Risk identification method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application disclose a risk identification method and device, electronic equipment and a storage medium, the method comprising: extracting a first low-dimensional transaction feature of real-time transaction data of a target transaction; inputting the first low-dimensional transaction feature into a target risk identification model comprising a long short-term memory network, a switched linear dynamic system and a deep hidden Markov model to obtain a risk identification result of the target transaction. Through the cooperation of the deep hidden Markov model and the switched linear dynamic system, embodiments of the present application can mine hidden states representing transaction behavior patterns, break through the bottleneck of traditional methods in complex money laundering mode identification, and achieve a risk identification effect of "high precision, low false alarm and strong adaptation", providing more reliable anti-money laundering decision support for financial institutions, effectively curbing the inflow of money laundering funds into the legal financial system, and protecting financial security and market order.

Description

Technical Field

[0001] This invention relates to the field of risk management technology, and in particular to a risk identification method, device, electronic device, and storage medium. Background Technology

[0002] Money laundering refers to the process of converting illegally obtained funds into seemingly legitimate funds or assets, thereby allowing these funds to enter the legitimate financial system and making it difficult to trace their criminal origins. Anti-Money Laundering (AML) refers to a set of regulations, policies, and procedures implemented to prevent money laundering. It plays a central role in ensuring the security of the financial system, especially for insurance companies and various financial institutions.

[0003] Currently, mainstream anti-money laundering risk identification technologies are mainly divided into two categories: rule-based identification methods and identification methods based on statistics and traditional machine learning. However, both types of methods have significant limitations: Rule-based identification methods suffer from extremely high false positive rates due to their relatively simple rule design, fixed applicable scenarios, and lack of adaptive learning capabilities. They are particularly ineffective in dealing with complex money laundering patterns such as layered money laundering, organized money laundering, and structured fund splitting, and their ability to identify new and evolving money laundering behaviors is especially weak.

[0004] Identification methods based on statistics and traditional machine learning rely heavily on manual feature engineering, making it difficult for models to uncover deep correlations in high-dimensional transaction data, time-series behavioral patterns, and account association networks. This results in a high rate of missed detection for highly concealed money laundering activities and an inability to effectively capture abnormal signals in complex money laundering patterns. Summary of the Invention

[0005] The purpose of this invention is to provide a risk identification method, device, electronic device, and storage medium to solve the problem of missed risk identification caused by the inability of existing risk identification technologies to accurately identify complex money laundering patterns.

[0006] To solve the above-mentioned technical problems, the embodiments of the present invention are implemented as follows: In a first aspect, an embodiment of the present invention provides a risk assessment method, the method comprising: Extract the first low-dimensional transaction features from the real-time transaction data of the target transaction; The first low-dimensional transaction feature is input into a target risk identification model that includes a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model to obtain the risk identification result of the target transaction. The target risk identification model is used to perform the following steps: The first hidden state feature of the first low-dimensional transaction feature is extracted by the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between real-time transaction data and historical transaction data. Based on the features of the first hidden state, the hidden state of the target transaction at the current moment is determined by a deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction. Based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction feature is optimized by switching the linear dynamic system to obtain the optimized first low-dimensional transaction feature. Based on the optimized first low-dimensional transaction features and first hidden state features, the risk identification result of the target transaction is determined.

[0007] In one implementation, the first low-dimensional transaction features of the target transaction's real-time transaction data are extracted, including: Obtain real-time transaction data for the target transaction; Real-time transaction data is preprocessed to obtain preprocessed data; The preprocessed data is encoded by an encoder to map the high-dimensional preprocessed data to a low-dimensional latent space, thus obtaining the first low-dimensional transaction features.

[0008] In one implementation, based on the features of the first hidden state, the hidden state of the target transaction at the current moment is determined by a deep hidden Markov model, including: The hidden state of the target transaction at the current moment is determined by reasoning based on the features of the first hidden state, the pre-trained state transition matrix, and the observation probability matrix using a deep hidden Markov model.

[0009] In one implementation, based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction features are optimized by switching the linear dynamic system to obtain the optimized first low-dimensional transaction features, including: Based on the hidden state, the linear dynamic system switches to the processing mode corresponding to the hidden state; where the processing mode corresponds to a set of pre-trained parameterized linear equations. In the processing mode, the switching linear dynamic system optimizes the first low-dimensional transaction feature based on the parameterized linear equation to obtain the optimized first low-dimensional transaction feature.

[0010] In one implementation, the switching linear dynamic system has multiple processing modes; the parameterized linear equations corresponding to each processing mode are determined through pre-training via the following steps: Key features are extracted from historical transaction data to obtain key features of historical transactions; Statistical characteristics to determine key features of historical transactions; Based on statistical characteristics, key features of historical transactions are divided into intervals to obtain multiple feature intervals; For each feature interval, a set of parameterized linear equations corresponding to a processing mode of the switching linear dynamic system are pre-trained.

[0011] In one implementation, the state transition matrix is ​​determined through pre-training using the following steps: Encode the data samples constructed based on historical transaction data to obtain the second low-dimensional transaction features of the data samples; The second hidden state features of the second low-dimensional transaction features are extracted using a long short-term memory network. By performing linear transformation and activation processing on the second hidden state features, the second state transition probability corresponding to the second hidden state features is determined. Based on the second state transition probability, the preset initial state transition probability is updated to obtain the state transition matrix.

[0012] In one implementation, the risk identification result of the target transaction is determined based on the optimized first low-dimensional transaction features and the first hidden state features, including: The optimized first low-dimensional transaction features and the first hidden state features are input into at least one fully connected layer to obtain the risk identification result of the target transaction output by the fully connected layer based on the first low-dimensional transaction features and the first hidden state features.

