A financial fraud detection method and device based on multi-scale spatio-temporal feature learning

Abstract

The present application relates to the field of financial risk control and anti-fraud technology, and provides a financial fraud detection method and device based on multi-scale space-time feature learning, comprising the following steps: S1, acquiring financial transaction data and performing basic preprocessing; S2, modeling structural features of input features through an encoder; S3, introducing multi-scale time modeling in a latent space; S4, constructing a training-stable latent representation learning model; S5, generating a risk prediction result through a decoder and a classifier; and S6, risk assessment output. The present application has enhanced structure relationship description capability, more sufficient multi-scale time feature expression, improved training stability and generalization capability, and strong engineering applicability.

Description

Technical Field

[0001] This invention relates to the field of financial risk control and anti-fraud technology, specifically to a financial fraud detection method and apparatus based on multi-scale spatiotemporal feature learning. Background Technology

[0002] With the digitalization, networking, and platformization of financial services, business scenarios such as electronic payment, mobile payment, cross-border settlement, and internet finance are constantly emerging, leading to a continuous increase in the scale, frequency, and complexity of financial transactions. While business efficiency has significantly improved, the risk of financial fraud has also intensified, with fraudulent activities showing a trend towards greater concealment, organization, and confrontation.

[0003] In early financial environments, fraud often manifested as anomalies in single accounts or transactions, easily detectable through manual review or simple rules. However, in today's complex business environment, fraudsters typically use multiple accounts to conduct diversified transactions over extended periods, deliberately controlling the amount and frequency of individual transactions to make their behavior appear normal in certain aspects, thus evading detection systems. While such fraudulent behavior may not show anomalies at the level of individual transactions, it exhibits distinct characteristics at the level of account relationship structures and long-term behavioral evolution.

[0004] Existing financial fraud detection technologies mainly include the following categories: First, rule-based fraud detection methods. These methods rely on human experience to set fixed thresholds or logical rules to judge transaction amounts, frequency, timing, or counterparties. While simple to implement and highly interpretable, rule updates depend on manual maintenance, making it difficult to adapt to new fraud patterns. Furthermore, an increase in the number of rules leads to a significant rise in false positive rates.

[0005] Second, there are detection methods based on traditional machine learning. These methods improve detection capabilities to some extent by manually constructing features and training classification models. However, their performance is highly dependent on feature engineering, making it difficult to automatically uncover complex relationships between accounts. Furthermore, they typically assume that transaction samples are independent, thus failing to effectively identify organized fraudulent activities.

[0006] Third, methods based on deep learning and generative models. While these methods possess strong feature learning capabilities, in real-world financial transactions, fraud samples are extremely scarce, data distribution is highly imbalanced, and model training is prone to instability and potential representation degradation. Furthermore, most existing methods use a single time scale for modeling, making it difficult to simultaneously capture both short-term anomalies and long-term risk accumulation characteristics.

[0007] Therefore, existing technologies still have significant shortcomings in structural relationship modeling, temporal evolution feature characterization, and training stability. There is an urgent need for a financial fraud detection technology that can comprehensively utilize structural and temporal information and maintain stable performance in imbalanced data scenarios. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention aims to provide a financial fraud detection method and apparatus based on multi-scale spatiotemporal feature learning. To solve these problems, this invention employs the following technical solution: A financial fraud detection method based on multi-scale spatiotemporal feature learning includes the following steps: S1. Acquire financial transaction data and perform basic preprocessing; S2. Perform structural feature modeling on the input features using the encoder; S3. Introduce multi-scale time modeling in the latent space; S4. Construct a training-stabilized latent representation learning model; S5. Generate risk prediction results through decoder and classifier; S6, Risk Assessment Output.

[0009] Optionally, S1 includes: acquiring financial transaction data, performing data cleaning, missing value handling, outlier handling, standardization, and formatting on the financial transaction data to form an original feature matrix. Where N represents the number of samples, d represents the feature dimension, and R is the set of real numbers; and the sample index and time order are organized according to account or transaction object to support subsequent structural modeling and multi-scale time series modeling.

