Intelligent decision analysis platform and method

Abstract

The invention provides an intelligent decision analysis platform and method, and relates to the field of intelligent data analysis, and the method comprises the steps: firstly obtaining a time queue of transaction data, storing the time queue, then carrying out the transaction mode recognition of the time queue of the transaction data, and if the transaction data of an abnormal transaction mode is recognized, determining that the transaction mode is abnormal; and if yes, generating a transaction abnormity early warning prompt, and generating a risk report. Therefore, transaction data can be managed and analyzed more efficiently and intelligently, abnormal transactions can be recognized in time, and corresponding decisions can be made.

Description

Technical Field

[0001] The present application relates to the field of intelligent data analysis, and more specifically, to an intelligent decision analysis platform and method. Background Art

[0002] In today's highly information-based society, the market and competitive environment are changing rapidly, placing higher demands on the decision-making of enterprises and departments at all levels. In order to cope with this rapidly changing environment, enterprises need to continuously improve their modern governance and operational capabilities to meet various challenges. Especially in the financial and economic fields, financial transaction data is highly complex and diverse. These data usually exist in the form of time series, containing a large amount of transaction information, and the relationship between these information is intricate. Therefore, how to efficiently manage and analyze these data has become the key to corporate decision-making.

[0003] However, traditional statistical methods and rule engines have some inherent limitations in financial decision-making. Specifically, traditional decision-making relies on pre-set rules, which are usually based on historical data and expert experience. But these rules are often static and cannot cover all possible abnormal transaction patterns, especially those new and complex frauds. For example, traditional rules may have difficulty identifying multi-stage, small-amount money laundering activities or highly disguised frauds. In addition, risk patterns in the financial field are constantly evolving, and new fraud methods and risk patterns continue to emerge. It is difficult for traditional rule engines to identify and respond to these emerging risks in a timely manner, because the update and adjustment of rules requires time and verification, which results in the platform being unable to effectively identify and prevent these new risks for a period of time before the new rules take effect.

[0004] Therefore, an intelligent decision analysis solution is desired. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiments of this application provide an intelligent decision analysis platform and method.

[0006] According to one aspect of the present application, an intelligent decision analysis platform is provided, which includes: a data acquisition module for acquiring a time queue of transaction data; a data storage module for storing the time queue of the transaction data; a transaction pattern recognition module for performing transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result, wherein the transaction pattern recognition module includes: a transaction data screening unit for screening the time queue of the transaction data to obtain a time queue of small transaction data; a small transaction data local encoding unit for performing sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; a small transaction full time domain aggregation unit for performing full time domain kernel feature dynamic aggregation on the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; a transaction pattern recognition result generation unit for obtaining the transaction pattern recognition result based on the full time domain kernel representation of small transactions; an early warning module for generating a transaction abnormality early warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern; and a risk reporting module for generating a risk report.

[0007] In the above-mentioned intelligent decision-making analysis platform, the transaction data screening unit is used to: screen the time queue of the transaction data based on a predetermined transaction amount to obtain the time queue of the small transaction data.

[0008] In the above-mentioned intelligent decision-making analysis platform, the local encoding unit for small transaction data includes: a small transaction time series data sampling subunit, which is used to perform sliding window sampling on the time queue of the small transaction data based on a preset time window to obtain a sequence of small transaction data sampling windows; a small transaction local time domain pattern feature extraction subunit, which is used to input each small transaction data sampling window in the sequence of small transaction data sampling windows into a small transaction sequence encoder based on a forward LSTM model to obtain a sequence of small transaction local time domain pattern feature vectors as the sequence of small transaction local time domain pattern features.

[0009] In the above-mentioned intelligent decision-making analysis platform, the small transaction full-time domain aggregation unit includes: a small transaction initial aggregation subunit, which is used to calculate the positional mean vector of the sequence of the small transaction local time domain pattern feature vectors to obtain the small transaction initial aggregation representation vector; a small transaction static energy potential factor calculation subunit, which is used to calculate the static energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vectors to obtain a sequence of small transaction static energy potential factors; a small transaction dynamic energy potential factor calculation subunit, which is used to calculate the static energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vector relative to the small transaction initial aggregation representation vector dynamic energy potential factors to obtain a sequence of dynamic energy potential factors of small transactions; a small transaction superposition energy potential factor calculation subunit, used to determine the superposition state energy potential factors of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors based on the static energy potential factors and dynamic energy potential factors of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors to obtain a sequence of superposition state energy potential factors of small transactions; a small transaction full time domain feature aggregation subunit, used to perform full time domain feature aggregation on the sequence of local time domain pattern feature vectors of small transactions based on the sequence of superposition state energy potential factors of small transactions to obtain the full time domain kernel representation of the small transactions.

[0010] In the above-mentioned intelligent decision-making analysis platform, the small-amount transaction static energy potential factor calculation subunit is used to: calculate the mean and standard deviation of the local time domain pattern feature vector of the small-amount transaction to obtain the local time domain mean and the local time domain standard deviation of the small-amount transaction; calculate the positional subtraction of the local time domain pattern feature vector of the small-amount transaction and the local time domain mean of the small-amount transaction, divide the small-amount transaction local time domain difference vector after the subtraction by the local time domain of the small-amount transaction, and calculate the fourth power of the eigenvalue of each position in the vector after the division to obtain the local time domain static energy potential vector of the small-amount transaction; calculate the positional mean of the local time domain static energy potential vector of the small-amount transaction to obtain the small-amount transaction static energy potential factor.

