FTransform-LightGBM-based imported commodity price fraud identification method

Abstract

The invention belongs to the field of intelligent algorithm fraud identification, and provides an imported commodity price fraud identification method based on FTransform-LightGBM, which improves the feature expression of price fraud data through the proposed double-branch architecture and dynamic feature fusion, and further improves the category identification accuracy of imported commodity price fraud. The method comprises the following steps: firstly, designing a double-branch architecture constructed based on a machine learning method and a deep learning method, and by fusing output features of LightGBM and FTModel models, better coping with data diversity and enhancing feature expression of fraudulent price data; secondly, a feature extraction branch based on FTModel is designed based on a full connection layer and Transform, feature expression of fraud price data is enhanced by focusing on extraction of key features of commodity price fraud data in a complex scene, and extraction of global features of the fraud data is enhanced; and finally, the proposed dynamic feature fusion can more autonomously select effective features through dynamic weight adjustment, deep mining of associated features is realized, and the model recognition performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent algorithm fraud identification, and in particular to an imported commodity price fraud identification method based on FTransformer-LightGBM. Background Art

[0002] Price fraud in imported goods refers to the intentional concealment of the true price of goods by importers when declaring the import of goods to the customs by falsely reporting the quantity or price, falsely reporting the name of the goods, and false certificates of origin, in order to evade or reduce the duties and other taxes payable. Machine learning models can analyze a large amount of historical data, extract the characteristics of fraudulent transactions, quickly identify and predict fraudulent behavior, and help customs to grasp the fraud situation in a timely manner, thereby reducing the loss of customs tax revenue. As the scale and complexity of imported goods data continue to grow, the data features are high in dimension and the potential relationships are fuzzy. In particular, a large amount of feature engineering needs to be extracted, which is a huge workload. The effectiveness of statistical and machine learning algorithms in detecting price fraud in imported goods is limited.

[0003] Compared with traditional machine learning methods, deep learning can automatically learn features and effectively process high-dimensional, nonlinear, and long-term dependent sequence data. It has a wide range of applications in time series prediction, image recognition, and natural language processing. At present, the features of imported goods data in big data scenarios are complex and hidden, and the manual extraction of features is labor-intensive. However, the existing shallow network learning methods are difficult to capture the deep-level feature correlations in imported goods data when identifying imported goods price fraud, resulting in a low accuracy rate in identifying imported goods price fraud categories. Summary of the invention

[0004] In view of the fact that the single model in the existing technology has weak ability to express the characteristics of imported commodity data; and the problem of fixed feature fusion weights in the multi-branch structure, which leads to a lack of flexibility and easily causes the loss of characteristic information of imported commodity data, and ultimately leads to the low accuracy of identifying price fraud transactions of imported commodities, an imported commodity price fraud identification method based on FTransformer-LightGBM is proposed to enhance the feature expression of imported commodity data, deeply mine related features, improve the robustness of the model, and improve the accuracy of identifying price fraud categories of imported commodities.

[0005] The specific plan is as follows:

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

[0007] S1: Data acquisition: Obtain the imported commodity price fraud identification data; the staff determines whether there is price fraud based on the basic information of the imported commodity in each data and labels it; the labeled data is used as the data set for the subsequent steps; the data format in the data set is [basic information of imported commodities, imported commodity price fraud category label];

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

[0009] S3: Dataset processing: Perform maximum and minimum normalization processing on the data set balanced in S2 to obtain the processed data set;

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

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

[0012] S6: Model training and tuning: Use the training set and validation set of S4 to train and tune the S5 imported commodity price fraud identification model to obtain the optimal imported commodity price fraud identification model;

[0013] S7: Model evaluation: Input the basic information of the imported goods in the S4 test set into the S6 optimal imported goods price fraud identification model for prediction, obtain the imported goods price fraud identification results, and compare them with the imported goods price fraud category labels, calculate the accuracy, recall rate, precision rate and F1 value of the imported goods price fraud identification results to evaluate the imported goods price fraud identification model.

[0014] Preferably, the specific method of the imported commodity price fraud identification model based on FTransformer-LightGBM is:

[0015] S51: Send the basic information of the imported goods to 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: Send the basic information of the imported goods to the feature extraction branch based on FTModel for feature extraction, and obtain the weight and extracted features of the feature extraction branch based on FTModel;

[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 sent to the dynamic feature fusion module for fusion to obtain fused features;

[0018] S54: Result output; send the S53 fusion feature to the sigmoid module to obtain the probabilities of the two categories of fraud and no fraud. The category with the largest probability value is the identification result of the imported goods price fraud model.

