A method and system for identifying price fraud in imported goods

Abstract

The application belongs to the field of intelligent algorithm fraud identification, and proposes a method and system for import commodity price fraud identification, which improves the kernel extreme learning machine parameters by improving the sea gull algorithm, and improves the import commodity price fraud identification performance of the kernel extreme learning machine. First, the Sinusoidal chaos thought is introduced into the population initialization, the population initialization mode is changed, and the quality of the initial individual and the population diversity are enhanced; second, the migration process of the sea gull population is simulated in a linear decreasing manner by adding variables, so that the global optimization ability is better improved; third, the Levy flight strategy is introduced to disturb and mutate at the optimal solution position, so that the algorithm can jump out of the local space, finally solve the problems of population diversity, slow convergence speed and easy falling into local optimum of the sea gull algorithm, realize automatic selection of the kernel extreme learning machine parameters, and improve the import commodity price fraud identification precision.

Description

Technical Field

[0001] The present invention relates to the field of intelligent algorithm fraud identification, and in particular to a method and system for identifying price fraud of imported goods. Background Art

[0002] Price fraud on imported goods occurs when importers intentionally conceal the true price of goods when declaring them to customs, through methods such as misdeclaring quantities or prices, misdeclaring product names, and using false certificates of origin, in order to evade or reduce payable duties and other taxes. Customs tax fraud identification faces the following difficulties: First, the booming development of new business models such as cross-border e-commerce, and the hidden nature of these emerging service forms, pose challenges to customs fraud detection; second, incomplete and non-standard data on imported goods declarations cannot meet customs classification and price management requirements; third, taxpayers employ diverse tax evasion methods, characterized by grouping, industry-specific, and specialized practices. For example, similar product names are used to string together tax declaration numbers, challenging customs fraud detection capabilities.

[0003] Current research on fraud detection primarily focuses on credit card and e-commerce transactions, while research on import price fraud detection is insufficient. Early approaches to import price fraud detection were mostly rule-based expert systems, with decision rules typically employing "if-then" or "if-else-then" triggering conditions. However, with technological advancements, fraud methods and patterns continue to evolve, requiring constant adjustment of triggering rules, increasing the difficulty of fraud detection. Data mining primarily employs machine learning methods to automatically capture fraud patterns from historical data, characterizing the probability of each transaction being "potentially fraudulent," thereby improving the accuracy of fraud detection in diverse and complex transactions. The Kernel Extreme Learning Machine (KELM) possesses excellent nonlinear learning capabilities and generalization performance, but determining the optimal regularization coefficient and kernel parameters is difficult. Optimizing the parameters of machine learning methods based on intelligent algorithms improves their adaptability across various fields. Compared with other optimization algorithms, the Seagull Optimization Algorithm (SOA) has a simple structure, requires few parameters, and is easy to implement, which provides a possibility for kernel limit learning parameter optimization. However, the Seagull Algorithm has insufficient population diversity and is prone to falling into local optimality. Summary of the Invention

[0004] In order to solve the problem that the gull algorithm in the prior art is easy to fall into local optimum when optimizing the parameter of the kernel extreme learning machine model, and further makes the kernel extreme learning machine model have greater volatility in the recognition accuracy of the import commodity price fraud, and affects the stability and reliability of the model, the present application provides a method and system for identifying import commodity price fraud, solves the problems of insufficient population diversity, weak global optimization ability and easy to fall into local optimum of the gull algorithm, and realizes the optimal selection of the parameter of the kernel extreme learning machine, obtains a stable and reliable kernel extreme learning machine, and improves the recognition accuracy of the kernel extreme learning machine on the import commodity price fraud category.

[0005] The specific scheme is as follows:

[0006] A method for identifying import commodity price fraud, comprising

[0007] S1: data acquisition: acquiring import commodity price fraud identification data; judging whether there is price fraud according to the basic information of each data of the import commodity by the staff, and labeling; the data labeled as the data set of the subsequent step; the data format in the data set is [basic information of the import commodity, import commodity price fraud category label];

[0008] S2: data set balancing: balancing the data set acquired in S1 by the SMOTE-CGAN data set balancing method to obtain the 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: data set processing: performing maximum and minimum normalization processing on the balanced data set in S2 to obtain the processed data set;

[0010] S4: determining the training set, the validation set and the test set: dividing the processed data set in S3 into the training set, the validation set and the test set according to the set proportion;

[0011] S5: model establishment and training: establishing an import commodity price fraud identification model based on the improved gull algorithm optimized kernel extreme learning machine; the import commodity price fraud identification model is a kernel extreme learning machine, the basic information of the import commodity in the training set is taken as the input of the kernel extreme learning machine, the import commodity price fraud category label is taken as the output, and the initialized kernel extreme learning machine parameter is trained; the kernel extreme learning machine parameter includes a regularization coefficient and a kernel parameter;

