Multi-profile fraud detection system for detecting suspicious airline ticket sales activity

The hybrid sampling mechanism with cost-weighted metrics and deep learning optimizes fraud detection in airline ticket sales by addressing dataset imbalances and transaction costs, enhancing detection accuracy and reducing financial losses.

JP2026537378APending Publication Date: 2026-11-09イェルディゼ テクニケ ウニベルシテシ ドネレ セレメイェ イシレテメ ムデ +1
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2025547510
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-10-01
Publication Date
2026-11-09

AI Technical Summary

Technical Problem

Existing fraud detection systems in airline ticket sales face challenges due to unbalanced transaction datasets, ignoring historical transaction links, dataset imbalance, and the inability to consider transaction costs, leading to inefficiencies and increased computing costs.

Method used

A hybrid sampling mechanism using Equilibrium Random Undersampling (BRUS) and oversampling techniques, combined with cost-weighted metrics and deep learning, to create profile-based datasets that consider historical data and transaction costs, optimizing fraud detection.

Benefits of technology

Enhances fraud detection by balancing dataset imbalances, considering transaction costs, and improving detection accuracy, thereby reducing financial losses for airlines and customers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2026537378000001
    Figure 2026537378000001
  • Figure 2026537378000002
    Figure 2026537378000002
  • Figure 2026537378000003
    Figure 2026537378000003
Patent Text Reader

Abstract

The present invention relates to a method for detecting fraudulent activity in sales transactions conducted via computer networks, particularly in the sale of tickets and services in the aviation sector. The invention makes it possible to overcome performance limitations caused by unbalanced distribution of transaction datasets during computer-implemented fraud detection. The invention also specifically contributes to the profitability of airlines by reducing the financial losses incurred by airlines due to fraudulent transactions and the potential losses to customers that may occur after such transactions.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present invention relates to a method that enables detection of fraudulent activities in sales transactions conducted via computer networks, in particular detection of fraudulent activities in the sale of tickets and services in the aviation field. BACKGROUND ART

[0002] On digital platforms where product sales and money transfers are carried out, cases of fraud using stolen credit cards are increasing. As a result of such an increase, serious revenue losses and damages are occurring for enterprises that must ensure security. Several solutions have been developed to prevent this. European Patent No. 1356438 describes a method for verification by matching sets of information during sales transactions. However, such methods are insufficient when extensive stolen information is used.

[0003] To prevent revenue losses arising from such fraud cases, several fraud detection systems using machine learning and deep learning techniques have been developed. However, due to the fact that these systems operate primarily on a per-transaction basis using payment / sales datasets, they may ignore the history of related transactions and transactions linked thereto. Since the fraud detection problem has historical and time-series characteristics, only a limited level of success can be achieved with this transaction dataset.

[0004] Another problem faced by fraud detection applications is dataset imbalance. The number of fraudulent transactions is significantly smaller than the number of legitimate transactions. This ratio is estimated to be around 0.02%. The problem with imbalanced datasets is that they lead to a tendency to make negative decisions about all records in an attempt to obtain the most accurate values ​​possible. To solve this problem, undersampling methods are often used. This can lead to inefficient negative sample selection, such as clustering of selected samples within arbitrary regions of the dataset, iterative selection of records, or selection of independent records, if the selection is completely random, thereby misleading the classification mechanism. Oversampling methods, used as an alternative to undersampling to solve the problem of imbalanced datasets, result in an increase in the number of records, increasing computing costs and therefore system requirements, and reducing performance efficiency in systems operating on large datasets.

[0005] To overcome the performance limitations imposed by transaction datasets, conventional techniques have employed clustering or deep learning methods, such as recurrent neural networks (RNNs) and long-shortened memory (LSTMs) that incorporate memory units at the algorithmic level. In clustering techniques, the dataset is divided into groups, but these groups are determined based on the similarity of transactions by specific parameters, rather than on the relevant transaction history. Nevertheless, such clustering techniques do not guarantee the provision of historical links crucial for fraud detection.

