in the multi-dimensional transfer data set, and directly deleting features according to the statistical missing rates or using the average value to fill in all missing samples.

[0018] Furthermore, the Chinese word embedding vector is input into the long short-term memory network for processing to generate a multi-dimensional feature representation value; the user-associated feature information is concatenated with the feature representation value generated by the long short-term memory network and then sent into the neural network for calculation.

[0019] The present invention also provides an anti-fraud system based on natural language processing technology, the system comprising a data receiving node, a transaction processing node, a model training node and a persistence node;

[0020] The data receiving node extracts the transfer transaction requests initiated by mobile banking and online banking users in the transfer business system in real time online, and then uses Kafka to collect the associated feature information of the initiating user of the transfer transaction request, and sends the transfer transaction decision data composed of the two to the transaction processing node;

[0021] The transaction processing node includes a transaction information processing module, a natural language processing module and an anti-fraud decision module; the transaction information processing module obtains the transfer transaction decision data required for processing from the data receiving node, extracts the user-related feature information and sends it to the anti-fraud decision module, and extracts the transfer note information and sends it to the natural language processing module; the natural language processing module processes the transfer note information and then sends it to the anti-fraud decision module; the anti-fraud decision module inputs the data obtained from the transaction information processing module and the natural language processing module to the model training node for training the neural network model after feature engineering and data preprocessing, obtains the neural network model deployed on the online anti-fraud server, and makes an anti-fraud transfer transaction decision, the output of the neural network model is a user fraud probability value, and the possibility of fraud in the transfer transaction request is judged according to the user fraud probability value, and then the result of the anti-fraud transaction decision is called through the online anti-fraud decision server calling interface, and the calling result is returned to the associated user of the transfer transaction request or the transaction platform where the user is located.

[0022] The model training node includes an experience collection module, a transfer note information module, an HDFS module, a neural network model module and a version management module; the experience collection module obtains the user-related feature information of each data from the historical financial anti-fraud transfer transaction data set of the persistent node, and inputs it into the HDFS module; the transfer note information module obtains the transfer note information of each data from the historical financial anti-fraud transfer transaction data set of the persistent node, and processes the transfer note information in the same way as the natural language processing module, and inputs the processed result into the HDFS module; the transfer note information processing method is as follows:

[0023] a. Split each transfer note into individual words, add a suffix to the end of each word, and count the frequency of occurrence of each word in the transfer note;

[0024] b. Sort the words by frequency, select the top m words in frequency and merge them with their neighbor words in the transfer note information into a string and put them into the subword word list;

[0025] c. Perform a deduplication operation on the subword vocabulary, calculate the mutual information between the individual words of the subwords, and delete the subwords whose mutual information is less than the threshold;

[0026] d. Use the above subword word list to segment the transfer transaction notes, train Chinese word embedding vectors using the Word2Vec method, and extract features from the Chinese word embedding vectors using a long short-term memory network;

[0027] The HDFS module receives data from the experience collection module and the transfer note information module and establishes multi-dimensional user feature data in combination with the sample labeling processing results of the historical financial anti-fraud transfer transaction data set to form a target sample data set, performs feature engineering and data preprocessing on the target sample data set and divides it into a training set and a test set, and stores the training set and the test set in the HDFS module. The training set data in the HDFS module is used for training the neural network model in the neural network model module. After the neural network model training is completed, it is deployed to the online anti-fraud decision server on the anti-fraud decision module of the transaction processing node; the version management module is used to store different neural network model versions to flexibly replace the online model, improve the online model decision effect and control the training status of the new model;

[0028] The persistent node is used to store historical financial anti-fraud transfer transaction data sets, including an SQL module and an archive storage module; the SQL module is used to provide users with data support for querying transfer transaction records, and is used by data developers to perform statistics and analysis on transfer transaction data; the archive storage module is used to back up process data collected by the server during the transaction process to achieve disaster recovery backup function.

[0029] The present invention also provides a device, comprising:

[0030] one or more processors;

[0031] A memory for storing one or more programs;

[0032] When the one or more programs are executed by the one or more processors, the one or more processors execute an anti-fraud method based on natural language processing technology.