[0012] S6: Optimization model based on the Seagull algorithm: The improved Seagull algorithm is used to optimize the kernel extreme learning machine parameters to obtain the optimal kernel extreme learning machine parameters; based on the sinusoidal chaotic map, the initialized kernel extreme learning machine parameters are set to the initial positions of the seagulls in the Seagull algorithm, and the migration process of the seagull population is simulated in a nonlinear decreasing manner; finally, the Levy flight strategy is introduced to perform perturbation mutation on the optimal seagull position, and the optimized seagull position is obtained as the optimized kernel extreme learning machine parameters, and the optimized imported commodity price fraud identification model is evaluated;

[0013] S7: Evaluate the model based on the test set: Input the S4 test set into the imported goods price fraud identification model optimized by S6 for prediction to obtain the imported goods price fraud category; and compare it with the real imported goods price fraud category label, calculate the accuracy, recall rate, precision rate, and F1 value as model performance evaluation indicators, and evaluate it.

[0014] Preferably, the specific method of the Seagull algorithm-based optimization model is:

[0015] S61: Initialization of parameters of the improved Seagull algorithm: Setting parameters in the improved Seagull algorithm, including population size, maximum number of iterations, search dimension, upper and lower limits of optimization target parameters, and spiral coefficient parameters;

[0016] S62: Initializing the initial positions of seagulls using chaotic mapping: Initializing the initial positions of seagulls using a sinusoidal chaotic mapping to obtain the initial positions of multiple seagulls;

[0017] S63: Determine the fitness of the seagulls: Input the S4 training set into the kernel extreme learning machine, and use the initial position of each seagull in S62 to represent the hyperparameters to train the kernel extreme learning machine. Then input the S4 validation set into the trained kernel extreme learning machine to obtain the imported goods price fraud category and calculate the recognition accuracy. The recognition accuracy is used as the fitness of each seagull. By comparing the fitness of all seagulls, the globally optimal seagull individual is determined.

[0018] S64: Update the seagull position according to the migration operation: calculate the new position of the seagull after the seagull individual migrates to the global optimal seagull individual; adjust the seagull position through the nonlinear control factor A during the migration process;

[0019] S65: Update the seagull position according to the attack operation: Update the new position of the seagull obtained in S64 again according to the attack operation integrated into Levi's flight to obtain the final updated position of the seagull;

[0020] S66: Find the optimal individual: Based on the final updated position of the seagull in S65, compare the fitness values ​​between the seagull individuals again and find the global optimal seagull individual;

[0021] S67: Determine the termination condition: Determine whether the termination condition is met. If so, use the global optimal seagull position as the optimal kernel extreme learning machine parameter; otherwise jump to S64; the termination condition is whether the number of iterations reaches the maximum number of iterations set in S61.

[0022] Preferably, the expression of the Sinusoidal chaotic map is as follows:

[0023]

[0024] Among them, the value range of a is [0,1], and the value range of x is [0,1].

[0025] Preferably, the nonlinear control factor A changes nonlinearly and is affected by the parameter h. The specific expression of the nonlinear control factor A is as follows:

[0026]

[0027] Where h is a random number between [0,1]; M represents the maximum number of iterations; C f represents a positive integer; t represents the number of iterations.

[0028] Preferably, the method for updating the seagull position according to the migration operation is:

[0029] A1: Determine the new position of the seagull to avoid collisions: The nonlinear control factor A is introduced to adjust the position of the seagull to avoid collisions between adjacent seagulls. The formula for determining the new position of the seagull is as follows:

[0030] C s (t) = A × P s (t) (3)

[0031] Among them, C s (t) represents the new position of the seagull; P s (t) represents the initial position of the seagull; t represents the number of iterations;

[0032] A2: Determine the direction of movement toward the optimal seagull individual: After determining the optimal seagull individual, on the basis of avoiding collisions with adjacent seagulls as in A1, the seagull moves toward the optimal seagull individual. The formula for determining the direction of movement toward the optimal seagull individual is as follows:

[0033] M s (t) = C B ×(P bs (t)-P s (t)) (4)

[0034] C B=2×A 2 ×rand (5)

[0035] Among them, M s (t) represents the direction in which the seagull individual moves toward the optimal seagull individual; C B is a random number used to balance global search and local search; P bs (t) represents the best individual seagull position; where rand is a random number in the range [0,1] that obeys a uniform distribution;

[0036] A3: Update the latest position of the seagull: Based on the new position of the seagull determined in A1 and the movement direction towards the optimal seagull determined in A2, the seagull individual reaches a new position during the continuous migration process. The final update formula for the new position is as follows:

[0037] D s (t) = C s (t)+M s (t)(6)

[0038] Among them, D s (t) represents the distance between the seagull and the optimal seagull.

