A Customs Risk Identification Method for Import and Export Commodities Based on Declaration Quality Assessment

Abstract

This invention relates to the field of data mining technology and discloses a method for identifying customs import and export commodity risks based on declaration quality assessment. The method includes the following steps: S1, Data preprocessing: extracting element fields and performing preliminary screening to obtain a customs declaration sample set; S2, Sampling the customs declaration sample set based on two-stage sampling: generating a candidate label sample set; S3, Reducing the candidate label sample set based on class centers: obtaining a training sample set after sample reduction; S4, Customs declaration quality assessment using short text binary classification: obtaining the probability of correct customs declaration and completing the customs declaration quality assessment. This invention neither overly relies on certain rules nor requires a large amount of manual review. It can identify loopholes in the regulatory system from massive amounts of declaration data, uncover abnormal customs declaration behavior patterns, efficiently and intelligently review customs declarations, identify risks in the declaration process, and achieve the function of automatically identifying customs declaration quality and providing real-time early warnings.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically a method for identifying risks in customs import and export commodities based on declaration quality assessment. Background Technology

[0002] In customs import and export operations, the quality assessment of customs declarations is one of the key issues in risk assessment for import and export activities. Currently, customs declaration review mainly includes electronic review and manual review. Electronic review uses fixed parameters and relies on past experience for fixed-parameter, mechanical review, which is too mechanical and experience-based, creating significant loopholes in customs declaration review and risk assessment. Some customs declarations rely on manual experience and are reviewed manually. Due to human resource limitations, it is impossible to handle the need to review a massive number of customs declarations one by one, making it difficult to accurately identify the low probability of non-compliant declarations among a large number of declarations. This makes it impossible to achieve accurate inspection, which allows criminals to use trial and error to find declaration loopholes, evade customs supervision, and fail to comply with declaration standards when declaring import and export goods.

[0003] A Chinese patent discloses a method for generating customs inspection rules based on knowledge graphs and tree models (publication number CN111753928A). This patented technology extracts several effective indicator features from traditional expert rule engines, uses multiple CART decision trees for feature information extraction, constructs a random forest, and learns the correlation between customs declarations based on knowledge graphs, applying it to the learning of decision tree models, thereby establishing a rule system based on effective indicator features. It then uses a new rule engine for customs inspection. However, it relies too heavily on expert rules, and this information is often difficult to keep up with the times. Criminals may try to figure out this rule system and seek to profit by designing new irregular declaration schemes.

[0004] Chinese patents also disclose customs inspection and monitoring methods, devices, computer equipment, and storage media (publication number CN111798092A). This patented technology obtains information on customs staff from a pre-constructed customs knowledge graph, as well as monitoring indicators associated with the inspection site. Then, it makes risk predictions for staff based on the data of these monitoring indicators and issues warnings to customs staff who are predicted to be at risk. However, this technology only considers the abnormalities of staff during inspections and does not investigate the customs declaration itself. In fact, many problems are attributed to the quality of the customs declaration itself. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying risks in import and export commodities by customs based on declaration quality assessment, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A customs risk identification method for import and export commodities based on declaration quality assessment includes the following steps:

[0008] S1. Data preprocessing: This includes element extraction and element quantity screening. First, the element data of the customs declaration is obtained through element extraction. Then, the non-standard customs declarations in the element data are screened out through element quantity screening to obtain a customs declaration sample set.

