Acquisition risk intelligent identification system based on deep learning

Abstract

The invention discloses an intelligent acquisition risk identification system based on deep learning, and the system comprises a time sequence sample construction module which is used for building a time sequence sample sequence; the hierarchical attention structural feature coding module is used for carrying out structural feature coding on the sequential sample sequence by utilizing a hierarchical attention network; the time sequence feature extraction module is used for inputting the structural feature vector into an ETSform model to perform time sequence feature extraction; the hierarchical time sequence collaborative attention adaptive fusion module is used for performing bidirectional attention interaction and dynamically generating a hierarchical weight and a time sequence weight through a meta-learning controller; the improved CatBoost risk identification module is used for outputting a risk score and a risk type label; and the system fusion module is used for summarizing and merging the risk scores and the risk type labels. The method and the device are suitable for merchant transaction risk identification in an acquiring business scene.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and artificial intelligence, and in particular to a deep learning-based intelligent identification system for acquiring risks. Background Technology

[0002] In modern electronic payment systems, acquiring institutions, as crucial nodes in payment services, bear the responsibility of receiving, processing, and managing transaction data and risks. With the rapid popularization of non-cash payment methods, the number of transactions and the scale of data have increased dramatically, placing higher demands on the real-time and accuracy of risk identification in transaction behavior. Traditional acquiring risk identification methods mostly rely on rule-based systems or shallow machine learning models, such as decision trees and logistic regression. While these methods offer some interpretability and computational efficiency, they are insufficient when processing high-dimensional, non-linear, and time-series dependent transaction data, often resulting in low identification accuracy and poor generalization ability, making them ill-suited to complex and ever-changing fraud scenarios.

[0003] In recent years, deep learning technology has gained widespread attention in the field of financial risk control, especially the introduction of recurrent neural networks (RNNs), attention mechanisms, and time series modeling methods, which have provided new ideas for the automatic identification of acquiring risks. However, existing models often focus on a certain feature dimension or time characteristic, lacking unified modeling of multi-level structural information (such as merchants, accounts, terminals, etc.) and failing to fully consider the evolution of transaction behavior over time. Some models use fixed weights or static fusion strategies in the feature fusion process, which cannot dynamically adapt to the diverse risk characteristics in the actual trading environment, resulting in limited ability to capture potential risk patterns.

[0004] While traditional ensemble learning methods have certain advantages in processing structured data, they fail to integrate time-series features and hierarchical structural information. During model training, they neglect temporal dependencies and feature hierarchy differences between samples, making it difficult to accurately reflect the evolution of merchant behavior and its risk correlations. Furthermore, the lack of adaptive weight adjustment mechanisms and model update strategies makes it difficult for the system to continuously optimize its risk identification performance in dynamic scenarios.

[0005] Therefore, how to provide a deep learning-based intelligent identification system for acquiring risks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based intelligent risk identification system for acquiring transactions. This invention fully utilizes algorithms such as hierarchical attention mechanisms, time series modeling, meta-learning controllers, and improved CatBoost ensemble models. It details the entire process of extracting structural and temporal features from multi-source transaction data, performing bidirectional attention interaction, generating fused feature vectors, and identifying risks. This system possesses advantages such as high risk identification accuracy, strong feature modeling capabilities, good model adaptability, and a robust dynamic update mechanism, effectively addressing complex and ever-changing acquiring risk scenarios.

[0007] According to an embodiment of the present invention, a deep learning-based intelligent identification system for acquiring risks includes: The time-series sample construction module is used to build time-series sample sequences based on merchant identifiers, transaction times, and transaction amounts. The hierarchical attention structure feature encoding module is used to encode the structure features of temporally sequenced samples using a hierarchical attention network. The time series feature extraction module is used to input structural feature vectors into the ETSformer model to extract time series features from transaction data. The hierarchical temporal collaborative attention adaptive fusion module is used to perform bidirectional attention interaction between structural feature vectors and temporal feature vectors, and dynamically generates hierarchical weights and temporal weights through a meta-learning controller; An improved CatBoost risk identification module is used to output risk scores and risk type labels; The system fusion module is used to summarize and merge risk scores and risk type labels to form merchant-level result records.

[0008] Optionally, modules can be integrated using the following methods: Step 1: Obtain the merchant's transaction data, account information, and terminal device information, and establish a time-series sample sequence based on the merchant's identifier, transaction time, and transaction amount; Step 2: Use a hierarchical attention network to encode the structural features of the temporal sample sequence, and extract and weight features for the merchant layer, terminal layer and account layer respectively to obtain structural feature vectors; Step 3: Input the structural feature vector into the ETSformer model to extract time series features from the transaction data and obtain the time feature vector; Step 4: In the hierarchical temporal collaborative attention adaptive fusion mechanism, the structural feature vector and the temporal feature vector are subjected to bidirectional attention interaction. The hierarchical weights and temporal weights are dynamically generated through the meta-learning controller, and the fusion ratio is adaptively adjusted to obtain the fused feature vector. Step 5: Input the fused feature vector into the improved CatBoost model. The temporally correlated ordered tree update module performs temporal correction by constructing a temporally correlated weight matrix; the feature hierarchical structure-aware tree splitting module performs splitting gain correction by constructing a hierarchical feature set and based on the hierarchical regularization constraint function. Step Six: The improved CatBoost model outputs risk scores and risk type labels to identify and classify the risks of merchant transaction behavior. The results are then returned to the system fusion module to update the hierarchical weights, time-series weights, and model parameters.

[0009] Optionally, step one includes the following specific steps: Obtain the original records related to merchant transactions; Extract the following fields from the original records: merchant identifier, account identifier, terminal device identifier, transaction time, transaction amount, transaction currency, transaction channel, merchant category code, geographic area code, and transaction serial number; Establish a standardized transaction record by using the transaction serial number as the primary key to link transaction records with merchant master data, account information, and terminal device information. Standardized transaction records are grouped by merchant identifiers. Within each merchant group, a timeline is constructed based on the transaction times arranged in ascending order. The timeline is then sliced ​​to generate a continuous sequence of time windows. For each window in the time window sequence, collect the ordered transaction records within the start and end time range of the current window to form a window-level sample fragment; Each window-level sample fragment corresponds one-to-one with a merchant identifier, generating a time-series sample sequence.