[0033] The present invention also provides a computer-readable storage medium for storing one or more computer programs, wherein the one or more computer programs include program codes. When the computer programs are run on a computer, the program codes are used to execute an anti-fraud method based on natural language processing technology.

[0034] Beneficial effects of the present invention:

[0035] 1. When making anti-fraud decisions on transfer transactions, natural language processing technology is used to extract features from transfer notes, and the user information and historical transactions of the transferor are considered. Compared with existing expert systems and methods based on machine learning and neural networks, it has richer feature information and is more in line with specific scenarios in the field of financial risk monitoring technology.

[0036] 2. When using natural language processing technology to process transfer notes, the method of constructing a subword word list in the transfer note information avoids the global dictionary and other supervised word segmentation methods, which is faster and more convenient. The subword word list is denoised using mutual information, omitting redundant information and improving the accuracy of word segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0038] Figure 1 A flow chart showing the method for modeling an anti-fraud model for a transfer transaction according to the present invention is shown;

[0039] Figure 2 A flow chart showing the method for processing transfer note information according to the present invention is shown;

[0040] Figure 3 A schematic diagram of the structure of the anti-fraud system of the present invention is shown;

[0041] Figure 4 A schematic diagram of the system risk control process when the anti-fraud system of the present invention is used is shown;

[0042] Figure 5 A schematic diagram of the device structure of a terminal device or a server suitable for the present invention is shown. Specific Embodiments

[0043] To more clearly illustrate the present invention, the present invention will be further described below with reference to the accompanying drawings. Similar structures in the drawings are denoted by the same reference numerals. Those skilled in the art should understand that the content specifically described below is illustrative rather than restrictive, and should not be used to limit the protection scope of the present invention.

[0044] The present invention provides an anti-fraud method, system, device and storage medium based on natural language processing technology, which can be applied to the anti-fraud identification system in various business operations of electronic banks to process transfer note information. For example, it is used to process transfer note information in the anti-fraud identification systems for online banking, mobile banking, direct banking and WeChat banking business operations.

[0045] Specifically, on the one hand, as Figure 1 shown, an anti-fraud method based on natural language processing technology provided by the present invention includes the following steps:

[0046] S101: Obtain the transfer note information of each transaction data from the financial anti-fraud transfer transaction dataset, and process it using the transfer note information processing method.

[0047] The transfer note information extracted from the financial anti-fraud transfer transaction dataset will first be sent to Figure 2 the transfer note information processing method shown in the figure for statistical processing, and the Chinese word segmentation model and Chinese word embedding vector of the transfer note information are obtained through the processing method.

[0048] The specific process of the transfer note information processing method is as Figure 2 shown, and mainly includes:

[0049] S1011: Split each transfer note information into single characters, add a suffix at the end of each single character, and count the occurrence frequency of each single character in the transfer note information.

[0050] S1012: Sort the single characters according to the frequency, and sequentially select the top m single characters with the highest frequency and their neighboring single characters in the transfer note information dataset to be merged into a string and put into the sub-word vocabulary. Among them, before sorting, all single characters are first processed to remove stop words. The stop word list includes but is not limited to words such as "de", "ge", "shi", etc.; m generally takes 5% - 20% of the total number of single characters as the value.

[0051] S1013: Perform a deduplication operation on the sub-word vocabulary, and calculate the mutual information amount between the single characters of the sub-words, and delete the sub-words with the mutual information amount less than the threshold.

[0052] The mutual information amount I(x; y) is used to measure the correlation degree between two single characters x and y in the sub-word.

[0053]

[0054] Where p(x,y) is the joint probability of the subword consisting of x and y appearing in the transfer note information dataset, and p(x) and p(y) are the marginal probabilities of the single word x and the single word y appearing in the transfer note information dataset.

[0055] S1014: Use the above subword vocabulary to segment the transfer transaction note information, and train the Chinese word embedding vector through the Word2Vec method, and then send the Chinese word embedding vector into the neural network model after feature extraction through the long short-term memory network.