[0009] S2. Sampling of the customs declaration sample set based on two-stage sampling: First, take the union of the data for each element in the customs declaration sample set to obtain the full data for each element, and sort them according to their frequency of occurrence; then, for each element, sample the top elements with the highest frequency. % of the samples are directly included in the training samples as the high-frequency sample set; the remaining 1- % of the samples were used as the low-frequency sample set. Then, using a two-stage sampling method with equal probability, the remaining 1- A percentage of the samples were clustered into text clusters, and then random samples were taken from each cluster. %, to obtain the training sample set; then, merge the training sample set with the high-frequency sample set to generate the candidate labeled sample set. ;

[0010] S3. Candidate label sample set reduction based on class center: First, calculate the sample set of each cluster. The cluster center; then calculate each cluster The distance from each sample point to the cluster center is used to obtain the candidate labeled sample set. The set of distances from each sample point to the cluster center is then arranged in order, and the closest distance is selected. Each sample yields the corresponding point set. Corresponding point samples To be reduced, each One sample is selected from the center and kept, the rest are discarded, resulting in... A set of sample points Customs inspectors perform specialized labeling, then take the union of the samples to obtain the training sample set after sample deletion. ;

[0011] S4. Quality assessment of customs declarations using short text binary classification: The FastText model is used to evaluate the quality of the final training sample set. The short text is segmented and stop word removed. The remaining words are processed into word feature vectors through word embedding. These word feature vectors are then used as the input layer of the model and fed into a neural network model for linear operations. The calculation results are used as the output of the hidden layer. A Softmax function is then used to establish a mapping from the hidden layer to the output layer. The Softmax function is normalized, and an unequal probability neuron dropout method is embedded to selectively discard neurons. The probabilities of compliant and non-compliant declarations are calculated, and the score of the customs declaration is determined to complete the customs declaration quality assessment.

[0012] As a further solution of the present invention: the method of element extraction in step S1 is as follows: First, extract the "commodity name and specifications" field from the customs declaration, then split the content of the field into a string, and then split the string by the '|' character. Each customs declaration obtains a set of element fields, which are the element data in the customs declaration.

[0013] The method for filtering the number of elements in step S1 is as follows: compare the number of element data in the customs declaration with the element list in the element table to determine whether the number of elements corresponds. If they match, store them in the customs declaration sample set; if they do not match, filter out the inconsistent customs declarations and determine them as non-standard customs declarations.

[0014] As a further aspect of the present invention: in step S2, a pond sampling method is used to randomly sample each cluster; the sampling method includes the following steps:

[0015] S201, Definition For the sample set, For the result set, for The size of the array, For each random number, The number of samples drawn. This represents the number of iterations.

[0016] S202, Take Before the set The number is filled into the set. ,from Start iterating and generating random numbers. , range ,like Then replace the value in R, that is ;

[0017] S203. Traversal ends, generating result array R; obtaining candidate labeled sample set. .

[0018] As a further aspect of the present invention, the text clustering method in step S2 includes the following steps:

[0019] S21. Dataset processing: Using the NLTK toolkit, the dataset is sequentially processed by tokenization, stop word removal, part-of-speech tagging, and stemming.

[0020] S22. Feature Extraction: First, the processed dataset is mapped to a vector space and converted into a TF-IDF word frequency matrix Q. Then, the Laplacian matrix of the word frequency matrix Q is calculated. Finally, the eigenvectors corresponding to the first K eigenvalues ​​of the Laplacian matrix are obtained and matrix R is formed.

[0021] S23. Graph Clustering: A short text clustering algorithm based on Laplacian graphs, which will output the clustering results.

[0022] As a further aspect of the present invention, the specific method for graph clustering in step S23 is as follows:

[0023] S231. Sort matrix R according to the size of its eigenvalues;

[0024] S232. Select the eigenvectors corresponding to the first K eigenvalues ​​of matrix R as the initial cluster centers to improve the clustering convergence speed.

[0025] S233. Assign each data object to the nearest cluster based on the cosine of the angle between the feature vectors.

[0026] S234. Recalculate cluster centers;

[0027] S235. If the algorithm converges, terminate the iteration and output the clustering results; otherwise, return to step S233.

