Automobile financial fraud detection method, system and equipment based on deep learning

Abstract

The invention discloses an automobile financial fraud detection method, system and device based on deep learning, and relates to the technical field of data processing, and the method comprises the steps: obtaining a transaction subject, obtaining historical financial transaction information and a historical detection time period according to the transaction subject, and obtaining the number of times of initiation, an initiation time point and an initiation amount; generating a time sequence, obtaining time interval data, obtaining a first trend index, and obtaining a first parameter; generating an amount sequence, obtaining amount difference data, obtaining a second trend index, and obtaining a second parameter; and obtaining an initial probability threshold, obtaining an adjustment parameter, correcting the initial probability threshold according to the adjustment parameter, obtaining a target probability threshold, obtaining a target fraud probability according to the probability model, the first parameter and the second parameter, and obtaining an automobile financial fraud detection result according to the target fraud probability and the target probability threshold. The method has the advantages of dynamic adjustment, high detection efficiency and good detection effect.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method, system and device for detecting automotive financial fraud based on deep learning. Background Art

[0002] With the rapid development of the automotive finance industry, automotive financial transactions are becoming increasingly frequent, providing convenient financing and payment means for consumers and enterprises. However, the accompanying financial fraud risks are also becoming increasingly prominent, becoming one of the key factors restricting the healthy development of the automotive finance industry. Financial fraud not only brings direct economic losses to financial institutions, but also may damage consumer trust and affect market stability. At the same time, in recent years, deep learning technology has made remarkable progress in various fields, and its advantages in feature extraction, pattern recognition and data mining provide new ideas for financial fraud detection. Deep learning can automatically extract useful feature representations from a large amount of transaction data without manual intervention, thus greatly improving the acquisition efficiency and accuracy of effective data; however, applying deep learning to automotive financial fraud detection still faces many challenges.

[0003] On the one hand, the automotive financial transaction data extracted by deep learning features has characteristics such as temporality, multi-dimensionality and heterogeneity. How to effectively utilize these characteristics to assist subsequent data processing is the key to its application in financial fraud detection. On the other hand, existing fraud detection methods often use fixed probability thresholds or rules to judge whether a transaction has fraud risks. This method ignores the differences under different transaction subjects and transaction scenarios, resulting in limited accuracy and reliability of fraud detection. Summary of the Invention

[0004] Aiming at the defects in the prior art, the present invention provides a method, system and device for detecting automotive financial fraud based on deep learning.

[0005] A method for detecting automotive financial fraud based on deep learning, comprising: obtaining a transaction subject associated with automotive financial transaction data, obtaining historical financial transaction information and a historical detection time period based on the transaction subject, and obtaining the number of initiation times, initiation time points, and initiation amounts of automotive financial transaction data associated with the transaction subject during the historical detection time period based on deep learning and the historical financial transaction information; sorting multiple initiation time points in chronological order to generate a time series, obtaining time interval data between adjacent initiation time points according to the time series, obtaining a first trend index based on the multiple time interval data, and obtaining a first parameter based on a first calculation model and the first trend index; sorting multiple initiation amounts in the order of the time series to generate an amount series, obtaining amount difference data between adjacent initiation amounts according to the amount series, obtaining a second trend index based on the multiple amount difference data, and obtaining a second parameter based on a second calculation model and the second trend index; obtaining an initial probability threshold according to the number of initiation times, obtaining an adjustment parameter based on a correction model, the first trend index, and the second trend index, correcting the initial probability threshold according to the adjustment parameter to obtain a target probability threshold, and obtaining a target fraud probability according to a probability model, the first parameter, and the second parameter, and obtaining an automotive financial fraud detection result according to the target fraud probability and the target probability threshold.

[0006] Optionally, obtaining historical financial transaction information and a historical detection time period based on the transaction subject includes: identifying identity identification information of the transaction subject; accessing a database or a data warehousing system using the identity identification information, retrieving and extracting historical financial transaction records associated with the transaction subject; and determining the historical detection time period according to the identity identification information.

[0007] Optionally, obtaining an automotive financial fraud detection result according to the target fraud probability and the target probability threshold includes: determining whether the target fraud probability exceeds the target probability threshold; if it exceeds, determining that there is a fraud risk for the transaction subject in the automotive financial transaction data; if it does not exceed, determining that there is no fraud risk for the automotive financial transactions of the transaction subject during the historical detection time period, and obtaining a risk interval according to the target probability threshold and the target fraud probability.

[0008] Optionally, obtaining a first trend index based on multiple time interval data is expressed as: wherein, TR 1 is the first trend index, ΔT i is the time interval data between the i-th adjacent initiation time points in the time series, ΔT i+1 is the time interval data between the (i + 1)-th adjacent initiation time points in the time series, n is the number of time interval data in the time series, and α is the influence degree.

[0009] Optionally, the first calculation model in the first parameter obtained based on the first calculation model and the first trend indicator is expressed as: where P 1 is the first parameter, TR 1 is the first trend indicator, ΔT i is the time interval data between the i-th adjacent initiation time points in the time series, and n is the number of time interval data in the time series.

[0010] Optionally, the correction model in the adjustment parameter obtained based on the correction model, the first trend indicator, and the second trend indicator is expressed as: where β is the adjustment parameter, t is the parameter standard value, TR 1 is the first trend indicator, and TR 2 is the second trend indicator.

