An import commodity price fraud identification method based on FTransformer-LightGBM

Abstract

The present application belongs to the field of intelligent algorithm fraud identification, and proposes an import commodity price fraud identification method based on FTransformer-LightGBM. Through the proposed double-branch architecture and dynamic feature fusion, the feature expression of price fraud data is improved, and the identification accuracy of import commodity price fraud categories is improved. First, a double-branch architecture based on machine learning method and deep learning method is designed. By fusing the output features of LightGBM and FTModel models, the diversity of data can be better handled, and the feature expression of fraudulent price data can be enhanced. Second, a feature extraction branch based on FTModel is designed based on full connection layer and Transformer, which focuses on the extraction of key features of commodity price fraud data in complex scenarios to enhance the feature expression of fraudulent price data and strengthen the extraction of global features of fraudulent data. Finally, the proposed dynamic feature fusion can more autonomously select effective features by dynamically adjusting the weight, deeply mine the associated features, and improve the model recognition performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent algorithm fraud detection, and in particular to a method for detecting price fraud of imported goods based on FTransformer-LightGBM. Background Technology

[0002] Import price fraud refers to the practice of importers deliberately concealing the true price of goods when declaring them to customs, such as misreporting quantity or price, misreporting product name, and using false certificates of origin, in order to evade or reduce applicable customs duties and other taxes. Machine learning models can analyze large amounts of historical data to extract characteristics of fraudulent transactions, quickly identify and predict fraudulent behavior, and help customs promptly grasp fraudulent situations, thereby reducing customs revenue losses. However, with the increasing scale and complexity of imported commodity data, the high dimensionality of data features, and the ambiguity of potential relationships, especially the need for extensive feature engineering, the workload is enormous, limiting the effectiveness of statistical and machine learning algorithms in detecting imported commodity price fraud.

[0003] Compared to traditional machine learning methods, deep learning can automatically learn features and effectively handle high-dimensional, non-linear, and long-term dependent sequence data, finding wide applications in time series prediction, image recognition, and natural language processing. Currently, imported goods data in big data scenarios are complex, with hidden correlations, making manual feature extraction laborious. However, existing shallow network learning methods struggle to capture the deep-seated feature relationships within imported goods data, resulting in low accuracy in identifying imported goods price fraud categories. Summary of the Invention

[0004] To address the issues of weak feature representation capabilities of single models in existing technologies for imported commodity data, and the fixed feature fusion weights in multi-branch structures leading to a lack of flexibility and potential loss of feature information in imported commodity data, ultimately resulting in low accuracy in identifying imported commodity price fraud transactions, this paper proposes an imported commodity price fraud identification method based on FTransformer-LightGBM. This method enhances the feature representation of imported commodity data, while deeply mining associated features to improve model robustness and increase the accuracy of imported commodity price fraud category identification.

[0005] The specific plan is as follows:

[0006] A method for identifying price fraud in imported goods based on FTransformer-LightGBM, comprising:

[0007] S1: Data Acquisition: Acquire imported commodity price fraud identification data; staff determine whether price fraud exists based on the basic information of the imported commodity in each data entry and label it; the labeled data is used as the dataset for subsequent steps; the data format in the dataset is [basic information of imported commodities, imported commodity price fraud category label];

[0008] S2: Dataset balancing: The dataset obtained in S1 is balanced using the SMOTE-CGAN dataset balancing method to obtain a balanced dataset; the SMOTE-CGAN dataset balancing method increases the number of minority class samples in the dataset by synthesizing new minority class samples;

[0009] S3: Dataset processing: Perform maximum and minimum value normalization on the balanced dataset after S2 to obtain the processed dataset;

[0010] S4: Determine the training set, validation set, and test set: Divide the dataset processed in S3 into the training set, validation set, and test set according to the set ratio;

[0011] S5: Design Model: Design and build an import commodity price fraud detection model based on FTransformer-LightGBM; the import commodity price fraud detection model consists of a LightGBM-based feature extraction branch, an FTModel-based feature extraction branch, a dynamic weight fusion module, and a sigmoid module; the FTModel-based feature extraction branch consists of two fully connected layers, one Transformer layer, and one fully connected layer in sequence;

[0012] S6: Model Training and Optimization: Use the training and validation sets of S4 to train and optimize the S5 imported goods price fraud detection model to obtain the optimal imported goods price fraud detection model.

