A cross-border e-commerce product credit mark and after-sales system

By applying the stochastic gradient enhancement tree algorithm and dynamic credit evaluation model in the cross-border e-commerce system, the problem of indistinguishable credit of cross-border e-commerce products is solved, and the overall system control of cross-border e-commerce products and product credit in different regions is realized. Tracking has improved operation and maintenance efficiency and reduced costs, and promoted the development of the cross-border free trade system.

CN117196776BActive Publication Date: 2024-11-08SUZHOU UNIV OF SCI & TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311167185.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-09-09
Publication Date
2024-11-08
Estimated Expiration
2043-09-09

AI Technical Summary

Technical Problem

There are problems in the existing cross-border e-commerce systems with complex manual operations, differentiated identity, inflexible system distribution and indistinguishable product credit, resulting in low operation and maintenance efficiency.

Method used

Using the credit marking and after-sales system for cross-border e-commerce products based on the stochastic gradient enhancement tree algorithm, the credit tracking and evaluation of cross-border product through information traceability, credit rating classification and dynamic credit evaluation model, combined with cloud platform and MongoDB database, cross-border product credit tracking and evaluation are realized, and malicious intentions are eliminated. Evaluation to ensure information security and objectivity.

Benefits of technology

It has realized the overall system control of cross-border e-commerce products and product credit tracking in different regions, reduced economic costs, improved operation and maintenance efficiency, enriched the cross-border free trade system, and brought good social and economic benefits.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117196776B_ABST
    Figure CN117196776B_ABST
Patent Text Reader

Abstract

A cross-border e-commerce product credit marking and after-sales system based on the stochastic gradient boosting tree algorithm, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module. Aiming at the problems of complex manual operation, differences in identity, inflexible system distribution and difficulty in distinguishing product credit in the existing international cross-border e-commerce integrated operation and maintenance problems, the cloud platform and intelligent algorithms are combined with cross-border e-commerce to achieve the overall system control of cross-border e-commerce and product credit tracking in different regions. It has a wide range of applications and low economic costs, and can be extended to social applications, enriching the cross-border free trade system and bringing good social and economic benefits.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of cross-border e-commerce, and in particular to a cross-border e-commerce product credit mark and after-sales system. Background Art

[0002] Cross-border e-commerce refers to an international business activity in which trading entities belonging to different customs territories reach transactions and conduct electronic payment settlements through e-commerce platforms, and deliver goods through cross-border e-commerce logistics and remote warehousing to complete transactions. The country has vigorously deployed and promoted the layout and construction of logistics hubs to promote the improvement of the quality and efficiency of the national economy. As the technical basis for promoting economic integration and trade globalization, cross-border e-commerce has very important strategic significance. Cross-border e-commerce has not only broken through the barriers between countries and made international trade move towards borderless trade, but it is also causing great changes in the world's economic trade. For enterprises, the open, multi-dimensional and three-dimensional multilateral economic and trade cooperation model constructed by cross-border e-commerce has greatly broadened the path to enter the international market, greatly promoted the optimal allocation of multilateral resources and mutual benefit and win-win between enterprises; for consumers, cross-border e-commerce makes it very easy for them to obtain information from other countries and buy high-quality and low-priced goods. However, the problems of complex manual operation, differences in identity, inflexible system distribution and difficulty in distinguishing product credit still exist in the existing international cross-border e-commerce integrated operation and maintenance problems. The present invention combines cloud platforms and intelligent algorithms with cross-border e-commerce to achieve overall system control of cross-border e-commerce and product credit tracking in different regions. It has a wide range of applications and low economic costs. It can be extended to social applications, enrich the cross-border free trade system, and bring good social and economic benefits. Summary of the invention

[0003] The purpose of the present invention is to provide a cross-border e-commerce product credit marking and after-sales system based on a stochastic gradient boosting tree algorithm to solve the problems raised in the above background technology.

[0004] To achieve the above objectives, a cross-border e-commerce product credit marking and after-sales system based on the stochastic gradient boosting tree algorithm is provided, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module.