[0011] Optionally, the probability model in the target fraud probability obtained based on the probability model, the first parameter, and the second parameter is expressed as: where P t is the target fraud probability, P 1 is the first parameter, and P 2 is the second parameter.

[0012] There is also provided a deep learning-based automotive finance fraud detection system for implementing any one of the deep learning-based automotive finance fraud detection methods described above. The system includes: an acquisition module for acquiring a transaction subject associated with automotive finance transaction data, obtaining historical financial transaction information and a historical detection time period according to the transaction subject, and obtaining the number of initiations, initiation time points, and initiation amounts of automotive finance transaction data associated with the transaction subject during the historical detection time period based on deep learning and historical financial transaction information; a first calculation module for sorting multiple initiation time points in chronological order to generate a time series, obtaining time interval data between adjacent initiation time points according to the time series, obtaining a first trend indicator according to multiple time interval data, and obtaining a first parameter based on the first calculation model and the first trend indicator; a second calculation module for sorting multiple initiation amounts in the order of the time series to generate an amount series, obtaining amount difference data between adjacent initiation amounts according to the amount series, obtaining a second trend indicator according to multiple amount difference data, and obtaining a second parameter based on the second calculation model and the second trend indicator; a fraud detection module for obtaining an initial probability threshold according to the number of initiations, obtaining an adjustment parameter based on the correction model, the first trend indicator, and the second trend indicator, correcting the initial probability threshold according to the adjustment parameter to obtain a target probability threshold, obtaining a target fraud probability according to the probability model, the first parameter, and the second parameter, and obtaining an automotive finance fraud detection result according to the target fraud probability and the target probability threshold.

[0013] An electronic device is also provided, including: a memory on which a computer program is stored; and a processor configured to execute the computer program in the memory to implement the above-mentioned deep learning-based automotive finance fraud detection method.

[0014] A non-transitory computer-readable storage medium is also provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned deep learning-based automotive finance fraud detection method is implemented.

[0015] The beneficial effects of the present invention are as follows:

[0016] In the entire deep learning-based automotive finance fraud detection method, first, through the application of deep learning technology, features closely related to fraud detection, such as the number of transaction initiations, initiation time points, and initiation amounts, can be automatically and quickly extracted from a large amount of transaction data, improving the efficiency and accuracy of data processing; further, in the solution, in-depth time series analysis and trend mining are performed on the initiation time points and initiation amounts. By calculating time interval data, amount difference data, and corresponding trend indicators, the behavioral trends of transaction entities in the time and amount dimensions are quantified, providing richer and more detailed information for subsequent fraud detection; further, a dynamically adjusted probability threshold mechanism is introduced. According to the historical behavior characteristics of transaction entities and the current transaction situation, adjustment parameters are calculated through a correction model to correct the initial probability threshold, obtaining a target probability threshold that more conforms to the actual transaction situation, effectively solving the problem that fixed probability thresholds or rules in traditional fraud detection methods ignore the differences between different transaction entities and transaction scenarios, and significantly improving the accuracy and reliability of fraud detection; further, the target fraud probability is calculated through a probability model and compared with the target probability threshold to obtain an accurate fraud detection result, providing strong support for risk management and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 It is a schematic diagram of the steps of the deep learning-based automotive finance fraud detection method of the present invention;

[0019] Figure 2 It is a partial schematic diagram of the steps of S1 in the deep learning-based automotive finance fraud detection method of the present invention;

[0020] Figure 3This is a schematic diagram of some steps of S4 in the vehicle finance fraud detection method based on deep learning of the present invention;

[0021] Figure 4 This is a block diagram of an electronic device shown in an embodiment of the present invention.

[0022] Reference numerals:

[0023] 700 - Electronic device, 701 - Processor, 702 - Memory, 703 - Multimedia component, 704 - I / O interface, 705 - Communication component. Detailed implementation manners

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the accompanying drawings here can be arranged and designed in various different configurations.

[0025] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0026] It should be noted that: Similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In addition, the terms "first", "second", etc. are only used for descriptive distinction and cannot be understood as indicating or implying relative importance.

[0027] As Figure 1 shown, a vehicle finance fraud detection method based on deep learning is provided, including:

[0028] S1. Obtain a transaction subject associated with vehicle finance transaction data, obtain historical financial transaction information and a historical detection time period according to the transaction subject, and obtain the initiation times, initiation time points, and initiation amounts of the vehicle finance transaction data associated with the transaction subject within the historical detection time period based on deep learning and the historical financial transaction information;

[0029] S2. Sort multiple initiation time points in chronological order to generate a time series, obtain time interval data between adjacent initiation time points according to the time series, obtain a first trend index according to the multiple time interval data, and obtain a first parameter based on a first calculation model and the first trend index;

[0030] S3. Sort multiple initiation amounts in chronological order to generate an amount sequence, obtain the amount difference data between adjacent initiation amounts according to the amount sequence, obtain a second trend indicator based on the multiple amount difference data, and obtain a second parameter based on the second calculation model and the second trend indicator;

[0031] S4. Obtain an initial probability threshold according to the number of initiations, obtain an adjustment parameter based on the correction model, the first trend indicator and the second trend indicator, correct the initial probability threshold according to the adjustment parameter to obtain a target probability threshold, and obtain a target fraud probability according to the probability model, the first parameter and the second parameter. Obtain the automotive finance fraud detection result according to the target fraud probability and the target probability threshold.