[0006] Algorithms that include memory units within deep network architectures such as LSTMs store this information in memory to represent desired information in long and short data types and use it during the learning phase. However, such architectures require independently managing statistics associated with each credit card, PNR (Passenger Name Record), and other objects within the fraud dataset. In other words, object-based memory must be used instead of global or local memory. In this case, object-based memory limits the maximum achievable success of memory-based deep learning techniques such as LSTMs.

[0007] Fraudulent transactions also cause financial problems for businesses. However, with conventional techniques, the success of fraud detection is determined based on the number of transactions detected, and does not reflect the costs associated with the fraudulent activity. [Prior art documents] [Patent Documents]

[0008] [Patent Document 1] European Patent No. 1356438 [Overview of the project] [Problems that the invention aims to solve]

[0009] The object of the present invention is to provide a method for detecting fraudulent transactions in the sale of airline tickets and services. In particular, the object is to provide a computer-implemented fraud detection method that uses historical data relationships, which enables overcoming performance limitations caused by unbalanced transaction datasets. [Means for solving the problem]

[0010] This invention makes it possible to reduce the financial losses incurred by airlines due to fraudulent transactions and the customer losses that may occur as a result of such transactions. Therefore, it makes a concrete contribution to the profitability of airlines.

[0011] In the fraud detection method of the present invention, machine learning and deep learning techniques are used in combination with profile-based methods to overcome the constraints of the transaction dataset. This can prevent events that could circumvent security measures by trying several different credit card, phone number, or email information.

[0012] Within the scope of this invention, not only user profiles but also profiles of credit cards and PNR objects can be generated to prevent the use of dangerous tools. In addition, suspicious transactions can also be detected by extracting meaningful information from the transactions.

[0013] The fraud detection method of the present invention develops a hybrid sampling mechanism that includes an Equilibrium Random Undersampling (BRUS) algorithm in combination with several oversampling techniques, thereby utilizing imbalanced datasets without resulting in inefficient negative sample selection. In the hybrid sampling mechanism, the number of samples is first reduced to a specific level using the BRUS algorithm, and then the ratio of fraudulent to legitimate transaction records is balanced by oversampling positive records. The range in which the decision mechanism's tendency to mark all records as negative in order to maximize success is avoided is defined as the equilibrium level range. Typically, this ratio is set to a fraudulent / legitimate transaction ratio between 1:5 and 1:10.

[0014] The present invention also provides a cost-weighted metric mechanism to take costs into account along with the number of transactions during fraud detection. According to the present invention, while the complexity matrix is ​​being generated, each transaction is weighted based on its respective price value and added to the complexity matrix. Thus, the metric generated using this complexity matrix can also be obtained in a cost-weighted form. [Modes for carrying out the invention]

[0015] The fraud detection method of the present invention is - A step of receiving a dataset containing a series of sales requests and sales transactions, - Steps include receiving these profiles, including historical transactions of all objects included within each transaction, into a dataset in chronological order, - The step of updating the received profile by adding a record of the last transaction, - Steps to generate a profile-based dataset, - A step of supplying a dataset to an equilibrium random undersampling algorithm in order to obtain a dataset in which legitimate and fraudulent transactions are distributed according to an equilibrium equation, - Preferably, the steps include oversampling an undersampled dataset, - Steps include supplying the dataset to the deep neural network, - A step of generating two complexity matrices, weighted based on the number of transactions and the cost of the transactions, for the output of a deep neural network, - Steps to generate performance metrics from a complexity matrix, - This primarily includes the step of training a deep neural network using feedback from performance metrics.

[0016] Each data object within a dataset that is thought to contribute to the decision-making mechanism by establishing relationships between transactions and their structure can be defined as a profile. These profiles typically contain statistical information. During the profiling process, information about past records is included in the statistical fields of the profile, regardless of the records in the dataset in which the transaction is currently being executed. In addition to statistical profiles, local profiles (in-record search profiles) are also generated, which reveal some hidden information within the record that belongs only to that record.