[0028] As a further aspect of the present invention: in step S3, the cluster The formula for calculating the cluster center is as follows:

[0029]

[0030] In the above formula (1), For clusters The candidate labeled sample set in for The dimension, For the i-th cluster center, for The dimension, for The number of data in the data;

[0031] cluster From each sample point to the cluster center The distance calculation formula is as follows:

[0032]

[0033] In the above formula (2), From sample point to cluster center The set of distances;

[0034] Final training sample set The union of the sets is as follows:

[0035]

[0036] In the above formula (3), for Each sample yields a corresponding set of points. for A set of sample points The final training sample set.

[0037] As a further aspect of the present invention, the calculation formula for the hidden layer in step S4 is as follows:

[0038]

[0039] The formula for calculating the loss function for the mapping from the hidden layer to the output layer in step S4 is as follows:

[0040]

[0041]

[0042] The calculation formula for the normalization process of the Softmax function in step S4 is as follows:

[0043]

[0044] The formula for calculating the probability of neuron dropping in step S4 is as follows:

[0045]

[0046] In the above formula (8), when the word is a non-strong feature word, the probability of it being discarded is 0.5; when the word is a strong feature word, the probability of it being discarded is less than 0.5, thus reducing the probability of strong feature words being discarded.

[0047] As a further aspect of the present invention, the method for calculating the score of the customs declaration in step S4 is as follows:

[0048] When the word feature vectors reach the output layer, dimensional vector Map to a vector whose values ​​are in the interval (0,1). Because this invention uses a binary classification method, When the probability of a predicted positive label is 1, it indicates that the element was declared in full compliance with regulations; when the probability is 0, it indicates that the element was not declared in full compliance with regulations. For each customs declaration data point, several vectors are calculated for each element. spliced ​​into a matrix The average probability that each element is predicted to be positive is calculated, which is the score of the customs declaration.

[0049] As a further aspect of the present invention, the formula for calculating the score of the customs declaration is as follows:

[0050]

[0051] In the above formula (9), e represents the number of elements corresponding to the customs declaration.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] This invention extracts element fields from customs declarations and performs preliminary screening of the number of elements to obtain a customs declaration sample set. A two-stage sampling method is then used to generate a candidate label sample set. Next, a candidate label sample set reduction method based on class centers is used to obtain a training sample set after sample reduction. Finally, the FastText model is used to perform short text binary classification on the final training sample set to obtain the probability of correct customs declaration, thus completing the customs declaration quality assessment. This method neither overly relies on certain rules nor requires a large amount of manual review. It can identify loopholes in the regulatory system from massive amounts of declaration data, uncover abnormal customs declaration behavior patterns, efficiently and intelligently review customs declarations, identify risks in the declaration process, and achieve the function of automatically identifying customs declaration quality and providing real-time early warnings. This provides important clues for customs risk assessment departments and meets the business needs of customs monitoring declaration risks. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the data preprocessing process in a customs risk identification method for import and export commodities based on declaration quality assessment;

[0055] Figure 2 This is a flowchart illustrating the process of generating a selected sample set in a customs import and export commodity risk identification method based on declaration quality assessment.

[0056] Figure 3 This is a flowchart illustrating the process of obtaining a training sample set after sample reduction in a customs import and export commodity risk identification method based on declaration quality assessment.

[0057] Figure 4This is a flowchart illustrating the short text clustering process in a customs import and export commodity risk identification method based on declaration quality assessment.

[0058] Figure 5 This is a flowchart illustrating the process of calculating the quality score of customs declarations in a customs risk identification method for import and export commodities based on declaration quality assessment. Detailed Implementation

[0059] Please see Figures 1-5 In this embodiment of the invention, a method for identifying risks in customs import and export commodities based on declaration quality assessment includes the following steps:

[0060] S1. Data preprocessing: This includes feature extraction and feature quantity filtering, such as... Figure 1 As shown, firstly, the customs declaration element data is obtained through element extraction, and then non-standard customs declarations are filtered out from the customs declaration element data through element quantity filtering to obtain a customs declaration sample set.