[0013] S7: Model Evaluation: Input the basic information of imported goods in the S4 test set into the optimal imported goods price fraud identification model in S6 for prediction, obtain the imported goods price fraud identification result, and compare it with the imported goods price fraud category label. Calculate the accuracy, recall, precision, and F1 score of the imported goods price fraud identification result to evaluate the imported goods price fraud identification model.

[0014] Preferably, the specific method for the imported goods price fraud identification model based on FTransformer-LightGBM is as follows:

[0015] S51: Input the basic information of the imported goods into the feature extraction branch based on LightGBM for feature extraction, and obtain the weight and extracted features of the feature extraction branch based on LightGBM.

[0016] S52: Input the basic information of the imported goods into the FTModel-based feature extraction branch for feature extraction, and obtain the weight and extracted features of the FTModel-based feature extraction branch;

[0017] S53: Dynamic Feature Fusion: The weights and extracted features of the feature extraction branch based on LightGBM in S51 and the weights and extracted features of the feature extraction branch based on FTModel in S52 are fed into the dynamic feature fusion module for fusion to obtain fused features;

[0018] S54: Output the results; feed the fused features from S53 into the sigmoid module to obtain the probabilities of two categories: fraud and no fraud. The category with the highest probability value is the result of the imported goods price fraud model.

[0019] Preferably, the feature extraction method based on the LightGBM feature extraction branch is as follows:

[0020] S511: A new dataset is obtained by sampling the sample data in the original dataset using a gradient-based one-sided sampling method; the new dataset includes continuous floating-point features;

[0021] S512: The continuous floating-point feature values ​​in the new dataset S511 are mapped and discretized into k discrete bins using a histogram algorithm; a histogram of width k is constructed using the discrete bins; the mapping and discretization process transforms the continuous floating-point features into discrete data features.

[0022] S513: Treat all discrete data features of S512 as a root node; split the root node according to the growth strategy of leaf nodes; extract the final extracted features according to the mutual exclusion feature binding algorithm; the selection of the split point is to find the best split point by traversing and sorting the discrete values ​​of the S512 histogram.

[0023] Preferably, the gradient-based unilateral sampling retains the basic information of imported goods with larger gradients through random sampling, while discarding the basic information of imported goods with smaller gradients; the mutually exclusive feature binding algorithm binds the basic information features of highly correlated imported goods and discards the basic information features of low-correlation imported goods.

[0024] Preferably, the feature extraction method based on the FTModel feature extraction branch is as follows:

[0025] S521: Input the basic information of imported goods into the first two fully connected layers to extract feature information;

[0026] S522: Input the feature information extracted in S521 into the Transformer layer; establish a dependency model between any two vectors in the feature information through the self-attention mechanism of the Transformer layer, and aggregate the features of global information; obtain the features fused with global information.

[0027] S523: The features fused with global information from S522 are fed into the fully connected layer for further extraction and compression to obtain the final extracted features.

[0028] Preferably, the Transformer layer includes a feedforward network, which consists of two linear layers and a ReLU activation function. The hidden representation at each position in the Transformer is input into the feedforward network to obtain a new hidden representation after nonlinear transformation and mapping. The output of the feedforward network is as follows:

[0029] FFN(x)=max(0,xW1+b1)W2+b2 (1)

[0030] Where x is the hidden representation at each position in the Transformer, and W1, W2, b1, b2 are learnable weight matrices and bias vectors.

[0031] Preferably, the dynamic feature fusion module achieves the fusion of features extracted from the two feature extraction branches by dynamically adjusting the weighting of each feature; the final expression of the fused features is shown below:

[0032] W all =W a ·W L +W β ·W F (2)

[0033] Among them, W a and W L These are the weights and extracted features of the feature extraction branch based on LightGBM, W. β and W F These are the weights and extracted features of the feature extraction branch based on FTModel.