[0011] In the above-mentioned intelligent decision-making analysis platform, the small transaction dynamic energy potential factor calculation subunit is used to: matrix multiply the small transaction aggregation weight matrix with the small transaction initial aggregation representation vector to obtain the small transaction initial aggregation weight representation vector; matrix multiply the small transaction weight matrix with the small transaction local time domain pattern feature vector to obtain the small transaction local time domain weight vector; after performing positional dot multiplication of the small transaction initial aggregation weight representation vector and the small transaction local time domain weight vector, the obtained small transaction fusion weight vector and the small transaction bias vector are added by position to obtain the small transaction local time domain dynamic energy potential vector; calculate the square of the Frobenius norm of the small transaction local time domain dynamic energy potential vector to obtain the small transaction dynamic energy potential factor.

[0012] In the above-mentioned intelligent decision-making analysis platform, the small-amount transaction superposition state energy potential factor calculation subunit is used to: calculate an exponential function with the natural constant e as the base and the static energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias static energy potential factor of the small-amount transaction; calculate an exponential function with the natural constant e as the base and the square of the dynamic energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias dynamic energy potential factor of the small-amount transaction; add the local time domain bias static energy potential factor of the small-amount transaction and the local time domain bias dynamic energy potential factor of the small-amount transaction to obtain the superposition state energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction.

[0013] In the above-mentioned intelligent decision-making analysis platform, the small-transaction full-time domain feature aggregation subunit is used to: input the sequence of small-transaction superposition state energy potential factors into the energy potential factor gated weighting network to obtain a sequence of small-transaction superposition state energy potential weight coefficients; based on the sequence of small-transaction superposition state energy potential weight coefficients, calculate the position-weighted sum of the sequence of small-transaction local time domain pattern feature vectors to obtain the small-transaction full-time domain kernel representation vector as the small-transaction full-time domain kernel representation.

[0014] In the above-mentioned intelligent decision analysis platform, the transaction pattern recognition result generation unit is used to: input the small transaction full-time domain kernel representation vector into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result.

[0015] According to another aspect of the present application, an intelligent decision analysis method is provided, which includes: obtaining a time queue of transaction data; storing the time queue of the transaction data; performing transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result, including: screening the time queue of the transaction data to obtain a time queue of small transaction data; performing sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; performing dynamic aggregation of full time domain kernel features on the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; obtaining the transaction pattern recognition result based on the full time domain kernel representation of small transactions; generating a transaction abnormality warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern; and generating a risk report.

[0016] Compared with the prior art, the intelligent decision analysis platform and method provided by the present application uses data analysis and encoding based on deep learning to screen the time queue of the transaction data, and then performs sliding window sampling and local time domain sequence encoding on the time queue of the screened small transaction data, so as to automatically obtain the transaction pattern recognition result based on the kernel aggregation representation of the local time domain pattern features of each small transaction in the full time domain. In this way, multi-stage, small-amount transaction patterns can be captured, which are often difficult to identify in traditional rule engines, such as small-amount, scattered money laundering activities. At the same time, it can automatically learn and identify new risk patterns without frequent manual updates of rules, which enables the platform to adapt to the continuous evolution of the financial field in a timely manner and improve its adaptive ability and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] In the attached picture: Figure 1 It is a block diagram of an intelligent decision analysis platform according to an embodiment of the present application.

[0019] Figure 2 It is a block diagram of a transaction pattern recognition module in an intelligent decision analysis platform according to an embodiment of the present application.

[0020] Figure 3 Schematic diagram of data flow of a transaction pattern recognition module in an intelligent decision-making analysis platform according to an embodiment of the present application.

[0021] Figure 4This is a block diagram of a local encoding unit for small transaction data in an intelligent decision-making analysis platform according to an embodiment of the present application.

[0022] Figure 5 It is a block diagram of a full-time domain aggregation unit for small transactions in an intelligent decision-making analysis platform according to an embodiment of the present application.

[0023] Figure 6 Flow chart of the intelligent decision-making analysis method according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0025] It should be noted that in this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization of the owner of the corresponding device.

[0026] In a highly information-based modern society, enterprises are facing rapid changes in the market and competitive environment, which requires them to improve their modern governance and operational capabilities to meet challenges. Especially in the field of finance and economics, due to the complexity and diversity of financial transaction data, enterprises need to efficiently manage and analyze these data to support decision-making. However, traditional statistical methods and rule engines have limitations in financial decision-making. They rely on static rules based on historical data and expert experience, and it is difficult to cover all abnormal transaction patterns, especially new types of complex fraud. In addition, due to the continuous evolution of financial risk patterns, traditional rule engines are difficult to identify and respond to emerging risks in a timely manner, because it takes time to update rules, resulting in a gap in risk identification and prevention before the new rules take effect.

[0027] Based on this, this application proposes an intelligent decision-making analysis platform. Figure 1 FIG. 1 is a block diagram of an intelligent decision analysis platform according to an embodiment of the present application. Figure 1 As shown, in the intelligent decision analysis platform 100, it includes: a data acquisition module 110, which is used to obtain the time queue of transaction data; a data storage module 120, which is used to store the time queue of the transaction data; a transaction pattern recognition module 130, which is used to perform transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result; an early warning module 140, which is used to generate a transaction abnormality early warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern; and a risk reporting module 150, which is used to generate a risk report.