[0032] In this embodiment, it should be noted that in S1, it first involves obtaining the transaction entity associated with the current automotive finance transaction data. In this process, specific individuals or entities conducting the transaction will be comprehensively collected and identified, which may be individual consumers, enterprises or any other legal entities participating in automotive finance transactions. Then, according to the unique identifier of this transaction entity, its historical financial transaction information will be deeply mined. This information includes but is not limited to past transaction records, transaction habits, credit status, etc. They provide rich background data for subsequent fraud detection. At the same time, a historical detection time period will be determined, which is set according to the characteristics and needs of the transaction entity, aiming to ensure that the analyzed data is both representative and can reflect the latest behavior patterns of the transaction entity. In this step, the application of deep learning technology enables the efficient extraction of features closely related to the transaction entity and crucial for fraud detection from massive data. For example, using the feature extraction layer in the deep learning model, key features related to fraud detection, such as the number of transaction initiations, initiation time points, and initiation amounts, are extracted from the preprocessed transaction records.

[0033] An example is given to improve the description of step S1: Suppose a transaction entity named "XX Automotive Finance Company" is applying for a new car loan. In step S1, "XX Automotive Finance Company" will be first identified and confirmed as the transaction entity, and then all automotive finance transaction records of this company in the past year (the historical detection time period obtained according to this transaction entity is one year) will be retrieved and extracted by accessing the internal database or external data warehouse. These records may include the number of loan applications initiated by this company, the submission time (i.e., the initiation time point) of each application, and the loan amount applied for (i.e., the initiation amount), etc. Through the collection and analysis of this information, a comprehensive and accurate data basis can be provided for subsequent fraud detection.

[0034] In S2, it is mainly responsible for performing time series analysis on the extracted initiation time points. First, multiple initiation time points are sorted in chronological order to generate a clear time series. This time series reflects the time distribution and frequency of the trading entity's automotive finance transactions during the historical detection period. Then, based on this time series, the time interval data between adjacent initiation time points is calculated. These time interval data can reveal the regularity and abnormality of the trading entity's transactions, such as whether the transactions are becoming more concentrated over time or whether there are sudden changes in the trading frequency. Next, based on these time interval data, a first trend indicator is calculated, which can quantify the trading behavior trend of the trading entity in the time dimension. Finally, using the first calculation model and combining this first trend indicator, a first parameter is calculated. This first parameter will serve as an important basis for subsequent correction of the probability threshold and calculation of the target fraud probability, helping to more accurately assess the fraud risk of the trading entity.

[0035] To illustrate and improve the description of step S2: Suppose "XX Automotive Finance Company" has initiated multiple car loan applications in the past year, and the submission time of each application has been recorded. In step S2, first, these submission times are sorted in chronological order to generate a time series. Then, the time intervals between adjacent submissions are calculated. For example, the interval between the first and second applications is 30 days, and the interval between the second and third applications is 20 days, and so on. Next, based on these time interval data, a first trend indicator is calculated, such as the average time interval, variance of the time intervals, etc. Finally, using the first calculation model and combining this first trend indicator, a first parameter is calculated. This first parameter will reflect the trading behavior trend of "XX Automotive Finance Company" in the time dimension and provide strong support for subsequent fraud detection.

[0036] In S3, focus on the serial analysis and trend mining of the initiated amount. First, multiple initiated amounts are sorted in the order of the time series to generate an amount sequence. This amount sequence clearly shows the amount distribution and change trend of the transaction entity in the historical detection time period for auto finance transactions. Then, based on this amount sequence, the amount difference data between adjacent initiated amounts is calculated. These amount difference data can reveal the regularity and abnormal fluctuations in the transaction amount of the transaction entity, such as whether the transaction amount gradually increases or decreases, or whether there are sudden jumps in the amount. Then, based on these amount difference data, a second trend indicator is calculated, which can quantify the transaction behavior trend of the transaction entity in the amount dimension. Finally, using the second calculation model and combining this second trend indicator, a second parameter is calculated. This second parameter will be an important reference for subsequent correction of the probability threshold and calculation of the target fraud probability, and helps to more comprehensively evaluate the fraud risk of the transaction entity.

[0037] Take an example to improve the description of step S3: Suppose "XX Auto Finance Company" has initiated multiple auto loan applications in the past year, and the loan amount of each application is recorded. In step S3, first, these loan amounts are sorted in chronological order to generate an amount sequence. Then, calculate the differences between adjacent loan amounts. For example, the amount difference between the first and second applications is 500,000 yuan, and the amount difference between the second and third applications is -200,000 yuan (indicating a decrease in amount), and so on. Then, based on these amount difference data, a second trend indicator is calculated, such as the average amount difference, the variance of the amount difference, etc. Finally, using the second calculation model and combining this second trend indicator, a second parameter is calculated. This second parameter will reflect the transaction behavior trend of "XX Auto Finance Company" in the amount dimension and provide more comprehensive data support for subsequent fraud detection.