[0034] Preferably, the model training and optimization involves iteratively optimizing the imported goods price fraud model using cross-entropy loss; the cross-entropy loss is obtained by calculating the logarithmic difference between the true label and the predicted probability, and its expression is as follows:

[0035]

[0036] Where L is the loss function, N is the number of samples, and y i p is the price fraud category label for the i-th sample of imported goods. i It is the maximum probability value predicted for the i-th sample between the two categories of fraud and no fraud.

[0037] Beneficial effects:

[0038] This invention proposes a method for identifying price fraud in imported goods based on FTransformer-LightGBM, addressing the problems of high workload and low accuracy in manually extracting features from high-dimensional, large-sample data. First, to address the simplistic feature representation of single-model imported goods price fraud identification, this invention designs a dual-branch architecture based on machine learning and deep learning methods to enhance the feature representation of fraudulent price data. Second, a feature extraction branch based on FTModel is designed using fully connected layers and Transformer to strengthen the extraction of global features from fraudulent data. Third, to address the issue of low model accuracy caused by fixed weighting, the proposed dynamic feature fusion can more autonomously select effective features, deeply mine related features, and improve model recognition performance. Specifically:

[0039] (1) To address the problem of weak feature expression capability of single models, this invention proposes a dual-branch architecture based on machine learning and deep learning methods to enhance the feature expression of imported commodity price fraud data. LightGBM serves as a feature extraction branch, which can efficiently process complex datasets. FTModel based on Transformer serves as another feature extraction branch, which focuses on extracting key features of commodity price fraud data in complex scenarios, thus solving the problem of weak feature expression capability of single models and improving the recognition accuracy of imported commodity price fraud models.

[0040] (2) In the deep learning branch, an FTModel feature extraction module based on fully connected layers and Transformer is proposed. The first two fully connected layers of FTModel are used to extract rich feature information from a large amount of imported goods price fraud data; while the Transformer layer is based entirely on the self-attention mechanism to establish a dependency model of any two vectors in a long sequence, which can more efficiently aggregate global information; the last fully connected layer further extracts and compresses the aggregated features; through the above operations, the extraction of global features of fraud data can be strengthened, and the recognition accuracy of the model can be improved. The FFN in Transformer can perform complex nonlinear transformations on the input through the combination of two linear layers and activation functions, thereby enhancing the expressive power of features. This helps the imported goods price fraud model to better capture the features and patterns in the input. The introduction of FFN can further increase the independence between positions, making the imported goods price fraud model more autonomous and independent in its representation of each position, reducing mutual interference between positions, thereby helping to improve the performance and representation power of the imported goods price fraud model.

[0041] (3) To address the issue of fixed feature fusion weights, dynamic feature fusion is proposed to deeply mine associated features, improve the robustness of the model, and increase the accuracy of identifying price fraud in imported goods. Dynamic feature fusion can automatically optimize the fusion method of the two models' features under different data samples and task scenarios by dynamically adjusting the weights, flexibly responding to the heterogeneity and complexity of the data, thereby improving the accuracy and adaptability of the imported goods price fraud model. Attached Figure Description

[0042] Figure 1 This is a flowchart of a method for identifying price fraud in imported goods based on FTransformer-LightGBM.

[0043] Figure 2 This is an architecture diagram of an imported goods price fraud detection model based on FTransformer-LightGBM.

[0044] Figure 3 This is an architecture diagram of Transformer.

[0045] Figure 4 The image shows the accuracy curves of the imported goods price fraud identification model based on FTransformer-LightGBM under different learning rates.

[0046] Figure 5 The image shows the loss curves of the imported goods price fraud detection model based on FTransformer-LightGBM under different learning rates.

[0047] Figure 6 The image shows the accuracy curves of the imported goods price fraud identification model based on FTransformer-LightGBM under different batches.

[0048] Figure 7 The image shows the loss curves of the imported goods price fraud detection model based on FTransformer-LightGBM under different batches.