[0061] S2. Sampling of the customs declaration sample set based on two-stage sampling: First, take the union of the data for each element in the customs declaration sample set to obtain the full data for each element, for example, element 1 sample set, element 2 sample set, and so on. The sample set was compiled and sorted by frequency of occurrence. However, manually labeling the entire dataset would be extremely costly; therefore, for each element, the most frequent elements were selected. % of the samples are directly included in the training samples as the high-frequency sample set; the remaining 1- % of the samples were used as the low-frequency sample set. Then, using a two-stage sampling method with equal probability, the remaining 1- A percentage of the samples are clustered into text clusters, such as cluster 1, cluster 2, ... cluster 3. Then, random sampling is performed in each cluster. %, to obtain the training sample set, where, %and The percentage can be adjusted between 0 and 1; then, the training sample set is merged with the high-frequency sample set to generate a candidate labeled sample set. Thus, a highly representative sample was successfully extracted;

[0062] S3. Candidate Label Sample Set Reduction Based on Class Centers: After sampling in step S2, it was found in practice that the data volume was still too large, and manual annotation remained a huge burden. Therefore, when the density of cluster centers is high, and these samples near the cluster centers are not very useful for classification decisions, these samples can be reduced. First, calculate the density of each cluster center... The cluster center; then calculate each cluster The distance from each sample point to the cluster center is used to obtain the candidate labeled sample set. The set of distances from each sample point to the cluster center is then arranged in order, and the closest distance is selected. Each sample yields the corresponding point set. Corresponding point samples To be reduced, each One sample is selected from the center and kept, the rest are discarded, resulting in... A set of sample points Customs inspectors perform specialized labeling, then take the union of the samples to obtain the training sample set after sample deletion. ,like Figure 3 As shown;

[0063] S4. Quality assessment of customs declarations using short text binary classification: The FastText model is used to evaluate the quality of the final training sample set. The short text is segmented and stop word removed. Word embedding is then used to process the remaining words into word feature vectors, for example... The word feature vectors are then used as the input layer of the model and fed into the neural network model for linear operations. The document information is represented by the mean of the word vectors in the input layer sequence, and the calculation result is used as the output of the hidden layer. Since classification is the ultimate goal, and predicting probabilities is more beneficial to the discrimination of customs declaration scoring, the common Softmax function is used to establish a mapping from the hidden layer to the output layer. The Softmax function is then normalized, and an unequal probability dropout method is embedded to suppress overfitting. Neurons are intentionally dropped to calculate the probabilities of compliant and non-compliant declarations, and the score of the customs declaration is calculated to complete the customs declaration quality assessment. The quality assessment index is the average of the correct declaration probabilities of all elements.

[0064] Preferably, the data sources for customs import and export commodities include element tables and customs declarations; the element table records the element list corresponding to each commodity number; in the customs declaration data, different types of commodities have different commodity numbers, and each commodity number corresponds to a standard element order. The useful fields on the customs declaration are "commodity name and specifications", which include the declaration elements of product name, brand and specifications. When making a standard declaration, the element order and quantity must be consistent with the element table.

[0065] The method for extracting elements in step S1 is as follows: First, extract the "commodity name and specifications" field from the customs declaration, then split the content of the field into strings, and then split the strings by the '|' character. Each customs declaration yields a set of element fields, which are the element data in the customs declaration. Due to the existence of irregular declarations, it cannot be guaranteed that the element data of the customs declaration corresponds completely with the element list in the element table. Therefore, it is necessary to filter the number of elements.

[0066] The method for filtering the number of elements in step S1 is as follows: Compare the number of element data in the customs declaration with the number of elements in the element table to determine if the number of elements corresponds. If they match, store them in the customs declaration sample set; if they do not match, directly filter out the customs declarations that do not match, and determine them as non-standard customs declarations. Since the number of declared elements in the vast majority of customs declarations is consistent with the number in the corresponding element list in the element table, it is only necessary to filter out the very few customs declarations with inconsistent numbers.

[0067] Preferably, in step S2, a pond sampling method is used to randomly sample from each cluster; the sampling method includes the following steps:

[0068] S201, Definition For the sample set, For the result set, for The size of the array, For each random number, The number of samples drawn. This represents the number of iterations.

