Abnormal financial payment transaction identification method based on neural network

Abstract

The present invention discloses a method for identifying abnormal financial payment transactions based on a neural network, which belongs to the technical field of identifying abnormal financial payment transactions. The present invention constructs an expenditure deviation sequence according to the difference between the transaction amount and daily expenditure of the user on the transaction platform; secondly, extracts the browsing records of the user on the platform, calculates the relevance of the content of each page, and thus constructs a user browsing relevance sequence; then extracts the transaction content of the user, calculates the fit of each transaction content with the browsing content, and forms a user transaction fit sequence; finally, based on a multi-input neural network, the expenditure deviation sequence, the user browsing relevance sequence, and the user transaction fit sequence are used as inputs to obtain a payment abnormality value. The present invention can effectively improve the recognition rate of abnormal transactions and reduce financial risks by comprehensively considering the user's expenditure behavior, browsing habits, and transaction matching.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal financial payment transaction identification, and in particular to a method for abnormal financial payment transaction identification based on a neural network. Background Art

[0002] With the rapid development of e-commerce and online payments, the identification of abnormal financial payment transactions has become an important task to ensure transaction security and user property protection. Existing technologies mainly rely on rule-driven monitoring mechanisms and traditional statistical analysis methods, which usually target specific types of transaction patterns and behaviors, while ignoring the diversity and complexity of individual users. For example, many existing systems identify suspicious transactions by setting fixed thresholds. This static threshold setting is often unable to adapt to the dynamically changing market environment and user behavior, and is prone to false alarms or missed events. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for identifying abnormal financial payment transactions based on a neural network, which solves the problem of low accuracy in identifying abnormal financial payment transactions in the prior art.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for identifying abnormal financial payment transactions based on neural network, comprising the following steps:

[0005] S1. Construct an expenditure deviation sequence based on the difference between the transaction amount and daily expenditure of the user on the trading platform;

[0006] S2, extracting the content of each page in the user's browsing history on the trading platform, calculating the relevance of the user's browsing content, and constructing a user browsing relevance sequence;

[0007] S3, extracting each transaction content of the user on the transaction platform, calculating the transaction fit between each transaction content and the content browsed by the user, and constructing a user transaction fit sequence;

[0008] S4. According to the expenditure deviation sequence, the user browsing association sequence and the user transaction compatibility sequence, a payment abnormality degree value is obtained based on a multi-input neural network.

[0009] Furthermore, S1 comprises the following sub-steps:

[0010] S11. Count the daily expenditure of users and obtain the daily average expenditure and the daily maximum expenditure;

[0011] S12. Record the user's multiple transaction amounts on the trading platform;

[0012] S13, calculating the expenditure deviation value according to the difference between each transaction amount and the daily average expenditure and the daily maximum expenditure;

[0013] S14. Use each expenditure deviation value as an element to construct an expenditure deviation sequence.

[0014] Furthermore, the formula for calculating the expenditure deviation value in S13 is: , where θ i is the i-th expenditure deviation value, P i is the i-th transaction amount, P day,avg is the average daily expenditure, P day,max is the maximum daily expenditure, and i is a positive integer.

[0015] Further, S2 includes the following sub-steps:

[0016] S21, extracting the browsing history of the user on the trading platform;

[0017] S22, separating each page content in the browsing record to obtain multiple page contents browsed by the user;

[0018] S23, arranging the contents of each page in chronological order to obtain a user browsing set;

[0019] S24, in the user browsing set, calculating the relevance of each user browsing content according to the content of each page and the content of adjacent pages;

[0020] S25. Use the relevance of each user's browsing content as an element to construct a user browsing relevance sequence.