[0049] Figure 8 This is a diagram of the architecture of an imported goods price fraud detection model based on LightGBM.

[0050] Figure 9 This is the confusion matrix for price fraud identification categories in the imported goods price fraud identification model based on FTransformer-LightGBM. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] like Figure 1 As shown, the specific steps of this invention include:

[0053] S1: Data Acquisition: Acquire imported commodity price fraud identification data; staff determine whether price fraud exists based on the basic information of the imported commodity in each data entry and label it; the labeled data is used as the dataset for subsequent steps; the data format in the dataset is [basic information of imported commodities, imported commodity price fraud category label];

[0054] S2: Dataset balancing: The dataset obtained in S1 is balanced using the SMOTE-CGAN dataset balancing method to obtain a balanced dataset; the SMOTE-CGAN dataset balancing method increases the number of minority class samples in the dataset by synthesizing new minority class samples;

[0055] S3: Dataset processing: Perform maximum and minimum value normalization on the balanced dataset after S2 to obtain the processed dataset;

[0056] S4: Determine the training set, validation set, and test set: Divide the dataset processed in S3 into the training set, validation set, and test set according to the set ratio;

[0057] S5: Design Model: Design and build an import commodity price fraud detection model based on FTransformer-LightGBM; the import commodity price fraud detection model consists of a LightGBM-based feature extraction branch, an FTModel-based feature extraction branch, a dynamic weight fusion module, and a sigmoid module; the FTModel-based feature extraction branch consists of two fully connected layers, one Transformer layer, and one fully connected layer in sequence;

[0058] S6: Model Training and Optimization: Use the training and validation sets of S4 to train and optimize the S5 imported goods price fraud detection model to obtain the optimal imported goods price fraud detection model.

[0059] S7: Model Evaluation: Input the basic information of imported goods in the S4 test set into the optimal imported goods price fraud identification model in S6 for prediction, obtain the imported goods price fraud identification result, and compare it with the imported goods price fraud category label. Calculate the accuracy, recall, precision, and F1 score of the imported goods price fraud identification result to evaluate the imported goods price fraud identification model.

[0060] The specific content includes:

[0061] (1) Data acquisition: Acquire the imported commodity price fraud identification dataset; the data format in the dataset is [basic information of imported commodities, imported commodity price fraud category label];

[0062] Specifically, the basic information of imported goods includes variables such as commodity code, commodity name, commodity specifications, commodity model, country of origin and destination, currency of transaction, declared unit price, declared total price, declared quantity, declared unit of measurement, total transaction price, and whether fraud has occurred.

[0063] (2) Dataset construction and expansion: This invention uses an optimized SMOTE-CGAN model to expand and construct the dataset.

[0064] (3) Dataset preprocessing: The maximum and minimum values ​​of the imported goods price fraud dataset are normalized.

[0065] (4) Selection of deep learning model and ensemble learning model: Select LightGBM, an ensemble learning model with fast and efficient machine learning algorithms; select Transformer, a deep learning model with strong ability to learn long-distance dependencies.

[0066] Specifically, Lightweight Gradient Boosting Machine (LightGBM) is an efficient gradient boosting decision tree algorithm that trains by constructing multiple decision trees and progressively optimizing them to capture complex patterns and nonlinear relationships in the data. LightGBM uses a histogram-based method to determine the optimal split points for features, eliminating the need to sort all values ​​for each feature. This method is more efficient than traditional sorting methods, especially when dealing with large-scale datasets, where LightGBM significantly reduces computation and memory consumption. Therefore, this invention uses LightGBM as a branch to represent the features of imported goods price fraud transactions.

[0067] LightGBM employs a histogram algorithm to bin feature values, effectively reducing the number of original feature values ​​and thus accelerating model training. Furthermore, LightGBM integrates gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB). GOSS retains samples with large gradients while discarding those with small gradients through random sampling, reducing the dataset size; EFB reduces the dimensionality of the feature space by bundling highly correlated features. The introduction of these two techniques further improves the model's training speed, performance, and efficiency, making LightGBM outstanding when handling large-scale datasets and complex features.