[0028] In the embodiment of the present application, the data acquisition module 110 is used to obtain the time queue of transaction data. It should be understood that transaction data usually includes transaction time information, transaction amount information, identity information of the transaction initiator and the transaction recipient, transaction type information, transaction status information, etc. The combination of these specific information forms a complete transaction portrait. In order to enable the platform to capture the changing trend of transaction behavior and more accurately identify different transaction patterns and their changes, it is first necessary to obtain the time queue of transaction data in the technical solution of the present application. By comprehensively analyzing the time queue of transaction data, it is helpful to more efficiently and accurately determine whether the transaction is abnormal, thereby improving the reliability of financial decision-making.

[0029] In the embodiment of the present application, the data storage module 120 is used to store the time queue of the transaction data. It should be understood that transaction data is an important asset in the financial field and needs to be stored for a long time for subsequent analysis and audit. By storing the time queue of transaction data in a database, data persistence can be achieved to ensure that data will not be lost due to platform failure or accident, which provides strong protection for business operations in the financial field.

[0030] In the embodiment of the present application, the transaction pattern recognition module 130 is used to perform transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result.

[0031] Accordingly, in the transaction pattern recognition module, the technical concept of the present application is to use data analysis and encoding methods based on deep learning to screen the time queue of the transaction data, and then perform sliding window sampling and local time domain sequence encoding on the time queue of the screened small transaction data, so as to automatically obtain the transaction pattern recognition result based on the kernel aggregation representation of the local time domain pattern features of each small transaction in the full time domain. In this way, multi-stage, small-amount transaction patterns can be captured, which are often difficult to identify in traditional rule engines, such as small-amount, scattered money laundering activities. At the same time, it is possible to automatically learn and identify new risk patterns without frequent manual updates of rules, which enables the platform to adapt to the continuous evolution of the financial field in a timely manner and improve its adaptive capabilities and intelligence.

[0032] Specifically, Figure 2 It is a block diagram of a transaction pattern recognition module in an intelligent decision analysis platform according to an embodiment of the present application. Figure 3 Schematic diagram of data flow in the transaction pattern recognition module in the intelligent decision analysis platform according to the embodiment of the present application. Figure 2 and Figure 3As shown, the transaction pattern recognition module 130 includes: a transaction data screening unit 131, used for screening the time queue of the transaction data to obtain the time queue of small transaction data; a small transaction data local encoding unit 132, used for sliding window sampling and sequence encoding the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; a small transaction full time domain aggregation unit 133, used for dynamically aggregating full time domain kernel features of the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; a transaction pattern recognition result generation unit 134, used to obtain the transaction pattern recognition result based on the full time domain kernel representation of small transactions.

[0033] In the embodiment of the present application, the transaction data screening unit 131 is used to screen the time queue of the transaction data to obtain the time queue of small transaction data. Specifically, in the embodiment of the present application, the transaction data screening unit is used to: screen the time queue of the transaction data based on the predetermined transaction amount to obtain the time queue of the small transaction data. It should be understood that small transactions are often easily overlooked, but in fact, some illegal activities (such as money laundering, fraud, etc.) may be carried out through a large number of small transactions to cover up their true purpose. Specifically, in order to circumvent the review of large transactions, illegal elements often split funds into multiple small transactions. This method makes each transaction look normal and not easy to attract attention, but it may pose a serious risk as a whole. Based on this, in the technical solution of the present application, the time queue of the transaction data is screened based on the predetermined transaction amount to obtain the time queue of small transaction data. In this way, by focusing on small transactions, these hidden risk points can be more effectively identified and these abnormal transaction patterns can be discovered in time.

[0034] In the embodiment of the present application, the small transaction data local encoding unit 132 is used to perform sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions. Specifically, Figure 4 FIG. 1 is a block diagram of a local encoding unit for small transaction data in an intelligent decision analysis platform according to an embodiment of the present application. Figure 4 As shown, the small transaction data local encoding unit 132 includes: a small transaction time series data sampling subunit 1321, which is used to perform sliding window sampling on the time queue of the small transaction data based on a preset time window to obtain a sequence of small transaction data sampling windows; a small transaction local time domain pattern feature extraction subunit 1322, which is used to input each small transaction data sampling window in the sequence of small transaction data sampling windows into a small transaction sequence encoder based on a forward LSTM model to obtain a sequence of small transaction local time domain pattern feature vectors as the sequence of small transaction local time domain pattern features.

[0035] In an embodiment of the present application, the small transaction time series data sampling subunit 1321 is used to perform sliding window sampling on the time queue of the small transaction data based on a preset time window to obtain a sequence of small transaction data sampling windows. Accordingly, considering that the original small transaction data may be very large and contain a large number of transaction records. In order to be able to more carefully analyze and process the time series change pattern of the small transaction data in each time window, in the technical solution of the present application, the time queue of the small transaction data is sampled based on the preset time window to obtain a sequence of small transaction data sampling windows. That is, through sliding window sampling, continuous small transaction data can be divided into multiple fixed-length time windows, and the data in each window forms a "sampling window", which helps to capture the trend and pattern of transaction data changing over time, such as changes in transaction frequency and transaction amount.