[0010] Optionally, S2 includes: inputting the feature matrix obtained in S1 into the encoder module, which includes a clustering embedding submodule and a structure-aware attention submodule; wherein, the clustering embedding submodule performs principal component analysis on the input features to reduce dimensionality and selects principal components whose cumulative variance contribution rate is not lower than a preset threshold; in the dimensionality-reduced feature space, a clustering algorithm is used to cluster the samples and the optimal number of clusters is determined based on the silhouette coefficient; the clustering labels of the samples are mapped to low-dimensional clustering embedding vectors to represent the semantic structure of the group to which the account belongs; on this basis, the latent representation of the sample and the context representation of its group are weighted and modulated through a structure-aware attention mechanism; The attention weights are calculated as follows:

[0011] in, The attention weight corresponding to the i-th sample is used to characterize the fusion ratio between the current sample's own latent representation and the aggregated representation of its group. Indicates the activation function; These represent the weight parameters used to map the latent representation of the samples; This represents the original latent representation of the i-th sample; Indicates the bias term; Generate a structure-enhanced latent representation:

[0012] This represents the structure-enhanced latent representation obtained after weighted fusion using a structure-aware attention mechanism. h̄ represents the aggregated representation of samples within the same group to which the current sample belongs, used to reflect the overall contextual features of the group; i represents the sample index; In the formula Indicates and The corresponding complementary weights are used to jointly determine the weighted combination relationship between the sample's own features and the group's contextual features.

[0013] in An aggregate representation of samples within the same group; Optionally, S3 includes: dividing the transaction behavior sequence into multiple different time scales and inputting them into multi-scale time modeling units for sequence modeling; In this process, a gated recurrent neural network is used to model the transaction sequence at each time scale to obtain the hidden state representation of the corresponding time scale. The hidden states obtained at different time scales are concatenated to form a multi-scale time representation, which is used to characterize the short-term abnormal fluctuations, medium-term behavioral changes and long-term risk evolution trends of trading behavior. For any given time scale, the GRU update process is as follows:

[0014]

[0015]

[0016]

[0017] in, The reset gate for time step t is used to control the degree of participation of the hidden state from the previous time step in the calculation of the current candidate state. The update gate represents the time step t and is used to control the fusion ratio of historical information and current input information. Indicates the activation function; and These represent the current input. Hidden state from the previous moment Reset door Weight parameters; and These represent the current input. Hidden state from the previous moment Update Gate Weight parameters; The input feature vector represents time step t; This indicates the hidden state at the previous time step; This represents the candidate hidden state at the current time step; Represents the hyperbolic tangent activation function; and These represent the weight parameters that map the current input and the historical state to the candidate hidden state, respectively. This indicates element-wise multiplication; This represents the final hidden state at time step t; t represents the time step index; The final output of the latent variables for time series augmentation is: .

[0018] This represents a joint representation formed by splicing hidden states at different time scales; Represents the mapping function of a multilayer perceptron; This represents the time series augmentation latent variable output after time series modeling by a recurrent neural network.

[0019] Optionally, S4 includes: fusing the structure-enhanced latent representation obtained in step S2 with the multi-scale temporal representation obtained in step S3, and inputting it into the latent representation learning module; the latent representation learning module adopts a variational autoencoder structure that introduces a channel attention mechanism to adaptively recalibrate the fused features; at the same time, a regularization constraint based on Wasserstein distance is introduced in the latent space to replace the traditional KL divergence constraint, thereby improving the training stability of the model under imbalanced sample conditions; The SENet channel attention mechanism is used to adaptively recalibrate latent features, and its computation process includes:

[0020]

[0021]

[0022] in, Indicates input features The channel statistics obtained after global compression; Indicates the number of dimensions or channels of the input features; ∑ represents the summation operation; i represents the feature dimension index; Representing input features The i-th component; This represents the weight vector generated by the channel attention mechanism, used for adaptive recalibration of features in each dimension; Indicates the activation function; and Represents the weight parameters in the channel attention mapping process; Represents the linear rectification activation function; This represents the output features after weighting by the SENet channel attention mechanism; This represents the original latent feature vector input to the channel attention module; This represents element-wise multiplication, used to implement weight vectors. For input features Dimensionally weighted; Meanwhile, Wasserstein-2 represents the second-order Wasserstein distance, which is used as a regularization constraint in the latent space to improve the stability of the model training process.