[0068] Specifically, the model architecture diagram of Transformer is as follows: Figure 3 As shown, the Transformer is a deep neural network based on self-attention and parallel data processing. As a sequence-to-sequence model, the Transformer architecture has become increasingly popular in natural language processing and computer vision. The Transformer architecture uses multi-head self-attention (MHA) and a feedforward neural network (FFN) to build the encoder and decoder structures. The Transformer has two main characteristics: first, it has a strong ability to learn long-distance dependencies, leveraging its inherent long-dependency characteristic and attention mechanism to capture global contextual information; second, it has strong multimodal fusion capabilities, typically allowing direct pixel manipulation to obtain initial embedding vectors, and the information from other modalities can be converted into vectors for direct fusion at the input. Therefore, this invention uses an optimized Transformer as a branch to express the characteristics of imported commodity price fraud transactions.

[0069] The optimized Transformer is named FTModel, which consists of two fully connected layers (FC), one Transformer layer, and one fully connected layer. The first two fully connected layers extract rich feature information from a large amount of imported goods price fraud data. The Transformer layer, based entirely on a self-attention mechanism, establishes a dependency model between any two vectors in a long sequence, more efficiently aggregating global information. The final fully connected layer further extracts and compresses the aggregated features.

[0070] A feedforward network (FFN) is a fully connected feedforward neural network and another important component of the Transformer model. FFN is primarily used to perform nonlinear transformations and mappings on the hidden representation at each location. An FFN consists of two linear layers and a ReLU activation function. The input is the hidden representation at each location in the Transformer, and the output is the new hidden representation after nonlinear transformation and mapping. The output FFN(x) is:

[0071] FFN(x)=max(0,xW1+b1)W2+b2(1)

[0072] FFN, through a combination of two linear layers and activation functions, can perform complex nonlinear transformations on the input, thereby enhancing the expressive power of features. This helps the model better capture features and patterns in the input. The introduction of FFN can further increase the independence between positions, making the model's representation of each position more autonomous and independent, reducing mutual interference between positions, and thus helping to improve the model's performance and representational power.

[0073] (5) Model Design: Design and build an import commodity price fraud detection model based on FTransformer-LightGBM; the architecture diagram of the import commodity price fraud detection model based on FTransformer-LightGBM is shown below. Figure 2 As shown.

[0074] To address the weaknesses of single models in representing features for identifying price fraud in imported goods, and the low accuracy and efficiency of single models in identifying price fraud in complex and dynamic scenarios, this invention proposes a large-sample dataset-based model for identifying price fraud in imported goods (FTModelwithLightGBM, FTransformer-LightGBM). First, the proposed model selects LightGBM as a feature extraction branch, enabling efficient processing of complex datasets. Second, it constructs an FTModel based on Transformer to focus on extracting key features of price fraud data in complex scenarios. Finally, it uses dynamic weights to fuse the features extracted by LightGBM and FTModel, adaptively capturing important information. The probability of whether a product price is fraudulent is output through a Softmax layer.

[0075] Specifically, in computer science, particularly in machine learning and deep learning applications, feature fusion strategies are effective ways to combine the feature extraction capabilities of different models to enhance their recognition performance. After feature extraction, a dynamic weighted fusion of features extracted by LightGBM and FTModel was designed to comprehensively consider the contributions of both models in feature extraction. However, the weighting is not fixed but dynamically adjusted. The introduction of this dynamic weighting mechanism aims to improve the model's generalization ability and robustness. By dynamically adjusting the weights, the fusion method of the two models' features can be automatically optimized under different data samples and task scenarios, flexibly addressing the heterogeneity and complexity of the data, thereby improving the model's accuracy and adaptability.

[0076] Features W after dynamic weighted fusion all for:

[0077] W all =W a ·W L +W β ·W F (2)

[0078] Among them, W a and W L These are the weights and extracted features of the feature extraction branch based on LightGBM, W. β and W F These are the weights and extracted features of the feature extraction branch based on FTModel.