[0038] In S4, it is responsible for determining whether there is a fraud risk in the transaction based on the features extracted in the previous steps and the calculated parameters. First, an initial probability threshold is obtained according to the number of initiations. This threshold is set based on historical data and security requirements and is used to judge whether there is fraud. Then, a correction model is used to calculate an adjustment parameter in combination with the first trend indicator (reflecting the transaction behavior trend in the time dimension) and the second trend indicator (reflecting the transaction behavior trend in the amount dimension). This adjustment parameter is a correction to the initial probability threshold. It takes into account the behavioral changes of the transaction subject in terms of time and amount, making the probability threshold more in line with the actual situation of the current transaction. Next, a probability model is used to calculate a target fraud probability in combination with the first parameter (calculated from the time dimension) and the second parameter (calculated from the amount dimension). This probability reflects the likelihood of the current transaction having a fraud risk. Finally, the target fraud probability is compared with the target probability threshold. If the target fraud probability exceeds the target probability threshold, it is determined that there is a fraud risk for the transaction subject in the automotive finance transaction data; otherwise, it is determined that there is no fraud risk for the transaction subject in the automotive finance transactions during the historical detection period, and a risk interval is calculated based on the target probability threshold and the target fraud probability for further risk assessment and management.

[0039] An example is given to improve the description of step S4: Suppose the automotive loan application data of "XX Automotive Finance Company" in the past year has been processed through steps S1 to S3, obtaining the number of initiations, the first parameter in the time dimension, and the second parameter in the amount dimension. In step S4, first, an initial probability threshold is obtained according to the number of initiations. For example, it is set to 0.5 (indicating that when the fraud probability exceeds 0.5, it is considered that there is a fraud risk). Then, a correction model is used to calculate an adjustment parameter in combination with the first trend indicator and the second trend indicator. For example, the adjustment parameter is 0.8. This adjustment parameter makes the probability threshold more in line with the actual transaction situation of "XX Automotive Finance Company". The initial probability threshold is corrected according to the adjustment parameter to obtain the target probability threshold. For example, 0.5 * 0.8 gives 0.4, and the target probability threshold is 0.4. Next, a probability model is used to calculate a target fraud probability in combination with the first parameter and the second parameter. For example, the target fraud probability is 0.6. Finally, the target fraud probability of 0.6 is compared with the target probability threshold of 0.4. Since 0.6 is greater than 0.4, it is determined that "XX Automotive Finance Company" has a fraud risk in the current transaction and further investigation and processing are required.

[0040] In summary, in the entire deep learning-based automotive finance fraud detection method, first, through the application of deep learning technology, features closely related to fraud detection, such as the number of transaction initiations, initiation time points, and initiation amounts, can be automatically and quickly extracted from a large amount of transaction data, improving the efficiency and accuracy of data processing. Further, in the solution, in-depth time series analysis and trend mining are performed on the initiation time points and initiation amounts. By calculating time interval data, amount difference data, and corresponding trend indicators, the behavioral trends of transaction entities in the time and amount dimensions are quantified, providing richer and more detailed information for subsequent fraud detection. Further, a dynamically adjusted probability threshold mechanism is introduced. According to the historical behavior characteristics of the transaction entity and the current transaction situation, adjustment parameters are calculated through a correction model to correct the initial probability threshold, obtaining a target probability threshold that better conforms to the actual transaction situation, effectively solving the problem that fixed probability thresholds or rules in traditional fraud detection methods ignore the differences between different transaction entities and transaction scenarios, and significantly improving the accuracy and reliability of fraud detection. Further, the target fraud probability is calculated through a probability model and compared with the target probability threshold to obtain an accurate fraud detection result, providing strong support for risk management and decision-making.

[0041] As Figure 2 shown, in one embodiment, in S1, obtaining historical financial transaction information and historical detection time periods according to the transaction entity includes:

[0042] S11. Identify the identity identification information of the transaction entity;

[0043] S12. Use the identity identification information to access the database or data warehousing system, retrieve and extract the historical financial transaction records associated with the transaction entity;

[0044] S13. Determine the historical detection time period according to the identity identification information.

[0045] In this embodiment, it should be noted that in S11, first, the identity identification information of the transaction entity needs to be identified. This step is the basis for ensuring that historical financial transaction records related to a specific transaction entity can be accurately extracted subsequently. The identity identification information includes, but is not limited to, the name of the transaction entity, the ID number (for individual consumers), the enterprise registration number (for enterprises), or any other information that can uniquely identify the transaction entity. Through these identity identification information, it can be determined who the subject of the current transaction is, thus providing an accurate positioning for subsequent data extraction and analysis. For example, if the transaction entity is an enterprise named "XX Automotive Finance Company", then its enterprise registration number is its unique identity identification information, and this registration number will be used to identify and confirm the transaction entity.

[0046] In S12, using the identity identification information identified in step S11, access the database or data warehouse, retrieve and extract historical financial transaction records associated with the transaction entity. This step is a process of obtaining key information such as the past transaction behaviors, transaction habits, and credit status of the transaction entity. A large amount of financial transaction data is stored in the database or data warehouse. By using the identity identification information as the retrieval condition, the data records related to a specific transaction entity can be quickly located. For example, for "XX Auto Finance Company", its enterprise registration number will be used to access the database and extract all the auto finance transaction records of the company in the past year, including details such as the number of loan applications, the submission time of each application, and the application amount.

[0047] In S13, determine the historical detection time period according to the identity identification information. This step is to ensure that the analyzed data is both representative and can reflect the latest behavior patterns of the transaction entity. The determination of the historical detection time period can be set according to the characteristics and needs of the transaction entity. For example, different time periods such as the past year, half year, or three months can be selected. In S13, factors such as the historical transaction records, transaction frequency, and current transaction situation of the transaction entity will be comprehensively considered to determine the most appropriate historical detection time period. For example, for "XX Auto Finance Company", the past year may be selected as the historical detection time period based on its past transaction records and the current market environment to ensure that the analyzed data can accurately reflect the latest transaction behavior patterns of the company.