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

[0020] S511: sampling the sample data in the original data set by a gradient-based single-side sampling method to obtain a new data set; the new data set includes continuous floating-point features;

[0021] S512: Mapping the continuous floating-point feature values ​​in the new data set in S511 into k discrete boxes, i.e., bins, through a histogram algorithm; constructing a histogram with a width of k through the discrete boxes; the mapping discretization process converts the continuous floating-point features into discrete data features;

[0022] S513: all discrete data features of S512 are regarded as a root node; the root node is split according to the growth strategy of the leaf node; the final extraction feature is extracted according to the mutually exclusive feature bundling algorithm; the split point is selected by traversing and sorting the discrete values ​​of the S512 histogram to find the best split point.

[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 bundling algorithm bundles the basic information features of imported goods with high correlation and discards the basic information features of imported goods with low correlation.

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

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

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

[0027] S523: Send the features of S522 fused with global information to 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 is composed of two linear layers and a ReLU activation function. The hidden representation of 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 expressed as:

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

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

[0031] Preferably, the dynamic feature fusion module realizes the fusion of the extracted features of the two feature extraction branches by dynamically adjusting the weight of each feature; the expression of the final fusion feature is as follows:

[0032] W all =W a ·W L +W β ·W F (2)

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

[0034] Preferably, the model training and tuning is to iteratively optimize the imported commodity price fraud model through 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 is the price fraud category label of the imported goods of the i-th sample, p i It is the maximum probability value of the i-th sample predicting the existence of fraud and no fraud.

[0037] Beneficial effects:

[0038] The present invention proposes a method for identifying price fraud of imported goods based on FTransformer-LightGBM, which solves the problems of large workload of manual feature extraction of high-dimensional large sample data and low fraud identification accuracy. First, in view of the simple expression of the single-model imported commodity price fraud identification features, the present invention designs a dual-branch architecture based on machine learning methods and deep learning methods to enhance the feature expression of fraudulent price data; second, based on the fully connected layer and Transformer, a feature extraction branch based on FTModel is designed to strengthen the extraction of global features of fraud data; third, in view of the problem 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 capabilities of a single model, the present invention proposes a dual-branch architecture based on machine learning methods and deep learning methods to enhance the feature expression of imported commodity fraud price data, where LightGBM is used as a feature extraction branch and can efficiently process complex data sets; the Transformer-based FTModel is used as another feature extraction branch, focusing on the extraction of key features of commodity price fraud data in complex scenarios, solving the problem of weak feature expression capabilities of a single model, and thus improving the recognition accuracy of the imported commodity price fraud model.

[0040] (2) In the deep learning branch, a 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 commodity price fraud data; the Transformer layer is completely based on the self-attention mechanism to establish a dependency model of any two vectors in a long sequence, and 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 the Transformer can perform complex nonlinear transformations on the input through a combination of two linear layers and an activation function, thereby enhancing the expressiveness of the features. This helps the imported commodity price fraud model better capture the features and patterns in the input. The introduction of FFN can further increase the independence between positions, making the imported commodity price fraud model more autonomous and independent in representing each position, reducing mutual interference between positions, and thus helping to improve the performance and representation ability of the imported commodity price fraud model.

[0041] (3) In order to solve the problem of fixed feature fusion weights, dynamic feature fusion is proposed to deeply mine related features, improve the robustness of the model, and improve the accuracy of identifying price fraud of imported goods. Dynamic feature fusion can automatically optimize the fusion method of the two model features under different data samples and task scenarios by dynamically adjusting the weights, flexibly deal with the heterogeneity and complexity in the data, and thus improve the accuracy and adaptability of the imported goods price fraud model. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

[0045] Figure 4 This is the accuracy curve of the imported commodity price fraud identification model based on FTransformer-LightGBM at different learning rates.

[0046] Figure 5 This is the loss value curve of the imported commodity price fraud identification model based on FTransformer-LightGBM at different learning rates.

[0047] Figure 6 This is the accuracy curve of the imported goods price fraud identification model based on FTransformer-LightGBM in different batches.

[0048] Figure 7 This is the loss value curve of the imported commodity price fraud identification model based on FTransformer-LightGBM in different batches.