[0039] Preferably, the updating of the seagull position again according to the attack operation integrated into Levi's flight is expressed as:

[0040] P s (t) = D s (t)×x×y×z×levy(a)+P bs (t) (7)

[0041] Where levy(a) is a random number vector based on the Levy distribution, and x, y, and z represent the coordinates of the seagull's spiral movement in the air during the attack operation. The specific expression is:

[0042] x=R×cos(i) (8)

[0043] y=R×sin(i) (9)

[0044] z=R×i (10)

[0045] R=u×e iv (11)

[0046] Where R is the spiral radius of the seagull; i is a random number in the range [0, 2π] that follows a uniform distribution, representing the seagull's attack angle; u and v are the spiral factors; e is the base of the natural logarithm;

[0047] The Lévy flight is a random method with random walk characteristics. The Lévy flight can be expressed as:

[0048]

[0049] Among them, parameter α represents a random number, Γ is the Gamma function, Γ(1+a)=a!; r1 and r2 are respectively r1 and σ r2 Normally distributed random numbers.

[0050] Preferably, the method for finding the optimal individual described in step S66 is: using the final updated position of each seagull as the new parameter of the kernel extreme learning machine to update the parameters of the kernel extreme learning machine; inputting the verification set into the kernel extreme learning machine after parameter update to obtain the category of price fraud of imported goods and calculate the recognition accuracy; using the obtained recognition accuracy as the new fitness of each seagull, and comparing them to determine the new global optimal seagull individual.

[0051] Preferably, the kernel extreme learning machine uses a kernel function mapping operation to replace the random mapping method used by the extreme learning machine in the hidden layer, wherein the output equation of the extreme learning machine is:

[0052]

[0053] Among them, I is the unit matrix, C is the regularization coefficient, T is the expected output, h(x) represents the output of the hidden layer, and H represents the output matrix of the hidden layer; through the Mercer condition, the kernel matrix Ω KELM Replace HH T , kernel matrix Ω KELM Expressed as:

[0054] Ω KELM =HH T (17)

[0055] Kernel matrix Ω KELM The element in row i and column j is:

[0056]

[0057] Among them, K(x i ·x j ) is the kernel function. From formula (17) and formula (18), we can know that HH T The expression is:

[0058]

[0059] h(x)H T Use the kernel function K(x i ,x j )The expression after replacement is:

[0060]

[0061] When h(x)H in formula (16) T Expressed in formula (20), HH T Expressed as (17), the KELM model f KELM (x) The output is:

[0062]

[0063] Among them, K(g) represents the kernel function, Ω KELM Represents the kernel function matrix, xLx N Represents training samples of price fraud on imported goods.

[0064] Preferably, the kernel function of the kernel extreme learning machine includes:

[0065] Linear kernel function:

[0066]

[0067] Gaussian kernel function:

[0068]

[0069] Among them, ‖·‖ represents the Euclidean distance; σ is the width parameter of the kernel function;

[0070] Polynomial kernel function:

[0071]

[0072] Where b represents the constant term; d represents the degree of the polynomial;

[0073] Sigmoid kernel function:

[0074]

[0075] Where tanh represents the hyperbolic tangent function, β represents the slope parameter, and γ represents the intercept parameter.

[0076] Preferably, including:

[0077] Data acquisition module: obtains import commodity price fraud identification data;

[0078] Data preprocessing module: normalize the data to its maximum and minimum values;

[0079] Imported commodity price fraud identification module: uses the imported commodity price fraud identification model to identify imported commodity price fraud on the acquired data. The commodity price fraud identification model is the optimal imported commodity price fraud identification model.

[0080] Beneficial effects:

[0081] The application provides a method and system for identifying import commodity price fraud, and solves the problems of poor population diversity, weak global optimization ability and easy falling into local optimum of a gull algorithm by fusing an improved gull optimization algorithm (SWLSOA) of Sinusoidal chaos initialization population, a nonlinear inertia weight strategy and a Lev flight strategy. Secondly, the regularization coefficient and the kernel parameter of a kernel extreme learning machine (KELM model) are optimized by introducing the SWLSOA algorithm, the optimal selection of the parameters of the kernel extreme learning machine is realized through simulation and comparison experiments, human factors are avoided, and the precision of the kernel extreme learning machine for identifying import commodity price fraud is improved. Specifically,

[0082] (1) The application adopts an SMOTE-CGAN-based import commodity price fraud identification data set balancing method, generates few-class samples, realizes effective expansion and balancing of the import commodity price fraud data set, and improves the accuracy of import commodity price fraud transaction identification.

[0083] (2) The application provides an import commodity price fraud identification method based on an improved gull algorithm optimized kernel extreme learning machine (SWLSOA-KELM), and the problems of poor population diversity and easy falling into local optimum of the gull algorithm are solved. Firstly, the Sinusoidal chaos initialization population is used to change the population initialization mode, enhance the quality of initial individuals and population diversity. Secondly, in the updating process of the gull individual position, the variable A is added in a linearly decreasing manner to simulate the migration process of the gull population, and the global optimization ability is better improved. Finally, the Lev flight strategy is introduced to disturb and mutate at the optimal solution position, and the ability of the algorithm to jump out of the local space is enhanced.