[0005] S1. Conduct information tracing on manufacturers and distributors of cross-border e-commerce products, and preliminarily assess the credit status of the manufacturers;

[0006] S2. Mark the sales comparison information of the same product in different international regions, build a system database of cross-border products, and build a distributed cloud platform and access rights of cross-border e-commerce network agents based on the system database, user interface and cloud service to ensure information security;

[0007] S3. Record the product processing problems in the transaction process in the system database, and classify the product manufacturer credit from multiple angles based on the random gradient boosting tree algorithm based on the transaction evaluation information of the same product in a unit time period;

[0008] S4. Based on the results of the credit classification of product manufacturers, a dynamic credit evaluation model is constructed in combination with the evaluation level of buyers. Buyer information is recorded in the system database, and after-sales evaluations and comments are tracked. The matching degree between the credit classification of product manufacturers and the credit evaluation model is verified, buyer information is inverted, and buyers are communicated with. If there is a phenomenon of malicious evaluation by buyers, the corresponding records are removed to ensure relatively objective credit information;

[0009] S5. Update system data in real time, update the information within a unit time period to the user interface, provide buyers with a basis for purchase, and provide credit constraints and sales services for cross-border e-commerce;

[0010] Furthermore, the information traceability in S1 includes one or more combinations of registered trademarks, brand information and operating qualifications of manufacturers and agent sellers of cross-border e-commerce products;

[0011] Furthermore, the system database in S2 is mainly built using MongoDB technology that processes non-relational and relational databases;

[0012] Furthermore, the transaction evaluation information of the same product in a unit time period in S3 is used to perform multi-angle product manufacturer credit classification based on the stochastic gradient boosting tree algorithm. The detailed process is as follows:

[0013] The improved stochastic gradient boosting tree algorithm (Grandient Boosting Decision Tree, GBDT for short) is used, i.e., gradient boosting decision tree, a machine learning algorithm for regression and classification research, to obtain a model that makes its predicted value F(x) of the input variable close to its true value y. A greedy strategy is adopted to train a weak classification model each time, so that the predicted value h of each base model m (x) approximates the actual value that it needs to predict, and then weights the predicted values ​​of these base models; the structure of each weak classification model is a binary decision tree. In the process of training the weak classification model, the model learns the actual value y and the predicted value F after the previous iteration. m-1 (x), that is, the fitting residual, the residual yF m-1 (x) is the reverse gradient of the cubic loss function, as shown below:

[0014]

[0015] Among them, x represents the attribute variable of the product database , after fitting the residual, the previous iteration prediction value F m-1(x) plus the residual h of this round of fitting m (x) The obtained F m (x), the square error loss function can be reduced, and finally the predicted value F(x) of the overall model can be close to the true value y. At this time, each base model fits the reverse gradient of the loss function. After input, training and iteration, the standard model is obtained. The present invention uses an improved stochastic gradient boosting tree algorithm, and the residual yF m-1 (x) Using the reverse gradient of the cubic loss function, the simulated prediction value iteration is faster than the traditional stochastic gradient boosting tree algorithm, which increases the efficiency of the algorithm.

[0016] In the process of solving the problem by gradient descent method, setting the learning rate v makes it easier to approach the true value. When v decreases, the number of base models will increase accordingly, thereby improving the accuracy of the model. The training algorithm of gradient boosting decision tree is as follows:

[0017] Input the given training set A differentiable loss function L(y, F(x)), the number of iterations M (i.e. the number of decision trees / the number of base models);

[0018] (1) Initialize the model with a constant γ0:

[0019]

[0020] Where F0(x) is the initialization model; for The variable point x corresponding to the minimum value; y i is the element in the given training set.

[0021] (2) For m = 1 to M:

[0022] Calculate the pseudo residual r min as follows:

[0023]

[0024] According to the training set Construct a weak learner h m (x) is used to fit the pseudo residual;

[0025] The multiplier γ is calculated by one-dimensional optimization as follows m :

[0026]

[0027] v represents the learning rate, and the calculation formula for updating the model is as follows:

[0028] F m =F m-1 (x)+vγm h m (x)0 <v≤1

[0029] (3) Output F M (x) represents the prediction of a strong classifier composed of a series of weak decision tree models.

[0030] A GBDT cascade classification prediction model for imbalanced data is used. During the model construction process, multiple learners f(x) are combined to form a hierarchical classifier g(x), and the positive example output results of multiple g(x) are concatenated to form a cascade classifier G(x).

[0031] The cascade classifier sets different thresholds in each layer of classifiers to divide samples and perform classification training. If the test sample of the previous layer classifier meets the threshold standard of the next layer, it can enter the next layer classifier test, and so on.

[0032] The input of the classification model is the original data set s = {x, y}, where x is the buyer's credit rating, x = {d, h, l, c}, and y is the binary label of the buyer's credit performance. The output of each layer of the classification model is the predicted product manufacturer credit z is the comprehensive level output The predicted product manufacturer credit obtained.