[0013] Secondly, embodiments of the present invention provide a risk identification device, the device comprising: The first extraction module is used to extract the first low-dimensional transaction features from the real-time transaction data of the target transaction. The identification module is used to input the first low-dimensional transaction features into a target risk identification model that includes a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model, to obtain the risk identification result of the target transaction; the target risk identification model is used to perform the following steps: The first hidden state feature of the first low-dimensional transaction feature is extracted by the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between real-time transaction data and historical transaction data. Based on the features of the first hidden state, the hidden state of the target transaction at the current moment is determined by a deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction. Based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction feature is optimized by switching the linear dynamic system to obtain the optimized first low-dimensional transaction feature. Based on the optimized first low-dimensional transaction features and first hidden state features, the risk identification result of the target transaction is determined.

[0014] In one implementation, the first extraction module is specifically used for: Obtain real-time transaction data for the target transaction; Real-time transaction data is preprocessed to obtain preprocessed data; The preprocessed data is encoded by an encoder to map the high-dimensional preprocessed data to a low-dimensional latent space, thus obtaining the first low-dimensional transaction features.

[0015] In one implementation, the identification module is specifically used for: The hidden state of the target transaction at the current moment is determined by reasoning based on the features of the first hidden state, the pre-trained state transition matrix, and the observation probability matrix using a deep hidden Markov model.

[0016] In one implementation, the identification module is specifically used for: Based on the hidden state, the linear dynamic system switches to the processing mode corresponding to the hidden state; where the processing mode corresponds to a set of pre-trained parameterized linear equations. In the processing mode, the switching linear dynamic system optimizes the first low-dimensional transaction feature based on the parameterized linear equation to obtain the optimized first low-dimensional transaction feature.

[0017] In one implementation, the device further includes a first pre-training module for: Key features are extracted from historical transaction data to obtain key features of historical transactions; Statistical characteristics to determine key features of historical transactions; Based on statistical characteristics, key features of historical transactions are divided into intervals to obtain multiple feature intervals; For each feature interval, a set of parameterized linear equations corresponding to a processing mode of the switching linear dynamic system are pre-trained.

[0018] In one implementation, the device further includes a second pre-training module for: Encode the data sample constructed based on historical transaction data to obtain the second low-dimensional transaction features of the data sample; The second hidden state features of the second low-dimensional transaction features are extracted using a long short-term memory network. By performing linear transformation and activation processing on the second hidden state features, the second state transition probability corresponding to the second hidden state features is determined. Based on the second state transition probability, the preset initial state transition probability is updated to obtain the state transition matrix.

[0019] In one implementation, the identification module is specifically used for: The optimized first low-dimensional transaction features and the first hidden state features are input into at least one fully connected layer to obtain the risk identification result of the target transaction output by the fully connected layer based on the first low-dimensional transaction features and the first hidden state features.

[0020] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the risk identification method provided in the above embodiments.

[0021] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the risk identification method provided in the above embodiments.

[0022] The above-described at least one technical solution adopted in the embodiments of the present invention can achieve the following beneficial effects: The risk identification method provided in this invention extracts first low-dimensional transaction features from real-time transaction data of the target transaction and inputs them into a target risk identification model comprising a Long Short-Term Memory (LSTM) network, a Switching Linear Dynamics (SLR) system, and a Deep Hidden Markov Model (DMR). The risk identification result of the target transaction is then obtained through this model. This method processes the first hidden state features extracted from the LSTM network using the DMR to determine the hidden state of the target transaction at the current moment. Then, the SLR optimizes the first low-dimensional transaction features based on this hidden state. Finally, the risk identification result of the target transaction is determined based on the optimized first low-dimensional transaction features and the first hidden state features. This invention, through the synergy of the DMR and the SLR, can uncover hidden states representing transaction behavior patterns, breaking through the bottleneck of traditional methods in identifying complex money laundering patterns. It achieves a risk identification effect of "high accuracy, low false alarm, and strong adaptability," providing financial institutions with more reliable anti-money laundering decision support, effectively curbing the inflow of money laundering funds into the legitimate financial system, and safeguarding financial security and market order. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating a risk identification method provided in an embodiment of the present invention. Figure 1 ; Figure 2 This is a schematic diagram of the structure of a target risk identification model provided in an embodiment of the present invention; Figure 3A flowchart illustrating a risk identification method provided in an embodiment of the present invention. Figure 2 ; Figure 4 This is a schematic diagram of the structure of a long short-term memory network provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a risk identification device provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a risk identification electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0026] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0027] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit this disclosure.

[0029] Explanation of relevant terms: 1. Self-encoder An autoencoder is an unsupervised deep neural network model designed to learn a low-dimensional representation (i.e., feature representation or latent space) of input data through the structure of an encoder and decoder, and to recover an output that is as similar as possible to the input data. Autoencoders are widely used for tasks such as dimensionality reduction, feature learning, denoising, and anomaly detection.

[0030] 2. Hidden Markov Model A Hidden Markov Model (HMM) is a statistical model used to describe stochastic processes with hidden states. In this model, the state of the system is implicit (i.e., not directly observable), but the system's output (observations) is observable, and there is a probabilistic relationship between state transitions and observations. HMMs are widely used in time series analysis, speech recognition, natural language processing, and other fields.

[0031] 3. Anti-money laundering AML stands for Anti-Money Laundering, which refers to a set of regulations, policies, and procedures implemented to prevent money laundering. Money laundering refers to the process of converting illegally obtained funds into seemingly legitimate funds or property, allowing these funds to enter the legitimate financial system and making it difficult to trace their criminal origins. AML aims to detect, prevent, and respond to this activity.

[0032] 4. Switching to a linear dynamic system Switching Linear Dynamical System (SLDS) is a complex probabilistic graphical model used to model piecewise stationary time series data. It integrates two core mechanisms: a discrete hidden Markov Model (HMM), responsible for modeling the stochastic, discrete switching of the system between different macroscopic modes or operating states; and a continuous linear dynamical system (LDS), responsible for modeling the linear evolution of its continuous latent variables when the system is in a specific discrete state.

[0033] 5. Long Short-Term Memory Neural Network Long Short-Term Memory (LSTM) neural networks are a special type of recurrent neural network (RNN) architecture, proposed in 1997. They were designed to address the vanishing or exploding gradient problems that occur in standard RNNs when processing long sequences of data, enabling them to effectively capture and learn long-term dependencies in sequential data.