[0021] Further, S24 includes the following sub-steps:

[0022] S241, extracting keywords from the content of each page and constructing a page keyword set;

[0023] S242, calculate the intersection-and-union ratio of the page keyword set of each page and the page keyword set of the left neighboring page to obtain a first intersection-and-union ratio: , where S 1,j is the jth first intersection-combination ratio, G j is the set of page keywords for the jth page, G j-1 is the page keyword set of the j-1th page, ∩ is the intersection, ∪ is the union, j is the number of the page in the user's browsing set, and f is the number of keywords in the calculation set;

[0024] S243, calculate the intersection-and-union ratio of each page keyword set and the keyword set of the neighboring page on the right, and obtain a second intersection-and-union ratio: , where S 2,j is the jth second intersection-combination ratio, G j+1 is the page keyword set of the j+1th page;

[0025] S244, according to the first intersection-over-union ratio and the second intersection-over-union ratio, obtain the relevance of each user's browsing content: , where γ j Browsing content relevance for the jth user.

[0026] Furthermore, S3 includes the following sub-steps:

[0027] S31, extracting each transaction content of the user on the transaction platform, extracting keywords in the transaction content, and constructing a transaction keyword set;

[0028] S32, calculating the intersection-and-union ratio of the transaction keyword set and each page keyword set to obtain a third intersection-and-union ratio;

[0029] S33, screen out a page keyword set whose third intersection-over-union ratio is greater than the intersection-over-union ratio threshold as a matching page keyword set, and use the corresponding page as a matching page;

[0030] S34, calculating the transaction compatibility according to the third intersection-over-union ratio corresponding to the matching page keyword set and the time spent on the matching page;

[0031] S35. Use the transaction compatibility corresponding to each transaction content as an element to construct a user transaction compatibility sequence.

[0032] Furthermore, the formula for calculating transaction compatibility in S34 is: , where F is the transaction fit, S 3,n is the third intersection-over-union ratio corresponding to the nth matching page keyword set, t n is the time the matching page corresponding to the nth matching page keyword set stays, n is the number of the matching page keyword set, and N is the number of matching page keyword sets.

[0033] Further, the multi-input neural network in S4 includes: a deviation feature extraction channel, a browsing feature extraction channel, a transaction feature extraction channel, a first fully connected layer, a second fully connected layer, and a classification layer;

[0034] The input end of the deviation feature extraction channel is used to input the expenditure deviation sequence; the input end of the browsing feature extraction channel is used to input the user browsing association degree sequence; the input end of the transaction feature extraction channel is used to input the user transaction fit degree sequence;

[0035] The input end of the first fully connected layer is respectively connected to the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, and the first output end of the transaction feature extraction channel; the input end of the second fully connected layer is respectively connected to the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, and the second output end of the transaction feature extraction channel; the input end of the classification layer is respectively connected to the output end of the first fully connected layer and the output end of the second fully connected layer.

[0036] Further, the deviation feature extraction channel, the browsing feature extraction channel and the transaction feature extraction channel all include: a first convolutional layer, a first scale feature extraction module and a second scale feature extraction module;

[0037] The input end of the first convolutional layer serves as the input end of the deviation feature extraction channel, the input end of the browsing feature extraction channel, or the input end of the transaction feature extraction channel, and its output end is respectively connected to the input end of the first scale feature extraction module and the input end of the second scale feature extraction module; the output end of the first scale feature extraction module serves as the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, or the first output end of the transaction feature extraction channel; the output end of the second scale feature extraction module serves as the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, or the second output end of the transaction feature extraction channel.

[0038] Furthermore, the first scale feature extraction module and the second scale feature extraction module both include: a second convolutional layer, a BN layer and a maximum pooling layer connected in sequence, the first convolutional layer and the second convolutional layer are both one-dimensional convolutional layers, the convolution kernel of the first convolutional layer is set to 1, the convolution kernel of the second convolutional layer in the first scale feature extraction module is set to 3, and the convolution kernel of the second convolutional layer in the second scale feature extraction module is set to 5.