[0069] S202, Take Before the set The number is filled into the set. ,from Start iterating and generating random numbers. , range ,like Then replace the value in R, that is ;

[0070] S203. Traversal ends, generating result array R; obtaining candidate labeled sample set. .

[0071] Preferably, the text clustering method in step S2 includes the following steps:

[0072] S21. Dataset processing: Using the NLTK toolkit, the dataset is sequentially processed by tokenization, stop word removal, part-of-speech tagging, and stemming.

[0073] S22. Feature Extraction: First, the processed dataset is mapped to a vector space and converted into a TF-IDF word frequency matrix Q. Then, the Laplacian matrix of the word frequency matrix Q is calculated. Finally, the eigenvectors corresponding to the first K eigenvalues ​​of the Laplacian matrix are obtained and matrix R is formed.

[0074] S23. Graph Clustering: A short text clustering algorithm based on Laplacian graphs, which will output the clustering results.

[0075] Preferably, the specific method for graph clustering in step S23 is as follows:

[0076] S231. Sort matrix R according to the size of its eigenvalues;

[0077] S232. Select the eigenvectors corresponding to the first K eigenvalues ​​of matrix R as the initial cluster centers to improve the clustering convergence speed.

[0078] S233. Assign each data object to the nearest cluster based on the cosine of the angle between the feature vectors.

[0079] S234. Recalculate cluster centers;

[0080] S235. If the algorithm converges, terminate the iteration and output the clustering results, such as cluster 1, cluster 2, ..., cluster n; otherwise, return to step S233.

[0081] Preferably, in step S3, the cluster The formula for calculating the cluster center is as follows:

[0082]

[0083] In the above formula (1), For clusters The candidate labeled sample set in for The dimension, For the i-th cluster center, for The dimension, for The number of data in the data;

[0084] cluster From each sample point to the cluster center The distance calculation formula is as follows:

[0085]

[0086] In the above formula (2), From sample point to cluster center The set of distances;

[0087] Final training sample set The union of the sets is as follows:

[0088]

[0089] In the above formula (3), for Each sample yields a corresponding set of points. for A set of sample points The final training sample set.

[0090] Preferably, the calculation formula for the hidden layer in step S4 is as follows:

[0091]

[0092] The formula for calculating the loss function for the mapping from the hidden layer to the output layer in step S4 is as follows:

[0093]

[0094]

[0095] The calculation formula for the normalization process using the Softmax function in step S4 is as follows:

[0096]

[0097] The formula for calculating the probability of a neuron being dropped in step S4 is as follows:

[0098]

[0099] In the above formula (8), when the word is a non-strong feature word, the probability of it being discarded is 0.5; when the word is a strong feature word, the probability of it being discarded is less than 0.5, which reduces the probability of strong feature words being discarded, thereby ensuring the status of strong feature words in the model; for example, the word "competition" is a strong feature word for sports.

[0100] Preferably, the calculation method for the score of the customs declaration in step S4 is as follows:

[0101] When the word feature vectors reach the output layer, dimensional vector Map to a vector whose values ​​are in the interval (0,1). Because this invention uses a binary classification method, When the probability of a predicted positive label is 1, it indicates that the element was declared in full compliance with regulations; when the probability is 0, it indicates that the element was not declared in full compliance with regulations. For each customs declaration data point, several vectors are calculated for each element. spliced ​​into a matrix The average probability that each element is predicted to be positive is calculated, which is the score of the customs declaration.

[0102] Preferably, the formula for calculating the score of the customs declaration is as follows:

[0103]

[0104] In the above formula (9), e represents the number of elements corresponding to the customs declaration.