[0034] Example 1 Embodiment 1 of the present invention provides a risk identification method to address the problem of missed risk identification caused by the difficulty of accurately identifying complex money laundering patterns in existing risk identification technologies.

[0035] like Figure 1 As shown, an embodiment of the present invention provides a risk identification method, the method comprising: Step S101: Extract the first low-dimensional transaction features from the real-time transaction data of the target transaction.

[0036] Specifically, the real-time transaction data of the target transaction includes basic transaction attribute information, transaction entity information, transaction behavior information, and environmental information.

[0037] Basic transaction attribute information refers to the core identifiers and static parameters of a transaction, directly reflecting the "surface characteristics" of the transaction itself, and serving as the fundamental input for risk identification. It mainly includes: transaction identifier, transaction time, transaction amount, transaction type, transaction status, and transaction channel.

[0038] Transaction entity information refers to the identity information of the accounts and related entities participating in the transaction, used to identify "account control relationships" and "abnormal fund flow" (such as multi-account coordination in money laundering). It mainly includes: entity account information and entity-related information. Entity account information further includes the entity's own account information and the counterparty's account information. Entity-related information further includes information on the actual controller / beneficiary and the network of related accounts.

[0039] Transaction behavior information refers to the dynamic performance of transactions in terms of time series, amount distribution, frequency, etc. For example, it can include recent transaction sequences, changes in transaction frequency, and periodic characteristics.

[0040] Environmental information refers to the external environment and technical carrier information in which a transaction occurs, used to identify abnormal scenarios such as "unauthorized operation" and "cross-regional collaboration" (e.g., group members logging in from different locations). It can include network and environmental information, device information, and operational behavior characteristics. Network and environmental information may include, for example, the transaction initiating IP address (public / private network), IP location (country / city), whether a proxy / VPN was used, GPS location (mobile device), and Wi-Fi hotspot MAC address; device information may include, for example, device ID (IMEI / MEID), device model, operating system version, whether jailbroken / rooted, device first use time, and the number of transactions on the device in the past N days (to determine if "device sharing" exists); operational behavior characteristics may include, for example, the transaction input method (manual input / copy and paste / biometric identification), page dwell time, and whether key transaction information was modified (e.g., temporarily changing the receiving account).

[0041] In one implementation, the first low-dimensional transaction features of the target transaction's real-time transaction data are extracted, including: Obtain real-time transaction data for the target transaction; Real-time transaction data is preprocessed to obtain preprocessed data; The preprocessed data is encoded by an encoder to map the high-dimensional preprocessed data to a low-dimensional latent space, thus obtaining the first low-dimensional transaction features.

[0042] Specifically, multi-dimensional data related to the target transaction, including basic transaction attributes, transaction entity information, transaction behavior information, and environmental information, can be collected in real time. The preprocessing objective is to remove noise, standardize the format, fill in missing information, and standardize the distribution, transforming the real-time transaction data into a numerical feature matrix suitable for encoding. The specific implementation is as follows: First, the timestamp field in the real-time transaction data is parsed, including string format (e.g., YYYY-MM-DD HH:MM:SS) and numerical Unix timestamp, and the format is standardized. Then, resampling is performed according to the preset time granularity, including upsampling and downsampling. Downsampling (reducing the time granularity, such as from 1 minute to 5 minutes): Mean aggregation is used. Let the t-th target time step cover [t1, t2, ..., tk] time points of the real-time transaction data. The aggregated value is: ; Upsampling (increasing time granularity, such as from 5 minutes to 1 minute): This involves using linear interpolation to fill in intermediate values. For example, The data in this column contains missing values. The mean of this column is... The calculation formula is: , For missing values, use Fill in the blanks.

[0043] The obtained preprocessed data is encoded by the encoder in the autoencoder, thereby mapping the high-dimensional preprocessed data to a low-dimensional latent space to obtain the first low-dimensional transaction features.

[0044] An autoencoder consists of two parts: an encoder and a decoder. The encoder's function is to process high-dimensional input data... Mapping to a low-dimensional latent space yields a low-dimensional representation. The decoder, on the other hand, represents the low-dimensional representation. Reconstructing back to the original high-dimensional data .

[0045] Encoder: Typically a multi-layer neural network whose output is a representation of the latent space. Assume the encoder has... Layer, number The input of the layer is The output is Then we have: , in, It is the first Layer weight matrix; It is the first Layer bias vector; It is an activation function, such as the ReLU function. ;go through After layer encoding, the final representation of the latent space is obtained. .

[0046] Decoder: Also a multi-layer neural network, representing the latent space Reconstruct the original data. Assume the decoder has... Layer, number The input of the layer is (Initial) The output is Then we have: , in, It is the first Layer weight matrix; It is the first Layer bias vector; It is an activation function, such as the Sigmoid function. ;go through After layer decoding, the reconstructed data is finally obtained. .

[0047] Loss function: The goal of an autoencoder is to reconstruct the data As close as possible to the original data The mean squared error (MSE) is typically used as the loss function. , in, It refers to the number of samples.

[0048] Step S102: Input the first low-dimensional transaction features into the target risk identification model, which includes a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model, to obtain the risk identification result of the target transaction.

[0049] Specifically, such as Figure 2As shown, the target risk identification model includes a Long Short-Term Memory (LSTM) network, a Switched Linear Dynamics System (SLDS), and a Deep Hidden Markov Model (DHMM). Building upon the original DHMM, the target risk identification model incorporates both LSTM and SLDS, enabling autonomous mining of dynamic patterns in time-series data. It combines the ability of an autoencoder to extract low-dimensional representations with the ability of a DHMM to capture temporal dependencies, thus automatically learning and discovering dynamic structures and potential clustering patterns in unlabeled data. The introduction of LSTM allows for more accurate modeling of long-term dependencies and nonlinear dynamics in time series data, while the addition of SLDS enhances the model's dynamic modeling capabilities on continuous and multimodal data.