[0048] As Figure 3 shown, in one embodiment, obtaining the auto finance fraud detection result according to the target fraud probability and the target probability threshold in S4 includes:

[0049] S41. Judge whether the target fraud probability exceeds the target probability threshold;

[0050] S42. If it exceeds, determine that there is a fraud risk for the transaction entity in the auto finance transaction data;

[0051] S43. If it does not exceed, determine that there is no fraud risk for the auto finance transactions of the transaction entity within the historical detection time period, and obtain the risk interval according to the target probability threshold and the target fraud probability.

[0052] In this embodiment, it should be noted that in S41, the target fraud probability calculated by the probability model before is directly compared with the target probability threshold adjusted by the correction model. After obtaining the comparison result, the subsequent fraud risk determination process will be triggered. For example, assume that the target fraud probability is 0.6 and the target probability threshold is 0.4. Since 0.6 is greater than 0.4, this difference will be identified and it will be ready to enter S42, that is, to determine that the transaction entity has a fraud risk.

[0053] In S42, when the result of S41 shows that the target fraud probability exceeds the target probability threshold, step S42 is immediately initiated, and it is officially determined that there is a fraud risk in the automotive finance transaction data of this transaction entity. This determination is based on the data processing and calculation results of all previous steps and is the conclusion after in-depth analysis of the behavior pattern of the transaction entity. Continuing with the above example, since the target fraud probability of "XX Automotive Finance Company" is 0.6, which exceeds the target probability threshold of 0.4, a fraud risk report will be generated, indicating that there is a fraud risk in this company and further investigation is required.

[0054] In S43, if the result of S41 shows that the target fraud probability does not exceed the target probability threshold, then S43 will be executed to determine that there is no fraud risk in the automotive finance transactions of this transaction entity during the historical detection time period. At the same time, in order to further refine risk management, a risk interval will also be calculated based on the difference between the target probability threshold and the target fraud probability. This risk interval can reflect the relative risk level of the current transaction behavior of the transaction entity. Even if there is no fraud risk in the transaction, the risk interval can provide additional reference information for risk managers. For example, if the target fraud probability is 0.3 and the target probability threshold is 0.4, it will not only be determined that "XX Automotive Finance Company" has no fraud risk in the current transaction, but also a risk interval will be calculated, such as 0.1 (i.e., 0.4 - 0.3), indicating that the company's transaction behavior has a certain safety distance from the fraud risk threshold, but risk managers still need to pay attention to the risk changes within this interval.

[0055] In one implementation, the first trend indicator obtained from multiple time interval data in S2 is expressed as:

[0056] Where

[0057] TR 1 is the first trend indicator, ΔT i is the time interval data between the i-th adjacent initiation time points in the time series, ΔT i+1 is the time interval data between the (i + 1)-th adjacent initiation time points in the time series, n is the number of time interval data in the time series, and α is the influence degree.

[0058] In this implementation, it should be noted that the sign function sgn(*) is used to judge the change direction of the time interval; when the variable is greater than 0, sgn(*) = 1; when the variable is equal to 0, sgn(*) = 0; when the variable is less than 0, sgn(*) = -1. If ΔT i >ΔT i+1 , it indicates that the time interval is decreasing and the transaction frequency may be increasing. At this time, sgn(ΔT i -ΔT i+1) = 1. Conversely, if ΔT i < ΔT i+1 , it indicates that the time interval is increasing and the trading frequency may be decreasing. At this time, sgn(ΔT i - ΔT i+1 ) = -1. Through the sign function, the directionality of the change in the time interval can be captured, which is important information for evaluating the trend of trading behavior.

[0059] |ΔT i - ΔT i+1 | α The absolute value difference represents the actual change amount of adjacent time intervals. By taking the power α, the influence degree of the difference on the trend index can be adjusted. When α > 1, the influence of the difference is amplified; when 0 < α < 1, the influence of the difference is reduced. This setting allows for flexible adjustment of the sensitivity of the trend index according to specific application scenarios and requirements. In most cases, α is equal to 1 to increase the influence of the difference.

[0060] represents summing up the differences of all adjacent time intervals and dividing by n - 1 (because n time intervals will generate n - 1 differences) to obtain the average trend index; at the same time, the averaging process can smooth out the influence of individual outliers and make the trend index more stable and reliable. At the same time, the larger ΔT i - ΔT i+1 , the higher TR 1 , which represents a faster trading frequency and higher risk.

[0061] Assume α = 1. XX Auto Finance Company initiated 5 car loan applications in the past year. The time intervals between the 5 car loan applications are: ΔT 1 = 45, ΔT 2 = 33, ΔT 3 = 21, ΔT 4 = 15. Then calculate the differences: ΔT 1 - ΔT 2 = 45 - 33 = 12; ΔT 2 - ΔT 3 = 33 - 21 = 12; ΔT 3 - ΔT 4 = 21 - 15 = 6. After substituting into the expression,

[0062] It should also be noted that in S3, obtaining the second trend index based on multiple amount difference data is similar to obtaining the first trend index, but there are differences. The first trend index obtained based on multiple time interval data in S3 is expressed as: where TR 2 is the second trend index, ΔA iis the amount difference data between the i-th adjacent initiated amounts in the amount sequence, ΔA i+1 is the amount difference data between the (i + 1)-th adjacent initiated amounts in the amount sequence, n is the number of amount difference data in the amount sequence, and α is the influence degree.