[0079] (6) Model training and optimization: The data is divided into training set, validation set and test set, which are used for training, optimization and evaluation of the imported commodity price fraud identification model to obtain the optimal parameters of the FTransformer-LightGBM model.

[0080] Specifically, binary cross-entropy is used to fine-tune the model during training. Binary cross-entropy (also known as log loss) is a commonly used loss function in machine learning. The binary cross-entropy loss function measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels. Specifically, it calculates the logarithmic difference between the true labels and the predicted probabilities.

[0081]

[0082] Where L is the loss function, N is the number of samples, and y i p is the true label (0 or 1) of the i-th sample. i It is the predicted probability of the i-th sample.

[0083] (7) Customs Imported Goods Price Fraud Detection: Imported goods price fraud detection is performed based on the optimal model. The model outputs the probability of two categories: fraud and no fraud. The category with the highest probability value is the model's detection result. The imported goods price fraud detection performance of the LightGBM model and FTModel is compared.

[0084] Specifically, the effectiveness of the proposed method is demonstrated through comparative experiments. Control group 1: an imported goods price fraud detection model based on LightGBM; Control group 2: an imported goods price fraud detection model based on FTModel; Control group 3: an imported goods price fraud detection model based on FTransformer-LightGBM. The collected imported goods price fraud dataset was balanced using SMOTE. The processed dataset was then used to train and optimize control groups 1, 2, and 3 respectively, obtaining the optimal models for the three control groups. The test machine was then input into these three optimal models for fraud detection to obtain the fraud identification types. The fraud identification types were then compared and evaluated with the actual values.

[0085] In terms of experimental setup, for the LightGBM model, the number of leaf nodes was set to 31, the learning rate to 0.05, and the feature score sampled for each tree during training was 0.9. For the FTransformer-LightGBM model, the batch size was set to 128, the iteration period to 100, the learning rate to 0.0002, and an early stopping strategy was used to prevent overfitting and save training time. Because price fraud detection is a binary classification task, all models used the binary cross-entropy loss function for parameter updates. This configuration aims to achieve a comprehensive update of all parameters in the algorithm through sufficient iterations. Finally, W in dynamic feature fusion... a and W β The values ​​are 0.4 and 0.6 respectively.

[0086] 1. Comparative Analysis of Model Parameters

[0087] (1) Comparative analysis of learning rates

[0088] The learning rate is a crucial hyperparameter in the deep learning model training process, representing the magnitude of parameter updates. It relates to whether and when the objective function converges to a local minimum; a suitable learning rate can accelerate model training and convergence. To improve the performance of the imported goods price fraud model, this chapter conducted multiple experiments during training, setting the learning rates to 0.0002, 0.002, and 0.02 respectively. The accuracy and loss curves of the training set obtained under different learning rates are shown below. Figure 4 and Figure 5 As shown.

[0089] When the learning rate is set to 0.02, the model achieves a minimum loss of 0.66 and a maximum accuracy of 69% on the training set. When the learning rate is set to 0.002, the model achieves a minimum loss of 0.52 and a maximum accuracy of 95% on the training set. The overall trend of the training process is consistent with that when the learning rate is set to 0.02. Figure 5 and Figure 6This indicates that under both learning rates, neither the loss value nor the accuracy during model training converged to a satisfactory value. Furthermore, the loss value fluctuated upwards, while the accuracy slightly decreased. This is because an excessively large learning rate caused gradient explosion in the loss value, resulting in excessively rapid parameter updates and disrupting the original weight information. When the learning rate was set to 0.0002, the model achieved a minimum loss value of 0.50 and a maximum accuracy of 99% on the training set. Overall, the model's loss value on the training set showed a continuous downward trend under this learning rate, eventually stabilizing. The accuracy showed a similar pattern, with the training set accuracy rising steadily and eventually stabilizing, and both figures were higher than the other two learning rates. Therefore, the model performed better when the learning rate was 0.0002.