[0050] In this implementation, such as Figure 3 As shown, the target risk identification model achieves complex money laundering pattern identification through a cascaded architecture of Long Short-Term Memory Network (LSTM), Deep Hidden Markov Model (DHMM), and Switched Linear Dynamical System (SLDS). Specifically, it performs the following steps: Step S1021: Extract the first hidden state feature of the first low-dimensional transaction feature through the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between real-time transaction data and historical transaction data.

[0051] like Figure 4 As shown, Long Short-Term Memory (LSTM) networks contain input gates. Forgotten Gate Output gate and memory unit .

[0052] The input to the LSTM is the current time of the encoder output. First low-dimensional transaction characteristics and memory units The hidden state at the previous moment is ,but: Forgotten Gate: , in, It is the Sigmoid function; It is the weight matrix of the forget gate; It is the bias vector of the forget gate.

[0053] Input Gate: , , in, It is the weight matrix of the input gate and candidate memory units; It is the corresponding bias vector.

[0054] Memory unit update: , in,( () is element-wise multiplication.

[0055] Output gate: , , in, This is the first hidden state feature, corresponding to the hidden state of the last time step; It is the weight matrix of the output gate; It is the bias vector of the output gate.

[0056] The LSTM gating mechanism (forget gate, input gate, output gate) captures the temporal correlation between real-time transaction data and historical transaction data (such as the continuity of fund flow and the trend of transaction frequency).

[0057] Step S1022: Based on the first hidden state features, determine the hidden state of the target transaction at the current moment using a deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction.

[0058] Specifically, a Deep Hidden Markov Model (DHMM) consists of a sequence of hidden states. and observation sequence Composition, its parameters include the initial state probability State transition matrix and observation probability matrix State transition probability Observation probability .

[0059] In one implementation, based on the features of the first hidden state, the hidden state of the target transaction at the current moment is determined by a deep hidden Markov model, including: The hidden state of the target transaction at the current moment is determined by reasoning based on the features of the first hidden state, the pre-trained state transition matrix, and the observation probability matrix using a deep hidden Markov model.

[0060] Specifically, the deep hidden Markov model in this embodiment of the invention includes a predefined set of hidden states (i.e., a sequence of hidden states), a pre-trained state transition matrix, and an observation probability matrix.

[0061] For example, setting The discrete hidden states (representing typical transaction behavior patterns) are as follows: Normal transaction status (such as daily consumption and periodic payment); : Trial splitting (e.g., multiple small transactions to test account limits); Group collaboration (such as simultaneous transfers between multiple accounts); Dispersed inflows and concentrated outflows (such as the classic money laundering pattern); Other abnormal situations (such as large transactions outside of business hours).

[0062] Features of the first hidden state at the current time The optimal hidden state sequence can be solved using the Viterbi algorithm, taking the hidden state at the current time step. As the output of the hidden state of the target transaction at the current moment. For example, the hidden state at the current moment. This represents a "dispersed inflow-concentrated outflow state," thereby quantitatively characterizing the deep trading behavior pattern of the target transaction.

[0063] In one implementation, the state transition matrix is ​​determined through pre-training using the following steps: Encode the data samples constructed based on historical transaction data to obtain the second low-dimensional transaction features of the data samples; The second hidden state features of the second low-dimensional transaction features are extracted using a long short-term memory network. By performing linear transformation and activation processing on the second hidden state features, the second state transition probability corresponding to the second hidden state features is determined. Based on the second state transition probability, the preset initial state transition probability is updated to obtain the state transition matrix.

[0064] Specifically, data samples are constructed based on historical transaction data. After dimensionality reduction processing by an encoder, the corresponding second hidden state features are extracted by a long short-term memory network. The second hidden state features have the same feature dimension as the first hidden state features.

[0065] set up This is a historical moment extracted by LSTM. The second hidden state feature, then the second state transition probability The second hidden state features can be linearly transformed and activated using a fully connected layer: , in, It is the weight matrix for state transitions; It is the bias vector.

[0066] The preset initial state transition probability is dynamically updated based on the calculated second state transition probability, and finally the pre-trained state transition matrix is ​​obtained.

[0067] Step S1023: Based on the hidden state of the target transaction at the current moment, optimize the first low-dimensional transaction feature by switching the linear dynamic system to obtain the optimized first low-dimensional transaction feature.

[0068] Switching linear dynamic systems are a type of "state-dependent linear dynamic model" that, at its core, achieves dynamic optimization of the first low-dimensional transaction features by switching different processing modes through hidden states.

[0069] In one implementation, based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction features are optimized by switching the linear dynamic system to obtain the optimized first low-dimensional transaction features, including: Based on the hidden state, the linear dynamic system switches to the processing mode corresponding to the hidden state; where the processing mode corresponds to a set of pre-trained parameterized linear equations. In the processing mode, the switching linear dynamic system optimizes the first low-dimensional transaction feature based on the parameterized linear equation to obtain the optimized first low-dimensional transaction feature.

[0070] Specifically, the pre-trained parameterized linear equations include state equations and observation equations.

[0071] Equations of state: , in, It is the system status; It is the control input (which can be simplified to a constant 1); Is it related to the hidden state? The relevant state transition matrix; It is a control input matrix, used to describe the control input. Regarding system status The impact; It is process noise.

[0072] An example, hidden states (e.g., "gang collaboration") This determines the evolutionary pattern of the system's state. When (In gang mode) This may strengthen the persistence of the "frequency of transactions by related accounts" dimension (the diagonal elements of the matrix are close to 1), reflecting the continuity of group transactions; when (In the normal state) It may weaken state memory (diagonal elements < 1), reflecting the randomness of normal transactions.

[0073] One example is controlling input. This usually represents external intervention (such as a "sudden increase in transaction amount" event). The weights are determined by the hidden state. Decision. When During the (distributed transfer-concentrated transfer-out state), It's possible to assign a high weight to the "cumulative difference amount" dimension, amplifying the impact of the control input; when (In the normal state) The lower weight reduces external interference.

[0074] Observation equation: , in, These are the observed values ​​(optimized first low-dimensional transaction features). Is and pattern The relevant observation probability matrix; It involves observing noise.