[0010] Optionally, step two includes the following specific steps: Using the temporally sequenced sample sequence as input to the hierarchical attention network, the set of time windows to be processed is determined; The Bi-GRU algorithm is used to encode the merchant layer, terminal layer and account layer at the feature coding layer to generate corresponding intra-layer coding representations; Attention weights are calculated for the merchant layer, terminal layer and account layer in the attention weighting layer, and the in-layer encoding representations of each layer are weighted and aggregated to obtain the weighted representations of the merchant layer, the terminal layer and the account layer. The merchant-level weighted representation, terminal-level weighted representation, and account-level weighted representation are hierarchically aggregated to generate a structural feature vector corresponding to the current time window; The structural feature vectors of each time window in the time window set are arranged in ascending order of transaction time to form a sequence of structural feature vectors that correspond one-to-one with the time-series sample sequence.

[0011] Optionally, step three includes the following specific steps: The structural feature vector sequence is input into the ETSformer model, and time series decomposition is performed within each time window to obtain the channel representations of trend components, seasonal components and residual components. Channel modeling is performed on the trend component, seasonal component, and residual component respectively. Exponential smoothing attention operation is applied to the trend component channel, frequency attention operation is applied to the seasonal component channel, and original fluctuation preservation operation is applied to the residual component channel to obtain the time series representation of each channel. The time series representations of the trend component channel, seasonal component channel, and residual component channel are fused at the channel level and aggregated at the time step level to generate a time feature vector corresponding to the current time window. For each time window in the time window set, the process involves repeated decomposition, modeling, fusion, and aggregation to obtain the time feature vector sequence corresponding to each time window.

[0012] Optionally, the hierarchical temporal collaborative attention adaptive fusion mechanism involves bidirectional attention interaction between structural feature vectors and temporal feature vectors, including the following specific steps: Arrange structural feature vectors and time feature vectors with the same time index according to the transaction time order to form corresponding pairs, and then aggregate them to form a set of time window pairs; Within each time window pair, the structural feature vector and the temporal feature vector are input into a hierarchical temporal collaborative attention adaptive fusion mechanism; Using structural feature vectors as query vectors and temporal feature vectors as key and value vectors, we calculate the attention weights from structure to time and generate a weighted representation. Using the time feature vector as the query vector and the structural feature vector as the key and value vectors, the attention weights from time to structure are calculated and a weighted representation is generated. The weighted representations from structure to time and from time to structure are concatenated and normalized to output the bidirectional attention interaction result.

[0013] Optionally, the step of dynamically generating hierarchical weights and temporal weights through a meta-learning controller and adaptively adjusting the fusion ratio to obtain a fused feature vector includes the following specific steps: In the meta-learning controller, a feedforward neural network algorithm is used to sequentially perform weighted summation, linear mapping, nonlinear activation, and output normalization operations on the bidirectional attention interaction results to generate hierarchical weights. In the inner update phase of the MAML transfer learning framework, the bidirectional attention interaction results of each time window pair are input into the gating network. The task loss of the current time window pair is calculated through forward propagation, and gradient descent is performed to update the parameters of the gating network to obtain the updated inner parameters. During the gradient update process, a gating signal is introduced to proportionally weight the gradient terms of the gated network parameters; The gating signal is calculated from the bidirectional attention interaction result using a nonlinear function; In the outer layer update phase of the MAML transfer learning framework, the updated inner layer parameters are summarized, the overall meta-loss is calculated based on the task loss of each time window, and the outer layer gradient update is performed on the meta-parameters of the gated network. The updated meta-parameters are mapped to a weighted coefficient vector for the time channels, and the output is used as the time series weights. Based on the hierarchical weights and temporal weights output by the meta-learning controller, weighted operations are performed on the structural feature vector and the temporal feature vector respectively; In the weighted calculation process, the weighting ratio of the corresponding features of the merchant layer, terminal layer and account layer in the structural feature vector is controlled by the hierarchical weight, and the weighting ratio of the corresponding features of the trend component, seasonal component and residual component in the time feature vector is controlled by the time series weight. After normalizing the weighted results, feature concatenation and linear combination are performed to generate a fused feature vector.

[0014] Optionally, the time-related ordered tree update module performs time-series correction by constructing a time-related weight matrix, including the following specific steps: The fused feature vector sequence is used as input to the improved CatBoost model, and the training sequence is constructed in ascending order of transaction time. An ordered prediction path without future sample information is established based on the time index of the training sequence. In the temporal correlation ordered tree update module, a temporal correlation weight matrix is ​​generated based on the temporal order of samples in the training sequence. During the ordered boosting iteration of CatBoost, the deviation between the current CatBoost model prediction and the true label is calculated as the sample residual for each time sample, and the corresponding second-order gradient term is calculated. The sample residuals and second-order gradient terms are weighted using a time-correlation weight matrix to obtain time-corrected weighted residuals and weighted second-order terms. The leaf node values ​​of the current base learner are then updated using the weighted residuals and weighted second-order terms. The training sequence is traversed using an ordered prediction path that does not contain information about future samples. The leaf node values ​​of the current base learner are updated in chronological order. At each time step, the leaf node values ​​are calculated using the weighted residual and the weighted second-order quantity. Optionally, the feature hierarchy perceptual tree splitting module constructs a hierarchical feature set and performs splitting gain correction based on a hierarchical regularization constraint function, including the following specific steps: In the feature hierarchy perceptual tree splitting module, the features from the merchant layer, account layer and terminal layer in the fused feature vector are divided into hierarchical feature sets; Perform candidate split point search: During the candidate split point search process, a candidate split threshold is generated for each level feature, and the level feature set is divided based on the split threshold to form candidate split sample pairs; Execute hierarchical order constraints: Perform hierarchical order determination operation on each candidate feature, determine the relationship between the hierarchical identifier of the candidate feature and the splitting hierarchical state of the current node, and call the hierarchical regular constraint function to calculate the constraint coefficient; The split gain function value of the candidate feature is adjusted according to the constraint coefficient. When the hierarchical identifier of the candidate feature is lower than the current node split hierarchical state, a penalty is applied to the candidate split gain. When the hierarchical identifier of the candidate feature is higher than or equal to the current node split hierarchical state, the split gain calculation result is retained. Perform gain evaluation: After hierarchical order constraint processing, evaluate and sort the split gain function values ​​of all candidate features, select the optimal split feature and split threshold according to the adjusted split gain calculation results to generate decision nodes, and update the leaf node values ​​of the current base learner. Based on the determined optimal splitting features and splitting threshold, the splitting structure of the current decision tree is generated. After splitting, candidate splitting point search, hierarchical order constraint and gain evaluation are performed on the child nodes until the adjusted splitting gain function value of all candidate splitting features is lower than the minimum gain threshold. The growth of the current iteration tree and the update of leaf node values ​​are completed under the alternating action of the time-related ordered tree update module and the feature hierarchical structure-aware tree splitting module, and the current iteration tree is added to the integration of the improved CatBoost model. Repeatedly iterate the training sequence in an orderly manner until the preset number of iterations is reached to obtain the ensemble output of the improved CatBoost model.