[0039] The beneficial effects of the present invention are:

[0040] 1. The present invention can dynamically monitor the difference between the user's transaction amount and his daily expenditure by constructing a spending deviation sequence, so as to timely discover atypical transaction behaviors. This method makes the identification process more sensitive and no longer relies on static threshold judgment, but can adapt to changes in the user's consumption habits.

[0041] 2. The present invention extracts the content of each page in the user's browsing history and calculates the browsing relevance, constructs a user browsing relevance sequence, and quantifies the user's behavior through the user browsing relevance sequence, which can more accurately quantify the relevance between pages, which means that the user's behavioral connection between different pages is effectively captured. This process ensures that the user's non-transactional behavior can be integrated into the risk assessment, thereby improving the ability to identify potential risky transactions.

[0042] 3. The present invention further enhances the accuracy of abnormal identification by analyzing the degree of fit between each transaction content and the user's browsing content and constructing a transaction fit sequence. This personalized analysis based on user behavior can identify transactions that are inconsistent with the user's regular behavior and reduce the false alarm rate.

[0043] 4. The present invention uses a multi-input neural network to integrate and analyze the above multi-dimensional data, thereby realizing a comprehensive analysis of transaction anomalies from multiple dimensions and improving the accuracy of identifying abnormal financial payment transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flowchart of a method for identifying abnormal financial payment transactions based on neural network;

[0045] Figure 2 It is a schematic diagram of the structure of a multi-input neural network;

[0046] Figure 3 This is a schematic diagram of the structures of the deviation feature extraction channel, the browsing feature extraction channel, and the transaction feature extraction channel. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, a method for identifying abnormal financial payment transactions based on a neural network includes the following steps:

[0049] S1. Construct an expenditure deviation sequence based on the difference between the transaction amount and daily expenditure of the user on the trading platform;

[0050] S2, extracting the content of each page in the user's browsing history on the trading platform, calculating the relevance of the user's browsing content, and constructing a user browsing relevance sequence;

[0051] S3, extracting each transaction content of the user on the transaction platform, calculating the transaction fit between each transaction content and the content browsed by the user, and constructing a user transaction fit sequence;

[0052] S4. According to the expenditure deviation sequence, the user browsing association sequence and the user transaction compatibility sequence, a payment abnormality degree value is obtained based on a multi-input neural network.

[0053] In this embodiment, S1 includes the following sub-steps:

[0054] S11. Count the daily expenditure of users and obtain the daily average expenditure and the daily maximum expenditure;

[0055] S12. Record the user's multiple transaction amounts on the trading platform;

[0056] S13, calculating the expenditure deviation value according to the difference between each transaction amount and the daily average expenditure and the daily maximum expenditure;

[0057] S14. Use each expenditure deviation value as an element to construct an expenditure deviation sequence.

[0058] In this embodiment, the formula for calculating the expenditure deviation value in S13 is: , where θ i is the ith expenditure deviation value, P i is the i-th transaction amount, P day,avg is the average daily expenditure, P day,max is the maximum daily expenditure, and i is a positive integer.

[0059] The present invention counts daily expenses, compares the amount of each transaction with the daily average expenditure and the daily maximum expenditure, and obtains a deviation value. The larger the deviation value is, the more abnormal the transaction amount is. The present invention also takes into account the daily average expenditure and the daily maximum expenditure at the same time, reducing the influence of accidental daily maximum expenditure on the deviation value.

[0060] In this embodiment, step S11 is specifically as follows: adding up the user's daily expenditures to obtain the daily expenditure amount, taking the average of all daily expenditure amounts within a period of time to obtain the daily average expenditure, taking the maximum value of all daily expenditure amounts within a period of time to obtain the daily maximum expenditure, and the length of the period of time can be selected as 1 month.

[0061] In this embodiment, S2 includes the following sub-steps:

[0062] S21, extracting the browsing history of the user on the trading platform;

[0063] S22, separating each page content in the browsing record to obtain multiple page contents browsed by the user;

[0064] S23, arranging the contents of each page in chronological order to obtain a user browsing set;

[0065] S24, in the user browsing set, calculating the relevance of each user browsing content according to the content of each page and the content of adjacent pages;

[0066] S25. Use the relevance of each user's browsing content as an element to construct a user browsing relevance sequence.