[0075] Accordingly, the switching linear dynamic system first infers and determines the system state at the current moment based on the state equation in the defined parameterized linear equation. Then, it is input into the observation equation, and a linear transformation is performed on it to obtain the optimized first low-dimensional transaction feature. .

[0076] For example, considering the aforementioned set of hidden states, the corresponding optimization logic is as follows: when (Normal State): SLDS weakens the weight of unconventional features (such as "non-business hours transactions" weight × 0.1) and strengthens periodic features (such as "fixed period transfers" weight × 1.5), outputting optimized features (close to the original features); when (Dispersed Inflow - Centralized Outflow): SLDS enhances the weight of features such as "number of associated accounts" and "cumulative split amount" (e.g., corresponding dimension ×2.0), suppresses the weight of "single transaction amount" (×0.5), and highlights the money laundering pattern signal.

[0077] In one implementation, the switching linear dynamic system has multiple processing modes; the parameterized linear equations corresponding to each processing mode are determined through pre-training via the following steps: Key features are extracted from historical transaction data to obtain key features of historical transactions; Statistical characteristics to determine key features of historical transactions; Based on statistical characteristics, key features of historical transactions are divided into intervals to obtain multiple feature intervals; For each feature interval, a set of parameterized linear equations corresponding to a processing mode of the switching linear dynamic system are pre-trained.

[0078] Specifically, statistical characteristics include at least one of the following: mean characteristics, variance characteristics, extreme value characteristics, and periodic characteristics.

[0079] Based on the anti-money laundering scenario, the following six key features (covering transaction behavior, correlation networks, and temporal patterns) can be extracted from the historical transaction database (including labeled normal / money laundering transaction samples), resulting in a multi-dimensional key feature vector for each historical transaction sample: Splitting characteristics: the percentage of transactions with a single amount less than the anti-money laundering threshold (e.g., 50,000 yuan) in the past N days (e.g., 7 days), and the percentage of the cumulative split amount to the total assets of the account; Related characteristics: Number of related accounts of the counterparty (same actual controller / IP / device), and cumulative transaction frequency with related accounts in the past N days; Time-series characteristics: trading frequency volatility (standard deviation / mean of the number of trades in the past 7 days), and the proportion of trading during non-business hours (0-6 am / 22-24 pm); Amount characteristics: Single transaction amount deviation (current amount - account average amount over the past 30 days) / average amount, percentage of large transactions (> 500,000 yuan); Channel characteristics: Number of transactions through uncommon channels (such as newly registered third-party payment accounts), and the proportion of cross-border foreign exchange transactions; Historical risk characteristics: the number of times the account has been warned by anti-money laundering authorities in the past year, and the number of accounts involved in the case among the related accounts.

[0080] Based on the obtained key historical transaction features, mean, variance, extreme value, and periodicity features are calculated to determine the statistical characteristics of these key historical transaction features. One or more of these statistical features can be selected for interval division, resulting in multiple feature intervals.

[0081] Each feature interval corresponds to a processing mode, and a set of parameterized linear equations is pre-trained for each processing mode. The specific pre-training steps are as follows: Samples whose key features fall within the target interval are selected from historical transaction data to form a training subset of that interval; the money laundering risk labels in the training subset are used as supervision signals, and the weight matrix included in the learned parameterized linear equation is used as the supervision signal. This allows the optimized key features to maximize the differentiation between money laundering and normal transactions within the specified interval. The training is then used as the pre-training target to obtain a set of parameterized linear equations for the processing mode corresponding to the target interval.

[0082] Step S1024: Determine the risk identification result of the target transaction based on the optimized first low-dimensional transaction features and the first hidden state features.

[0083] Specifically, based on the first low-dimensional transaction features optimized by SLDS (risk signals focusing on the current behavior pattern) and the first hidden state features extracted by LSTM (containing long-term behavior patterns with temporal correlations), the risk probability is output by the classifier, and the risk level is determined by combining the threshold, thereby achieving accurate identification of complex money laundering patterns.

[0084] In one implementation, the risk identification result of the target transaction is determined based on the optimized first low-dimensional transaction features and the first hidden state features, including: The optimized first low-dimensional transaction features and the first hidden state features are input into at least one fully connected layer to obtain the risk identification result of the target transaction output by the fully connected layer based on the first low-dimensional transaction features and the first hidden state features.

[0085] Accordingly, a direct concatenation method can be used to fuse the optimized first low-dimensional transaction features with the first hidden state features. The fused features can cover the full-dimensional risk signals from "micro-features to macro-time series". Then, a lightweight multilayer perceptron (MLP) is used as a classifier to output the predicted probability that the target transaction is a money laundering risk.

[0086] Specifically, the classifier consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The first hidden layer is a fully connected layer (e.g., 32-dimensional), with the ReLU activation function (introducing non-linearity to capture feature interactions) and a dropout ratio of 0.2 (to prevent overfitting); the second hidden layer is also a fully connected layer (e.g., 16-dimensional), with the ReLU activation function (further refining core risk signals); the output layer is a 1-dimensional fully connected layer with the Sigmoid activation function (outputting probability values ​​of 0-1).

[0087] The risk identification method provided in this invention extracts first low-dimensional transaction features from real-time transaction data of a target transaction and inputs them into a target risk identification model comprising a Long Short-Term Memory (LSTM) network, a Switching Linear Dynamics (SLR) system, and a Deep Hidden Markov Model (DMR). The model then yields the risk identification result for the target transaction. Specifically, the method processes the first hidden state features extracted from the LSTM network using the DMR to determine the hidden state of the target transaction at the current moment. The SLR then optimizes the first low-dimensional transaction features based on this hidden state. Finally, the risk identification result of the target transaction is determined based on the optimized first low-dimensional transaction features and the first hidden state features. This invention, through the synergy of the DMR and the SLR, can uncover hidden states representing transaction behavior patterns, overcoming the bottlenecks of traditional methods in identifying complex money laundering patterns. It achieves a risk identification effect of "high accuracy, low false alarm, and strong adaptability," providing financial institutions with more reliable anti-money laundering decision support, effectively curbing the inflow of money laundering funds into the legitimate financial system, and safeguarding financial security and market order.