[0015] Optionally, step six includes the following specific steps: The integrated output is compared with a preset set of risk thresholds to generate a risk score and risk type label corresponding to each time window; the risk score and risk type label are organized according to merchant identifier and time index to form a time window level result set. Merge the time window-level result sets according to the merchant identifier to generate merchant-level result records; The merchant-level results are returned to the system fusion module, which then registers the results and triggers the update process. In the system fusion module, update flags for hierarchical weights and temporal weights are set based on the returned results, and update instructions are sent to the meta-learning controller and the improved CatBoost model. The hierarchical weight update process is executed by the meta-learning controller, and the model parameter update process is executed by the improved CatBoost model. After the update is completed, the update status will be written to the status record table of the system fusion module.

[0016] The beneficial effects of this invention are: This invention constructs a hierarchical attention network to achieve refined modeling of multi-dimensional transaction data at the merchant, terminal, and account levels. This effectively mines the structural features of different actors within the transaction sequence, enhancing the model's ability to identify potential risky behaviors. This hierarchical extraction mechanism of structural features not only strengthens the expressive power of the representation but also preserves the hierarchical semantic information of the data, avoiding semantic loss during information fusion.

[0017] By utilizing the ETSformer model to model trading behavior over time series, the model captures long-term trends, cyclical fluctuations, and local anomalies in three channels: trend, seasonality, and residual components, significantly enhancing its ability to represent complex time-series signals. Furthermore, a hierarchical temporal collaborative attention mechanism enables information complementarity and dynamic fusion between structural and temporal features, ensuring the model's stability and generalization performance under high-dimensional inputs.

[0018] This invention introduces a meta-learning controller that dynamically generates hierarchical and temporal weights during the fusion phase. This controller can adaptively adjust the feature fusion ratio according to different merchant transaction scenarios, enhancing the model's robustness to sample heterogeneity. Simultaneously, it utilizes the MAML transfer learning framework to implement gated meta-learning computation, enabling the system to rapidly adapt to changes in task distribution and effectively improving the efficiency of model transfer and retraining.

[0019] In the risk prediction stage, an improved CatBoost model combining time correlation and feature hierarchy perception mechanism was designed. The temporal correction of sample gradients is realized through the time correlation ordered tree update module, and the hierarchical regularization constraint is introduced through the feature hierarchy structure perception tree splitting module. This effectively enhances the model's adaptability to time logic and business structure, and improves the accuracy and stability of risk scoring.

[0020] The intelligent acquisition risk identification system proposed in this invention has achieved significant improvements in feature modeling, temporal representation, risk identification, and model adaptability, and can identify risky behaviors in acquisition transactions more accurately and efficiently. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0022] Figure 1 This is an overall flowchart of a deep learning-based intelligent identification system for acquiring risks proposed in this invention. Figure 2 This is an overall flowchart of a deep learning-based intelligent identification method for acquiring risks proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figure 1-2 A deep learning-based intelligent identification system for acquiring risks includes: The time-series sample construction module is used to build time-series sample sequences based on merchant identifiers, transaction times, and transaction amounts. The hierarchical attention structure feature encoding module is used to encode the structure features of temporally sequenced samples using a hierarchical attention network. The time series feature extraction module is used to input structural feature vectors into the ETSformer model to extract time series features from transaction data. The hierarchical temporal collaborative attention adaptive fusion module is used to perform bidirectional attention interaction between structural feature vectors and temporal feature vectors, and dynamically generates hierarchical weights and temporal weights through a meta-learning controller; An improved CatBoost risk identification module is used to output risk scores and risk type labels; The system fusion module is used to summarize and merge risk scores and risk type labels to form merchant-level result records.

[0025] This invention provides a deep learning-based intelligent risk identification system for acquiring merchants, comprising, in sequence, a time-series sample construction module, a hierarchical attention structure feature encoding module, a time-series feature extraction module, a hierarchical time-series collaborative attention adaptive fusion module, an improved CatBoost risk identification module, and a system fusion module. This system constructs a time-series sample sequence, fuses structural and temporal features, utilizes a meta-learning mechanism to achieve adaptive feature weighting, and performs risk scoring and type identification based on an improved CatBoost model. Finally, it completes merchant-level aggregation and output of risk results, possessing advantages such as high identification accuracy, strong adaptability, and high intelligence.

[0026] In this embodiment, the modules are interconnected using the following method: Step 1: Obtain the merchant's transaction data, account information, and terminal device information, and establish a time-series sample sequence based on the merchant's identifier, transaction time, and transaction amount; Step 2: Use a hierarchical attention network to encode the structural features of the temporal sample sequence, and extract and weight features for the merchant layer, terminal layer and account layer respectively to obtain structural feature vectors; Step 3: Input the structural feature vector into the ETSformer model to extract time series features from the transaction data and obtain the time feature vector; Step 4: In the hierarchical temporal collaborative attention adaptive fusion mechanism, the structural feature vector and the temporal feature vector are subjected to bidirectional attention interaction, and the hierarchical weights and temporal weights are dynamically generated through the meta-learning controller. The fusion ratio is adaptively adjusted to obtain the fused feature vector. Step 5: Input the fused feature vector into the improved CatBoost model. During the gradient update process, the temporal correlation ordered tree update module constructs a temporal correlation weight matrix based on the temporal order between transaction samples and performs temporal correction on the sample gradient. The feature hierarchy structure-aware tree splitting module sets hierarchical regularization constraints based on the hierarchical semantics of the features during feature splitting, so that the splitting path of the tree follows the hierarchical structure of merchant layer, account layer and terminal layer. Step Six: The improved CatBoost model outputs risk scores and risk type labels to identify and classify the risks associated with merchant transaction behavior; Step 7: Return the recognition results to the system fusion module and update the hierarchical weights, temporal weights, and model parameters.

[0027] This invention provides a deep learning-based intelligent identification method for acquiring risks. It constructs a time-series sample sequence based on multi-source data, integrates a hierarchical attention mechanism and time series modeling methods, and extracts multi-layered structural features such as merchants, terminals, and accounts, as well as temporal features such as trends, seasonality, and residuals. Through a hierarchical time-series collaborative attention mechanism and a meta-learning controller, it dynamically generates hierarchical weights and time-series weights, achieving adaptive fusion of structural and temporal features. By combining an improved CatBoost model with a time-related weight matrix and feature hierarchy regularization constraints, it enhances the model's ability to perceive time sequence and hierarchical semantics. Finally, it outputs a risk score and risk type label for transaction behavior, and through a system fusion module, it achieves feedback updates and iterative optimization, effectively enhancing the accuracy and intelligence of acquiring risk identification.