[0036] In an embodiment of the present application, the small transaction local time domain pattern feature extraction subunit 1322 is used to input each small transaction data sampling window in the sequence of the small transaction data sampling window into a small transaction sequence encoder based on a forward LSTM model to obtain a sequence of small transaction local time domain pattern feature vectors as the sequence of the small transaction local time domain pattern features. It should be understood that each small transaction data sampling window in the sequence of the small transaction data sampling window has a local time domain association pattern between different time segments, and the forward LSTM model can effectively retain important information in the input data through a special gating structure, and can capture the long-term dependency relationship between different positions in the input data. Therefore, in the technical solution of the present application, each small transaction data sampling window in the sequence of the small transaction data sampling window is input into a small transaction sequence encoder based on a forward LSTM model to utilize the long-term dependency capture characteristics of the forward LSTM model to effectively extract the time dependency characteristics of the transaction data in each time window, and obtain a sequence of small transaction local time domain pattern feature vectors.

[0037] In the embodiment of the present application, the small transaction full-time domain aggregation unit 133 is used to dynamically aggregate the full-time domain kernel features of the sequence of the local time domain pattern features of the small transaction to obtain the full-time domain kernel representation of the small transaction. Accordingly, considering that each small transaction local time domain pattern feature vector only reflects the transaction pattern and behavior characteristics within a specific time window, and there are intricate connections in the time domain between the local time domain pattern features of each small transaction, and each feature contains important semantic information and core essential features from different sources, in order to integrate the information of multiple time windows and generate a full-time domain kernel representation that can reflect the global characteristics of the entire time series, so as to capture the changing trends and laws of the transaction pattern over a long time span, in the technical solution of the present application, the full-time domain kernel features of the sequence of the local time domain pattern features of the small transaction are dynamically aggregated to obtain the full-time domain kernel representation vector of the small transaction as the full-time domain kernel representation of the small transaction.

[0038] In detail, first, in order to find a global representation that integrates the local features in all time windows, that is, to determine a core representative point or reference vector, the positional mean vector of the sequence of the local time-domain pattern feature vectors of the small transactions is calculated to obtain the initial aggregate representation vector of the small transactions. In other words, the initial aggregate representation vector compresses multiple local feature vectors into a single vector, greatly reducing the dimension of the data, which helps to capture the overall characteristics of the entire time series. Next, the static potential factor of each local time-domain pattern feature vector of the small transaction is calculated to quantify the importance of each local time-domain pattern feature vector in the overall transaction pattern, focusing on those feature vectors that have a greater impact on the global pattern, and obtaining a sequence of static potential factors of small transactions. Then, in order to evaluate the semantic connection and interaction between each independent local time-domain vector and the initial aggregation point, the dynamic potential factor of each local time-domain pattern feature vector of the small transaction is calculated relative to the initial aggregate representation vector of the small transaction to obtain a sequence of dynamic potential factors of the small transaction, which helps to identify which transaction patterns in which time windows are significantly different from the overall pattern, which may be abnormal or noteworthy patterns. Subsequently, considering that the static potential factor reflects the importance of each local time-domain pattern feature vector in the overall transaction pattern, and the dynamic potential factor reflects the degree of change of each local time-domain pattern feature vector relative to the global pattern, in order to comprehensively evaluate the importance and degree of change of each local time-domain pattern feature vector in both global and local dimensions, the static potential factor and dynamic potential factor of each small transaction local time-domain pattern feature vector are superimposed to obtain a sequence of small transaction superposition state potential factors, so as to more accurately capture the complexity and diversity of the transaction pattern. Next, according to the importance and degree of change of the transaction pattern, different weights are given, and the sequence of small transaction superposition state potential factors is input into the potential factor gated weighting network to dynamically adjust the weight of each feature, so as to obtain a sequence of small transaction superposition state potential weight coefficients. Finally, based on the sequence of small transaction superposition state potential weight coefficients, the sequence of small transaction local time-domain pattern feature vectors is weighted by position and then calculated to obtain a small transaction full-time domain kernel representation vector that integrates the key local features in all time windows.

[0039] Specifically, Figure 5 : is a block diagram of a small transaction full-time domain aggregation unit in an intelligent decision analysis platform according to an embodiment of the present application. Figure 5As shown, the small transaction full time domain aggregation unit 133 includes: a small transaction initial aggregation subunit 1331, which is used to calculate the positional mean vector of the sequence of the small transaction local time domain pattern feature vectors to obtain the small transaction initial aggregation representation vector; a small transaction static energy potential factor calculation subunit 1332, which is used to calculate the static energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vectors to obtain a sequence of small transaction static energy potential factors; a small transaction dynamic energy potential factor calculation subunit 1333, which is used to calculate the static energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vector relative to the small transaction initial aggregation representation vector. Dynamic energy potential factor to obtain a sequence of dynamic energy potential factors of small transactions; a small transaction superposition energy potential factor calculation subunit 1334 is used to determine the superposition state energy potential factor of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors based on the static energy potential factor and dynamic energy potential factor of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors to obtain a sequence of superposition state energy potential factors of small transactions; a small transaction full time domain feature aggregation subunit 1335 is used to perform full time domain feature aggregation on the sequence of small transaction local time domain pattern feature vectors based on the sequence of superposition state energy potential factors of small transactions to obtain the full time domain kernel representation of the small transaction.