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

[0050] Fig. 9 This is the confusion matrix of the price fraud identification category of the imported goods price fraud identification model based on FTransformer-LightGBM. DETAILED DESCRIPTION

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

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

[0053] S1: Data acquisition: Obtain the imported commodity price fraud identification data; the staff determines whether there is price fraud based on the basic information of the imported commodity in each data and labels it; the labeled data is used as the data set for the subsequent steps; the data format in the data set is [basic information of imported commodities, imported commodity price fraud category label];

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

[0055] S3: Dataset processing: Perform maximum and minimum normalization processing on the data set balanced in S2 to obtain the processed data set;

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

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

[0058] S6: Model training and tuning: Use the training set and validation set of S4 to train and tune the S5 imported commodity price fraud identification model to obtain the optimal imported commodity price fraud identification model;

[0059] S7: Model evaluation: Input the basic information of the imported goods in the S4 test set into the S6 optimal imported goods price fraud identification model for prediction, obtain the imported goods price fraud identification results, and compare them with the imported goods price fraud category labels, calculate the accuracy, recall rate, precision rate and F1 value of the imported goods price fraud identification results to evaluate the imported goods price fraud identification model.

[0060] Specific content includes:

[0061] (1) Data acquisition: Obtain a dataset for identifying price fraud of imported goods; the data format in the dataset is [basic information of imported goods, category label of price fraud of imported goods];

[0062] Specifically, the basic information of imported goods includes variables such as product number, product name, product specifications, product model, country of production and sales, transaction currency, declared unit price, declared total price, declared quantity, declared measurement unit, total transaction price and whether there is fraud.

[0063] (2) Constructing and expanding the dataset: The present invention uses an optimized SMOTE-CGAN model to expand and construct the dataset.

[0064] (3) Dataset preprocessing: The import commodity price fraud dataset is normalized to its maximum and minimum values.

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

[0066] Specifically, the Lightweight Gradient Boosting Machine (LightGBM) is an efficient gradient boosting decision tree algorithm that builds multiple decision trees and trains them in a step-by-step optimization manner to capture complex patterns and nonlinear relationships in the data. LightGBM uses a histogram-based method to determine the optimal split point of the feature, without the need to sort all the values ​​of each feature. This method is more efficient than traditional sorting methods, especially when processing large-scale data sets. LightGBM significantly reduces the amount of calculation and reduces memory consumption. Therefore, the present invention uses LightGBM as a branch to express the expression of the transaction characteristics of price fraud of imported goods.

[0067] LightGBM introduces a histogram algorithm to bin eigenvalues, effectively reducing the number of original eigenvalues, thereby speeding up the training of the model. In addition, LightGBM also integrates gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB). GOSS retains samples with larger gradients through random sampling, while discarding samples with smaller gradients to reduce the size of the data set; EFB reduces the dimension of the feature space by bundling highly correlated features. The introduction of these two technologies further improves the training speed, performance, and efficiency of the model, making LightGBM perform well in processing large-scale data sets and complex features.

[0068] Specifically, the Transformer model architecture is as follows Figure 3 As shown, Transformer is a deep neural network based on self-attention mechanism and parallel data processing. As a sequence-to-sequence model, Transformer architecture has set off a craze in the fields of natural language processing and computer vision. The Transformer architecture uses a multi-head self-attention mechanism (multi-head self-attention, MHA) and a feedforward neural network (feedforward neural network, FFN) to build the encoder and decoder structure. Transformer has two characteristics. One is that it has a strong ability to learn long-distance dependencies. The long dependency characteristics of Transformer itself use the attention mechanism to capture global context information; the second is that it has a strong multimodal fusion ability, and can usually directly operate on pixels to obtain the initial embedding vector. The information of other modes can be converted into vectors and directly fused at the input end. Therefore, the present invention adopts the optimized Transformer as a branch to express the characteristics of price fraud transactions of imported goods.

[0069] The optimized Transformer is named FTModel, which consists of two fully connected layers (FC), a Transformer layer and a fully connected layer. The first two fully connected layers are used to extract rich feature information from a large amount of imported commodity price fraud data, while the Transformer layer is completely based on the self-attention mechanism to establish a dependency model between any two vectors in a long sequence, and aggregate global information more efficiently. The last fully connected layer further extracts and compresses the aggregated features.