[0056] S102: Obtain the user-related feature information required for each transaction data and the transfer note information processing result obtained from S101 from the financial anti-fraud transfer transaction data set, and label each transaction data in the financial anti-fraud transfer transaction data set, establish multi-dimensional user feature data to form a target sample data set, perform feature engineering and data preprocessing on the target sample data set and divide it into a training set and a test set, use the training set to train a neural network, and obtain a neural network model that is subsequently used to predict the possibility of fraud in real-time transfer transaction requests.

[0057] The user association information of the 10,000 transfer transaction samples is associated and matched from the data set, and the user association information includes the user authorization data and user behavior data corresponding to the sample data. The user authorization data includes features such as user age and user credit record, and the user behavior data includes features such as the cumulative amount of transfers made by the user in the last 12 hours.

[0058] When labeling the financial anti-fraud transfer transaction data set, first determine the scope of the selected data, for example, select 10,000 transfer transactions with transaction time from June to December 2020 and containing transaction notes information at the time of the transaction as sample data. Then, according to the user's transaction record, each of the sample data is labeled, that is, a user label is established according to the user's historical credit situation corresponding to the sample data. For example, the good or bad label of the user is defined by the performance of the user's transfer transaction. Taking the sample data of a user as an example, if a user has no abnormal transfer transaction record within 180 days, it is defined as a good user, that is, a high-quality user, represented by 0; conversely, if a user has an abnormal transfer transaction record within 180 days, it is defined as a bad user, that is, a user who needs to be monitored, represented by 1; in this way, the labeling of a sample is completed, and the label is used as the target variable.

[0059] The target sample data set is subjected to data preprocessing, including data standardization, missing value filling, etc. The data standardization process refers to converting the specified features in all samples into a standard normal distribution with a mean of 0 and a variance of 1 using a standardization process. The missing value filling process refers to separately counting the missing rates of all variables in the multi-dimensional transfer data set, and performing processing based on the statistical missing rates. When the missing rate is too high, such as exceeding 60%, it is considered that the feature has lost its reference value, so the feature is directly deleted. When the missing rate is in a lower range, the average value of the missing values ​​of the feature can be first counted, and all missing samples can be filled with the average value.

[0060] The neural network model of the present invention includes two input sources, one of which is a word embedding vector generated after being processed by the transfer note information processing method. This part of the input will be sent to the long short-term memory network for processing to generate a multi-dimensional feature representation value. The other part of the input is the sample association feature after data preprocessing and feature engineering. This part of the feature data is spliced ​​with the feature representation value generated by the long short-term memory network and sent to the neural network for calculation.

[0061] The neural network model can be a fully connected neural network, a convolutional neural network, a residual neural network, etc.

[0062] S103: When making online decisions, the transfer transaction decision data consisting of the transfer transaction request data initiated by the user in real time and the user's associated characteristic information is input into the neural network model pre-trained in step S102 to make anti-fraud transfer transaction decisions. The output of the neural network model is a user fraud probability value. The possibility of fraud in the transfer transaction request is judged based on the user fraud probability value, and the anti-fraud transaction decision result is returned to the associated user of the transfer transaction request or the transaction and platform where the user is located.

[0063] The following is an introduction to an anti-fraud system based on natural language processing technology proposed by the present invention. Figure 3 shown. Figure 3 This is a schematic diagram of the architecture of the anti-fraud system 300 described in the present invention. The anti-fraud system 300 is mainly divided into four parts: a data receiving node 310, a transaction processing node 320, a model training node 330 and a persistence node 340.

[0064] The data receiving node 310 includes a mobile banking user 311, an online banking user 312 and Kafka 313;

[0065] Specifically, the transfer transaction request submitted by the mobile banking user 311 and the online banking user 312 in the transfer business system. The data receiving node 310 uses Kafka 313 to collect the required user-related feature information. The two parts constitute the transfer transaction decision data and are sent to the transaction information processing module 321 of the transaction processing node 320 for transaction information processing.