[0088] In this embodiment of the invention, the training of the target risk identification model is completed in stages, including component pre-training and overall joint training. It learns temporal dependencies, latent states, and dynamic feature optimization capabilities through historical transaction data, ultimately achieving accurate identification of complex money laundering patterns. The following are the detailed training steps, including the design of the loss function for each stage: Step S201: Collect historical transaction data of labeled normal / money laundering transaction samples, and preprocess the historical transaction data; Historical transaction data includes multi-dimensional data such as basic transaction attribute information, transaction entity information, transaction behavior information, and environmental information.

[0089] Step S202: Input the preprocessed historical transaction data into an autoencoder that includes an encoder and a decoder for training to obtain a pre-trained encoder.

[0090] The specific training process includes encoder forward propagation, decoder reconstruction, loss function, and Adam optimizer optimization.

[0091] (1) Encoder forward propagation: The preprocessed historical transaction data is input into the encoder, the encoder performs forward propagation calculation, and outputs low-dimensional transaction features (second low-dimensional transaction features).

[0092] Encoder: , in, For encoder number The output features of the layer; For encoder number The output features of the layer (at the initial input) , (High-dimensional preprocessed historical transaction data); For the first Layer weight matrix; For the first Layer bias vector; This is the activation function.

[0093] (2) Decoder reconstruction: Input the low-dimensional transaction features into the decoder for reconstruction to obtain the reconstructed high-dimensional transaction features.

[0094] Decoder: , in, For decoder number The output features of the layer; For encoder number The output features of the layer (at the initial input) , (Low-dimensional transaction features output by the encoder). For the first Layer weight matrix; For the first Layer bias vector; This is the activation function.

[0095] (3) Autoencoder loss function: measures the difference between the reconstructed data and the original data, and enables the encoder to learn an effective low-dimensional representation by minimizing the loss.

[0096] Loss function: , in, The number of samples; For the first The original high-dimensional preprocessed data of each sample; For the first Reconstructed data for each sample; It is the square of the L2 norm.

[0097] (4) Optimizer: The Adam optimizer is used to minimize the time. After pre-training, the encoder parameters are fixed.

[0098] Step S203: Input the low-dimensional transaction features output by the pre-trained encoder into the long short-term memory network for training to obtain the pre-trained long short-term memory network.

[0099] The specific training process includes forward propagation, auxiliary loss, and Adam optimizer optimization.

[0100] (1) Forward propagation: Calculate the hidden state features at the current time step based on the LSTM gating mechanism. (Characterizing temporal correlation features). The specific calculation process can be found in step S1021.

[0101] (2) LSTM auxiliary loss function: based on the weak labels of transaction behavior pattern categories, cross-entropy is used to constrain the matching degree between hidden state features and real patterns.

[0102] Loss function: , in, This represents the total number of time steps. This represents the number of hidden state categories (such as the 5 transaction behavior patterns mentioned above). For a moment The true pattern label (one-hot encoded, 1 if it is the target category, 0 otherwise); for Mode probabilities after Softmax (the first) (Probability of class).

[0103] (3) Minimize using the Adam optimizer After pre-training, the LSTM parameters are fixed.

[0104] Step S204: Based on the second low-dimensional transaction features and the second hidden state features of historical transaction data, obtain the pre-trained state transition matrix and observation probability matrix; and construct the hidden state loss using Viterbi path log-likelihood.

[0105] First, the second hidden state features are extracted using LSTM. ; Secondly, for The dynamic state transition probability is calculated using linear transformations of the fully connected layer and Softmax. , in, For a moment From hidden state Transition to hidden state The probability ( ); Softmax is the normalized exponential function; This is the weight matrix for the state transition; Historical moments extracted by LSTM The second hidden state feature; For hidden state index; This is the state transition bias vector.

[0106] Next, the KL divergence constraint is used to ensure consistency between the dynamic state transition probability and the initial state transition probability, and the state transition matrix is ​​trained to obtain the state transition matrix. , in, Initial state probability (hidden state) (probability of being the initial state); The initial state transition probability is preset (such as uniform distribution or empirical value). It is the core term of the KL divergence.

[0107] Then, the observation probability matrix is ​​obtained by fitting the distribution of observations in the hidden state using a Gaussian mixture model (GMM). , in, This represents the total number of time steps. For a moment The hidden state is ; Hidden state The prior probability; The probability density function of a multivariate Gaussian distribution (observations) In state The distribution below, It is the mean vector. (where covariance matrix) The observation sequence is (here, the hidden state features output by the LSTM); R is the total number of hidden states.

[0108] Finally, based on the state transition matrix, the observation probability matrix, and the optimal hidden state obtained by the Viterbi algorithm, a DHMM hidden state loss is constructed for training.

[0109] DHMM hidden state loss: , in, The optimal hidden state sequence obtained by the Viterbi algorithm; Observation sequence (the hidden state feature sequence output by LSTM); Here are the state transition matrix and observation probability matrix of the DHMM; Given an observation sequence, the conditional probability of the hidden state sequence.

[0110] Step S205: Calculate statistical features for key features of historical transactions, divide feature intervals according to statistical features; screen historical samples that fall into the target feature interval, use money laundering risk labels as monitoring signals; use the optimized features to distinguish between money laundering and normal transactions, train parameterized linear equations, and obtain a pre-trained switching linear dynamic system.

[0111] Specifically, a pre-trained loss function for the switching linear dynamic system is constructed using binary cross-entropy.

[0112] , in, This represents the total number of time steps. This is a true risk label; Based on optimized features The money laundering probability output by a simple classifier.

[0113] Step S206: Jointly train the target risk identification model.

[0114] The total loss function for joint training: , in, and The hyperparameters are defined; stochastic gradient descent is used to minimize the total loss to complete the training of the target risk identification model.

[0115] In this embodiment of the invention, the specific parameter settings for the target risk identification model are as follows: the total number of training iterations is 20,000; the initial learning rate is set to 0.001, and when the number of iterations reaches 8,000, the learning rate is reduced to one-twentieth of the previous rate; the momentum parameter is set to 0.85 to accelerate the model's convergence speed. After repeated experiments and comparisons, the above parameter values ​​are found to be optimal.