[0084] (3) The application divides the import commodity price fraud data set into a training set, a validation set and a test set. The training set and the test set are used to train and optimize the kernel extreme learning machine, respectively, and the regularization coefficient and the kernel parameter of the kernel extreme learning machine are updated. The performance of the kernel extreme learning machine after updating the parameters is evaluated through the validation set, the import commodity price fraud identification accuracy is taken as a standard, and finally the kernel extreme learning machine with the highest identification accuracy is determined as the optimal model. The test set is used to test and compare the optimal model, and the kernel extreme learning machine and the gull algorithm optimized kernel extreme learning machine (SOA-KELM) are compared, and it is found that the SWLSOA-KELM has the best recognition effect.

[0085] (4) The present invention optimizes the regularization coefficient and kernel parameters of KELM through SWLSOA, regards the optimized regularization coefficient and kernel parameters as the position of the seagull, regards the optimization process of the regularization coefficient and kernel parameters of the kernel extreme learning machine as the behavior of the seagull migrating and attacking prey in search of abundant food sources, takes the classification accuracy of the verification set as the fitness function, and judges whether the individual position is optimal through the fitness. When the algorithm ends, the optimal regularization coefficient and kernel parameters of KELM are obtained, and the optimal SWLSOA-KELM import commodity price fraud transaction identification method is constructed to achieve accurate identification of import commodity price fraud transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 Flowchart of a method for identifying price fraud on imported goods.

[0087] Figure 2 A flowchart of the overall implementation of a method for identifying price fraud on imported goods.

[0088] Figure 3 Improved Seagull algorithm flowchart.

[0089] Figure 4 Import commodity price fraud identification model based on SWLSOA-KELM Import commodity price fraud identification confusion matrix.

[0090] Figure 5 An architecture diagram of a system for identifying price fraud on imported goods. DETAILED DESCRIPTION

[0091] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0092] like Figure 1 As shown, the specific steps of the present invention include:

[0093] A method for identifying price fraud of imported goods, comprising:

[0094] S1: Data Acquisition: Obtain data identifying price fraud on imported goods. Staff will determine whether each piece of data involves price fraud based on the basic information of the imported goods and assign a label. The labeled data will serve as the dataset for subsequent steps. The data format in the dataset is [basic information of the imported goods, price fraud category label for the imported goods].

[0095] 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;

[0096] S3: Dataset processing: Perform maximum and minimum normalization on the data set balanced in S2 to obtain the processed data set;

[0097] S4: Determine the training set, validation set, and test set: Divide the dataset processed in S3 into a training set, validation set, and test set according to the set ratio;

[0098] S5: Model establishment and training: Establishing an imported commodity price fraud identification model based on an improved Seagull algorithm-optimized kernel extreme learning machine; the imported commodity price fraud identification model is a kernel extreme learning machine, which uses basic information of imported commodities in a training set as input to the kernel extreme learning machine and an imported commodity price fraud category label as output, and trains to obtain initialized kernel extreme learning machine parameters; the kernel extreme learning machine parameters include a regularization coefficient and a kernel parameter;

[0099] S6: Optimization model based on the Seagull algorithm: The improved Seagull algorithm is used to optimize the kernel extreme learning machine parameters to obtain the optimal kernel extreme learning machine parameters; based on the sinusoidal chaotic map, the initialized kernel extreme learning machine parameters are set to the initial positions of the seagulls in the Seagull algorithm, and the migration process of the seagull population is simulated in a nonlinear decreasing manner; finally, the Levy flight strategy is introduced to perform perturbation mutation on the optimal seagull position, and the optimized seagull position is obtained as the optimized kernel extreme learning machine parameters, and the optimized imported commodity price fraud identification model is evaluated;

[0100] S7: Evaluate the model based on the test set: Input the S4 test set into the imported goods price fraud identification model optimized by S6 for prediction to obtain the imported goods price fraud category; and compare it with the real imported goods price fraud category label, calculate the accuracy, recall rate, precision rate, and F1 value as model performance evaluation indicators, and evaluate it.

[0101] like Figure 2 The specific contents include:

[0102] 1. Obtain a dataset for identifying price fraud on imported goods. The data format in the dataset is [basic information of imported goods, category label of price fraud on imported goods].

[0103] 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.

[0104] 2. Import commodity price fraud data balance: The acquired dataset is balanced using the SMOTE-CGAN dataset balancing method to obtain a balanced dataset;

[0105] 3. Normalization of the imported goods price fraud dataset:

[0106] Normalize the data of the imported commodity price fraud indicator to the range [0,1]. The processing process is as follows:

[0107]

[0108] Among them, x(k) is the original data, x(k) min is the minimum value of the original data, x(k) max is the maximum value of the original data, and y(k) is the normalized data.

[0109] 4. Divide the dataset into training set, validation set and test set in a certain ratio, with the ratio set to 7:2:1;

[0110] 5. Use the training set and validation set of imported product price fraud to train the kernel extreme learning machine and obtain the optimal kernel extreme learning machine model;

[0111] A model for identifying price fraud in imported goods was established based on an extreme learning machine optimized with the improved Seagull algorithm, namely the kernel extreme learning machine. The kernel extreme learning machine was trained using the training and validation sets of imported goods price fraud to obtain the optimal kernel extreme learning machine model.