[0066] The transaction processing node 320 includes a transaction information processing module 321, a natural language processing module 322, and an anti-fraud decision module 323;

[0067] Specifically, after the transaction information processing module 321 collects the required user transfer transaction decision data, it extracts the user-related feature information and sends it to the anti-fraud decision module, extracts the transfer note information and sends it to the natural language processing module; the natural language processing module 322 performs word segmentation and word embedding vector processing on the transfer note information, and then sends it to the anti-fraud decision module. The anti-fraud decision module 323 inputs the data obtained from the transaction information processing module and the natural language processing module into the neural network model obtained by the model training node 330 and deployed on the online anti-fraud decision server for anti-fraud decision after feature engineering and data preprocessing, and the output of the neural network model is the user fraud probability value, and the possibility of fraud in the transfer transaction request is judged according to the user fraud probability value, and then the result of the anti-fraud transaction decision is called through the online anti-fraud decision server call interface, and the call result is returned to the associated user of the transfer transaction request or the transaction platform where the user is located.

[0068] The model training node includes an experience collection module 331, a transfer note information module 332, an HDFS module 333, a neural network model 334 and a version management module 335;

[0069] Specifically, the experience collection module 331 obtains the user-related feature information of each piece of data from the historical financial anti-fraud transfer transaction data set of the persistent node 340, and inputs it into the HDFS module 333; the transfer note information module 332 obtains the transfer note information of each piece of data from the historical financial anti-fraud transfer transaction data set of the persistent node 340, and uses the transfer note information processing method to process and input it into the HDFS module 333. The details are as follows:

[0070] a. Split each transfer note into individual words, add a suffix to the end of each word, and count the frequency of occurrence of each word in the transfer note.

[0071] b. Sort the single characters by frequency, and sequentially select the top m single characters with the highest frequencies and merge them with their neighboring single characters in the transfer note information into strings, which are then put into the sub-word vocabulary. Before sorting, all single characters are first processed to remove stop words, and the stop word list includes but is not limited to "de", "ge", "shi"; m generally takes 5% - 20% of the total number of single characters as the value.

[0072] c. Perform a deduplication operation on the sub-word vocabulary, calculate the mutual information between the single characters of the sub-words, delete the sub-words with mutual information less than the threshold, and perform a denoising operation on the sub-word vocabulary.

[0073] d. Use the above sub-word vocabulary to segment the transfer transaction note information, and train Chinese word embedding vectors through the Word2Vec method. After extracting features from the Chinese word embedding vectors through a long short-term memory network, they are sent to the HDFS module 333.

[0074] The HDFS module 333 receives the data from the experience collection module 331 and the transfer note information module 332, and combines the sample labeling processing results of the historical financial anti-fraud transfer transaction dataset to establish multi-dimensional user feature data to form a target sample dataset. Perform feature engineering and data preprocessing on the target sample dataset and divide it into a training set and a test set, and store the training set and the test set in the HDFS module 333. The data in the HDFS module 333 is used for the training of the neural network model module 334. After the neural network model 334 is trained, it is deployed to the online anti-fraud decision server on the anti-fraud decision module 323 of the transaction processing node 320. The version management module 335 is used to control different neural network model versions to flexibly replace the online model, improve the decision-making effect of the online model, and control the training status of the new model.

[0075] The persistent node 340 is used to store the historical financial anti-fraud transfer transaction dataset, including the SQL module 341 and the archive storage module 342;

[0076] Specifically, the persistent node 340 is mainly used for the historical financial anti-fraud transfer transaction dataset, including SQL 341 and the archive storage module 342. SQL 341 is used to provide data support for users to query transfer transaction records and facilitate data developers to perform statistics and analysis on transfer transaction data. The archive storage module is used to back up the process data collected by the server during the transaction process to achieve the disaster recovery backup function.

[0077] In the present invention, the specific implementation of the experience collection module 331 is a key-value pair storage map, where the key is a string of user ID and timestamp, and the value is the specific features and label value of the sample. The main reason for using map as the experience collection module is that map, as an associative container, can provide one-to-one data processing capabilities, the key in the map can only appear once inside the container, and its query speed is very fast, and the corresponding value of the specified key can be found within a constant time complexity.