[0063] Among them, what is different between the expression for obtaining the second trend indicator and the expression for obtaining the first trend indicator is that the sign function is sgn(ΔA i+1 -ΔA i ), because it is necessary to ensure that the higher the TR 2 represents the higher the risk. In the dimension of transaction amount, the higher the risk, the trend of the transaction amount shows a gradual increase. Therefore, it is necessary to ensure that the variable of the sign function is ΔA i+1 -ΔA i ; this is different from the dimension of transaction time. In the dimension of transaction time, the higher the risk, the trend of the transaction time interval shows a gradual decrease.

[0064] In one embodiment, the first calculation model in obtaining the first parameter based on the first calculation model and the first trend indicator in S2 is expressed as:

[0065] Among them,

[0066] P 1 is the first parameter, TR 1 is the first trend indicator, ΔT i is the time interval data between the i-th adjacent initiated time points in the time sequence, and n is the number of time interval data in the time sequence.

[0067] In this embodiment, it should be noted that TR 1 quantifies the change trend of the transaction frequency by considering the differences and directions of adjacent time intervals; a higher TR 1 value indicates a larger change in the time interval, which may mean the instability or abnormality of transaction behaviors, thus being related to fraud risks. The denominator calculates the average value of all time intervals; the average time interval provides a benchmark for the transaction frequency, which is used to standardize the first trend indicator; by dividing by the average time interval, the first parameter P 1 can reflect the intensity of the trend change relative to the average transaction frequency. P 1 combines the trend indicator and the average time interval, providing a standardized metric for evaluating the behavioral changes of the transaction entity in the time dimension.

[0068] It should also be noted that in S3, the second calculation model in the second parameter obtained based on the second calculation model and the second trend indicator is exactly the same as the first calculation model. The second calculation model in the second parameter obtained based on the second calculation model and the second trend indicator in S3 is expressed as: where P 2 is the second parameter, TR 2 is the second trend indicator, ΔA i is the amount difference data between the i-th adjacent initiating amounts in the amount sequence, and n is the number of amount difference data in the amount sequence.

[0069] In one embodiment, the correction model in the adjustment parameter obtained based on the correction model, the first trend indicator, and the second trend indicator in S4 is expressed as:

[0070] where

[0071] β is the adjustment parameter, t is the parameter standard value, TR 1 is the first trend indicator, and TR 2 is the second trend indicator.

[0072] In this embodiment, it should be noted that TR 1 and TR 2 represent the transaction behavior trends in the time and amount dimensions respectively; by introducing these two indicators, the correction model can consider the behavioral changes of the transaction subject in different dimensions, so as to more comprehensively evaluate the fraud risk.

[0073] max(TR 1 , 1) and max(TR 2 , 1) ensure that the trend indicator is at least 1; this is to screen out the first trend indicator or the second trend indicator less than 1. If the first trend indicator or the second trend indicator is negative, it means there is no risk and does not need to be brought into the calculation of the adjustment parameter, which does not affect the judgment of the threshold; if the first trend indicator or the second trend indicator approaches 0, it will cause the denominator to be too small and the adjustment parameter to be abnormally amplified. By setting the minimum value to 1, the stability of the adjustment parameter can be guaranteed.

[0074] and are to convert the trend indicator into an influence factor on the adjustment parameter. The larger the trend indicator, the smaller its reciprocal, the smaller it is, and the smaller the adjustment parameter; conversely, the smaller the trend indicator, the larger its reciprocal, and the larger the adjustment parameter.

[0075] The structure of t(1+*) is used to amplify the influence of the trend indicator on the adjustment parameter. When the trend indicator is small, the denominator is small, the amplification factor is equal to 2, and the adjustment parameter is equal to 2t; when the trend indicator is large, the denominator is large, the amplification factor approaches 1, and the adjustment parameter is equal to t. Among them, t is used as the parameter standard value to control the benchmark range of the adjustment parameter. By adjusting the value of t, the sensitivity and range of the adjustment parameter can be flexibly set to adapt to different application scenarios and requirements; preferably, t = 0.5.

[0076] Suppose TR 1 = 10 (calculated based on time interval data), TR 2 = 15 (calculated based on amount difference data), and the parameter standard value t = 0.5.

[0077] Substituting into the expression of the correction model, it is calculated that β = 0.5 * (1 + 0.167) = 0.5835.

[0078] It should also be noted that to correct the initial probability threshold according to the adjustment parameter and obtain the target probability threshold, the adjustment parameter can be directly multiplied by the initial probability threshold. For example, if the initial probability threshold is 0.5 and the calculated adjustment parameter β is 0.5835, the target probability threshold is 0.5 * 0.5835 = 0.29175.

[0079] In one embodiment, the probability model for obtaining the target fraud probability according to the probability model, the first parameter, and the second parameter in S4 is expressed as:

[0080] Among them,

[0081] P t is the target fraud probability, P 1 is the first parameter, and P 2 is the second parameter.