[0023] Meanwhile, a regularization term based on Wasserstein-2 distance is introduced into the latent space.

[0024] Optionally, S5 includes: inputting the trained and stabilized latent representation into the decoder network to reconstruct the original features; and inputting the latent representation into the classifier network to output the predicted probability that the sample belongs to the fraud category; wherein, the reconstruction loss, Wasserstein regularization loss and classification loss together constitute the joint optimization objective of the model.

[0025] Optionally, S6 includes: generating corresponding financial fraud risk assessment information based on the predicted probability output by the classifier network; comparing the predicted probability with a preset risk threshold to determine the risk level of an account or transaction; and outputting suspicious account identifiers, suspicious transaction identifiers, and corresponding structural and temporal feature explanations for risk warning, transaction interception, manual review, and subsequent business processing.

[0026] Optionally, financial transaction data may include any one or more of the following: account identifier, counterparty, transaction amount, transaction time, transaction type, transaction channel, and geographic information.

[0027] A financial fraud detection device based on multi-scale spatiotemporal feature learning, used to implement the aforementioned financial fraud detection method based on multi-scale spatiotemporal feature learning, includes: The data acquisition and preprocessing module is used to acquire financial transaction data and perform basic preprocessing. The structural feature modeling module is used to model the structural features of transaction data through the encoder and generate a structure-enhanced latent representation. The multi-scale time modeling module is used to model transaction behavior sequences at multiple time scales and generate multi-scale time representations. A stable latent representation learning module is trained to fuse structural and temporal features, and a stable latent representation is generated by introducing a variational autoencoder with Wasserstein regularization. The risk assessment module is used to predict fraud risk based on potential representations and output risk scores, risk levels, or suspicious object identifiers.

[0028] Optionally, the structural feature modeling module includes a clustering embedding unit and a structure-aware attention unit; The multi-scale time modeling module includes sequence modeling units at at least two time scales; The training of the stable latent representation learning module introduces Wasserstein distance constraints. The risk assessment module includes a decoding submodule and a classification submodule.

[0029] The present invention has the following beneficial effects: Enhanced ability to characterize structural relationships: It can systematically characterize the relationships and fund interaction patterns between accounts, improve the ability to identify gang-related, chain-like, and covert fraudulent activities, and reduce the risk of missed detections caused by judging based on a single transaction; More comprehensive expression of multi-scale time features: By modeling multiple time scales, it can simultaneously capture short-term anomalies and long-term risk accumulation processes, thereby improving the ability to identify the evolution trend of fraudulent activities and reducing misjudgments caused by a single time window; Improved training stability and generalization ability: By training a stable latent representation learning mechanism, the training instability and performance fluctuations caused by sample imbalance and scenario changes are mitigated, thereby improving the robustness and deployability of the model in real financial business environments. It has strong engineering applicability: the methods and steps are clear and the modules are well-defined, making it easy to deploy into actual financial anti-fraud systems and providing stable technical support for risk warning, alarm handling and business review. Attached Figure Description

[0030] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating a financial fraud detection method based on multi-scale spatiotemporal feature learning according to the present invention. Figure 2 This is a schematic diagram of the structure of a financial fraud detection device based on multi-scale spatiotemporal feature learning according to the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] In the description of this invention, it should be noted that the terms "vertical," "upper," "lower," and "horizontal," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0034] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or a connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] like Figure 1 As shown, a financial fraud detection method based on multi-scale spatiotemporal feature learning includes the following steps: S1. Acquire financial transaction data and perform basic preprocessing; S2. Perform structural feature modeling on the input features using the encoder; S3. Introduce multi-scale time modeling in the latent space; S4. Construct a training-stabilized latent representation learning model; S5. Generate risk prediction results through decoder and classifier; S6, Risk Assessment Output.