[0040] More specifically, in an embodiment of the present application, the small-amount transaction static energy potential factor calculation subunit is used to: calculate the mean and standard deviation of the small-amount transaction local time domain pattern feature vector to obtain the small-amount transaction local time domain mean and the small-amount transaction local time domain standard deviation; calculate the positional subtraction of the small-amount transaction local time domain pattern feature vector and the small-amount transaction local time domain mean, divide the small-amount transaction local time domain difference vector after the subtraction by the small-amount transaction local time domain, and calculate the fourth power of the eigenvalue of each position in the vector after the division to obtain the small-amount transaction local time domain static energy potential vector; calculate the positional mean of the small-amount transaction local time domain static energy potential vector to obtain the small-amount transaction static energy potential factor.

[0041] More specifically, in an embodiment of the present application, the small-amount transaction dynamic energy potential factor calculation subunit is used to: matrix multiply the small-amount transaction aggregation weight matrix with the small-amount transaction initial aggregation representation vector to obtain the small-amount transaction initial aggregation weight representation vector; matrix multiply the small-amount transaction weight matrix with the small-amount transaction local time domain pattern feature vector to obtain the small-amount transaction local time domain weight vector; perform positional dot multiplication of the small-amount transaction initial aggregation weight representation vector and the small-amount transaction local time domain weight vector, and then perform positional addition of the obtained small-amount transaction fusion weight vector and the small-amount transaction bias vector to obtain the small-amount transaction local time domain dynamic energy potential vector; calculate the square of the Frobenius norm of the small-amount transaction local time domain dynamic energy potential vector to obtain the small-amount transaction dynamic energy potential factor.

[0042] More specifically, in an embodiment of the present application, the small-amount transaction superposition state energy potential factor calculation subunit is used to: calculate an exponential function with the natural constant e as the base and the static energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias static energy potential factor of the small-amount transaction; calculate an exponential function with the natural constant e as the base and the square of the dynamic energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias dynamic energy potential factor of the small-amount transaction; add the local time domain bias static energy potential factor of the small-amount transaction and the local time domain bias dynamic energy potential factor of the small-amount transaction to obtain the superposition state energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction.

[0043] More specifically, in an embodiment of the present application, the small-transaction full-time-domain feature aggregation subunit is used to: input the sequence of small-transaction superposition state energy potential factors into the energy potential factor gated weighting network to obtain a sequence of small-transaction superposition state energy potential weight coefficients; based on the sequence of small-transaction superposition state energy potential weight coefficients, calculate the position-weighted sum of the sequence of small-transaction local time-domain pattern feature vectors to obtain the small-transaction full-time-domain kernel representation vector as the small-transaction full-time-domain kernel representation.

[0044] More specifically, in an embodiment of the present application, the sequence of small-transaction superposition state energy potential factors is input into an energy factor gated weighting network to obtain a sequence of small-transaction superposition state energy potential weight coefficients, including: normalizing the sequence of small-transaction superposition state energy potential factors to obtain a sequence of normalized small-transaction superposition state energy potential factors; inputting the sequence of normalized small-transaction superposition state energy potential factors into a mask function for gated mask screening to obtain a sequence of small-transaction superposition state energy potential weight coefficients, wherein, in response to each normalized small-transaction superposition state energy potential factor in the sequence of normalized small-transaction superposition state energy potential factors being greater than a preset threshold, the normalized small-transaction superposition state energy potential factor is set to the original value, and the rest are set to zero to obtain a sequence of small-transaction superposition state energy potential weight coefficients.

[0045] More specifically, in an embodiment of the present application, the sequence of small-amount transaction superposition state energy potential factors is normalized to obtain a sequence of normalized small-amount transaction superposition state energy potential factors, including: calculating an exponential function value with the natural constant e as the base and the negative number of each small-amount transaction superposition state energy potential factor in the sequence of small-amount transaction superposition state energy potential factors as the exponent to obtain a sequence of small-amount transaction superposition state energy potential index factors; adding the sequence of small-amount transaction superposition state energy potential index factors to the constant one by position, and then calculating the reciprocal of each small-amount transaction superposition state energy potential index in the obtained sequence of small-amount transaction superposition state energy potential index modulation factors to obtain the sequence of normalized small-amount transaction superposition state energy potential factors.

[0046] In the embodiment of the present application, the small transaction full-time domain aggregation unit is used to: perform full-time domain kernel feature dynamic aggregation processing on the sequence of the small transaction local time domain pattern feature vectors according to the following formula; wherein the formula is: ;in, is the sequence of local time-domain pattern feature vectors of the small transaction, , , and are the first, second, and third in the sequence of the local time domain pattern feature vectors of the small transaction. and Small transaction local time domain pattern feature vector, is the number of vectors in the sequence of the local time-domain pattern feature vectors of small transactions, is the initial aggregate representation vector of small transactions, yes The eigenvalues ​​at each position in , and They are The mean and standard deviation of yes The number of eigenvalues ​​in , yes The corresponding small transaction static energy potential factor, is the small transaction aggregation weight matrix, is the small transaction weight matrix, It is the point multiplication by position. is the small transaction bias vector, is the square of the Frobenius norm of the calculated vector, yes The corresponding small transaction dynamic potential factor, yes The corresponding small transaction superposition state energy potential factor, yes The corresponding normalized small transaction superposition state energy potential factor, For gated mask operation, is the preset threshold, yes The corresponding small transaction superposition state energy potential weight coefficient, It is the full-time domain kernel representation vector of the small transaction.