[0017] The process steps of receiving profiles, updating profiles, and generating profile-based datasets are performed by a multi-profiling mechanism. The multi-profiling mechanism creates fields to represent the profile history within each transaction. The entire dataset is scanned chronologically to fill these fields with appropriate values. During the scan, credit card profiles are stored in memory, and each transaction uses the information currently available in memory. After all historical information associated with a profile has been processed, the scan is resumed by calculating new statistics and overwriting previous records to update memory. When the scan process is complete and all profiled fields are filled, a profile-based dataset is obtained; that is, the dataset is converted to a profile-based format. Some examples of profiles that can be created by the multi-profiling mechanism and the fields associated with these profiles are given below.

[0018] The first profile created by the multi-profiling mechanism is the credit card profile. All transactions are profiled according to their respective credit card numbers. The data fields that can be used in relation to the credit card profile are listed below. ·Number of past payment cancellations ·Number of transactions accepted in the past ·Number of transactions rejected in the past ·Number of transactions submitted for manual review in the past ·Number of domestic flights purchased to date with this credit card ·Number of international flights purchased to date with this credit card ·Number of unique phone numbers used in the past with this credit card ·Number of unique email addresses used in the past with this credit card ·Average price spent with this credit card in the past ·Highest price spent with this credit card in the past

[0019] The second profile created by the multi-profiling mechanism is a PNR profile. All transactions are profiled according to their respective PNR codes. Data fields that can be used in connection with the PNR profile are listed below. ·Number of past payment cancellations ·Number of transactions accepted in the past ·Number of transactions rejected in the past ·Number of transactions submitted for manual review in the past ·Number of domestic flights attempted to be purchased under this PNR code to date ·Number of international flights attempted to be purchased under this PNR code to date ·Number of unique credit cards used in the past with this PNR code ·Number of unique phone numbers used in the past with this PNR code ·Number of unique email addresses used in the past with this PNR code

[0020] The third profile created by the multi-profiling mechanism is the in-record search profile. This profile contains information about current transactions, rather than past transactions. The data fields that can be used in connection with the in-record search profile are listed below. • The time remaining until the flight, for example in minutes, at the time the request for the current transaction is received. • Information on whether the credit card holder's last name matches the last name of a passenger on the related flight.

[0021] An example of a dataset profiling algorithm that can be used by the multi-profiling mechanism is provided below. [Table 1]

[0022] In this invention, the BRUS algorithm is used to prevent the generation of imbalanced datasets and inefficient negative sample selection, thereby maximizing the representation of the dataset and eliminating imbalances resulting from random selection. The BRUS algorithm ensures that records selected from the majority class are evenly distributed within the dataset, that chronological order is guaranteed within the newly created dataset, and that the same record is not selected more than once.

[0023] A hybrid sampling algorithm can also be used in this invention by oversampling an undersampled dataset after an undersampling process performed by the BRUS algorithm. Oversampling performed within the scope of the hybrid sampling algorithm can increase the illicit / legal transaction ratio to a desired ratio, thereby training a deep neural network.

[0024] A cost-focused metric is implemented to generate a complexity matrix weighted based on transaction costs, and therefore performance metrics. The performance metrics derived from the complexity matrix are precision, precision, recall, and FI score. Unlike solutions known in the art where success is calculated based on the number of transactions, thereby yielding results unrelated to actual revenue loss, the cost-focused metric ensures that fraud detection success is achieved by considering transaction costs. In this way, it is possible to avoid failure-based decisions, such as blocking legitimate transactions that could generate high revenue or enabling transactions that could result in high financial losses. In addition, comparing cost-based success with transaction-count-based success allows for an evaluation of the financial characterization of fraudulent and legitimate transactions within the dataset.

[0025] In exemplary embodiments of the present invention, the main steps of the cost-focused measurement mechanism are also described below.

[0026] Airline sales transactions can be conducted in many different currencies. Since fraud risks can also differ depending on the currency, currency is also used as a separate data field. On the other hand, exchange rate standardization is performed to assess cost-based success due to differences in exchange rates. In this process, price information for each transaction is converted to a specific currency according to its value at the time of performance.

[0027] To reduce the cost of computer processing, preferably after standardizing the exchange rate, the price is normalized between 0 and 1 using a minimum-maximum normalization process.