[0105] To better illustrate the technical effects of the present invention, the following experiments are conducted:

[0106] In customs declarations, 50,000 data points were selected to construct training and test sets for repeated testing. To evaluate the effectiveness of the risk identification and classification method, the main metric is the accuracy of the model's label predictions. Therefore, this invention's experimental scheme uses the same amount of training data to train the model, and the test data uses accuracy and F1 score to evaluate the method's effectiveness. Accuracy refers to the proportion of correctly predicted data samples out of the total number of samples in the test set (or validation set). The F1 score, as an evaluation metric for the model, is a weighted average of precision and recall. In this experiment, 70% of the total data was used as the training dataset, and 30% as the test dataset. In the training dataset, other methods use simple random sampling to select 20% of the training dataset for training; the sampling data used in this invention is also set to 20% of the total training data. The average results of 10 repeated experiments are shown in Table 1.

[0107] Table 1. Performance comparison of each model under the condition of 20% total training samples.

[0108] Model Accuracy (ACC) F1 value CART decision tree algorithm 0.833 0.827 Naive Bayes 0.783 0.771 Logistic regression analysis 0.811 0.803 Support Vector Machine Method 0.759 0.743 The method of the present invention 0.878 0.877

[0109] As can be seen from Table 1 above, due to the selection of a more representative dataset, the method used in this invention has significantly better prediction performance than traditional classification algorithms.

[0110] The above description is merely 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 method for identifying risks in import and export commodities based on declaration quality assessment by customs, characterized in that, Includes the following steps: S1. Data preprocessing: This includes element extraction and element quantity screening. First, the element data of the customs declaration is obtained through element extraction. Then, the non-standard customs declarations in the element data are screened out through element quantity screening to obtain a customs declaration sample set. S2. Sampling of the customs declaration sample set based on two-stage sampling: First, take the union of the data for each element in the customs declaration sample set to obtain the full data for each element, and sort them according to their frequency of occurrence; then, for each element, sample the top elements with the highest frequency. % of the samples are directly included in the training samples as the high-frequency sample set; the remaining 1- % of the samples were used as the low-frequency sample set. Then, using a two-stage sampling method with equal probability, the remaining 1- A percentage of the samples were clustered into text clusters, and then random samples were taken from each cluster. %, to obtain the training sample set; Then, the training sample set is merged with the high-frequency sample set to generate a candidate labeled sample set. ; S3. Candidate label sample set reduction based on class center: First, calculate the sample set of each cluster. The cluster center; then calculate each cluster The distance from each sample point to the cluster center is used to obtain the candidate labeled sample set. The set of distances from each sample point to the cluster center is then arranged in order, and the closest distance is selected. Each sample yields the corresponding point set. Corresponding point samples To be reduced, each One sample is selected from the center and kept, the rest are discarded, resulting in... A set of sample points Customs inspectors perform specialized labeling, then take the union of the samples to obtain the training sample set after sample deletion. ; S4. Quality assessment of customs declarations using short text binary classification: The FastText model is used to evaluate the quality of the final training sample set. The short text is segmented and stop word removed. The remaining words are processed into word feature vectors through word embedding. These word feature vectors are then used as the input layer of the model and fed into a neural network model for linear operations. The calculation results are used as the output of the hidden layer. A Softmax function is then used to establish a mapping from the hidden layer to the output layer. The Softmax function is normalized, and an unequal probability neuron dropout method is embedded to selectively discard neurons. The probabilities of compliant and non-compliant declarations are calculated, and the score of the customs declaration is determined to complete the customs declaration quality assessment.

2. The customs risk identification method for import and export commodities based on declaration quality assessment according to claim 1, characterized in that, The method for extracting elements in step S1 is as follows: First, extract the "commodity name and specifications" field from the customs declaration, then split the content of the field into strings, and then split the strings by the '|' character. Each customs declaration yields a set of element fields, which are the element data in the customs declaration. The method for filtering the number of elements in step S1 is as follows: compare the number of element data in the customs declaration with the element list in the element table to determine whether the number of elements corresponds. When the two are consistent, store them in the customs declaration sample set. When the two are inconsistent, the inconsistent customs declaration will be directly screened out and judged as an irregular declaration.