[0090] (2) Batch size comparison analysis

[0091] Batch size is another important hyperparameter in deep learning training. Its size affects model training speed and stability. Smaller batch sizes can improve training stability because the model is updated multiple times per epoch, helping to avoid local optima, but the training process can be unstable. Larger batch sizes can process training data faster, thus reducing training time. However, larger batch sizes can also lead to insufficient GPU memory and higher memory consumption, thus affecting training speed. Therefore, to choose an appropriate batch size to balance training speed, stability, memory consumption, and model performance, this section conducted multiple experiments during training with batch sizes of 32, 64, and 128. The accuracy and loss curves of the training set obtained under different batch sizes are shown in the figure below. Figure 6 and Figure 7 As shown.

[0092] Figure 7 and Figure 8 This reflects the relationship between recognition performance and batch size, as observed in experiments on an imported goods price fraud identification dataset. Recognition performance does not significantly increase with increasing batch size. Under all three batch size settings, the accuracy generally reaches a maximum of 0.99, and the minimum loss is around 0.50. However, when the batch size is set to 32 and 64, the model exhibits significant fluctuations. This is because excessively small batch sizes lead to frequent weight updates, resulting in unstable training. When the batch size is set to 128, the model's loss during training shows a stable decreasing trend, while the accuracy continuously improves, eventually stabilizing. Overall, the model training process becomes increasingly smooth as the batch size increases. Furthermore, the model performs better when the batch size is set to 128.

[0093] 2. Model Comparison Analysis

[0094] Precision and recall accurately reflect the accuracy and comprehensiveness of the identification model in classifying the two categories of fraud and non-fraud. F1 score and precision reflect the overall performance of the model. Table 1 shows the overall identification results of different models on the imported goods price fraud dataset. Figure 8 The architecture diagram of the imported goods price fraud detection model based on LightGBM is shown. Figure 9 The confusion matrix for identifying price fraud in imported goods using the FTransformer-LightGBM model is presented.

[0095] Table 1 Comparison of recognition accuracy of different evaluation models

[0096]

[0097]

[0098] The proposed import commodity price fraud detection model based on FTransformer-LightGBM achieves an overall accuracy of 98.24%, which is 10.18%, 48.80%, and 4.44% higher than the overall accuracy of the LightGBM, DeepSpectra, and FTModel models, respectively. Furthermore, the proposed model has an average F1 score of 0.98, which is superior to other comparative models. Experimental results show that the FTransformer-LightGBM-based price fraud detection model exhibits strong overall classification performance and the best detection effect for commodity price fraud. This is because the model design simultaneously considers the extraction of local and global features and employs dynamic weights for feature fusion, improving the model's recognition performance and robustness.

[0099] Finally, it should be noted that the above description is only used to illustrate the technical solution of the present invention and not to limit it. 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 solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for identifying price fraud in imported goods based on FTransformer-LightGBM, characterized in that, S1: Data Acquisition: Acquire imported goods price fraud identification data; staff determine whether price fraud exists based on the basic information of the imported goods in each data entry and label it; the labeled data is used as the dataset for subsequent steps; the data format in the dataset is: basic information of imported goods, imported goods price fraud category label; S2: Dataset balancing: The dataset obtained in S1 is balanced using the SMOTE-CGAN dataset balancing method to obtain a balanced dataset; the SMOTE-CGAN dataset balancing method increases the number of minority class samples in the dataset by synthesizing new minority class samples; S3: Dataset processing: Perform maximum and minimum value normalization on the balanced dataset after S2 to obtain the processed dataset; S4: Determine the training set, validation set, and test set: Divide the dataset processed in S3 into the training set, validation set, and test set according to the set ratio; S5: Design Model: Design and build an import commodity price fraud detection model based on FTransformer-LightGBM; the import commodity price fraud detection model consists of a LightGBM-based feature extraction branch, an FTModel-based feature extraction branch, a dynamic feature fusion module, and a sigmoid module; the FTModel-based feature extraction branch consists of two fully connected layers, one Transformer layer, and one fully connected layer in sequence; The dynamic feature fusion module achieves the fusion of features extracted from the two feature extraction branches by dynamically adjusting the weighting of each feature; the final expression of the fused features is shown below: in, and These are the weights and extracted features of the feature extraction branch based on LightGBM. and These are the weights and extracted features of the feature extraction branch based on FTModel, respectively. S6: Model Training and Optimization: The basic information of imported goods in the training and validation sets is used as the input to the S5 imported goods price fraud identification model, and the imported goods price fraud category label is used as the output to train the model; the imported goods price fraud identification model is optimized by combining cross-entropy loss to obtain the optimal imported goods price fraud identification model; S7: Model Evaluation: Input the basic information of imported goods in the S4 test set into the optimal imported goods price fraud identification model in S6 for prediction, obtain the imported goods price fraud identification result, and compare it with the imported goods price fraud category label. Calculate the accuracy, recall, precision, and F1 score of the imported goods price fraud identification result to evaluate the imported goods price fraud identification model.

2. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 1, characterized in that, The specific method of the imported goods price fraud detection model based on FTransformer-LightGBM is as follows: S51: Input the basic information of the imported goods into the feature extraction branch based on LightGBM for feature extraction, and obtain the weight and extracted features of the feature extraction branch based on LightGBM. S52: Input the basic information of the imported goods into the FTModel-based feature extraction branch for feature extraction, and obtain the weight and extracted features of the FTModel-based feature extraction branch; S53: Dynamic Feature Fusion: The weights and extracted features of the feature extraction branch based on LightGBM in S51 and the weights and extracted features of the feature extraction branch based on FTModel in S52 are fed into the dynamic feature fusion module for fusion to obtain fused features; S54: Output the results; feed the fused features from S53 into the sigmoid module to obtain the probabilities of two categories: fraud and no fraud. The category with the highest probability value is the result of the imported goods price fraud model.

3. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 2, characterized in that, The feature extraction method based on the LightGBM feature extraction branch is as follows: S511: A new dataset is obtained by sampling the sample data in the original dataset using a gradient-based one-sided sampling method; the new dataset includes continuous floating-point features; S512: The continuous floating-point feature values ​​in the new dataset S511 are mapped and discretized into k discrete bins using a histogram algorithm; a histogram of width k is constructed using the discrete bins; the mapping and discretization process transforms the continuous floating-point features into discrete data features. S513: Treat all discrete data features of S512 as a root node; split the root node according to the growth strategy of leaf nodes; extract the final extracted features according to the mutual exclusion feature binding algorithm; the selection of the split point is to find the best split point by traversing and sorting the discrete values ​​of the S512 histogram.

4. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 3, characterized in that, The gradient-based unilateral sampling retains the basic information of imported goods with larger gradients through random sampling, while discarding the basic information of imported goods with smaller gradients; the mutually exclusive feature binding algorithm binds the basic information features of highly correlated imported goods and discards the basic information features of low-correlation imported goods.

5. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 2, characterized in that, The feature extraction method based on the FTModel feature extraction branch is as follows: S521: Input the basic information of imported goods into the first two fully connected layers to extract feature information; S522: Input the feature information extracted in S521 into the Transformer layer; establish a dependency model between any two vectors in the feature information through the self-attention mechanism of the Transformer layer, and aggregate the features of global information; obtain the features fused with global information. S523: The features fused with global information from S522 are fed into the fully connected layer for further extraction and compression to obtain the final extracted features.

6. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 5, characterized in that, The Transformer layer includes a feedforward network, which consists of two linear layers and a ReLU activation function. The hidden representation at each position in the Transformer is input into the feedforward network, resulting in a new hidden representation after nonlinear transformation and mapping. The output of the feedforward network is as follows: Where x is the hidden representation at each position in the Transformer, and W1, W2, b1, b2 are learnable weight matrices and bias vectors.

7. A method for identifying price fraud of imported goods based on FTransformer-LightGBM as described in claim 1, characterized in that, The cross-entropy loss is obtained by calculating the logarithmic difference between the true label and the predicted probability, and its expression is as follows: in, It is a loss function. It is the number of samples. It is the first Price fraud category labels for imported goods in a sample. It is the first The maximum probability value predicted for each sample is between the two categories of fraud and no fraud.

Images