[0047] In an embodiment of the present application, the transaction pattern recognition result generation unit 134 is used to obtain the transaction pattern recognition result based on the full-time domain kernel representation of the small transaction. Specifically, in an embodiment of the present application, the transaction pattern recognition result generation unit is used to: input the full-time domain kernel representation vector of the small transaction into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result. That is, the full-time domain kernel representation of the small transaction obtained by full-time domain dynamic aggregation of the sequence of the local time domain pattern feature vectors of the small transaction is used for classification processing, so as to automatically obtain the transaction pattern recognition result. In this way, multi-stage, small-amount transaction patterns can be captured, which are often difficult to identify in traditional rule engines. For example, small-amount, decentralized money laundering activities. At the same time, it is possible to automatically learn and identify new risk patterns without frequent manual updates of rules, which enables the platform to adapt to the continuous evolution of the financial field in a timely manner and improve its adaptive ability and intelligence. In particular, in a specific embodiment of the present application, the small-amount transaction full-time domain kernel representation vector is input into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result, including: using the fully connected layer of the classifier to fully connect the small-amount transaction full-time domain kernel representation vector to obtain a small-amount transaction full-time domain kernel representation fully connected encoding feature vector; inputting the small-amount transaction full-time domain kernel representation fully connected encoding feature vector into the Softmax classification function of the classifier to obtain the probability value of the small-amount transaction full-time domain kernel representation vector belonging to each classification label, the classification label including one for indicating the presence of abnormal transaction patterns in the transaction data and one for indicating the absence of abnormal transaction patterns in the transaction data; the classification label corresponding to the largest of the probability values ​​is determined as the transaction pattern recognition result.

[0048] In particular, when each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors respectively represents the short-range-long-range bidirectional temporal correlation characteristics of the small transaction data in the local time domain, when performing kernel feature extraction based on feature dynamic analysis, the small transaction full time domain kernel representation vector will also have a temporal correlation feature force field dynamic difference based on the local time domain source temporal distribution difference, thereby causing kernel feature distribution diversity based on force field dynamics. As a result, when the small transaction full time domain kernel representation vector is subjected to classification probability regression through a classifier-based transaction pattern classifier, the accuracy of the classification result will be affected.

[0049] Preferably, inputting the small transaction full-time domain kernel representation vector into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result includes: determining the feature mean and feature variance of the small transaction full-time domain kernel representation vector, and dividing the feature mean by the feature variance to obtain the small transaction full-time domain kernel probability statistics value: wherein the process can be expressed as: ;in, is the characteristic mean of the kernel representation vector of the small transaction in the entire time domain, is the characteristic variance of the kernel representation vector of the small transaction in the entire time domain, It is the kernel probability statistic value of the small transaction in the entire time domain.

[0050] The small transaction full time domain kernel representation vector is point-multiplied with the inverse of the maximum eigenvalue in the small transaction full time domain kernel representation vector to obtain the small transaction full time domain kernel probability constraint vector: wherein the process can be expressed as: ;in, is the small transaction full-time domain kernel representation vector, It means point multiplication by position. represents the inverse of the maximum eigenvalue in the small transaction full-time domain kernel representation vector, is the full-time domain kernel probability constraint vector of the small transaction; After adding the small transaction full-time domain kernel probability constraint vector and the small transaction full-time domain kernel probability statistic value, the logarithmic value with base 2 is calculated to obtain the small transaction full-time domain kernel information interaction vector: wherein, the process can be expressed as: ;in, is the kernel probability constraint vector of the small transaction in the entire time domain, Indicates adding by position point. represents the kernel probability statistics value of the small transaction in the whole time domain, represents the logarithmic function with base 2, It is the full-time domain kernel information interaction vector of the small transaction.

[0051] After performing point subtraction on the small transaction full-time domain kernel probability constraint vector with the small transaction full-time domain kernel probability statistics value, the inverse of each eigenvalue is calculated to obtain the small transaction full-time domain kernel sequence constraint vector: wherein, the process can be expressed as: ;in, is the kernel probability statistic value of the small transaction in the entire time domain, It means to subtract by position point. is the kernel probability constraint vector of the small transaction in the entire time domain, is the full-time domain kernel sequence constraint vector of the small transaction.

[0052] The small transaction full-time domain kernel information interaction vector is interpolated with the small transaction full-time domain kernel sequence constraint vector to obtain an optimized small transaction full-time domain kernel representation vector; the optimized small transaction full-time domain kernel representation vector is input into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result.

[0053] That is, considering that the weight matrix during classification probability regression acts on the small-amount transaction full-time domain kernel representation vector, the distribution diversity of the small-amount transaction full-time domain kernel representation vector causes the uncertainty of the weight stimulus parameters of the weight matrix, resulting in the lack of posterior probability density inference of the small-amount transaction full-time domain kernel representation vector through the weight matrix, which will affect the accuracy of the probability regression results.

[0054] Therefore, the mesoscale interaction structure under probability constraints is simulated between the eigenvalue scale and the eigenvector scale of the small-amount transaction full-time domain kernel representation vector through the probability statistical characteristics of the small-amount transaction full-time domain kernel representation vector, so as to construct a bidirectional latent variable motif based on the mesoscale short sequence relative to the probability statistical value to perform mesoscale bidirectional migration, and perform a posteriori recovery based on the sequence constraints on the interactive information, thereby improving the convergence effect in the probability density domain and improving the accuracy of the transaction pattern recognition result obtained by the transaction pattern classifier based on the classifier input into the small-amount transaction full-time domain kernel representation vector.