[0070] Feedforward network (FFN) is a fully connected feedforward neural network and another important component in the Transformer model. FFN is mainly used to perform nonlinear transformation and mapping on the hidden representation of each position. FFN consists of two linear layers and a ReLU activation function. The input is the hidden representation of each position 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 can perform complex nonlinear transformations on the input through the combination of two linear layers and activation functions, thereby enhancing the expressiveness of features. This helps the model better capture the 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 performance and representation of the model.

[0073] (5) Model design: Design and build an import commodity price fraud identification model based on FTransformer-LightGBM; The architecture diagram of the import commodity price fraud identification model based on FTransformer-LightGBM, such as Figure 2 shown.

[0074] Aiming at the problem that a single model has weak feature expression ability for price fraud identification of imported goods, and a single model has low accuracy and low efficiency in identifying price fraud of goods in complex dynamic scenarios, this paper proposes an imported goods price fraud transaction identification model (FTModelwithLightGBM, FTransformer-LightGBM) based on LightGBM and Transformer methods for large sample data sets. First, the proposed model selects LightGBM as a feature extraction branch, which can efficiently process complex data sets; secondly, a FTModel is constructed based on Transformer to focus on the extraction of key features of commodity price fraud data in complex scenarios; finally, dynamic weights are used to fuse the features extracted by LightGBM and FTModel to adaptively capture important information; the probability of whether the commodity price is fraudulent is output through the Softmax layer.

[0075] Specifically, in the field of computer science, especially in the application of machine learning and deep learning, the feature fusion strategy combines the feature extraction capabilities of different models, which is an effective strategy to enhance the recognition performance of the model. After the feature extraction is completed, a dynamic weighted fusion of the features extracted by the LightGBM and FTModel models is designed to comprehensively consider the contribution of the two models in feature extraction. However, the weighting here is not fixed, but dynamically adjusted. The introduction of the dynamic weighting mechanism aims to improve the generalization ability and robustness of the model. By dynamically adjusting the weights, it is possible to automatically optimize the fusion method of the two model features under different data samples and task scenarios, flexibly deal with the heterogeneity and complexity in the data, and thus improve the accuracy and adaptability of the model.

[0076] Dynamic weighted fusion feature W all for:

[0077] W all =W a ·W L +W β ·W F (2)

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

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

[0080] Specifically, binary cross entropy is used to tune the model during model training. Binary Cross Entropy (also known as log loss) is a loss function commonly used 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 label. Specifically, it calculates the logarithmic difference between the true label and the predicted probability.

[0081]

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

[0083] (7) Customs import price fraud identification: Based on the optimal model, import price fraud is identified. The model outputs the probabilities of two categories: fraud and no fraud. The category with the largest probability value is the model identification result. The import price fraud identification effects of the LightGBM model and the FTModel are compared.

[0084] Specifically, the effectiveness of the proposed method is demonstrated through comparative experiments, where control group 1: an import commodity price fraud identification model based on LightGBM; control group 2: an import commodity price fraud identification model based on FTModel; control group 3: an import commodity price fraud identification model based on FTransformer-LightGBM. The collected import commodity price fraud dataset is balanced by SMOTE, and then the processed dataset is trained and optimized for control groups 1, 2, and 3 respectively to obtain the optimal models of the three control groups, and the test machine is input into the three optimal models for identification to obtain the fraud identification type; the fraud identification type is evaluated and analyzed and compared with the actual value.

[0085] In terms of experimental settings, for the LightGBM model, the number of leaf nodes is set to 31, the learning rate is 0.05, and the feature fraction sampled when training each tree is 0.9. For the FTransformer-LightGBM model, the batch value is set to 128, the iteration cycle is set to 100, the learning rate is set to 0.0002, and the early stopping strategy is used to prevent overfitting and save training time. Because price fraud identification is a binary classification task, each model uses the binary cross entropy loss function (BinaryCrossEntropyloss) 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 rate

[0088] The learning rate is an important hyperparameter in the deep learning model training process, which indicates the magnitude of each parameter update. It is related to whether the objective function can converge and when it converges to the local minimum. The appropriate learning rate can speed up the convergence of model training. In order to make the imported commodity price fraud model have better performance, this chapter conducted multiple sets of experiments during the training process, and the learning rates were set to 0.0002, 0.002, and 0.02 respectively. The accuracy curves and loss value curves of the training set obtained at different learning rates are as follows: Figure 4 and Figure 5 shown.