[0067] The present invention separates user browsed pages and arranges them in chronological order to obtain a user browsed set, which is convenient for calculating the relevance of adjacent page contents, judging the relevance between user behaviors, and reflecting the user's screening of content.

[0068] In this embodiment, S24 includes the following sub-steps:

[0069] S241, extracting keywords from the content of each page and constructing a page keyword set;

[0070] S242, calculate the intersection-and-union ratio of the page keyword set of each page and the page keyword set of the left neighboring page to obtain a first intersection-and-union ratio: , where S 1,j is the jth first intersection-combination ratio, G j is the set of page keywords for the jth page, G j-1 is the page keyword set of the j-1th page, ∩ is the intersection, ∪ is the union, j is the number of the page in the user's browsing set, and f is the number of keywords in the calculation set;

[0071] S243, calculate the intersection-and-union ratio of each page keyword set and the keyword set of the neighboring page on the right, and obtain a second intersection-and-union ratio: , where S 2,j is the jth second intersection-combination ratio, G j+1 is the page keyword set of the j+1th page;

[0072] S244, according to the first intersection-over-union ratio and the second intersection-over-union ratio, obtain the relevance of each user's browsing content: , where γ j Browsing content relevance for the jth user.

[0073] When there is only one neighborhood page, there is only one intersection-and-union ratio, and the relevance of the user's browsing content is obtained by subtracting the intersection-and-union ratio from 1.

[0074] In the present invention, keywords are words excluding connecting words.

[0075] f(G j-1 ∩G j ) is the number of keywords in the intersection, f(G j-1 ∪G j ) is the number of keywords in the union. j+1 ∩G j ) is the number of keywords in the intersection, f(G j+1 ∪G j ) is the number of keywords in the union.

[0076] The present invention extracts keywords in a page and calculates the intersection-and-union ratio with adjacent pages to evaluate the similarity between different pages and reflect the user's behavior. The greater the relevance of the content browsed by the user, the lower the similarity between the page and the adjacent pages, and the higher the abnormality of the user's behavior.

[0077] Take the specific process of step S2 as an example: the user browses the following three pages on the trading platform, and the browsing records are as follows:

[0078] Page 1: "New winter white goose down workwear style fur collar parka down jacket", time: 2024-11-10 10:00; Page 2: "Light and windproof jacket", time: 2024-11-10 10:05, Page 3: "Black long down coat", time: 2024-11-10 10:10; sorted by time, the user browsing set is organized into: User browsing set: [page 1, page 2, page 3], and keywords are extracted for each page: Page 1: Keyword set G1={winter, new style, white goose down, workwear, fur collar, parka, down jacket}; Page 2: Keyword set G2={light, windproof, jacket}, Page 3: Keyword set G3={black, long, down, coat}.

[0079] In this embodiment, S3 includes the following sub-steps:

[0080] S31, extracting each transaction content of the user on the transaction platform, extracting keywords in the transaction content, and constructing a transaction keyword set;

[0081] S32, calculating the intersection-and-union ratio of the transaction keyword set and each page keyword set to obtain a third intersection-and-union ratio;

[0082] The formula for calculating the intersection-and-union ratio in S32 is: , where S 3,p is the intersection-and-union ratio of the transaction keyword set and the jth page keyword set, P is the transaction keyword set, G j is the set of keywords for the jth page;

[0083] S33, screen out a page keyword set whose third intersection-over-union ratio is greater than the intersection-over-union ratio threshold as a matching page keyword set, and use the corresponding page as a matching page;

[0084] S34, calculating the transaction compatibility according to the third intersection-over-union ratio corresponding to the matching page keyword set and the time spent on the matching page;

[0085] S35. Use the transaction compatibility corresponding to each transaction content as an element to construct a user transaction compatibility sequence.