[0055] Therefore, the transaction pattern recognition module 130 is explained, which uses data analysis and encoding based on deep learning to screen the time queue of the transaction data, and then performs sliding window sampling and local time domain sequence encoding on the time queue of the screened small transaction data, so as to automatically obtain the transaction pattern recognition result based on the kernel aggregation representation of the local time domain pattern features of each small transaction in the full time domain. In this way, multi-stage, small-amount transaction patterns can be captured, which are often difficult to identify in traditional rule engines, such as small-amount, scattered money laundering activities. At the same time, it can automatically learn and identify new risk patterns without frequent manual updates of rules, which enables the platform to adapt to the continuous evolution of the financial field in a timely manner and improve its adaptive ability and intelligence.

[0056] In the embodiment of the present application, the early warning module 140 is used to generate an abnormal transaction early warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern. It should be understood that by real-time monitoring of transaction data, once an abnormal transaction pattern is found, an early warning prompt is immediately generated. This helps companies to promptly discover potential risks of fraud, money laundering or other illegal activities, so as to quickly take measures to intervene, such as freezing suspicious accounts, further investigating transaction details, etc. Through an effective early warning mechanism, companies can better protect their own and their customers' interests in a complex financial environment, and further enhance their risk management capabilities and customer trust.

[0057] In an embodiment of the present application, the risk reporting module 150 is used to generate a risk report. It should be understood that the risk report is intended to record in detail the specific information of the abnormal transaction pattern, including but not limited to key data such as transaction time, transaction parties, transaction type, transaction amount, etc. The risk report is not only a simple record of abnormal transaction patterns, but also provides an important reference for decision makers. Through the report, decision makers can more accurately judge the degree of risk and formulate effective response strategies and measures accordingly. In this way, through effective risk reporting, enterprises can significantly improve risk management efficiency, reduce decision-making errors, enhance compliance, and continuously improve their risk management strategies, thereby maintaining competitiveness and stability in a complex financial environment.

[0058] In summary, the intelligent decision analysis platform 100 based on the embodiment of the present application is explained, which first obtains the time queue of transaction data and stores it, and then performs transaction pattern recognition on the time queue of the transaction data. If transaction data with abnormal transaction patterns are recognized, a transaction abnormality warning prompt is generated, and a risk report is generated. In this way, more efficient and intelligent management and analysis of transaction data is achieved, so as to timely identify abnormal transactions and make corresponding decisions.

[0059] Figure 6 Flowchart of the intelligent decision analysis method according to the embodiment of the present application. Figure 6 As shown, in the intelligent decision-making analysis method, it includes: S110, obtaining the time queue of transaction data; S120, storing the time queue of the transaction data; S130, performing transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result; S140, in response to the transaction pattern recognition result being an abnormal transaction pattern, generating a transaction abnormality warning prompt; S150, generating a risk report; wherein, step S130 includes: screening the time queue of the transaction data to obtain a time queue of small transaction data; performing sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; performing full time domain kernel feature dynamic aggregation on the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; and obtaining the transaction pattern recognition result based on the full time domain kernel representation of small transactions.

[0060] Here, those skilled in the art can understand that the specific operations of each step in the above intelligent decision analysis method have been referred to above. Figures 1 to 5 The intelligent decision analysis platform has been introduced in detail, and therefore, its repeated description will be omitted.

[0061] In summary, the intelligent decision-making analysis method based on the embodiment of the present application is explained, which first obtains the time queue of transaction data and stores it, and then identifies the transaction pattern of the time queue of the transaction data. If transaction data with abnormal transaction patterns are identified, an abnormal transaction warning prompt is generated, and a risk report is generated. In this way, more efficient and intelligent management and analysis of transaction data is achieved, so as to timely identify abnormal transactions and make corresponding decisions.

Claims

1. An intelligent decision analysis platform, characterized in that: include: Data collection module, used to obtain the time queue of transaction data; A data storage module, used for storing the time queue of the transaction data; A transaction pattern recognition module is used to perform transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result, wherein the transaction pattern recognition module includes: a transaction data screening unit, used to screen the time queue of the transaction data to obtain a time queue of small transaction data; a small transaction data local encoding unit, used to perform sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; a small transaction full time domain aggregation unit, used to perform full time domain kernel feature dynamic aggregation on the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; a transaction pattern recognition result generation unit, used to obtain the transaction pattern recognition result based on the full time domain kernel representation of small transactions; an early warning module, used to generate a transaction abnormality early warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern; a risk reporting module, used to generate a risk report.

2. The intelligent decision analysis platform according to claim 1, characterized in that: The transaction data screening unit is used to screen the time queue of the transaction data based on a predetermined transaction amount to obtain the time queue of the small transaction data.

3. The intelligent decision analysis platform according to claim 2 is characterized in that: The small transaction data local encoding unit includes: a small transaction time series data sampling subunit, which is used to perform sliding window sampling on the time queue of the small transaction data based on a preset time window to obtain a sequence of small transaction data sampling windows; a small transaction local time domain pattern feature extraction subunit, which is used to input each small transaction data sampling window in the sequence of small transaction data sampling windows into a small transaction sequence encoder based on a forward LSTM model to obtain a sequence of small transaction local time domain pattern feature vectors as the sequence of small transaction local time domain pattern features.