[0116] Example 2 The above describes the risk identification method provided by embodiments of the present invention. Based on the same idea, embodiments of the present invention also provide a risk identification device, such as... Figure 5 As shown.

[0117] The risk identification device includes: The first extraction module 501 is used to extract the first low-dimensional transaction features of the real-time transaction data of the target transaction; The identification module 502 is used to input the first low-dimensional transaction features into a target risk identification model that includes a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model, to obtain the risk identification result of the target transaction; the target risk identification model is used to perform the following steps: The first hidden state feature of the first low-dimensional transaction feature is extracted by the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between real-time transaction data and historical transaction data. Based on the features of the first hidden state, the hidden state of the target transaction at the current moment is determined by a deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction. Based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction feature is optimized by switching the linear dynamic system to obtain the optimized first low-dimensional transaction feature. Based on the optimized first low-dimensional transaction features and first hidden state features, the risk identification result of the target transaction is determined.

[0118] In one implementation, the first extraction module 501 is specifically used for: Obtain real-time transaction data for the target transaction; Real-time transaction data is preprocessed to obtain preprocessed data; The preprocessed data is encoded by an encoder to map the high-dimensional preprocessed data to a low-dimensional latent space, thus obtaining the first low-dimensional transaction features.

[0119] In one implementation, the identification module 502 is specifically used for: The hidden state of the target transaction at the current moment is determined by reasoning based on the features of the first hidden state, the pre-trained state transition matrix, and the observation probability matrix using a deep hidden Markov model.

[0120] In one implementation, the identification module 502 is specifically used for: Based on the hidden state, the linear dynamic system switches to the processing mode corresponding to the hidden state; where the processing mode corresponds to a set of pre-trained parameterized linear equations. In the processing mode, the switching linear dynamic system optimizes the first low-dimensional transaction feature based on the parameterized linear equation to obtain the optimized first low-dimensional transaction feature.

[0121] In one implementation, the device further includes a first pre-training module 503, used for: Key features are extracted from historical transaction data to obtain key features of historical transactions; Statistical characteristics to determine key features of historical transactions; Based on statistical characteristics, key features of historical transactions are divided into intervals to obtain multiple feature intervals; For each feature interval, a set of parameterized linear equations corresponding to a processing mode of the switching linear dynamic system are pre-trained.

[0122] In one implementation, the device further includes a second pre-training module 504, used for: Encode the data sample constructed based on historical transaction data to obtain the second low-dimensional transaction features of the data sample; The second hidden state features of the second low-dimensional transaction features are extracted using a long short-term memory network. By performing linear transformation and activation processing on the second hidden state features, the second state transition probability corresponding to the second hidden state features is determined. Based on the second state transition probability, the preset initial state transition probability is updated to obtain the state transition matrix.

[0123] In one implementation, the identification module 502 is specifically used for: The optimized first low-dimensional transaction features and the first hidden state features are input into at least one fully connected layer to obtain the risk identification result of the target transaction output by the fully connected layer based on the first low-dimensional transaction features and the first hidden state features.

[0124] The risk identification device provided in this invention, through the collaboration of a deep hidden Markov model and a switching linear dynamic system, can uncover the hidden states that characterize transaction behavior patterns. This breaks through the bottleneck of traditional methods in identifying complex money laundering patterns and achieves a risk identification effect of "high accuracy, low false alarm, and strong adaptability." It provides financial institutions with more reliable anti-money laundering decision support, effectively curbs the inflow of money laundering funds into the legitimate financial system, and safeguards financial security and market order.

[0125] Example 3 Figure 6 A schematic diagram of the hardware structure of a risk identification electronic device for implementing various embodiments of the present invention.

[0126] The electronic device includes a processor 601 and a memory 602 storing computer program instructions. Specifically, the processor 601 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0127] Memory 602 may include mass storage for data or instructions. For example, and not limitingly, memory 602 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where suitable, memory 602 may include removable or non-removable (or fixed) media. Where suitable, memory 602 may be internal or external to an electronic device. In a particular embodiment, memory 602 may be a non-volatile solid-state memory.

[0128] In one embodiment, memory 602 may be read-only memory (ROM). In one embodiment, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0129] The processor 601 implements any of the risk identification methods described in the above embodiments by reading and executing computer program instructions stored in the memory 602.

[0130] In one example, the electronic device may also include a communication interface 603 and a bus 610. The processor 601, memory 602, and communication interface 603 are connected via the bus 610 and communicate with each other.

[0131] The communication interface 603 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of the present invention.

[0132] Bus 610 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 610 may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0133] Furthermore, in conjunction with the risk identification methods in the above embodiments, this invention can be implemented using a computer-readable storage medium. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the risk identification methods in the above embodiments.

[0134] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0135] The above description is merely a specific implementation of the present invention. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0136] Secondly, those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0141] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0142] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0143] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0144] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

[0145] The entire process of model training and application involved in this application strictly follows Article 5 of the Patent Law of the People's Republic of China (hereinafter referred to as "A5") and the relevant examination standards of the Patent Examination Guidelines (2023 Edition) to ensure that the technical solution does not violate the law, social morality and public interest, and that the data acquisition and utilization and model training process all comply with ethical compliance requirements. Specific details are as follows: I. Explanation of whether the data used in the model meets A5 requirements 1. Data Source Legality: All datasets used in training this model were obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, with complete copyright attribution and authorization scope clearly marked, and no unauthorized open-source code or data reuse. Data authorized by partners has a formal data usage agreement clearly defining the scope, duration, and confidentiality obligations, with complete authorization chain proof. For self-collected data involving personal information, strict informed consent procedures have been followed, and anonymization processes (including but not limited to field masking, feature anonymization, and differential privacy technology application) have been used to remove personally identifiable information, fully complying with the requirements of the "Interim Measures for the Administration of Generative Artificial Intelligence Services," the "Personal Information Protection Law," and other relevant laws and regulations.