[0086] In this embodiment, the intersection-over-union ratio threshold is a threshold set by the third intersection-over-union ratio.

[0087] The formula for calculating transaction compatibility in S34 is: , where F is the transaction fit, S 3,n is the third intersection-over-union ratio corresponding to the nth matching page keyword set, t n is the time the matching page corresponding to the nth matching page keyword set stays, n is the number of the matching page keyword set, and N is the number of matching page keyword sets.

[0088] The present invention extracts keywords from the transaction content, constructs a transaction keyword set, and then compares the transaction keyword set with each page keyword set in the browsing record, calculates the intersection-and-union ratio, and extracts the page keyword set of the browsing page related to the transaction content. The transaction fit is calculated based on the third intersection-and-union ratio corresponding to the matching page keyword set screened out for each transaction content, and the corresponding page residence time. The smaller the transaction fit, the larger the third intersection-and-union ratio corresponding to the matching page keyword set, and the longer the residence time, indicating that the user is carefully screening the related content rather than accidental events, and the smaller the probability of payment abnormality.

[0089] like Figure 2 As shown, the multi-input neural network in S4 includes: a deviation feature extraction channel, a browsing feature extraction channel, a transaction feature extraction channel, a first fully connected layer, a second fully connected layer and a classification layer;

[0090] The input end of the deviation feature extraction channel is used to input the expenditure deviation sequence; the input end of the browsing feature extraction channel is used to input the user browsing association degree sequence; the input end of the transaction feature extraction channel is used to input the user transaction fit degree sequence;

[0091] The input end of the first fully connected layer is respectively connected to the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, and the first output end of the transaction feature extraction channel; the input end of the second fully connected layer is respectively connected to the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, and the second output end of the transaction feature extraction channel; the input end of the classification layer is respectively connected to the output end of the first fully connected layer and the output end of the second fully connected layer.

[0092] The present invention can extract user behavior features from different dimensions through deviation feature extraction channels, browsing feature extraction channels and transaction feature extraction channels. This multi-channel design enables the model to fully understand the user's behavior pattern, thereby improving the accuracy of prediction. Each channel has two output ends, which output the first feature sequence and the second feature sequence respectively. The first fully connected layer integrates the three first feature sequences to obtain the first abnormal evaluation value, and the second fully connected layer integrates the three second feature sequences to obtain the second abnormal evaluation value. The classification layer is used to obtain the payment abnormality value based on the first abnormal evaluation value and the second abnormal evaluation value.

[0093] In this embodiment, the classification layer may adopt a fully connected layer, and the first abnormality assessment value and the second abnormality assessment value may be directly weighted to obtain a payment abnormality degree value.

[0094] like Figure 3 As shown, the deviation feature extraction channel, the browsing feature extraction channel and the transaction feature extraction channel all include: a first convolution layer, a first scale feature extraction module and a second scale feature extraction module;

[0095] The input end of the first convolutional layer serves as the input end of the deviation feature extraction channel, the input end of the browsing feature extraction channel, or the input end of the transaction feature extraction channel, and its output end is respectively connected to the input end of the first scale feature extraction module and the input end of the second scale feature extraction module; the output end of the first scale feature extraction module serves as the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, or the first output end of the transaction feature extraction channel; the output end of the second scale feature extraction module serves as the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, or the second output end of the transaction feature extraction channel.

[0096] Both the first scale feature extraction module and the second scale feature extraction module include: a second convolution layer, a BN layer and a maximum pooling layer connected in sequence, the first convolution layer and the second convolution layer are both one-dimensional convolution layers, the convolution kernel of the first convolution layer is set to 1, the convolution kernel of the second convolution layer in the first scale feature extraction module is set to 3, and the convolution kernel of the second convolution layer in the second scale feature extraction module is set to 5.