[0028] In this embodiment, step one includes the following specific steps: It accesses four types of data sources: payment transaction records, merchant master data, account information, and terminal device information, to obtain original records related to merchant transactions; Extract the following fields from the original records: merchant identifier, account identifier, terminal device identifier, transaction time, transaction amount, transaction currency, transaction channel, merchant category code, geographic area code, and transaction serial number; Establish a mapping between transaction records and merchant master data, account information, and terminal device information using the transaction serial number as the primary key, forming a standardized transaction record with associated keys; The transaction times in standardized transaction records are formatted uniformly and sorted in ascending order of transaction time; The merchants are grouped by their identifiers. Within each merchant group, a timeline is constructed based on the transaction times arranged in ascending order. The time granularity, window length, and sliding step size are set, and the timeline is sliced ​​to generate a continuous sequence of time windows. Within each time window, orderly transaction records between the start and end times of the window are collected, and the merchant identifier, the start and end times of the window, the list of transaction records within the window, the set of associated account identifiers, and the set of associated terminal device identifiers are retained to form a window-level sample fragment. The sample segments corresponding to each time window are concatenated in chronological order to generate a time-series sample sequence that corresponds one-to-one with the merchant identifier, and a sequence number and sequence index are assigned to each time-series sample sequence.

[0029] In the data preprocessing stage, this invention extracts key fields and establishes standardized associated transaction records based on transaction serial numbers by accessing multi-source data such as payment transaction records, merchant master data, account information, and terminal device information. These records are then grouped according to merchant identifiers and a timeline is constructed based on transaction time. Time slicing is performed by setting time granularity, window length, and sliding step size to generate a continuous time window sequence. Within each time window, ordered transaction records within the corresponding time range are summarized to form a window-level sample fragment containing merchant identifiers, window boundary times, a list of transaction records, and a set of associated account and terminal identifiers. Finally, these fragments are concatenated in chronological order to form a time-series sample sequence that corresponds one-to-one with the merchant identifier, laying the foundation for subsequent structural feature encoding and time series modeling.

[0030] In this embodiment, step two includes the following specific steps: Using the temporally sequenced sample sequence as input to the hierarchical attention network, the set of time windows to be processed is determined; Within each time window, a hierarchical sample unit is constructed, which includes a set of serialized records for the merchant layer, terminal layer, and account layer. The Bi-GRU algorithm is used to encode the merchant layer, terminal layer and account layer at the feature coding layer to generate corresponding intra-layer coding representations; Attention weights are calculated for the merchant layer, terminal layer and account layer in the attention weighting layer, and the in-layer encoding representations of each layer are weighted and aggregated to obtain the weighted representations of the merchant layer, the terminal layer and the account layer. The merchant-level weighted representation, terminal-level weighted representation, and account-level weighted representation are hierarchically aggregated to generate a structural feature vector corresponding to the current time window; The structural feature vectors of each time window in the time window set are arranged in ascending order of transaction time to form a sequence of structural feature vectors that correspond one-to-one with the time-series sample sequence.

[0031] In the structural feature extraction stage, this invention employs a hierarchical attention network to structurally process temporal sample sequences. First, a set of time windows to be processed is determined, and within each window, hierarchical sample units comprising merchant, terminal, and account layers are constructed. In the feature encoding layer, the serialized record sets of each layer are encoded using the Bi-GRU algorithm to obtain intra-layer encoded representations. In the attention weighting layer, attention weights for each layer are calculated, and the intra-layer encoded representations are weighted and aggregated to obtain weighted representations for the merchant, terminal, and account layers. Finally, the weighted representations of each layer are hierarchically aggregated to generate a structural feature vector corresponding to the current time window. The structural feature vectors of all time windows are arranged in ascending order of transaction time to construct a complete sequence of structural feature vectors, providing a structured input foundation for time feature extraction and fusion modeling.

[0032] In this embodiment, step three includes the following specific steps: The structural feature vector sequence is input into the ETSformer model, and time series decomposition is performed within each time window to obtain the channel representations of trend components, seasonal components and residual components. Channel modeling is performed on the trend component, seasonal component, and residual component respectively. Exponential smoothing attention operation is applied to the trend component channel, frequency attention operation is applied to the seasonal component channel, and original fluctuation preservation operation is applied to the residual component channel to obtain the time series representation of each channel. The time series representations of the trend component channel, seasonal component channel, and residual component channel are fused at the channel level and aggregated at the time step level to generate a time feature vector corresponding to the current time window. Each time window in the time window set is repeatedly decomposed, modeled, merged, and aggregated to obtain the time feature vector corresponding to each time window. These vectors are then connected sequentially according to the transaction time order to obtain a time feature vector sequence arranged in ascending order of transaction time.

[0033] This invention introduces the ETSformer model in the time feature extraction stage. Taking the structural feature vector sequence as input, it performs time series decomposition within each time window, splitting the trading data into three independent channels: trend component, seasonal component, and residual component. Each channel is modeled separately. In the trend component channel, an exponential smoothing attention mechanism is applied to extract long-term evolution features; in the seasonal component channel, a frequency attention mechanism is applied to capture periodic fluctuation patterns; and in the residual component channel, an original fluctuation preservation strategy is used to retain abnormal disturbance features. The time series representations of the three components are fused at the channel level and aggregated at the time step level to generate a time feature vector corresponding to the current time window. This process is repeated for all time windows, and the time feature vectors are concatenated sequentially according to the trading time order to construct a complete time feature vector sequence, thereby achieving a multi-dimensional time series feature expression of the trading data.

[0034] In this embodiment, the bidirectional attention interaction between the structural feature vector and the temporal feature vector in the hierarchical temporal collaborative attention adaptive fusion mechanism includes the following specific steps: Arrange structural feature vectors and time feature vectors with the same time index according to the transaction time order to form corresponding pairs, and then aggregate them to form a set of time window pairs; Within each time window pair, the structural feature vector and the temporal feature vector are input into a hierarchical temporal collaborative attention adaptive fusion mechanism; The hierarchical temporal collaborative attention adaptive fusion mechanism includes a structural feature attention submodule, a temporal feature attention submodule, a collaborative fusion submodule, and a meta-learning controller; In the structural feature attention submodule, the structural feature vector is used as the query vector, and the temporal feature vector is used as the key vector and value vector. The attention weights from the structure to the time direction are calculated and a weighted representation is generated. In the temporal feature attention submodule, the temporal feature vector is used as the query vector, and the structural feature vector is used as the key vector and value vector. The attention weights from time to structure are calculated and a weighted representation is generated. In the collaborative fusion submodule, the weighted representations from structure to time and from time to structure are concatenated and normalized to output the bidirectional attention interaction results.