[0112] Specifically, the basic idea of ​​identifying price fraud of imported goods based on the improved seagull algorithm and optimized kernel extreme learning machine is to regard the regularization coefficient and kernel parameters of the extreme learning machine as the seagull individuals in the seagull algorithm, and to regard the optimization process of the regularization coefficient and kernel parameters of the kernel extreme learning machine as the behavior of seagulls migrating and attacking prey in search of abundant food sources.

[0113] The kernel extreme learning machine replaces the random mapping method used by the extreme learning machine (ELM) to the hidden layer by using a kernel function mapping operation. This can improve the robustness of the model. In addition, this method can make the ELM more efficient and reduce the computational overhead during the training process. The output equation of the ELM is:

[0114]

[0115] Among them, I is the unit matrix, C is the regularization coefficient, T is the expected output, h(x) represents the output of the hidden layer, and H represents the output matrix of the hidden layer; through the Mercer condition, the kernel matrix Ω KELM Replace HH T , kernel matrix Ω KELM Expressed as:

[0116] Ω KELM =HH T (17)

[0117] Kernel matrix Ω KELM The element in row i and column j is:

[0118]

[0119] Among them, K(x i ·x j ) is the kernel function. From formula (17) and formula (18), we can know that HH T The expression is:

[0120]

[0121] h(x)H T Use the kernel function K(x i ,x j )The expression after replacement is:

[0122]

[0123] When h(x)H in formula (16) T Expressed in formula (20), HH T Expressed as (17), the KELM model f KELM (x) The output is:

[0124]

[0125] Among them, K(g) represents the kernel function, Ω KELM Represents the kernel function matrix, xLx N Represents training samples of price fraud on imported goods.

[0126] For kernel extreme learning machines, different kernel functions are selected to effectively balance model accuracy and computational complexity. Currently, the commonly used kernel functions for extreme learning machines include linear kernel function, Gaussian kernel function, polynomial kernel function, and Sigmoid kernel function.

[0127] Linear kernel function:

[0128]

[0129] Gaussian kernel function:

[0130]

[0131] Polynomial kernel function:

[0132]

[0133] Sigmoid kernel function:

[0134]

[0135] At present, the usual practice of selecting kernel functions is to select them according to the actual application effect. The identification of price fraud of imported goods has the characteristics of multivariate, nonlinearity, and concealment. The Gaussian function has the advantages of good ability to handle nonlinearity and low computational complexity. Therefore, the Gaussian kernel function is used as the kernel function of the kernel extreme learning machine, which gives it good learning and generalization capabilities.

[0136] Specifically, the improved Seagull algorithm implementation process is as follows Figure 3 Shown, including:

[0137] Step 1: Initialize the relevant parameters of the Seagull optimization algorithm. Specifically, set the population size to 20, the maximum number of iterations to 50, the search dimension to 2, and the upper and lower limits of the optimization target parameters: lower limit: lb = [0.01, 1]; upper limit: ub = [50, 10]; spiral coefficient parameters: u and v are used to control the spiral shape, and u and v are set to 1;

[0138] According to formula (1), the population is initialized based on the Sinusoidal chaotic map;

[0139] Specifically, the Seagull Algorithm uses a random generation method to initialize the population, which will make the distribution of the seagull population uneven, affecting the subsequent iterative optimization. Chaotic mapping has the characteristics of randomness, ergodicity, and regularity, and can traverse all states without repeating within a certain range. The chaotic sequence can be used to initialize the position of individual seagulls. Sinusoidal mapping is a typical representative of chaotic mapping. The mathematical form of Sinusoidal mapping is as follows:

[0140]

[0141] Among them, the value range of a is [0,1], and the value range of x is [0,1].

[0142] Step 2: Calculate the fitness value of each seagull and mark the best individual;

[0143] Specifically, the classification accuracy of the training set is used as the fitness function, and the calculation formula is as follows (27): by improving the Seagull algorithm to find the optimal iteration and satisfy the convergence condition or the maximum number of iterations, the search is terminated, and the optimal regularization coefficient and kernel parameter of the kernel extreme learning machine are obtained, so that the performance of the imported commodity price fraud identification model optimized by the improved Seagull algorithm and the kernel extreme learning machine is optimal.

[0144]

[0145] Among them, TP represents the number of correctly identified positive examples, FP represents the number of incorrectly classified positive examples, FN represents the number of incorrectly classified negative examples, and TN represents the number of correctly classified negative examples.

[0146] Step 3: Update the seagull migration process according to equations (2), (3), (4), (5) and (6), and calculate the new position of the seagull individual after it moves to the optimal position;

[0147] Specifically, the seagull migration model is designed by imitating the migration behavior of seagull groups. At this stage, three conditions need to be met: avoiding collisions between seagulls, approaching the best seagull individual, and updating the position through the best seagull individual.