[0036] In a preferred embodiment of the present invention, S1 includes: acquiring financial transaction data, performing data cleaning, missing value processing, outlier processing, standardization processing, and formatting processing on the financial transaction data to form an original feature matrix. Where N represents the number of samples, d represents the feature dimension, and R is the set of real numbers; and the sample index and time order are organized according to account or transaction object to support subsequent structural modeling and multi-scale time series modeling.

[0037] In a preferred embodiment of the present invention, S2 includes: inputting the feature matrix obtained in S1 into an encoder module, the encoder module including a clustering embedding submodule and a structure-aware attention submodule; wherein, the clustering embedding submodule performs principal component analysis on the input features to reduce dimensionality, and selects principal components whose cumulative variance contribution rate is not lower than a preset threshold; in the dimensionality-reduced feature space, a clustering algorithm (such as the K-means algorithm) is used to cluster the samples, and the optimal number of clusters is determined according to the silhouette coefficient; the clustering labels of the samples are mapped to low-dimensional clustering embedding vectors to represent the semantic structure of the group to which the account belongs; on this basis, the latent representation of the sample and the context representation of its group are weighted and modulated through a structure-aware attention mechanism; The attention weights are calculated as follows:

[0038] Generate a structure-enhanced latent representation:

[0039] in This represents the aggregate representation of samples within the same group.

[0040] This step enables the model to explicitly focus on group-level structural anomalies during the encoding phase.

[0041] In a preferred embodiment of the present invention, S3 includes: dividing the transaction behavior sequence according to multiple different time scales, and inputting them into a multi-scale time modeling unit for sequence modeling; In this process, a gated recurrent neural network is used to model the transaction sequence at each time scale to obtain the hidden state representation of the corresponding time scale. The hidden states obtained at different time scales are concatenated to form a multi-scale time representation, which is used to characterize the short-term abnormal fluctuations, medium-term behavioral changes and long-term risk evolution trends of trading behavior. For any given time scale, the GRU update process is as follows:

[0042]

[0043]

[0044]

[0045] The final output of the latent variables for time series augmentation is: .

[0046] This enhanced latent representation effectively enriches the latent space and time dependencies of the variational autoencoder, overcoming the limitation of the standard VAE encoder that processes input features independently without considering their sequential properties.

[0047] In a preferred embodiment of the present invention, S4 includes: fusing the structure-enhanced latent representation obtained in step S2 with the multi-scale temporal representation obtained in step S3, and inputting it into the latent representation learning module; the latent representation learning module adopts a variational autoencoder structure that introduces a channel attention mechanism to adaptively recalibrate the fused features; at the same time, a regularization constraint based on Wasserstein distance is introduced into the latent space to replace the traditional KL divergence constraint, thereby improving the training stability of the model under imbalanced sample conditions; The SENet channel attention mechanism is used to adaptively recalibrate latent features, and its computation process includes:

[0048]

[0049]

[0050] Meanwhile, a regularization term based on Wasserstein-2 distance is introduced into the latent space.

[0051] This replaces the traditional KL divergence constraint, improving training stability under imbalanced sample conditions.

[0052] In a preferred embodiment of the present invention, S5 includes: inputting the trained and stabilized latent representation into the decoder network to reconstruct the original features; and inputting the latent representation into the classifier network to output the predicted probability that the sample belongs to the fraud category; wherein, the reconstruction loss, the Wasserstein regularization loss and the classification loss together constitute the joint optimization objective of the model.

[0053] Used to simultaneously guarantee the discriminability and stability of the latent representation.