[0035] This invention enhances the collaborative expressive power of structural and temporal features by constructing a bidirectional attention interaction path between structural and temporal features in a hierarchical temporal collaborative attention adaptive fusion mechanism. Specifically, structural feature vectors and temporal feature vectors are aligned according to transaction time order to form a set of time window pairs, which are then sequentially input into the fusion mechanism. The fusion mechanism includes a structural feature attention submodule, a temporal feature attention submodule, a collaborative fusion submodule, and a meta-learning controller. In the structural feature attention submodule, the structural feature vector is used as the query vector, and the temporal feature vector is used as the key-value pair to calculate the weighted attention representation from structure to time. In the temporal feature attention submodule, an attention path with the temporal feature as the query is constructed in reverse to generate a weighted representation from time to structure. Finally, in the collaborative fusion submodule, the weighted representations from the two directions are concatenated and normalized to output a bidirectional attention interaction result representing the deep coupling relationship between structural and temporal features, providing basic support for subsequent fusion weight generation and representation fusion.

[0036] In this embodiment, the step of dynamically generating hierarchical weights and temporal weights through a meta-learning controller, adaptively adjusting the fusion ratio, and obtaining a fused feature vector includes the following specific steps: In the meta-learning controller, a feedforward neural network algorithm is used to sequentially perform weighted summation, linear mapping, nonlinear activation, and output normalization operations on the bidirectional attention interaction results to generate hierarchical weights. Gated meta-learning computation is performed on the same bidirectional attention interaction results to generate temporal weights; The gated meta-learning computation is based on the MAML transfer learning framework to perform parameter updates. In the inner update phase of the MAML transfer learning framework, the bidirectional attention interaction results of each time window pair are input into the gating network. The task loss for the current time window pair is calculated through forward propagation, and the parameters of the gated network are updated by gradient descent to obtain the updated inner layer parameters. During the gradient update process, a gating signal is introduced to proportionally weight the gradient terms of the gated network parameters in order to achieve dynamic adjustment of the sample update magnitude for different time windows. The gating signal is calculated from the bidirectional attention interaction result through a nonlinear function and is used to dynamically adjust the contribution of the window sample to the parameter update. In the outer layer update phase of the MAML transfer learning framework, the updated inner layer parameters are summarized, the overall meta-loss is calculated based on the task loss of each time window, and the outer layer gradient update is performed on the meta-parameters of the gated network to correct the initialization parameters of the gated network in subsequent tasks. The updated meta-parameters are mapped to a weighted coefficient vector for the time channels, and the output is used as a temporal weight to control the weight ratio of temporal features in the fusion stage.

[0037] Based on the hierarchical weights and temporal weights output by the meta-learning controller, weighted operations are performed on the structural feature vector and the temporal feature vector respectively; In the weighted calculation process, the weighting ratio of the corresponding features of the merchant layer, terminal layer and account layer in the structural feature vector is controlled by the hierarchical weight, and the weighting ratio of the corresponding features of the trend component, seasonal component and residual component in the time feature vector is controlled by the time series weight. After normalizing the weighted results, feature concatenation and linear combination are performed to generate a fused feature vector.

[0038] In the feature generation stage, this invention achieves dynamic fusion of structural and temporal features by constructing an adaptive weight allocation mechanism based on a meta-learning controller. Specifically, in the meta-learning controller, a feedforward neural network is used to sequentially perform weighted summation, linear mapping, nonlinear activation, and output normalization operations on the bidirectional attention interaction results to generate hierarchical weights for adjusting the weighting ratio of structural features. Simultaneously, a gated meta-learning method based on the MAML transfer learning framework is employed. The bidirectional attention interaction results are used to perform forward propagation and gradient update of the task loss for each time window. During the inner-layer update process, a gated signal is introduced to dynamically weight the gradient, controlling the contribution of different time windows to the model parameter update. In the outer-layer update stage, the inner-layer parameters of all time windows are aggregated, and the meta-parameters are updated based on the overall meta-loss. By mapping the updated meta-parameters to a weighted coefficient vector of the time channel, temporal weights are output. Finally, structural and temporal features are weighted according to the hierarchical weights and temporal weights, respectively, and the results are normalized, feature concatenated, and linearly combined to generate a highly representative fusion feature vector, effectively improving the expressive power and decision accuracy of the risk identification model.

[0039] In this embodiment, step five includes the following specific steps: The feature vector sequence is fused as the input to the improved CatBoost model, and the training sequence is constructed in ascending order of transaction time, and an ordered boosting iteration process is established. An ordered prediction path without future sample information is established based on the time index of the training sequence. The ordered prediction path only includes historical sample information that is earlier than the current sample in time. In the temporal correlation ordered tree update module, a temporal correlation weight matrix is ​​generated based on the temporal order of samples in the training sequence. An initial temporal weight matrix is ​​generated by calculating the time difference between any pair of samples and using an exponential decay kernel function. A temporal embedding representation is introduced based on the initial temporal weight matrix, and a self-attention weighting method is used to normalize and correct the temporal correlation to obtain the temporal correlation weight matrix. The temporal correlation weight matrix is ​​used to perform temporal weighting on the sample residuals and second-order statistics during the gradient update process.

[0040] Based on the temporal order of samples in the training sequence, the time difference between any pair of samples is calculated, and a time correlation weight matrix is ​​generated by weighting with an exponential decay kernel function and temporal embedding attention. During the ordered boosting iteration of CatBoost, the deviation between the current CatBoost model prediction and the true label is calculated as the sample residual for each time sample, and the corresponding second-order gradient term is calculated. The sample residuals and second-order gradient terms are weighted using a time-correlation weight matrix to obtain time-corrected weighted residuals and weighted second-order terms. The leaf node values ​​of the current base learner are then updated using the weighted residuals and weighted second-order terms. The base learner is a decision tree newly generated in a certain iteration of the improved CatBoost model.