[0078] Taking into account the problem of low sample processing efficiency of TensorFlow Serving in real-time environment, the present invention adopts some improvement methods: 1) The characteristics of user behavior data in online real-time scenarios will change over time. For dense type features, the present invention maintains an incremental algorithm based on streaming computing to pre-process the features. 2) Considering the problem that TensorFlow Serving has poor feature embedding effect, the present invention directly processes the feature embedding function in the data preprocessing module to reduce the model calculation cost and improve the online model training speed. 3) In order to improve the experimental efficiency of data mining personnel in feature engineering, the method proposed in the present invention supports feature configuration to generate model structure.

[0079] According to the anti-fraud system for transfer transactions proposed by the present invention, the dynamic characteristics of each user request operation in the financial risk control scenario can be obtained online at the moment, and used for online model decision-making and offline model training, so that the model can promptly reflect the real-time changes online and improve the accuracy of online predictions. The transaction processing nodes and model training nodes in the present invention can use the visual operation interface of the monitoring module to quickly add new task modules according to business needs. Without changing the overall framework of the system, new data processing logic can be added to the anti-fraud system for transfer transactions according to business needs.

[0080] Figure 4 A schematic diagram of the risk control process of the anti-fraud system of the present invention is shown. Here, two business scenarios, mobile banking and online banking, are specifically cited as examples to illustrate the risk control process of the anti-fraud system of the present invention. In actual application, it can be expanded to include various other banking scenarios, such as online banking, mobile banking, direct banking and / or WeChat banking.

[0081] The anti-fraud system of the present invention is the core module of the entire risk control process. In the figure, the anti-fraud system is connected to the business system of the mobile banking channel and the business system of the online banking channel respectively. By accepting transfer transaction requests from mobile banking or online banking, the risk value of the transfer transaction request is evaluated, and the transfer result is displayed on the mobile banking transfer transaction interface or the online banking transfer transaction interface. The bank transfer transaction system can use the decision results of the anti-fraud system to perform anti-fraud interception operations, for example, including interception operations or release operations.

[0082] The following will generally describe the intelligent judgment process of the anti-fraud system 400 for transfer transactions, and describe the functions of each functional module and the data types transmitted between these functional modules:

[0083] 1. First, a mobile banking user 401 or an online banking user 402 initiates a transfer transaction request using a mobile banking transfer transaction interface 403 or an online banking transfer transaction interface 404. The transfer transaction operation requests are relatively similar under the bank's transfer business system. Generally speaking, these operation requests of different operation types correspond to different business scenarios of different bank channels and have their own specific business models.

[0084] 2. The mobile banking channel transfer transaction interface 403 sends the user's transfer transaction request to the anti-fraud system 405. The online banking user's transfer transaction interface also sends the user's transfer transaction request to the anti-fraud system 405. The system uses the anti-fraud system based on natural language processing technology of the present invention to analyze the feature data corresponding to the user's operation request. The specific anti-fraud model decision method is as described above.

[0085] 3. After the anti-fraud system 405 completes the decision on the user's transfer transaction request, it will store the transfer transaction record and the decision result in the transfer transaction database 406, and return the decision result to the mobile banking transfer transaction interface 403 and the online banking transfer transaction interface 404.

[0086] The experimental environment of the anti-fraud method and system based on natural language processing technology of the present invention is:

[0087] Operating system: Ubuntu 16.04.1LTS

[0088] CPU: Intel(R)Xeon(R)Gold 6226CPU@2.70GHz

[0089] GPU: NVIDIA GeForce RTX 2080Ti

[0090] Memory: 64GB

[0091] Hard disk: 4TB

[0092] Programming language: Python 3.7

[0093] Anaconda: 5.3.0

[0094] TensorFlow: 2.3.0

[0095] Figure 5 FIG. 1 shows a schematic diagram of a device structure suitable for a terminal device or a server of the present invention. Figure 5 As shown, the computer system 500 includes a removable medium 501, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., which is installed on a drive 502 as needed, so that the computer program read therefrom is installed into a storage part 506 as needed. The drive 502 is also connected to an I / O interface 507 as needed. The communication part 503 performs communication processing via a network such as the Internet. The following components are connected to an output part 505 of a line tube (CRT), a display, etc., and a speaker, etc.; a storage part 506 including a hard disk, etc.; and a communication part 503 including a network interface card such as a LAN card, a modem, etc.