[0146] 2. Data Content Compliance: This dataset has undergone multiple screening and cleaning processes to remove all content that may violate social morality or harm public interests, and does not involve the illegal acquisition or use of genetic resources. For data in sensitive fields (such as medical and financial fields), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technology) is used to ensure that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0147] 3. Data Governance Compliance: Establish a complete data traceability system to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure data is verifiable throughout its entire lifecycle. The dataset annotation process is completed by a professional human R&D team, clearly defining the proportion of human creative contributions, avoiding reliance on AI-generated data that has not undergone substantial human modification, and complying with the "human main contribution" examination requirements in AI patent applications.

[0148] II. Explanation of Model Training Process Meeting A5 Requirements 1. Compliance of training objectives and scheme: The training objective of this model focuses on risk identification. The training scheme and the final output results do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risk of being used for illegal activities, privacy infringement, or public safety disruption.

[0149] 2. Compliance Management of Training Process: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the first neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on expert system feedback to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop management system of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0150] 3. Compliance of Training Environment and Tools: Model training is implemented on a compliant training platform. All open-source frameworks and components used in the training process have obtained the corresponding licenses, and copyright statements and patent citation information are fully retained, with no infringement or reuse. The training environment is built using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. At the same time, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0151] 4. Ethical verification of training results: After the model is trained, it will undergo an additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism will be established to ensure that the model always complies with A5 and relevant laws and regulations in practical applications.

[0152] In summary, the data and training process used in this application model strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there are no violations of laws, social ethics, public interests, or illegal use of genetic resources. Therefore, it fully meets the compliance requirements for patent authorization.

Claims

1. A risk identification method, characterized in that, The method includes: Extract the first low-dimensional transaction features from the real-time transaction data of the target transaction; The first low-dimensional transaction feature is input into a target risk identification model comprising a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model to obtain the risk identification result of the target transaction; the target risk identification model is used to perform the following steps: The first hidden state feature of the first low-dimensional transaction feature is extracted through the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between the real-time transaction data and the historical transaction data. Based on the first hidden state feature, the hidden state of the target transaction at the current moment is determined by the deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction. Based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction feature is optimized through the switching linear dynamic system to obtain the optimized first low-dimensional transaction feature. Based on the optimized first low-dimensional transaction features and the first hidden state features, the risk identification result of the target transaction is determined.

2. The method according to claim 1, characterized in that, The first low-dimensional transaction features extracted from the real-time transaction data of the target transaction include: Obtain the real-time transaction data of the target transaction; The real-time transaction data is preprocessed to obtain preprocessed data; The preprocessed data is encoded by an encoder to map the high-dimensional preprocessed data to a low-dimensional latent space, thereby obtaining the first low-dimensional transaction feature.

3. The method according to claim 1, characterized in that, The step of determining the hidden state of the target transaction at the current moment based on the first hidden state feature using the deep hidden Markov model includes: The hidden state of the target transaction at the current moment is determined by reasoning based on the first hidden state features, the pre-trained state transition matrix, and the observation probability matrix using the deep hidden Markov model.

4. The method according to claim 3, characterized in that, The step of optimizing the first low-dimensional transaction feature based on the hidden state of the target transaction at the current moment through the switching linear dynamic system to obtain the optimized first low-dimensional transaction feature includes: Based on the hidden state, the switching linear dynamic system switches to the processing mode corresponding to the hidden state; wherein, the processing mode corresponds to a set of pre-trained parameterized linear equations; In the processing mode, the switching linear dynamic system optimizes the first low-dimensional transaction feature based on the parameterized linear equation to obtain the optimized first low-dimensional transaction feature.

5. The method according to claim 4, characterized in that, The switching linear dynamic system has multiple processing modes; the parameterized linear equations corresponding to each processing mode are determined through pre-training using the following steps: Key features are extracted from historical transaction data to obtain key features of historical transactions; Determine the statistical characteristics of the key features of the historical transactions; Based on the statistical characteristics, the key features of the historical transactions are divided into intervals to obtain multiple feature intervals; For each feature interval, a set of parameterized linear equations corresponding to a processing mode of the switching linear dynamic system are pre-trained.

6. The method according to claim 3, characterized in that, The state transition matrix is ​​determined through pre-training using the following steps: Encode the data sample constructed based on historical transaction data to obtain the second low-dimensional transaction features of the data sample; The second hidden state feature of the second low-dimensional transaction feature is extracted through the Long Short-Term Memory network; By performing linear transformation and activation processing on the second hidden state feature, the second state transition probability corresponding to the second hidden state feature is determined. Based on the second state transition probability, the preset initial state transition probability is updated to obtain the state transition matrix.

7. The method according to claim 1, characterized in that, The step of determining the risk identification result of the target transaction based on the optimized first low-dimensional transaction features and the first hidden state features includes: The optimized first low-dimensional transaction feature and the first hidden state feature are input into at least one fully connected layer to obtain the risk identification result of the target transaction output by the fully connected layer based on the first low-dimensional transaction feature and the first hidden state feature.

8. A risk identification device, characterized in that, The device includes: The first extraction module is used to extract the first low-dimensional transaction features from the real-time transaction data of the target transaction. The identification module is used to input the first low-dimensional transaction features into a target risk identification model that includes a long short-term memory network, a switching linear dynamic system, and a deep hidden Markov model, to obtain the risk identification result of the target transaction; the target risk identification model is used to perform the following steps: The first hidden state feature of the first low-dimensional transaction feature is extracted through the Long Short-Term Memory Network; the first hidden state feature represents the temporal correlation between the real-time transaction data and the historical transaction data. Based on the first hidden state feature, the hidden state of the target transaction at the current moment is determined by the deep hidden Markov model; the hidden state represents the transaction behavior pattern of the target transaction. Based on the hidden state of the target transaction at the current moment, the first low-dimensional transaction feature is optimized through the switching linear dynamic system to obtain the optimized first low-dimensional transaction feature. Based on the optimized first low-dimensional transaction features and the first hidden state features, the risk identification result of the target transaction is determined.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the risk identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the risk identification method as described in any one of claims 1 to 7.

Images