[0089] When the learning rate is set to 0.02, the minimum loss value obtained by the model on the training set is 0.66, and the maximum accuracy is 69%. When the learning rate is set to 0.002, the minimum loss value obtained by the model on the training set is 0.52, and the maximum accuracy is 95%. The overall trend of the training process is consistent with that when the learning rate is set to 0.02. Figure 5 and Figure 6It shows that under these two learning rates, the loss value and accuracy of the model training process did not converge to a good value. And the loss value fluctuated and increased, while the accuracy decreased slightly. The reason is that the excessive learning rate caused the loss value gradient to explode, the parameters were updated too quickly, and the original weight information was destroyed. When the learning rate was set to 0.0002, the minimum loss value obtained by the model on the training set was 0.50, and the maximum accuracy was 99%. Overall, the loss value of the training set of the model showed a continuous downward trend under this learning rate, and finally stabilized. The change in accuracy is similar. The accuracy of the training set rose steadily and finally stabilized, and both were higher than the other two learning rates. Therefore, when the learning rate is 0.0002, the performance of the model is better.

[0090] (2) Batch size comparison analysis

[0091] Batch size is another important hyperparameter in deep learning training. Its size affects the training speed and stability of the model. Smaller batches can improve the training stability of the model, because the model will be updated multiple times in each epoch, which helps to avoid local optimal solutions, but the training process is unstable. Larger batches can process training data faster, thereby reducing training time. However, a larger batch size may also lead to insufficient video memory and occupy more memory, thus affecting the training speed. Therefore, in order to select a suitable batch size to balance the training speed, stability, memory consumption and model performance of the model, this chapter conducted multiple sets of experiments during the training process, and the batch sizes were set to 32, 64, and 128 respectively. The accuracy curves and loss value curves of the training set obtained under different batch sizes are as follows Figure 6 and Figure 7 shown.

[0092] Figure 7 and Figure 8 It reflects the relationship between recognition performance and batch size obtained through experiments on the imported goods price fraud identification dataset. The recognition performance does not increase significantly with the increase of batch size. Under the three batch size settings, the highest accuracy can basically reach 0.99, and the minimum loss value is about 0.50. However, when the batch size is set to 32 and 64, it can be seen that the model fluctuates greatly. The reason is that the weights are frequently updated with a batch size that is too small, resulting in an unstable training process. When the batch size is set to 128, the loss value of the model in training shows a steady downward trend, the accuracy continues to improve, and finally stabilizes. Overall, as the batch size increases, the training process of the model becomes more and more stable. And when the batch size is set to 128, the performance of the model is better.

[0093] 2. Model Comparison Analysis

[0094] Precision and recall can accurately reflect the classification accuracy and comprehensiveness of the recognition model for the two categories of fraud or non-fraud. F1 value and accuracy can reflect the overall performance of the model. Table 1 shows the overall recognition results of different models on the imported goods price fraud dataset. Figure 8 The architecture diagram of the imported goods price fraud identification model based on LightGBM is shown. Fig. 9 The confusion matrix of imported goods price fraud identification using the FTransformer-LightGBM model is shown.

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

[0096]

[0097]

[0098] The overall accuracy of the imported commodity price fraud identification model based on FTransformer-LightGBM proposed in the present invention is 98.24%, which is 10.18%, 48.80% and 4.44% higher than the overall accuracy of the LightGBM model, DeepSpectra model and FTModel model, respectively, and the recognition accuracy is higher than the comparison method. In addition, the average F1 value of the proposed model is 0.98, which is superior to the values ​​of other comparison models. The experimental results show that a price fraud identification model based on FTransformer-LightGBM has strong comprehensive classification performance and the best recognition effect for commodity price fraud. This is because when designing the model structure, the extraction of local features and global features is simultaneously considered, and dynamic weights are used for feature fusion to improve the recognition performance and robustness of the model.