[0096] A central processing unit (CPU) 509 can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 510 or a program loaded from a storage portion 506 into a random access memory (RAM) 511. In the RAM 511, various programs and data required for the operation of the system 500 are also stored. The CPU 509, the ROM 510, and the RAM 511 are connected to each other via a bus 508. An input / output (I / O) interface 507 is also connected to the bus 508.

[0097] In particular, according to the present invention, the process described above with reference to the flowchart can be implemented as computer application software. For example, the present invention includes a computer application, which includes a computer program carried on a computer readable medium. The computer program can be downloaded and installed from a network through a communication part 503, and / or installed from a removable medium 501. When the computer program is executed by a central processing unit (CPU) 509, the above-mentioned functions defined in the system of the present invention are performed.

[0098] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a random access memory (RAM), a read-only memory (ROM), a computer disk, a hard disk, 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 of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0099] The above contents are used to explain the present invention, rather than to limit the present invention. For those skilled in the art, there are many ways of implementation, deformation and improvement without departing from the concept of the present invention, without being limited to the details of the exemplary implementation cases described in this application. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An anti-fraud method based on natural language processing technology, characterized in that: The method comprises the following steps: (1) Obtain the transfer note information of each transaction data from the financial anti-fraud transfer transaction data set and process it; the details are as follows: (1.1) Split each transfer note into individual words, add a suffix to the end of each word, and count the frequency of occurrence of each word in the transfer note; (1.2) Sort the words by frequency, select the top m words in frequency and merge them with their neighbor words in the transfer notes into a string and put them into the subword list; (1.3) performing a deduplication operation on the subword word list, calculating the mutual information between the individual words of the subwords, and deleting the subwords whose mutual information is less than a threshold; the mutual information I(x; y) is used to measure the correlation between two individual words x and y in the subword; ; Where p(x, y) is the joint probability of the subword x and y appearing in the transfer note information dataset, and p(x) and p(y) are the marginal probabilities of the single word x and the single word y appearing in the transfer note information dataset; (1.4) Using the above subword word list to segment the transfer transaction note information, and using the Word2Vec method to train Chinese word embedding vectors, the Chinese word embedding vectors are used as input for subsequent neural networks after feature extraction through a long short-term memory network; the Chinese word embedding vectors are input into the long short-term memory network for processing to generate multi-dimensional feature representation values; the user-related feature information is concatenated with the feature representation values ​​generated by the long short-term memory network and then sent to the neural network for calculation; (2) obtaining the user-related feature information required for each transaction data and the processing result of the transfer note information obtained from step (1) from the financial anti-fraud transfer transaction data set, wherein the user-related feature information includes the user authorization data and user behavior data corresponding to the sample data; the user authorization data includes the user age and the user credit record characteristics, and the user behavior data includes the user's cumulative transfer amount characteristics in the last 12 hours; and labeling each transaction data in the financial anti-fraud transfer transaction data set to establish multi-dimensional user feature data to form a target sample data set, specifically: labeling the sample data according to the user's transaction record, establishing a user label according to the user's historical credit status corresponding to the sample data, that is, defining users with abnormal transfer transaction records within a specified period as high-quality users and key monitored users respectively; performing feature engineering and data preprocessing on the target sample data set and dividing it into a training set and a test set, using the training set to train a neural network to obtain a neural network model for subsequently predicting the possibility of fraud in real-time transfer transaction requests; (3) When making online decisions, the real-time transfer transaction request data initiated by the user is processed according to the method of step (1) and the user's associated characteristic information is used to form the transfer transaction decision data, which is input into the neural network model trained in step (2) to make anti-fraud transfer transaction decisions. The output of the neural network model is the user fraud probability value. The possibility of fraud in the transfer transaction request is judged based on the user fraud probability value, and the result of the anti-fraud transaction decision is returned to the associated user of the transfer transaction request or the transaction platform where the user is located.

2. The anti-fraud method according to claim 1, characterized in that: The financial anti-fraud transfer transaction dataset is obtained from a bank and is a transfer transaction dataset containing transfer remark information.