[0148] (1) Avoid collision

[0149] To prevent collisions between individual seagulls, the Seagull Algorithm introduces a control factor, A, which adjusts the positions of the seagulls to avoid collisions between adjacent gulls. This factor, to a certain extent, balances global and local search. During seagull migration, if the additional control factor A were to simulate the gulls' migration in a linearly decreasing manner, the algorithm's optimization capability would be reduced. Therefore, the proposed Seagull Optimization Algorithm with a nonlinearly decreasing control factor offers two key improvements. First, the nonlinearly varying additional variable, A, achieves a sharp decrease in the early stages and a slow decrease in the later stages, accelerating the algorithm's convergence and better matching the actual migration process of seagull populations, resulting in more accurate search results. Second, the value of the additional variable A is influenced by h, increasing the randomness of its value. The random exponent is similar to the mutation factor in a genetic algorithm, which helps maintain population diversity.

[0150] In this method, the value of A shows a nonlinear trend in the decreasing process, which can better improve the global optimization ability. Each iteration can avoid position conflicts between seagulls and better balance exploration and development.

[0151] The formula for updating the seagull position is as follows:

[0152]

[0153] C s (t) = A × P s (t) (3)

[0154] Among them, C s (t) represents the new position of the seagull; P s (t) represents the initial position of the seagull; the value of A is obtained by C f Linearly regulated, its value is from C f Decrease linearly to 0, Cf The value of t is usually set to 2; t represents the number of iterations; M represents the maximum number of iterations; h represents a random index, which is a random number between 0 and 1.

[0155] (2) Optimal position direction

[0156] After determining the migration site, the seagull moves in the optimal position direction while avoiding collision with other individuals. At this time, the seagull motion behavior update formula is as follows:

[0157] M s (t)=C B ×(P bs (t)-P s (t)) (4)

[0158] C B =2×A 2 ×rand (5)

[0159] Wherein, M s (t) represents the direction of the seagull individual moving towards the best seagull individual; C B is a random number, used to balance global search and local search; P bs (t) represents the position of the best seagull individual. Where rand is a random number in the range of [0, 1] subject to uniform distribution.

[0160] (3) Optimal position update

[0161] During the continuous migration movement of the seagull individual, it will continuously approach the best position direction and reach a new position. The new position update formula is as follows:

[0162] D s (t)=C s (t)+M s (t) (6)

[0163] Wherein, D s (t) represents the distance between the seagull and the optimal seagull.

[0164] Step 4: Update the attack behavior of the seagull according to formula (7), formula (8), formula (9), formula (10) and formula (11), and calculate the final updated position of the seagull;

[0165] Specifically, the seagull will use its weight and wing beat frequency to reach the desired height during continuous migration movement, and will continuously change the attack angle and speed according to the actual needs. If the seagull performs an attack behavior, it will take a spiral shape motion in the air. At this time, the seagull motion behavior update formula in three-dimensional space is as follows:

[0166] x=R×cos(i) (8)

[0167] y=R×sin(i) (9)

[0168] z=R×i (10)

[0169] R=u×e iv (11)

[0170] Where R is the spiral radius of the seagull; i is a random number in the range [0, 2π] that follows a uniform distribution, representing the seagull's attack angle; u and v are spiral factors, usually set to 1; and e is the base of the natural logarithm.

[0171] To enhance the Seagull Optimization Algorithm's ability to escape local optima, the Lévy flight strategy was incorporated into the Seagull Optimization Algorithm. Lévy flight was incorporated into the process of the seagull population moving toward the optimal individual position, allowing the seagulls to randomly wander near the optimal individual position. This mitigates the Seagull Optimization Algorithm's tendency to converge prematurely, effectively improving population diversity and enhancing the algorithm's ability to escape local optima. The updated position algorithm for the seagull population after adding Lévy flight is as follows:

[0172] P s (t) = D s (t)×x×y×z×levy(a)+P bs (t) (7)

[0173] Where levy(a) is a random number vector based on the Levy distribution.

[0174] Lévy flight refers to a random method with random walk characteristics, which is often used in the optimization process of metaheuristic algorithms. In uncertain environments, the Lévy flight strategy is beneficial to improving the search efficiency of the algorithm. The Lévy flight can be expressed as:

[0175]

[0176]

[0177] Wherein, parameter α = 1.5, Γ is the Gamma function, Γ(1+a) = a!; r1 and r2 are the standard deviation σ r1 and σ r2 Normally distributed random number, σ r2 =1.

[0178] Step 5: Compare the fitness values ​​of each seagull individual again and find the global optimal seagull individual;

[0179] Step 6: Determine whether the algorithm has reached the maximum number of iterations. If so, the algorithm terminates; otherwise, loop steps 3 to 6 until the algorithm reaches the maximum number of iterations.

[0180] Step 7: Output the best position and fitness value of the seagull;

[0181] Specifically, the final parameters of the best position of Seagull are: regularization coefficient is 49.84, kernel parameter is 1.05

[0182] 6. Input the test set into the optimal kernel extreme learning machine to obtain the classification results of imported commodity price fraud identification;

[0183] 7. Evaluate the fraud classification of the test set: Use four statistical indicators, including accuracy, recall, precision, and F1-score, to evaluate the fraud classification results; use the evaluation results to evaluate the recognition performance of the model.