[0082] In this embodiment, it should be noted that exp[-(P 1 + P 2 )] converts the linear input into a non-linear output. In financial fraud detection, the changes in transaction behaviors are not linear, so using an exponential function can more accurately reflect this non-linear relationship. The entire expression maps P 1 and P 2 to the interval (0, 1), which enables the output P t to be interpreted as a probability. In financial fraud detection, a probability value is required to represent the possibility of a transaction having a fraud risk. P 1 + P 2It means adding the parameters of the time dimension and the amount dimension. This setting is because fraudulent behaviors often manifest in multiple dimensions simultaneously. For example, an abnormal increase in transaction frequency and an abnormal fluctuation in transaction amount. By summing them up, the information of these two dimensions can be comprehensively considered, thus more comprehensively evaluating the fraud risk. -(P 1 +P 2 ) The negative sign is used to adjust the range of the input P 1 P 2 to a suitable probability output range. When the input is 0, P t outputs 0.5; when P 1 +P 2 is relatively large (indicating a relatively high fraud risk), -(P 1 +P 2 ) is relatively small, and P t outputs close to 1; conversely, when P 1 +P 2 is relatively small, -(P 1 +P 2 ) is relatively large, and the P t function output is close to 0.

[0083] A deep learning-based automotive finance fraud detection system is also provided. The system is used to implement the deep learning-based automotive finance fraud detection method in any of the above real-time manners. The system includes:

[0084] An acquisition module, which is used to acquire the transaction subject associated with the automotive finance transaction data, obtain the historical financial transaction information and the historical detection time period according to the transaction subject, and obtain the number of initiations, the initiation time points, and the initiation amounts of the automotive finance transaction data associated with the transaction subject during the historical detection time period based on deep learning and the historical financial transaction information;

[0085] A first calculation module, which is used to sort multiple initiation time points in chronological order to generate a time series, obtain the time interval data between adjacent initiation time points according to the time series, obtain a first trend indicator according to multiple time interval data, and obtain a first parameter based on the first calculation model and the first trend indicator;

[0086] A second calculation module, which is used to sort multiple initiation amounts in the order of the time series to generate an amount series, obtain the amount difference data between adjacent initiation amounts according to the amount series, obtain a second trend indicator according to multiple amount difference data, and obtain a second parameter based on the second calculation model and the second trend indicator;

[0087] A fraud detection module, which is configured to obtain an initial probability threshold according to the number of initiations, obtain an adjustment parameter based on a correction model, a first trend indicator, and a second trend indicator, correct the initial probability threshold according to the adjustment parameter to obtain a target probability threshold, obtain a target fraud probability according to a probability model, a first parameter, and a second parameter, and obtain an automotive finance fraud detection result according to the target fraud probability and the target probability threshold.

[0088] In one embodiment, the obtaining module is further configured to: identify the identity identification information of the transaction subject; access a database or a data warehousing system by using the identity identification information, retrieve and extract historical financial transaction records associated with the transaction subject; and determine a historical detection time period according to the identity identification information.

[0089] In one embodiment, the fraud detection module is further configured to: determine whether the target fraud probability exceeds the target probability threshold; if it exceeds, determine that the transaction subject has a fraud risk in the automotive finance transaction data; if it does not exceed, determine that there is no fraud risk in the automotive finance transaction of the transaction subject during the historical detection time period, and obtain a risk interval according to the target probability threshold and the target fraud probability.

[0090] In this embodiment, it should be noted that regarding the above-mentioned automotive finance fraud detection system based on deep learning, the specific manner of performing operations has been described in detail in the embodiments of the automotive finance fraud detection method based on deep learning, and will not be elaborated herein.

[0091] Figure 3 is a block diagram of an electronic device for an automotive finance fraud detection method shown according to an exemplary embodiment. As Figure 3 shown, the electronic device 700 may include: a processor 701, a memory 702. The electronic device 700 may further include one or more of a multimedia component 703, an I / O interface 704 (input / output interface), and a communication component 705.

[0092] Among them, the processor 701 is used to control the overall operation of the electronic device 700 to complete all or part of the steps in the above-mentioned automotive finance fraud detection method based on deep learning. The memory 702 is used to store various types of data to support the operation of the electronic device 700. These data may include, for example, instructions for any application or method operating on the electronic device 700, as well as application-related data, such as contact data, sent and received messages, pictures, audio, video, and so on. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc. The multimedia component 703 may include a screen and an audio component. Among them, the screen may be a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 702 or sent through the communication component 705. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 704 provides an interface between the processor 701 and other interface modules, and the above-mentioned other interface modules may be a keyboard, a mouse, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited herein. Therefore, the corresponding communication component 705 may include: a Wi-Fi module, a Bluetooth module, an NFC module, and so on.

[0093] In an exemplary embodiment, the electronic device 700 can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components, and is used to execute the above-mentioned deep learning-based automotive finance fraud detection method.

[0094] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-mentioned deep learning-based automotive finance fraud detection method are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 702 including program instructions, and the above-mentioned program instructions can be executed by the processor 701 of the electronic device 700 to complete the above-mentioned deep learning-based automotive finance fraud detection method.

[0095] In another exemplary embodiment, a computer program product is further provided. The computer program product includes a computer program that can be executed by a programmable device, and the computer program has a code part for executing the above-mentioned deep learning-based automotive finance fraud detection method when executed by the programmable device.

[0096] The preferred embodiments of the present disclosure have been described in detail above in conjunction with the accompanying drawings. However, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept scope of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0097] In addition, it should be noted that, in the above specific embodiments, the various specific technical features described can be combined in any suitable manner without conflict. To avoid unnecessary repetition, the present disclosure does not separately describe various possible combination methods.

[0098] Furthermore, any combination can be made between various different embodiments of the present disclosure as long as it does not violate the idea of the present disclosure, and it should also be regarded as the content disclosed by the present disclosure.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and the description of the present invention.