[0054] In a preferred embodiment of the present invention, S6 includes: generating corresponding financial fraud risk assessment information based on the predicted probability output by the classifier network; comparing the predicted probability with a preset risk threshold to determine the risk level of an account or transaction; and outputting suspicious account identifiers, suspicious transaction identifiers, and corresponding structural and temporal feature explanations for risk warning, transaction interception, manual review, and subsequent business processing.

[0055] In a preferred embodiment of the present invention, the financial transaction data includes any one or more of the following: account identifier, counterparty, transaction amount, transaction time, transaction type, transaction channel, and geographical information.

[0056] like Figure 2 As shown, a financial fraud detection device based on multi-scale spatiotemporal feature learning is used to implement the aforementioned financial fraud detection method based on multi-scale spatiotemporal feature learning, comprising: The data acquisition and preprocessing module is used to acquire financial transaction data and perform basic preprocessing. The structural feature modeling module is used to model the structural features of transaction data through the encoder and generate a structure-enhanced latent representation. The multi-scale time modeling module is used to model transaction behavior sequences at multiple time scales and generate multi-scale time representations. A stable latent representation learning module is trained to fuse structural and temporal features, and a stable latent representation is generated by introducing a variational autoencoder with Wasserstein regularization. The risk assessment module is used to predict fraud risk based on potential representations and output risk scores, risk levels, or suspicious object identifiers.

[0057] In a preferred embodiment of the present invention, the structural feature modeling module includes a clustering embedding unit and a structure-aware attention unit; The multi-scale time modeling module includes sequence modeling units at at least two time scales; The training of the stable latent representation learning module introduces Wasserstein distance constraints. The risk assessment module includes a decoding submodule and a classification submodule.

[0058] Example 1: Specific implementation of the financial fraud detection method: In this embodiment, the financial fraud detection method provided by the present invention is applied to actual financial business scenarios, such as, but not limited to, bank transfer business, payment and clearing systems, internet finance platforms, third-party payment systems, or cross-border fund transfer scenarios.

[0059] First, financial transaction data is acquired. This data can originate from a financial institution's transaction system, accounting system, or risk management system. The financial transaction data includes at least the following fields: account identifier, counterparty account, transaction amount, transaction time, transaction type, transaction channel, and transaction region. Next, preprocessing operations are performed on the acquired financial transaction data. These preprocessing operations include, but are not limited to: identifying and correcting or removing outliers; filling in or marking missing data; normalizing or standardizing numerical features of different dimensions; and uniformly formatting and aligning the time field. Through these preprocessing operations, a standardized transaction dataset with a unified structure is formed, which can be directly used for subsequent modeling. A sample set indexed by account or transaction object is constructed, and the samples are organized sequentially according to the transaction occurrence time, providing a data foundation for subsequent structural modeling and multi-scale time modeling.

[0060] After data preprocessing, structural feature modeling is performed on the standardized transaction data. Specifically, the preprocessed feature data is input into the encoder module for processing. The encoder module is used to characterize the relationship structure between accounts, and it can include a clustering embedding submodule and a structure-aware attention submodule. The clustering embedding submodule performs dimensionality reduction and clustering analysis on the input features, mapping account samples with similar transaction behavior patterns to the same group, thereby obtaining a low-dimensional embedding representation that reflects the semantic structure of the account group. Based on this, the structure-aware attention submodule weights and modulates the latent representation of the sample with the contextual representation of its group, enabling the model to explicitly focus on structural differences and anomalous features within and between account groups during the encoding stage. In this way, the model can not only learn the behavioral characteristics of individual accounts, but also capture group-level risk signals hidden in the account relationship structure.

[0061] Considering that financial fraud exhibits different dynamic characteristics across different time spans, this embodiment models transaction behavior using a multi-scale approach. Specifically, transaction records for the same account or trading entity are divided into multiple time scales according to a preset time window, and transaction behavior sequences are constructed for each time scale. Subsequently, the transaction sequences at each time scale are input into a sequence modeling unit for learning to extract multi-scale time features. These multi-scale time features are used to characterize abnormal fluctuations in transaction behavior in the short term, mid-term phased behavioral changes, and long-term risk accumulation and evolution trends. By introducing multi-scale time modeling, the model can simultaneously focus on instantaneous abnormal behavior and long-term potential risk patterns, thereby improving its ability to identify complex fraudulent activities.