3. The anti-fraud method according to claim 1, characterized in that: Before training the neural network model, the target sample data set is preprocessed, including data standardization and missing value filling; the missing value filling process includes separately counting the missing rates of all variables in the multi-dimensional transfer data set, and directly deleting features or using the average value to fill all missing samples according to the statistical missing rates.

4. An anti-fraud system based on natural language processing technology, characterized in that: The system includes data receiving nodes, transaction processing nodes, model training nodes and persistence nodes; The data receiving node extracts the transfer transaction requests initiated by mobile banking and online banking users in the transfer business system in real time online, and then uses Kafka to collect the associated feature information of the initiating user of the transfer transaction request, and sends the transfer transaction decision data composed of the two to the transaction processing node; The transaction processing node includes a transaction information processing module, a natural language processing module and an anti-fraud decision module; the transaction information processing module obtains the transfer transaction decision data required for processing from the data receiving node, extracts the user-related feature information and sends it to the anti-fraud decision module, and extracts the transfer note information and sends it to the natural language processing module; The natural language processing module processes the transfer note information and then sends it to the anti-fraud decision module; the anti-fraud decision module inputs the data obtained from the transaction information processing module and the natural language processing module into the model training node for training the neural network model after feature engineering and data preprocessing, obtains the neural network model deployed on the online anti-fraud server, and makes anti-fraud transfer transaction decisions. The output of the neural network model is a user fraud probability value, and the possibility of fraud in the transfer transaction request is judged according to the user fraud probability value, and then the result of the anti-fraud transaction decision is called through the online anti-fraud decision server calling interface, and the call result is returned to the associated user of the transfer transaction request or the transaction platform where the user is located; The model training node includes an experience collection module, a transfer note information module, an HDFS module, a neural network model module and a version management module; the experience collection module obtains the user-related feature information of each data from the historical financial anti-fraud transfer transaction data set of the persistent node, and inputs it into the HDFS module; the transfer note information module obtains the transfer note information of each data from the historical financial anti-fraud transfer transaction data set of the persistent node, and processes the transfer note information in the same way as the natural language processing module, and inputs the processed result into the HDFS module; the transfer note information processing method is as follows: a. Split each transfer note into individual words, add a suffix to the end of each word, and count the frequency of occurrence of each word in the transfer note; b. Sort the words by frequency, select the top m words in frequency and merge them with their neighbor words in the transfer note information into a string and put them into the subword vocabulary; c. Perform a deduplication operation on the subword vocabulary, calculate the mutual information between the words in the subwords, and delete the subwords whose mutual information is less than the threshold; d. Use the above subword word list to segment the transfer transaction notes, train Chinese word embedding vectors using the Word2Vec method, and extract features from the Chinese word embedding vectors using a long short-term memory network; The HDFS module receives data from the experience collection module and the transfer note information module and establishes multi-dimensional user feature data in combination with the sample labeling processing results of the historical financial anti-fraud transfer transaction data set to form a target sample data set, performs feature engineering and data preprocessing on the target sample data set and divides it into a training set and a test set, and stores the training set and the test set in the HDFS module. The training set data in the HDFS module is used for training the neural network model in the neural network model module. After the neural network model training is completed, it is deployed to the online anti-fraud decision server on the anti-fraud decision module of the transaction processing node; the version management module is used to store different neural network model versions to flexibly replace the online model, improve the online model decision effect and control the training status of the new model; The persistent node is used to store historical financial anti-fraud transfer transaction data sets, including an SQL module and an archive storage module; the SQL module is used to provide users with data support for querying transfer transaction records, and is used by data developers to perform statistics and analysis on transfer transaction data; the archive storage module is used to back up process data collected by the server during the transaction process to achieve disaster recovery backup function.

5. A device, characterized in that: The device comprises: one or more processors; A memory 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 are caused to perform the method according to any one of claims 1 to 4.

6. A computer-readable storage medium for storing one or more computer programs, wherein the one or more computer programs include program codes. When the computer programs are run on a computer, the program codes are used to execute the anti-fraud method based on natural language processing technology as described in any one of claims 1 to 4.

Citation Information

Patent Citations

  • Machine learning anti-fraud monitoring system based on transaction data

    CN106682067A

  • Real-time anti-fraud method and system based on network data and storage medium

    CN109300028A