[0184] Specifically, in order to verify the recognition effect of the SWLSOA-KELM model, two models were selected for comparison with the SWLSOA-KELM model. Control group 1: Import commodity price fraud recognition model based on KELM; Control group 2: Import commodity price fraud recognition model based on SOA-KELM; Control group 3: Import commodity price fraud recognition model based on SWLSOA-KELM. The collected import commodity price fraud data set was balanced by SMOTE-CGAN, and then according to the technical solution of the import commodity price fraud recognition method based on the improved Seagull algorithm provided by the present invention, control group 1 and control group 2 were used respectively to obtain the import commodity price fraud recognition type, and the fraud recognition type was evaluated and analyzed with the actual value. The results are shown in Table 1.

[0185] Table 1 Comparison of the accuracy of price fraud identification for imported goods by different evaluation models

[0186]

[0187]

[0188] From Table 1 and Figure 4 The SWLSOA-KELM model achieved the highest overall accuracy for identifying price fraud on imported goods, at 92.50%, while the KELM model achieved the lowest overall accuracy, at 80.50%. Compared to the standard KELM model, the SOA-KELM and SWLSOA-KELM models improved their overall accuracy by 14.91% and 3.45%, respectively. This demonstrates that optimizing KELM parameters using the SOA and SWLSOA algorithms can effectively improve the models' accuracy for identifying price fraud on imported goods.

[0189] Training time varied across models. Compared to the KELM model, the SOA-KELM and SWLSOA-KELM models all took varying degrees longer to train on the sample data, with training times of 225.73 and 223.0 seconds, respectively. This suggests that SOA and SWLSOA both increase training time to some extent when optimizing KELM parameters.

[0190] A system for identifying price fraud of imported goods, such as Figure 5 Shown, including:

[0191] Data acquisition module: obtains import commodity price fraud identification data;

[0192] Data preprocessing module: normalize the data to its maximum and minimum values;

[0193] Imported commodity price fraud identification module: uses the imported commodity price fraud identification model to identify imported commodity price fraud on the acquired data. The commodity price fraud identification model is the optimal imported commodity price fraud identification model.

[0194] Finally, it should be noted that the above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, those skilled in the art should understand that the technical solution of the present invention may 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, characterized in that: S1: Data Acquisition: Obtain data identifying price fraud on imported goods. Staff will determine whether each piece of data involves price fraud based on the basic information of the imported goods and assign a label. The labeled data will serve as the dataset for subsequent steps. The data format in the dataset is [basic information of the imported goods, price fraud category label for the imported goods]. 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 normalization 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 dataset processed in S3 into a training set, validation set, and test set according to the set ratio; S5: Model establishment and training: Establishing an imported commodity price fraud identification model based on an improved Seagull algorithm-optimized kernel extreme learning machine; the imported commodity price fraud identification model is a kernel extreme learning machine, which uses basic information of imported commodities in a training set as input to the kernel extreme learning machine and an imported commodity price fraud category label as output, and trains to obtain initialized kernel extreme learning machine parameters; the kernel extreme learning machine parameters include a regularization coefficient and a kernel parameter; S6: Optimization model based on improved Seagull algorithm: S61: Initialization of parameters of the improved Seagull algorithm: Setting parameters in the improved Seagull algorithm, including population size, maximum number of iterations, search dimension, upper and lower limits of optimization target parameters, and spiral coefficient parameters; S62: Initializing the initial positions of seagulls using chaotic mapping: Initializing the initial positions of seagulls using a sinusoidal chaotic mapping to obtain the initial positions of multiple seagulls; S63: Determine the fitness of the seagulls: Input the S4 training set into the kernel extreme learning machine, and use the initial position of each seagull in S62 to represent the hyperparameters to train the kernel extreme learning machine. Then input the S4 validation set into the trained kernel extreme learning machine to obtain the imported goods price fraud category and calculate the recognition accuracy. The recognition accuracy is used as the fitness of each seagull. By comparing the fitness of all seagulls, the globally optimal seagull individual is determined. S64: Update the seagull position according to the migration operation: calculate the new position of the seagull after the seagull individual migrates to the global optimal seagull individual; During the migration process, the position of the seagull is adjusted by the nonlinear control factor A; S65: Update the seagull position according to the attack operation: Update the new position of the seagull obtained in S64 again according to the attack operation integrated into Levi's flight to obtain the final updated position of the seagull; S66: Find the optimal individual: Based on the final updated position of the seagull in S65, compare the fitness values ​​between the seagull individuals again and find the global optimal seagull individual; S67: Determine termination condition: Determine whether the termination condition is met. If so, use the global optimal seagull position as the optimal kernel extreme learning machine parameter; otherwise, jump to S64; the termination condition is whether the number of iterations reaches the maximum number of iterations set in S61; and evaluate the optimized imported commodity price fraud identification model; S7: Evaluate the model based on the test set: Input the S4 test set into the imported goods price fraud identification model optimized in S6 to make predictions and obtain the imported goods price fraud category; The results were compared with the price fraud category labels of real imported goods, and the accuracy, recall, precision and F1 value were calculated as the model performance evaluation indicators and evaluated.

2. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The expression of the Sinusoidal chaotic map is as follows: Among them, the value range of a is [0,1], and the value range of x is [0,1].

3. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The nonlinear control factor A changes nonlinearly and is affected by the parameter h. The specific expression of the nonlinear control factor A is as follows: Where h is a random number between [0,1]; M represents the maximum number of iterations; C f represents a positive integer; t represents the number of iterations.

4. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The method for updating the seagull position according to the migration operation is: A1: Determine the new position of the seagull to avoid collisions: The nonlinear control factor A is introduced to adjust the position of the seagull to avoid collisions between adjacent seagulls. The formula for determining the new position of the seagull is as follows: C s (t)=A×P s (t) (3) Among them, C s (t) represents the new position of the seagull; P s (t) represents the initial position of the seagull; t represents the number of iterations; A2: Determine the direction of movement toward the optimal seagull individual: After determining the optimal seagull individual, on the basis of avoiding collisions with adjacent seagulls as in A1, the seagull moves toward the optimal seagull individual. The formula for determining the direction of movement toward the optimal seagull individual is as follows: M s (t)=C B ×(P bs (t)-P s (t)) (4) C B =2×A 2 ×rand (5) Among them, M s (t) represents the direction in which the seagull individual moves toward the optimal seagull individual; C B is a random number used to balance global search and local search; P bs (t) represents the best individual seagull position; where rand is a random number in the range [0,1] that obeys a uniform distribution; A3: Update the latest position of the seagull: Based on the new position of the seagull determined in A1 and the movement direction towards the optimal seagull determined in A2, the seagull individual reaches a new position during the continuous migration process. The final update formula for the new position is as follows: D s (t)=|C s (t)+M s (t)| (6) Among them, D s (t) represents the distance between the seagull and the optimal seagull.

5. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The above-mentioned updating of the seagull position according to the attack operation integrated into Levi's flight is expressed as: P s (t)=D s (t)×x×y×z×levy(a)+P bs (t) (7) Where levy(a) is a random number vector based on the Levy distribution, and x, y, and z represent the coordinates of the seagull's spiral movement in the air during the attack operation. The specific expression is: x=R×cos(i) (8) y=R×sin(i) (9) z=R×i (10) R=u×e iv (11) Where R is the spiral radius of the seagull; i is a random number in the range [0, 2π] that follows a uniform distribution, representing the seagull's attack angle; u and v are the spiral factors; e is the base of the natural logarithm; The Lévy flight is a random method with random walk characteristics. The Lévy flight can be expressed as: Among them, parameter α represents a random number, Γ is the Gamma function, Γ(1+a)=a!; r1 and r2 are respectively r1 and σ r2 Normally distributed random numbers.

6. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The method for finding the optimal individual described in step S66 is: using the final updated position of each seagull as the new parameter of the kernel extreme learning machine to update the parameters of the kernel extreme learning machine; inputting the verification set into the kernel extreme learning machine after parameter update to obtain the price fraud category of imported goods and calculate the recognition accuracy; using the obtained recognition accuracy as the new fitness of each seagull, and comparing them to determine the new global optimal seagull individual.

7. A method for identifying price fraud of imported goods according to claim 1, characterized in that: The kernel extreme learning machine uses a kernel function mapping operation to replace the random mapping method used by the extreme learning machine in the hidden layer, where the output equation of the extreme learning machine is: Among them, I is the unit matrix, C is the regularization coefficient, T is the expected output, h(x) represents the output of the hidden layer, and H represents the output matrix of the hidden layer; through the Mercer condition, the kernel matrix Ω KELM Replace HH T , kernel matrix Ω KELM Expressed as: Oh KELM =HH T (17) Kernel matrix Ω KELM The element in row i and column j is: Among them, K(x i ·x j ) is the kernel function. From formula (17) and formula (18), we can know that HH T The expression is: h(x)H T Use the kernel function K(x i ,x j )The expression after replacement is: When h(x)H in formula (16) T Expressed in formula (20), HH T Expressed as (17), the KELM model f KELM (x) The output is: Among them, K(g) represents the kernel function, Ω KELM represents the kernel function matrix, xL x N Represents training samples of price fraud on imported goods.

8. A method for identifying price fraud of imported goods according to claim 7, characterized in that: The kernel function of the kernel extreme learning machine includes: Linear kernel function: Gaussian kernel function: Among them, ‖·‖ represents the Euclidean distance; σ is the width parameter of the kernel function; Polynomial kernel function: Where b represents the constant term; d represents the degree of the polynomial; Sigmoid kernel function: Where tanh represents the hyperbolic tangent function; β represents the slope parameter, and γ represents the intercept parameter.

9. An imported commodity price fraud identification system formed based on the method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module: obtains import commodity price fraud identification data; Data preprocessing module: normalize the data to its maximum and minimum values; Imported commodity price fraud identification module: uses the imported commodity price fraud identification model to identify imported commodity price fraud on the acquired data. The commodity price fraud identification model is the optimal imported commodity price fraud identification model.

Images