[0062] After obtaining structural and multi-scale temporal features, these features are fused and enhanced. Through feature concatenation, mapping, or modulation, structural and temporal information are comprehensively expressed to form a unified feature representation. Based on this unified feature representation, a training-stabilized latent representation learning mechanism is introduced to further model the fused features. This latent representation learning mechanism is used to learn stable and discriminative latent risk representations to alleviate model training instability caused by factors such as scarce fraud samples, imbalanced sample distribution, changes in business scenarios, or concept drift. By introducing a training stabilization mechanism, model fluctuations during training can be effectively reduced, improving the robustness and generalization ability of the latent representation, enabling the model to maintain stable risk identification performance across different business scenarios.

[0063] After completing the latent representation learning, the obtained stable latent representation is input into the risk prediction module. The original features are reconstructed through the decoder network, and the latent representation is used by the classifier network to determine fraud risk, outputting the predicted probability that the sample belongs to the fraud category.

[0064] Finally, based on the output of the financial fraud risk model, corresponding fraud risk assessment information is generated. This risk assessment information includes, but is not limited to, risk scores, risk levels, suspicious account identifiers, or suspicious transaction identifiers. By comparing the prediction results with preset risk thresholds, risk classification of transactions or accounts can be achieved, supporting risk warnings, transaction interception, manual review, or subsequent business processing.

[0065] Through the above steps, this embodiment can effectively identify highly concealed and structurally complex financial fraud activities in complex financial transaction environments.

[0066] Example 2: Specific Implementation of a Financial Fraud Detection Device: The present invention also provides a financial fraud detection device for implementing the above-described financial fraud detection method. The financial fraud detection device includes a data acquisition and preprocessing module, a structural feature modeling module, a multi-scale temporal modeling module, a training and stabilizing latent representation learning module, and a risk assessment module.

[0067] The data acquisition and preprocessing module is used to acquire financial transaction data from the financial business system and perform cleaning, standardization, and formatting on the transaction data; the structural feature modeling module is used to construct the relationship between accounts based on the standardized transaction data and extract structural features; the multi-scale time modeling module is used to perform multi-time-scale analysis on transaction behavior and extract multi-scale time features; the training and stabilizing latent representation learning module is used to construct a stable risk representation based on a unified feature representation and generate a financial fraud risk model; and the risk assessment module is used to generate fraud risk assessment information based on the model output results.

[0068] The modules described above can be deployed on the same computing device or distributed across multiple computing nodes. The modules communicate and work together through interfaces, and their specific implementation does not constitute a limitation on the present invention.

[0069] The components, modules, mechanisms, and devices in this invention that are not described in detail are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A financial fraud detection method based on multi-scale spatiotemporal feature learning, characterized in that, Includes the following steps: S1. Acquire financial transaction data and perform basic preprocessing; S2. Perform structural feature modeling on the input features using the encoder; S3. Introduce multi-scale time modeling in the latent space; S4. Construct a training-stabilized latent representation learning model; S5. Generate risk prediction results through decoder and classifier; S6, Risk Assessment Output.

2. The financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 1, characterized in that, S1 include: Acquire financial transaction data, and perform data cleaning, missing value handling, outlier handling, standardization, and formatting on the financial transaction data to form the original feature matrix. Where N represents the number of samples, d represents the feature dimension, and R is the set of real numbers; and the sample index and time order are organized according to account or transaction object to support subsequent structural modeling and multi-scale time series modeling.