[0041] The training sequence is traversed using an ordered prediction path that does not contain information about future samples. The leaf node values ​​of the current base learner are updated in chronological order. At each time step, the leaf node values ​​are calculated using the weighted residual and the weighted second-order quantity. In the feature hierarchy perceptual tree splitting module, the features from the merchant layer, account layer and terminal layer in the fused feature vector are divided into hierarchical feature sets; Perform candidate split point search: During the candidate split point search process, a candidate split threshold is generated for each level feature, and the level feature set is divided based on the split threshold to form candidate split sample pairs; Execute hierarchical order constraints: Perform hierarchical order determination operation on each candidate feature, determine the relationship between the hierarchical identifier of the candidate feature and the splitting hierarchical state of the current node, and call the hierarchical regular constraint function to calculate the constraint coefficient; The hierarchical regularization constraint function calculates the constraint coefficient based on the difference between the candidate feature hierarchical identifier and the current node split hierarchical state; taking the current node split hierarchical state as the upper reference level, when the candidate feature level is lower than the reference level, the penalty weight is calculated based on the hierarchical difference; the hierarchical difference is weighted using an exponential decay function, and the calculated output value is used as the constraint coefficient. The split gain function value of the candidate feature is adjusted according to the constraint coefficient. When the hierarchical identifier of the candidate feature is lower than the current node split hierarchical state, a penalty is applied to the candidate split gain. When the hierarchical identifier of the candidate feature is higher than or equal to the current node split hierarchical state, the split gain calculation result is retained. Perform gain evaluation: After processing the hierarchical order constraints, evaluate and sort the split gain function values ​​of all candidate features, select the optimal split feature and split threshold to generate decision nodes according to the adjusted split gain calculation results, and update the leaf node values ​​of the current base learner; to ensure that the splitting path of the tree follows the hierarchical order of merchant layer, account layer and terminal layer; Based on the determined optimal splitting features and splitting threshold, the splitting structure of the current decision tree is generated. After splitting, candidate splitting point search, hierarchical order constraint and gain evaluation are performed on the child nodes until the adjusted splitting gain function value of all candidate splitting features is lower than the minimum gain threshold. The growth of the current iteration tree and the update of leaf node values ​​are completed under the alternating action of the time-related ordered tree update module and the feature hierarchical structure-aware tree splitting module, and the current iteration tree is added to the integration of the improved CatBoost model. Repeatedly iterate the training sequence in an orderly manner until the preset number of iterations is reached to obtain the ensemble output of the improved CatBoost model; The integrated output is used as the score vector of the improved CatBoost model.

[0042] This invention proposes an improved CatBoost algorithm that integrates time-related modeling and feature hierarchy perception in the construction of a risk identification model. First, a training sequence arranged in ascending order of transaction time is constructed using the fused feature vector sequence, establishing an ordered prediction path without future sample information. In the time-related ordered tree update module, the time difference between samples is calculated and a time embedding representation is introduced. An exponential decay kernel function and a self-attention weighting mechanism are used to generate a time-related weight matrix, achieving temporal correction of sample gradient residuals and second-order terms. Second, during feature splitting, a feature hierarchy perception mechanism is introduced, dividing the fused feature vector into hierarchical feature sets such as merchant layer, account layer, and terminal layer. A hierarchy order constraint function is set based on the hierarchy identifier. Combining the hierarchy differences of candidate features, the constraint coefficient is calculated using an exponential decay method to correct the splitting gain, ensuring that the model splitting path follows the hierarchy order. Through the alternating collaboration of the ordered tree update and the hierarchy perception module, the training and updating of the base learner are completed. Multiple iterations are used to improve and build an improved CatBoost model ensemble output, achieving accurate risk scoring and classification of merchant transaction behavior.

[0043] In this embodiment, step six includes the following specific steps: The integrated output is compared with a preset set of risk thresholds to generate a risk score and risk type label corresponding to each time window; the risk score and risk type label are organized according to merchant identifier and time index to form a time window level result set. The time window level result sets are merged according to the merchant identifier to generate a merchant level result record. The merchant level result record includes the merchant identifier, time index, risk score and risk type label. The merchant-level results are returned to the system fusion module, which then registers the results and triggers the update process. In the system fusion module, update flags for hierarchical weights and temporal weights are set based on the returned results, and update instructions are sent to the meta-learning controller and the improved CatBoost model. The hierarchical weight update process is executed by the meta-learning controller, and the model parameter update process is executed by the improved CatBoost model. After the update is completed, the update status will be written to the status record table of the system fusion module.

[0044] In the output and feedback stages of the risk identification process, this invention compares the integrated model output with a preset set of risk thresholds to generate a risk score and risk type label for each time window. These are then organized into a time window-level result set based on merchant identifiers and time indices, and further merged to form a merchant-level result record containing merchant identifiers, time indices, risk scores, and risk type labels. This result record is returned to the system fusion module to register the identification results and trigger subsequent adaptive update processes. The system fusion module sets update markers for hierarchical weights and temporal weights based on the risk identification results and issues update instructions to the meta-learning controller and the improved CatBoost model, executing update operations for the hierarchical weights and model parameters respectively. After the update is completed, the system fusion module records the update status, thereby realizing a closed-loop feedback mechanism from identification results to model adjustment, improving the model's adaptability to dynamic transaction behavior and risk response efficiency.

[0045] Example 1: To verify the feasibility of this invention in practice, it was applied to the actual business of a third-party payment platform in a merchant transaction risk identification scenario. This platform processes over 2 million transactions daily, serving merchants across multiple industries including retail, catering, services, and education. Transaction terminals include POS devices, mobile applications, and web interfaces. The platform has consistently faced the challenge of accurately identifying counterfeit transactions, abnormal transaction behavior, and high-risk merchants. Especially given the highly dynamic nature of transaction samples and the significant time-varying nature of data distribution, traditional risk models suffer from insufficient accuracy, poor generalization ability, and slow response speed, making it difficult to meet the demands of real-time and efficient risk control.

[0046] In this scenario, the "Deep Learning-Based Intelligent Acquiring Risk Identification System" described in this invention is deployed within the risk control system of the payment platform, primarily used for risk scoring and classification of merchants' daily transaction behavior. The system accesses four types of data sources on the platform: payment records, merchant basic information, terminal device registration records, and account information. The data covers key indicators such as transaction amount, transaction time, account identifier, terminal identifier, geolocation code, and transaction channel. By constructing standardized transaction records and time-series sample sequences, the system divides each merchant's continuous transactions into several time windows, performs hierarchical modeling of the structural information within each window, and then extracts multi-level structural features.