[0097] The present invention sets convolution kernels of different sizes in the second convolution layer in the first scale feature extraction module and the second scale feature extraction module to extract features of different scales, thereby facilitating the maximum pooling layer to extract significant features from different scales.

[0098] In this embodiment, the size of the pooling window of the one-dimensional maximum pooling layer is set to 2 or 3.

[0099] The multi-input neural network is trained using the existing gradient descent method.

[0100] In the present invention, the greater the payment anomaly degree value is, the greater the probability that there is transaction anomaly in these transaction behaviors.

[0101] By constructing a spending deviation sequence, the present invention can dynamically monitor the difference between the user's transaction amount and his daily spending, thereby timely discovering atypical transaction behaviors. This method makes the identification process more sensitive and no longer relies on static threshold judgments, but can adapt to changes in the user's consumption habits.

[0102] The present invention extracts the content of each page in the user's browsing history and calculates the browsing relevance, constructs a user browsing relevance sequence, and quantifies the user's behavior through the user browsing relevance sequence, which can more accurately quantify the relevance between pages, which means that the user's behavioral connection between different pages is effectively captured. This process ensures that the user's non-transactional behavior can be integrated into the risk assessment, thereby improving the ability to identify potential risky transactions.

[0103] The present invention further enhances the accuracy of abnormal identification by analyzing the degree of fit between each transaction content and the user's browsing content and constructing a transaction fit sequence. This personalized analysis based on user behavior can identify transactions that are inconsistent with the user's regular behavior and reduce the false alarm rate.

[0104] The present invention uses a multi-input neural network to integrate and analyze the above multi-dimensional data, thereby realizing a comprehensive analysis of transaction anomalies from multiple dimensions and improving the accuracy of identifying abnormal financial payment transactions.

[0105] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying abnormal financial payment transactions based on neural networks, characterized in that: The following steps are involved: S1. Construct an expenditure deviation sequence based on the difference between the transaction amount and daily expenditure of the user on the trading platform; S2, extracting the content of each page in the user's browsing history on the trading platform, calculating the relevance of the user's browsing content, and constructing a user browsing relevance sequence; S3, extracting each transaction content of the user on the transaction platform, calculating the transaction fit between each transaction content and the content browsed by the user, and constructing a user transaction fit sequence; S4, obtaining a payment abnormality value based on a multi-input neural network according to the expenditure deviation sequence, the user browsing association sequence and the user transaction compatibility sequence; The S2 comprises the following sub-steps: S21, extracting the browsing history of the user on the trading platform; S22, separating each page content in the browsing record to obtain multiple page contents browsed by the user; S23, arranging the contents of each page in chronological order to obtain a user browsing set; S24, in the user browsing set, calculating the relevance of each user browsing content according to the content of each page and the content of adjacent pages; S25, using the relevance of each user's browsing content as an element to construct a user browsing relevance sequence; The S24 comprises the following sub-steps: S241, extracting keywords from the content of each page and constructing a page keyword set; S242, calculate the intersection-and-union ratio of the page keyword set of each page and the page keyword set of the left neighboring page to obtain a first intersection-and-union ratio: , where S 1,j is the jth first intersection-combination ratio, G j is the set of page keywords for the jth page, G j-1 is the page keyword set of the j-1th page, ∩ is the intersection, ∪ is the union, j is the number of the page in the user's browsing set, and f is the number of keywords in the calculation set; S243, calculate the intersection-and-union ratio of each page keyword set and the keyword set of the neighboring page on the right, and obtain a second intersection-and-union ratio: , where S 2,j is the jth second intersection-combination ratio, G j+1 is the page keyword set of the j+1th page; S244, according to the first intersection-over-union ratio and the second intersection-over-union ratio, obtain the relevance of each user's browsing content: , where γ j Browsing content relevance for the jth user; The S3 comprises the following sub-steps: S31, extracting each transaction content of the user on the transaction platform, extracting keywords in the transaction content, and constructing a transaction keyword set; S32, calculating the intersection-and-union ratio of the transaction keyword set and each page keyword set to obtain a third intersection-and-union ratio; S33, screen out a page keyword set whose third intersection-over-union ratio is greater than the intersection-over-union ratio threshold as a matching page keyword set, and use the corresponding page as a matching page; S34, calculating the transaction compatibility according to the third intersection-over-union ratio corresponding to the matching page keyword set and the time spent on the matching page; S35, taking the transaction compatibility corresponding to each transaction content as an element, and constructing a user transaction compatibility sequence; The multi-input neural network in S4 includes: a deviation feature extraction channel, a browsing feature extraction channel, a transaction feature extraction channel, a first fully connected layer, a second fully connected layer and a classification layer; The input end of the deviation feature extraction channel is used to input the expenditure deviation sequence; the input end of the browsing feature extraction channel is used to input the user browsing association degree sequence; the input end of the transaction feature extraction channel is used to input the user transaction fit degree sequence; The input end of the first fully connected layer is respectively connected to the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, and the first output end of the transaction feature extraction channel; the input end of the second fully connected layer is respectively connected to the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, and the second output end of the transaction feature extraction channel; the input end of the classification layer is respectively connected to the output end of the first fully connected layer and the output end of the second fully connected layer.