Claims

1. A method for detecting automobile financial fraud based on deep learning, characterized in that: include: Obtain the transaction subject associated with the automobile financial transaction data, and obtain historical financial transaction information and historical detection time periods based on the transaction subject, and obtain the number of initiations, initiation time points, and initiation amounts of automobile financial transaction data associated with the transaction subject within the historical detection time period based on deep learning and historical financial transaction information; Arrange the multiple initiation time points in chronological order and generate a time series, obtain time interval data between adjacent initiation time points according to the time series, obtain a first trend indicator according to the multiple time interval data, and obtain a first parameter based on the first calculation model and the first trend indicator; Sort the multiple initiated amounts in time series order and generate an amount sequence, obtain amount difference data between adjacent initiated amounts according to the amount sequence, obtain a second trend indicator according to the multiple amount difference data, and obtain a second parameter based on the second calculation model and the second trend indicator; An initial probability threshold is obtained according to the number of initiations, an adjustment parameter is obtained based on the correction model, the first trend indicator and the second trend indicator, the initial probability threshold is corrected according to the adjustment parameter and a target probability threshold is obtained, and a target fraud probability is obtained according to the probability model, the first parameter and the second parameter, and an automobile finance fraud detection result is obtained according to the target fraud probability and the target probability threshold.

2. The automobile finance fraud detection method based on deep learning according to claim 1 is characterized in that: The acquisition of historical financial transaction information and historical detection time periods according to the transaction subject includes: Identify the identity information of the transaction subject; Using identity information to access a database or data storage system to retrieve and extract historical financial transaction records associated with the transaction subject; Determine the historical detection time period based on the identity identification information.

3. The automobile finance fraud detection method based on deep learning according to claim 1 is characterized in that: The obtaining of automobile finance fraud detection results according to the target fraud probability and the target probability threshold comprises: Determine whether the target fraud probability exceeds the target probability threshold; If it exceeds, it is determined that the transaction subject has a fraud risk in the automobile financial transaction data; If it does not exceed, it is determined that there is no fraud risk in the automobile financial transactions of the transaction subject during the historical detection period, and the risk interval is obtained based on the target probability threshold and the target fraud probability.

4. The automobile finance fraud detection method based on deep learning according to claim 1 is characterized in that: The first trend indicator is obtained according to the data of multiple time intervals as follows: in, TR1 is the first trend indicator, ΔT i is the time interval data between the ith adjacent initiation time points in the time series, ΔT i+1 is the time interval data between the i+1th adjacent initiation time points in the time series, n is the number of time interval data in the time series, and α is the degree of influence.

5. The automobile finance fraud detection method based on deep learning according to claim 1 is characterized in that: The first calculation model in obtaining the first parameter based on the first calculation model and the first trend indicator is expressed as: in, P1 is the first parameter, TR1 is the first trend indicator, ΔT i is the time interval data between the i-th adjacent initiation time points in the time series, and n is the number of time interval data in the time series.

6. The automobile finance fraud detection method based on deep learning according to claim 1, characterized in that: The correction model in obtaining the adjustment parameter based on the correction model, the first trend indicator and the second trend indicator is expressed as: in, β is the adjustment parameter, t is the standard value of the parameter, TR1 is the first trend indicator, and TR2 is the second trend indicator.

7. The automobile finance fraud detection method based on deep learning according to claim 1 is characterized in that: The probability model in obtaining the target fraud probability according to the probability model, the first parameter and the second parameter is expressed as: in, P t is the target fraud probability, P1 is the first parameter, and P2 is the second parameter.

8. A deep learning-based automobile finance fraud detection system, characterized in that: The system is used to implement the automobile financial fraud detection method based on deep learning as described in any one of claims 1 to 7, and the system includes: An acquisition module is used to acquire a transaction subject associated with automobile financial transaction data, and acquire historical financial transaction information and historical detection time periods according to the transaction subject, and acquire the number of initiations, initiation time points, and initiation amounts of automobile financial transaction data associated with the transaction subject within the historical detection time period based on deep learning and historical financial transaction information; A first calculation module, used to sort the multiple initiation time points in chronological order and generate a time series, obtain time interval data between adjacent initiation time points according to the time series, obtain a first trend indicator according to the multiple time interval data, and obtain a first parameter based on the first calculation model and the first trend indicator; A second calculation module is used to sort the multiple initiated amounts in time series order and generate an amount sequence, obtain amount difference data between adjacent initiated amounts according to the amount sequence, obtain a second trend indicator according to the multiple amount difference data, and obtain a second parameter based on the second calculation model and the second trend indicator; The fraud detection module is used to obtain an initial probability threshold based on the number of initiations, obtain an adjustment parameter based on a correction model, a first trend indicator and a second trend indicator, correct the initial probability threshold according to the adjustment parameter and obtain a target probability threshold, and obtain a target fraud probability according to the probability model, a first parameter and a second parameter, and obtain an automobile financial fraud detection result according to the target fraud probability and the target probability threshold.

9. The deep learning-based automobile finance fraud detection system according to claim 8, characterized in that: The acquisition module is also used for: Identify the identity information of the transaction subject; Using identity information to access a database or data storage system to retrieve and extract historical financial transaction records associated with the transaction subject; Determine the historical detection time period based on the identity identification information.

10. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the automobile financial fraud detection method based on deep learning as described in any one of claims 1 to 7.

Images