[0047] The system utilizes the ETSformer model to decompose trading time series data into trend, seasonal, and residual components. It extracts time series features using attention mechanisms across different channels and interacts bidirectionally with the structural feature vectors through a hierarchical temporal collaborative attention adaptive fusion module. Guided by a meta-learning controller, a fused feature vector is generated. This fused feature is then input into an improved CatBoost model for risk identification. The model introduces a time-related weight matrix to correct the gradient calculation process temporally. Furthermore, a perceptual splitting module is designed based on the hierarchical structure of features at the merchant, account, and terminal levels, enabling more accurate modeling of risk features. The system's fusion module categorizes the model outputs by risk level and triggers an adaptive update process, further enhancing the system's real-time performance and adaptability.

[0048] Within one calendar month of the system's launch, 10,000 typical small and medium-sized merchants on the platform were selected as pilot samples, and the system's identification performance was compared and analyzed. Compared with the platform's original risk identification system, the system of this invention improved the high-risk transaction identification rate by 13.2% and reduced the false positive rate by 7.8%. Meanwhile, the model's daily adaptive iterative updates ensured a sensitive response capability when new fraudulent transaction methods emerged, effectively blocking three potential counterfeit transactions and recovering economic losses exceeding 150,000 yuan.

[0049] Table 1 summarizes the identification results of some high-risk merchants during the pilot period, including information such as merchant identification, window start and end time, number of transactions, number of high-risk transactions, system score, original model score, and final label comparison.

[0050] Table 1 Comparison of High-Risk Identification Results for Pilot Merchants

[0051] As can be seen from Table 1, the present invention demonstrates a higher ability to identify high-risk transactions, and the scoring results are more discriminative. At the same time, combined with the result feedback mechanism of the system fusion module, it realizes the closed-loop operation of risk identification and model optimization, which significantly improves the intelligence level and business practical value of the acquiring risk control system.

[0052] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based acquirer risk intelligent identification system, characterized in that, include: The time-series sample construction module is used to build time-series sample sequences based on merchant identifiers, transaction times, and transaction amounts. The hierarchical attention structure feature encoding module is used to encode the structure features of temporally sequenced samples using a hierarchical attention network. The time series feature extraction module is used to input structural feature vectors into the ETSformer model to extract time series features from transaction data. The hierarchical temporal collaborative attention adaptive fusion module is used to perform bidirectional attention interaction between structural feature vectors and temporal feature vectors, and dynamically generates hierarchical weights and temporal weights through a meta-learning controller; An improved CatBoost risk identification module is used to output risk scores and risk type labels; The system fusion module is used to summarize and merge risk scores and risk type labels to form merchant-level result records. 2.The deep learning-based acquirer risk intelligent identification system of claim 1, wherein, The modules are connected in the following way: Step 1: Obtain the merchant's transaction data, account information, and terminal device information, and establish a time-series sample sequence based on the merchant's identifier, transaction time, and transaction amount; Step 2: Use a hierarchical attention network to encode the structural features of the temporal sample sequence, and extract and weight features for the merchant layer, terminal layer and account layer respectively to obtain structural feature vectors; Step 3: Input the structural feature vector into the ETSformer model to extract time series features from the transaction data and obtain the time feature vector; Step 4: In the hierarchical temporal collaborative attention adaptive fusion mechanism, the structural feature vector and the temporal feature vector are subjected to bidirectional attention interaction. The hierarchical weights and temporal weights are dynamically generated through the meta-learning controller, and the fusion ratio is adaptively adjusted to obtain the fused feature vector. Step 5: Input the fused feature vector into the improved CatBoost model. The temporally correlated ordered tree update module performs temporal correction by constructing a temporally correlated weight matrix; the feature hierarchical structure-aware tree splitting module performs splitting gain correction by constructing a hierarchical feature set and based on the hierarchical regularization constraint function. Step Six: The improved CatBoost model outputs risk scores and risk type labels to identify and classify risks in merchant transaction behavior. The results are then returned to the system fusion module to update the hierarchical weights, time-series weights, and model parameters. 3.The deep learning-based acquirer risk intelligent identification system of claim 2, wherein, Step one includes the following specific steps: Obtain the original records related to merchant transactions; Extract the following fields from the original records: merchant identifier, account identifier, terminal device identifier, transaction time, transaction amount, transaction currency, transaction channel, merchant category code, geographic area code, and transaction serial number; Establish a standardized transaction record by using the transaction serial number as the primary key to link transaction records with merchant master data, account information, and terminal device information. Standardized transaction records are grouped by merchant identifiers. Within each merchant group, a timeline is constructed based on the transaction times arranged in ascending order. The timeline is then sliced ​​to generate a continuous sequence of time windows. For each window in the time window sequence, collect the ordered transaction records within the start and end time range of the current window to form a window-level sample fragment; Each window-level sample fragment corresponds one-to-one with a merchant identifier, generating a time-series sample sequence. 4.The deep learning-based acquirer risk intelligent identification system of claim 2, wherein, Step two includes the following specific steps: Using the temporally sequenced sample sequence as input to the hierarchical attention network, the set of time windows to be processed is determined; The Bi-GRU algorithm is used to encode the merchant layer, terminal layer and account layer at the feature coding layer to generate corresponding intra-layer coding representations; Attention weights are calculated for the merchant layer, terminal layer and account layer in the attention weighting layer, and the in-layer encoding representations of each layer are weighted and aggregated to obtain the weighted representations of the merchant layer, the terminal layer and the account layer. The merchant-level weighted representation, terminal-level weighted representation, and account-level weighted representation are hierarchically aggregated to generate a structural feature vector corresponding to the current time window; The structural feature vectors of each time window in the time window set are arranged in ascending order of transaction time to form a sequence of structural feature vectors that correspond one-to-one with the time-series sample sequence. 5.The deep learning-based acquirer risk intelligent identification system of claim 2, wherein, Step three includes the following specific steps: The structural feature vector sequence is input into the ETSformer model, and time series decomposition is performed within each time window to obtain the channel representations of trend components, seasonal components and residual components. Channel modeling is performed on the trend component, seasonal component, and residual component respectively. Exponential smoothing attention operation is applied to the trend component channel, frequency attention operation is applied to the seasonal component channel, and original fluctuation preservation operation is applied to the residual component channel to obtain the time series representation of each channel. The time series representations of the trend component channel, seasonal component channel, and residual component channel are fused at the channel level and aggregated at the time step level to generate a time feature vector corresponding to the current time window. For each time window in the time window set, the process involves repeated decomposition, modeling, fusion, and aggregation to obtain the time feature vector sequence corresponding to each time window. 6.The deep learning-based acquirer risk intelligent identification system of claim 2, wherein, The hierarchical temporal collaborative attention adaptive fusion mechanism involves bidirectional attention interaction between structural feature vectors and temporal feature vectors, including the following specific steps: Arrange structural feature vectors and time feature vectors with the same time index according to the transaction time order to form corresponding pairs, and then aggregate them to form a set of time window pairs; Within each time window pair, the structural feature vector and the temporal feature vector are input into a hierarchical temporal collaborative attention adaptive fusion mechanism; Using structural feature vectors as query vectors and temporal feature vectors as key and value vectors, we calculate the attention weights from structure to time and generate a weighted representation. Using the time feature vector as the query vector and the structural feature vector as the key and value vectors, the attention weights from time to structure are calculated and a weighted representation is generated. The weighted representations from structure to time and from time to structure are concatenated and normalized to output the bidirectional attention interaction result.