4. The intelligent decision analysis platform according to claim 3 is characterized in that: The small transaction full time domain aggregation unit includes: a small transaction initial aggregation subunit, which is used to calculate the positional mean vector of the sequence of the small transaction local time domain pattern feature vectors to obtain the small transaction initial aggregation representation vector; a small transaction static energy potential factor calculation subunit, which is used to calculate the static energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vectors to obtain a sequence of small transaction static energy potential factors; a small transaction dynamic energy potential factor calculation subunit, which is used to calculate the dynamic energy potential factor of each small transaction local time domain pattern feature vector in the sequence of the small transaction local time domain pattern feature vector relative to the small transaction initial aggregation representation vector. The sub-unit is used to obtain a sequence of dynamic energy potential factors of small transactions; the small transaction superposition energy potential factor calculation sub-unit is used to determine the superposition state energy potential factor of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors based on the static energy potential factor and dynamic energy potential factor of each small transaction local time domain pattern feature vector in the sequence of small transaction local time domain pattern feature vectors to obtain a sequence of superposition state energy potential factors of small transactions; the small transaction full time domain feature aggregation sub-unit is used to perform full time domain feature aggregation on the sequence of local time domain pattern feature vectors of small transactions based on the sequence of superposition state energy potential factors of small transactions to obtain the full time domain kernel representation of the small transactions.

5. The intelligent decision analysis platform according to claim 4, characterized in that: The small-amount transaction static energy potential factor calculation subunit is used to: calculate the mean and standard deviation of the small-amount transaction local time domain pattern feature vector to obtain the small-amount transaction local time domain mean and the small-amount transaction local time domain standard deviation; calculate the positional subtraction of the small-amount transaction local time domain pattern feature vector and the small-amount transaction local time domain mean, divide the small-amount transaction local time domain difference vector after the subtraction by the small-amount transaction local time domain, and calculate the fourth power of the eigenvalue of each position in the vector after the division to obtain the small-amount transaction local time domain static energy potential vector; calculate the positional mean of the small-amount transaction local time domain static energy potential vector to obtain the small-amount transaction static energy potential factor.

6. The intelligent decision analysis platform according to claim 5, characterized in that: The small transaction dynamic energy potential factor calculation subunit is used to: matrix multiply the small transaction aggregation weight matrix with the small transaction initial aggregation representation vector to obtain the small transaction initial aggregation weight representation vector; matrix multiply the small transaction weight matrix with the small transaction local time domain pattern feature vector to obtain the small transaction local time domain weight vector; perform positional dot multiplication of the small transaction initial aggregation weight representation vector and the small transaction local time domain weight vector, and then add the obtained small transaction fusion weight vector and the small transaction bias vector by position to obtain the small transaction local time domain dynamic energy potential vector; calculate the square of the Frobenius norm of the small transaction local time domain dynamic energy potential vector to obtain the small transaction dynamic energy potential factor.

7. The intelligent decision analysis platform according to claim 6, characterized in that: The small-amount transaction superposition state energy potential factor calculation subunit is used to: calculate an exponential function with the natural constant e as the base and the static energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias static energy potential factor of the small-amount transaction; calculate an exponential function with the natural constant e as the base and the square of the dynamic energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction as the exponent to obtain the local time domain bias dynamic energy potential factor of the small-amount transaction; add the local time domain bias static energy potential factor of the small-amount transaction and the local time domain bias dynamic energy potential factor of the small-amount transaction to obtain the superposition state energy potential factor of the local time domain pattern characteristic vector of the small-amount transaction.

8. The intelligent decision analysis platform according to claim 7, characterized in that: The small transaction full-time domain feature aggregation subunit is used to: input the sequence of small transaction superposition state energy potential factors into the energy potential factor gated weighting network to obtain a sequence of small transaction superposition state energy potential weight coefficients; Based on the sequence of small-amount transaction superposition state energy potential weight coefficients, the position-weighted sum of the sequence of small-amount transaction local time-domain pattern feature vectors is calculated to obtain the small-amount transaction full-time-domain kernel representation vector as the small-amount transaction full-time-domain kernel representation.

9. The intelligent decision analysis platform according to claim 8, characterized in that: The transaction pattern recognition result generating unit is used to: input the small transaction full-time domain kernel representation vector into a classifier-based transaction pattern classifier to obtain the transaction pattern recognition result.

10. An intelligent decision analysis method, characterized in that: include: Get the time queue of transaction data; A time queue storing the transaction data; Performing transaction pattern recognition on the time queue of the transaction data to obtain a transaction pattern recognition result, including: screening the time queue of the transaction data to obtain a time queue of small transaction data; performing sliding window sampling and sequence encoding on the time queue of the small transaction data to obtain a sequence of local time domain pattern features of small transactions; dynamically aggregating full time domain kernel features on the sequence of local time domain pattern features of small transactions to obtain a full time domain kernel representation of small transactions; obtaining the transaction pattern recognition result based on the full time domain kernel representation of small transactions; generating a transaction abnormality warning prompt in response to the transaction pattern recognition result being an abnormal transaction pattern; and generating a risk report.

Images