[0099] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for identifying price fraud of imported goods based on FTransformer-LightGBM, characterized in that: S1: Data acquisition: Obtain the imported commodity price fraud identification data; the staff determines whether there is price fraud based on the basic information of the imported commodity in each data and labels it; the labeled data is used as the data set for the subsequent steps; the data format in the data set is [basic information of imported commodities, imported commodity price fraud category label]; S2: Dataset balancing: The data set obtained in S1 is balanced by the SMOTE-CGAN data set balancing method to obtain a balanced data set; the SMOTE-CGAN data set balancing method is to increase the number of minority class samples in the data set by synthesizing new minority class samples; S3: Dataset processing: Perform maximum and minimum normalization processing on the data set balanced in S2 to obtain the processed data set; S4: Determine the training set, validation set, and test set: Divide the data set processed in S3 into a training set, a validation set, and a test set according to the set ratio; S5: Design model: Design and build an import commodity price fraud identification model based on FTransformer-LightGBM; the import commodity price fraud identification model consists of a feature extraction branch based on LightGBM, a feature extraction branch based on FTModel, a dynamic weight fusion module and a sigmoid module; the feature extraction branch based on FTModel consists of two fully connected layers, a Transformer layer and a fully connected layer in sequence; S6: Model training and tuning: The basic information of the imported goods in the training set and the validation set is used as the input of 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 tuned by combining the cross entropy loss to obtain the optimal imported goods price fraud identification model; S7: Model evaluation: Input the basic information of the imported goods in the S4 test set into the S6 optimal imported goods price fraud identification model for prediction, obtain the imported goods price fraud identification results, and compare them with the imported goods price fraud category labels, calculate the accuracy, recall rate, precision rate and F1 value of the imported goods price fraud identification results to evaluate the imported goods price fraud identification model.

2. A method for identifying price fraud of imported goods based on FTransformer-LightGBM according to claim 1, characterized in that: The specific method of the imported commodity price fraud identification model based on FTransformer-LightGBM is: S51: Send the basic information of the imported goods to 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: Send the basic information of the imported goods to the feature extraction branch based on FTModel for feature extraction, and obtain the weight and extracted features of the feature extraction branch based on FTModel; 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 sent to the dynamic feature fusion module for fusion to obtain fused features; S54: Result output; send the S53 fusion feature to the sigmoid module to obtain the probabilities of the two categories of fraud and no fraud. The category with the largest probability value is the identification result of the imported commodity price fraud model.

3. A method for identifying price fraud of imported goods based on FTransformer-LightGBM according to claim 2, characterized in that: The feature extraction method based on the feature extraction branch of LightGBM is: S511: sampling the sample data in the original data set by a gradient-based single-side sampling method to obtain a new data set; the new data set includes continuous floating-point features; S512: Mapping the continuous floating-point feature values ​​in the new data set in S511 into k discrete boxes, i.e., bins, through a histogram algorithm; constructing a histogram with a width of k through the discrete boxes; the mapping discretization process converts the continuous floating-point features into discrete data features; S513: all discrete data features of S512 are regarded as a root node; the root node is split according to the growth strategy of the leaf node; the final extraction feature is extracted according to the mutually exclusive feature bundling algorithm; the split point is selected by traversing and sorting the discrete values ​​of the S512 histogram to find the best split point.

4. A method for identifying price fraud of imported goods based on FTransformer-LightGBM according to 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 bundling algorithm bundles the basic information features of imported goods with high correlation and discards the basic information features of imported goods with low correlation.

5. A method for identifying price fraud of imported goods based on FTransformer-LightGBM according to claim 2, characterized in that: The feature extraction method of the feature extraction branch based on FTModel is: S521: Input the basic information of imported goods into the first two fully connected layers to extract feature information; S522: Send the feature information extracted in S521 to the Transformer layer; establish a dependency model of any two vectors in the feature information through the self-attention mechanism of the Transformer layer, aggregate the features of the global information; and obtain the features of the fused global information; S523: Send the features of S522 fused with global information to 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 according to 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 of 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 expressed as: FFN(x)=max(0,xW1+b1)W2+b2 (1) Where x is the hidden representation of 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 according to claim 2, characterized in that: The dynamic feature fusion module realizes the fusion of the extracted features of the two feature extraction branches by dynamically adjusting the weight of each feature; the expression of the final fusion feature is as follows: IN all =In a ·IN L +W β ·IN F (2) Among them, W a and W L are the weights and extracted features of the feature extraction branch based on LightGBM, W β and W F They are the weight and extracted features of the feature extraction branch based on FTModel respectively.

8. A method for identifying price fraud of imported goods based on FTransformer-LightGBM according to 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: Where L is the loss function, N is the number of samples, and y i is the price fraud category label of the imported goods of the i-th sample, p i It is the maximum probability value of the i-th sample predicting the existence of fraud and no fraud.

Images