7. The deep learning-based acquirer risk intelligent identification system of claim 2, wherein, The process of dynamically generating hierarchical weights and temporal weights through a meta-learning controller, adaptively adjusting the fusion ratio, and obtaining a fused feature vector includes the following specific steps: In the meta-learning controller, a feedforward neural network algorithm is used to sequentially perform weighted summation, linear mapping, nonlinear activation, and output normalization operations on the bidirectional attention interaction results to generate hierarchical weights. In the inner update phase of the MAML transfer learning framework, the bidirectional attention interaction results of each time window pair are input into the gating network. The task loss of the current time window pair is calculated through forward propagation, and gradient descent is performed to update the parameters of the gating network to obtain the updated inner parameters. During the gradient update process, a gating signal is introduced to proportionally weight the gradient terms of the gated network parameters; The gating signal is calculated from the bidirectional attention interaction result using a nonlinear function; In the outer layer update phase of the MAML transfer learning framework, the updated inner layer parameters are summarized, the overall meta-loss is calculated based on the task loss of each time window, and the outer layer gradient update is performed on the meta-parameters of the gated network. The updated meta-parameters are mapped to a weighted coefficient vector for the time channels, and the output is used as the time series weights. Based on the hierarchical weights and temporal weights output by the meta-learning controller, weighted operations are performed on the structural feature vector and the temporal feature vector respectively; In the weighted calculation process, the weighting ratio of the corresponding features of the merchant layer, terminal layer and account layer in the structural feature vector is controlled by the hierarchical weight, and the weighting ratio of the corresponding features of the trend component, seasonal component and residual component in the time feature vector is controlled by the time series weight. After normalizing the weighted results, feature concatenation and linear combination are performed to generate a fused feature vector.

8. The intelligent acquisition risk identification system based on deep learning according to claim 2, characterized in that, The time-correlation ordered tree update module performs time-series correction by constructing a time-correlation weight matrix, including the following specific steps: The fused feature vector sequence is used as input to the improved CatBoost model, and the training sequence is constructed in ascending order of transaction time. An ordered prediction path without future sample information is established based on the time index of the training sequence. In the temporal correlation ordered tree update module, a temporal correlation weight matrix is ​​generated based on the temporal order of samples in the training sequence. During the ordered boosting iteration of CatBoost, the deviation between the current CatBoost model prediction and the true label is calculated as the sample residual for each time sample, and the corresponding second-order gradient term is calculated. The sample residuals and second-order gradient terms are weighted using a time-correlation weight matrix to obtain time-corrected weighted residuals and weighted second-order terms. The leaf node values ​​of the current base learner are then updated using the weighted residuals and weighted second-order terms. The training sequence is traversed using an ordered prediction path that does not contain information about future samples. The leaf node values ​​of the current base learner are updated in chronological order, and the leaf node values ​​are calculated at each time step using weighted residuals and weighted second-order quantities.

9. The intelligent acquisition risk identification system based on deep learning according to claim 2, characterized in that, The feature-hierarchical perceptual tree splitting module constructs a hierarchical feature set and performs splitting gain correction based on a hierarchical regularization constraint function, including the following specific steps: In the feature hierarchy perceptual tree splitting module, the features from the merchant layer, account layer and terminal layer in the fused feature vector are divided into hierarchical feature sets; Perform candidate split point search: During the candidate split point search process, a candidate split threshold is generated for each level feature, and the level feature set is divided based on the split threshold to form candidate split sample pairs; Execute hierarchical order constraints: Perform hierarchical order determination operation on each candidate feature, determine the relationship between the hierarchical identifier of the candidate feature and the splitting hierarchical state of the current node, and call the hierarchical regular constraint function to calculate the constraint coefficient; The split gain function value of the candidate feature is adjusted according to the constraint coefficient. When the hierarchical identifier of the candidate feature is lower than the current node split hierarchical state, a penalty is applied to the candidate split gain. When the hierarchical identifier of the candidate feature is higher than or equal to the current node split hierarchical state, the split gain calculation result is retained. Perform gain evaluation: After processing the hierarchical order constraints, evaluate and sort the split gain function values ​​of all candidate features, select the optimal split feature and split threshold according to the adjusted split gain calculation results to generate decision nodes, and update the leaf node values ​​of the current base learner. Based on the determined optimal splitting features and splitting threshold, the splitting structure of the current decision tree is generated. After splitting, candidate splitting point search, hierarchical order constraint and gain evaluation are performed on the child nodes until the adjusted splitting gain function value of all candidate splitting features is lower than the minimum gain threshold. The growth of the current iteration tree and the update of leaf node values ​​are completed under the alternating action of the time-related ordered tree update module and the feature hierarchical structure-aware tree splitting module, and the current iteration tree is added to the integration of the improved CatBoost model. Repeatedly iterate the training sequence in an orderly manner until the preset number of iterations is reached to obtain the ensemble output of the improved CatBoost model.

10. The intelligent acquisition risk identification system based on deep learning according to claim 2, characterized in that, Step six includes the following specific steps: The integrated output is compared with a preset set of risk thresholds to generate a risk score and risk type label corresponding to each time window; the risk score and risk type label are organized according to merchant identifier and time index to form a time window level result set. Merge the time window-level result sets according to the merchant identifier to generate merchant-level result records; The merchant-level results are returned to the system fusion module, which then registers the results and triggers the update process. In the system fusion module, update flags for hierarchical weights and temporal weights are set based on the returned results, and update instructions are sent to the meta-learning controller and the improved CatBoost model. The hierarchical weight update process is executed by the meta-learning controller, and the model parameter update process is executed by the improved CatBoost model. After the update is completed, the update status will be written to the status record table of the system fusion module.

Images