3. The financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 2, characterized in that, S2 include: The feature matrix obtained from S1 is input into the encoder module, which includes a clustering embedding submodule and a structure-aware attention submodule. The clustering embedding submodule performs principal component analysis on the input features to reduce dimensionality and selects principal components whose cumulative variance contribution rate is not lower than a preset threshold. In the dimensionality-reduced feature space, a clustering algorithm is used to cluster the samples, and the optimal number of clusters is determined based on the silhouette coefficient. The clustering labels of the samples are mapped to low-dimensional clustering embedding vectors to represent the semantic structure of the group to which the account belongs. Based on this, a structure-aware attention mechanism is used to weight and modulate the latent representation of the sample with the contextual representation of its group. The attention weights are calculated as follows: Generate a structure-enhanced latent representation: in This represents the aggregate representation of samples within the same group.

4. The financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 3, characterized in that, S3 include: The transaction behavior sequence is divided into multiple different time scales and then input into a multi-scale time modeling unit for sequence modeling. In this process, a gated recurrent neural network is used to model the transaction sequence at each time scale to obtain the hidden state representation of the corresponding time scale. The hidden states obtained at different time scales are concatenated to form a multi-scale time representation, which is used to characterize the short-term abnormal fluctuations, medium-term behavioral changes and long-term risk evolution trends of trading behavior. For any given time scale, the GRU update process is as follows: The final output of the latent variables for time series augmentation is: 。 5. A financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 4, characterized in that, S4 include: The structure-enhanced latent representation obtained in step S2 is fused with the multi-scale temporal representation obtained in step S3 and then input into the latent representation learning module. The latent representation learning module adopts a variational autoencoder structure that introduces a channel attention mechanism to adaptively recalibrate the fused features. At the same time, a regularization constraint based on Wasserstein distance is introduced into the latent space to replace the traditional KL divergence constraint, thereby improving the training stability of the model under imbalanced sample conditions. The SENet channel attention mechanism is used to adaptively recalibrate latent features, and its computation process includes: Meanwhile, a regularization term based on Wasserstein-2 distance is introduced into the latent space.

6. A financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 5, characterized in that, S5 include: The stabilized latent representations trained are input into the decoder network to reconstruct the original features; The latent representation is input into the classifier network, and the output is the predicted probability that the sample belongs to the fraud category; where the reconstruction loss, Wasserstein regularization loss and classification loss together constitute the joint optimization objective of the model.

7. A financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 6, characterized in that, S6 include: The corresponding financial fraud risk assessment information is generated based on the predicted probability output by the classifier network; the predicted probability is compared with the preset risk threshold to determine the risk level of the account or transaction. It outputs suspicious account identifiers, suspicious transaction identifiers, and corresponding structural and temporal feature explanations for risk warning, transaction interception, manual review, and subsequent business processing.

8. The financial fraud detection method based on multi-scale spatiotemporal feature learning according to claim 7, characterized in that, Financial transaction data includes any one or more of the following: account identifier, counterparty, transaction amount, transaction time, transaction type, transaction channel, and geographical information.

9. A financial fraud detection device based on multi-scale spatiotemporal feature learning, characterized in that, A financial fraud detection method based on multi-scale spatiotemporal feature learning as described in any one of claims 1-8 includes: The data acquisition and preprocessing module is used to acquire financial transaction data and perform basic preprocessing. The structural feature modeling module is used to model the structural features of transaction data through the encoder and generate a structure-enhanced latent representation. The multi-scale time modeling module is used to model transaction behavior sequences at multiple time scales and generate multi-scale time representations. A stable latent representation learning module is trained to fuse structural and temporal features, and a stable latent representation is generated by introducing a variational autoencoder with Wasserstein regularization. The risk assessment module is used to predict fraud risk based on potential representations and output risk scores, risk levels, or suspicious object identifiers.

10. A financial fraud detection device based on multi-scale spatiotemporal feature learning according to claim 9, characterized in that, The structural feature modeling module includes a clustering embedding unit and a structure-aware attention unit; The multi-scale time modeling module includes sequence modeling units at at least two time scales; The training of the stable latent representation learning module introduces Wasserstein distance constraints. The risk assessment module includes a decoding submodule and a classification submodule.

Images