2. The method for identifying abnormal financial payment transactions based on neural networks according to claim 1 is characterized in that: The S1 comprises the following sub-steps: S11. Count the daily expenditure of users and obtain the daily average expenditure and the daily maximum expenditure; S12. Record the user's multiple transaction amounts on the trading platform; S13, calculating the expenditure deviation value according to the difference between each transaction amount and the daily average expenditure and the daily maximum expenditure; S14. Use each expenditure deviation value as an element to construct an expenditure deviation sequence.

3. The method for identifying abnormal financial payment transactions based on neural networks according to claim 2 is characterized in that: The formula for calculating the expenditure deviation value in S13 is: , where θ i is the ith expenditure deviation value, P i is the i-th transaction amount, P day,avg is the average daily expenditure, P day,max is the maximum daily expenditure, and i is a positive integer.

4. The method for identifying abnormal financial payment transactions based on neural networks according to claim 1 is characterized in that: The formula for calculating the transaction compatibility in S34 is: , where F is the transaction fit, S 3,n is the third intersection-over-union ratio corresponding to the nth matching page keyword set, t n is the time the matching page corresponding to the nth matching page keyword set stays, n is the number of the matching page keyword set, and N is the number of matching page keyword sets.

5. The method for identifying abnormal financial payment transactions based on neural networks according to claim 1 is characterized in that: The deviation feature extraction channel, the browsing feature extraction channel and the transaction feature extraction channel all include: a first convolution layer, a first scale feature extraction module and a second scale feature extraction module; The input end of the first convolutional layer serves as the input end of the deviation feature extraction channel, the input end of the browsing feature extraction channel, or the input end of the transaction feature extraction channel, and the output end of the first convolutional layer is respectively connected to the input end of the first scale feature extraction module and the input end of the second scale feature extraction module; the output end of the first scale feature extraction module serves as the first output end of the deviation feature extraction channel, the first output end of the browsing feature extraction channel, or the first output end of the transaction feature extraction channel; the output end of the second scale feature extraction module serves as the second output end of the deviation feature extraction channel, the second output end of the browsing feature extraction channel, or the second output end of the transaction feature extraction channel.

6. The method for identifying abnormal financial payment transactions based on neural networks according to claim 5 is characterized in that: The first scale feature extraction module and the second scale feature extraction module both include: a second convolution layer, a BN layer and a maximum pooling layer connected in sequence, the first convolution layer and the second convolution layer are both one-dimensional convolution layers, the convolution kernel of the first convolution layer is set to 1, the convolution kernel of the second convolution layer in the first scale feature extraction module is set to 3, and the convolution kernel of the second convolution layer in the second scale feature extraction module is set to 5.

Images