3. The method for identifying customs import and export commodity risks based on declaration quality assessment according to claim 1, characterized in that, In step S2, a pond sampling method is used to randomly sample each cluster; the sampling method includes the following steps: S201, Definition For the sample set, For the result set, for The size of the array, For each random number, The number of samples drawn. This represents the number of iterations. S202, Take Before the set The number is filled into the set. ,from Start iterating and generating random numbers. , range ,like Then replace the value in R, that is ; S203. Traversal ends, generating result array R; obtaining candidate labeled sample set. .

4. The customs risk identification method for import and export commodities based on declaration quality assessment according to claim 1, characterized in that, The text clustering method in step S2 includes the following steps: S21. Dataset processing: Using the NLTK toolkit, the dataset is sequentially processed by tokenization, stop word removal, part-of-speech tagging, and stemming. S22. Feature Extraction: First, the processed dataset is mapped to a vector space and converted into a TF-IDF word frequency matrix Q. Then, the Laplacian matrix of the word frequency matrix Q is calculated. Finally, the eigenvectors corresponding to the first K eigenvalues ​​of the Laplacian matrix are obtained and matrix R is formed. S23. Graph Clustering: A short text clustering algorithm based on Laplacian graphs, which will output the clustering results.

5. The customs risk identification method for import and export commodities based on declaration quality assessment according to claim 4, characterized in that, The specific method for graph clustering in step S23 is as follows: S231. Sort matrix R according to the size of its eigenvalues; S232. Select the eigenvectors corresponding to the first K eigenvalues ​​of matrix R as the initial cluster centers to improve the clustering convergence speed. S233. Assign each data object to the nearest cluster based on the cosine of the angle between the feature vectors. S234. Recalculate cluster centers; S235. If the algorithm converges, terminate the iteration and output the clustering results; otherwise, return to step S233.

6. The method for customs risk identification of import and export commodities based on declaration quality assessment according to claim 1, characterized in that, In step S3, the cluster The formula for calculating the cluster center is as follows: In the above formula (1), For clusters The candidate labeled sample set in for The dimension, For the i-th cluster center, for The dimension, for The number of data in the data; cluster From each sample point to the cluster center The distance calculation formula is as follows: In the above formula (2), From sample point to cluster center The set of distances; Final training sample set The union of the sets is as follows: In the above formula (3), for Each sample yields a corresponding set of points. for A set of sample points The final training sample set.

7. The customs risk identification method for import and export commodities based on declaration quality assessment according to claim 1, characterized in that, The formula for calculating the hidden layer in step S4 is as follows: The formula for calculating the loss function for the mapping from the hidden layer to the output layer in step S4 is as follows: The calculation formula for the normalization process of the Softmax function in step S4 is as follows: The formula for calculating the probability of neuron dropping in step S4 is as follows: In the above formula (8), when the word is a non-strong feature word, the probability of it being discarded is 0.5; when the word is a strong feature word, the probability of it being discarded is less than 0.5, thus reducing the probability of strong feature words being discarded.

8. The customs risk identification method for import and export commodities based on declaration quality assessment according to claim 1, characterized in that, The calculation method for the score of the customs declaration in step S4 is as follows: When the word feature vectors reach the output layer, dimensional vector Map to a vector whose values ​​are in the interval (0,1). Because a binary classification method is used, When the probability of a predicted positive label is 1, it indicates that the element was declared in full compliance with regulations; when the probability is 0, it indicates that the element was not declared in full compliance with regulations. For each customs declaration data point, several vectors are calculated for each element. spliced ​​into a matrix The average probability that each element is predicted to be positive is calculated, which is the score of the customs declaration.

9. A method for identifying customs import and export commodity risks based on declaration quality assessment according to claim 8, characterized in that, The formula for calculating the score of the customs declaration is as follows: In the above formula (9), e represents the number of elements corresponding to the customs declaration.

Images