[0028] After exchange rate standardization and price normalization, a cost-based complexity matrix is ​​generated. To generate the cost-based complexity matrix, TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative) are updated by adding the standardized and normalized price weights (eScore) of the processed transactions. The C value for each transaction from which the matrix is ​​generated is calculated as follows: c=currencyUSDRatio*price eScore=(c i -min(c)) / (max(c)-min(c))

[0029] The deep neural network used in the method of the present invention is optimized for use in the airline sector. For this purpose, optimization and calibration tests are performed on the network architecture and hyperparameters to determine the configuration that provides the best success in the airline dataset. The deep neural network is then integrated into a multi-profiling mechanism and a cost-focused measurement mechanism. Thus, fraudulent activity can be detected by utilizing profiling and deep learning techniques.

[0030] In an exemplary embodiment of the present invention, the following steps are performed. • Steps to obtain a 15-month dataset containing fraudulent information regarding airline payments within a single file. The data collector deletes or corrects corrupted lines in the file, corrects unexpected tokens, standardizes data types, and fills in missing GMT values ​​by matching with city information. The data divider performs the step of splitting the data file into monthly data files according to the transaction data. The data cleaner performs the following steps: delete records containing missing or incorrect values ​​in key fields, standardize numerical representations, fill in missing values, delete dirty or invalid columns, and perform numerical type checks and corrections. The dataset combiner performs the step of combining monthly data files into a single data file in the same chronological order. • The profiler performs the step of profiling data files according to profiling rules. - A step of undersampling the primary records on the dataset according to the equilibrium equation using an equilibrium random undersampler. The dynamic data processor performs steps such as eliminating reusable data fields, eliminating data fields with too many categories, converting different currencies to a single currency, and performing minimum-maximum normalization of numeric fields. The step of applying one-hot encoding to categorical data fields in a dataset using a one-hot encoder. • Function selector, CHI 2 Steps to write the value, • After the transactions are sorted chronologically by date and time, the training test set organizer distributes the transactions. • Steps to supply the dataset to the deep neural network, • A step to obtain transaction-based and transaction cost-based measurement results using a success measurement tool.

[0031] Additionally, optional support mechanisms such as visualizers, file reporters, and loggers capable of generating test results in tabular format can be used.

Claims

1. A computer implementation method for detecting fraudulent activity in the sale of airline tickets and services, A step of receiving a dataset containing a series of airline ticket and service sales requests and sales transactions, The steps include receiving a profile in the dataset containing past transactions in chronological order for all objects included in each transaction, The steps include updating the received profile by adding a record of the last transaction, Steps to generate a profile-based dataset, To obtain a dataset in which legitimate and fraudulent transactions are distributed according to an equilibrium equation, the steps include supplying the dataset to an equilibrium random undersampling algorithm, The steps include supplying the aforementioned dataset to a deep neural network, The steps include generating a complexity matrix for the output of the deep neural network, The steps include generating performance indicators from the aforementioned complexity matrix, A computer implementation method comprising the step of training the deep neural network using feedback of the performance indicators.

2. The method according to claim 1, comprising the step of generating the complexity matrix weighted based on the cost of the transaction.

3. The method according to claim 1 or 2, comprising the step of generating the complexity matrix weighted based on the number of transactions.

4. The method according to any one of claims 1 to 3, comprising the step of oversampling an undersampled dataset, wherein the step of supplying the dataset to an equilibrium random undersampling algorithm in order to obtain a dataset having equilibrium variances of legitimate and fraudulent transactions.

5. The method according to any one of claims 1 to 4, wherein the time series order within the dataset is obtained by a balanced random undersampling algorithm.

6. The method according to claim 2, further comprising the step of performing exchange rate standardization before the complexity matrix, which is weighted based on the cost of the transaction, is generated.

7. The method according to claim 2 or 4, further comprising the step of performing price normalization before the complexity matrix, weighted according to the cost of the transaction, is generated.

Citation Information

Patent Citations

  • System and method for verifying a financial instrument

    EP1356438A2