[0033] The transaction evaluation information of the same product is divided into negative samples and positive samples according to the hierarchical time standard. The evaluation data shows the problem of label imbalance. The random undersampling method is used to achieve the balance of positive and negative samples. The specific method is: for each level i (i∈[1,m]), the data is randomly undersampled to obtain k independent positive and negative sample balanced data sets, each of which is denoted as s i,j (j is the data set number), k GBDT learners f are trained i,j (x)(i∈[1,m],j∈[1,k]), combine the results of k GBDT learners to obtain the final classification result.

[0034] GBDT is used as the learner f(x). GBDT is an integrated learning model based on CART (Classification and Regression Decision Tree). This model trains a group of weak learners (CART decision tree) in series to gradually fit the predicted delay time to the true value. For the binary classification model, the samples are classified into positive and negative categories, and the sigmoid function is used to calculate the category.

[0035] The input sample set is s i,j =(x,y i )i∈[1,m],j∈[1,k]. Where x is the input feature, x={d,h,l,c}; y iis the actual delay label corresponding to sample x. The learner is trained on a dataset containing n samples at the i-th level. The construction steps of the GBDT model f(x) are as follows:

[0036] Step 1: Initialize the learner f(x), use the logarithmic loss function, and adjust the decision tree parameters so that the loss function L(y i ,f(x)) reaches its minimum.

[0037]

[0038] in, is the model's predicted delay label for sample x; θ1 is the decision tree parameter;

[0039] Step 2: Use the negative gradient γ of the loss function i Fit the residuals, adjust the parameter target of the decision tree to minimize the loss function, and update the model f(x).

[0040]

[0041] Among them, θ2 is the decision tree parameter;

[0042] Step 3: Repeat step 2 to complete L-1 iterations, and use the sigmoid function to complete the probability calculation to achieve category discrimination:

[0043]

[0044] Among them, σ is a constant; θ1 is the decision tree parameter;

[0045] GBDT multi-classification is to classify samples with multiple classification labels. Compared with GBDT multi-classification, the GBDT cascade classification prediction model performs balanced processing of category data at each level. Each level is independent of each other and can be trained simultaneously. When the predicted product manufacturer credit is greater than the level standard, the hierarchical classifier predicts and outputs a positive example, and enters the next level classifier for prediction; if the hierarchical classifier predicts and outputs a negative example, the calculation is terminated. The prediction results of each level classifier are connected in series through positive examples to obtain the predicted product manufacturer credit.

[0046] Furthermore, in S4, a dynamic credit evaluation model is constructed in combination with the buyer's evaluation level. The detailed process is as follows:

[0047] The dynamic credit assessment model is constructed using the Logistic regression model. In the practice of corporate credit assessment, the linear structure contained in the model makes it have good stability and interpretability. Assume that in a cross-border e-commerce product transaction, there are n buyers, X = (x1, x2, ..., x m) is the characteristic variable that affects the buyer’s credit performance, m is the number of variables, and each buyer is given a label v according to his credit performance, where v is a binary variable indicating whether the buyer’s credit performance is normal. If we want to assess the probability of a buyer’s credit performance defaulting, we need to calculate the probability p of the model’s prediction result v = 1, where p = f(v = 1|x1,x2,…,x m ), the specific expression of logistic regression logit(p) is:

[0048]

[0049] Among them, (β0,β1,β2,…,β m ) is the unknown coefficient of the model, exp(β0+β1x1+β2x2+…+β m x m ) represents the desired calculation;

[0050] The maximum likelihood estimation method is used to solve the unknown coefficients in the formula, from which the probability of buyer credit performance p(v=1|X) can be obtained:

[0051]

[0052] p(v=0|X)=1-θ

[0053] The probability function of v can be combined into:

[0054]

[0055] According to the Bernoulli distribution, the maximum likelihood function l(β,X) is written as:

[0056]

[0057] Then the log-likelihood function ln(l(β,X)) is:

[0058]

[0059] Among them, θ i Represents the unknown parameter to be calculated. For Logistic regression, we get ln(l(β,X)):

[0060]

[0061] Among them, x1, x2, x i is the parameter of Bernoulli distribution.

[0062] In the actual problem of building a dynamic credit assessment model based on buyer ratings, most transaction process records are difficult to meet the premise of probability distribution, while the Logistic regression model does not have overly strict restrictions and has good interpretability.

[0063] Furthermore, the matching degree between the credit classification of the product manufacturer and the credit assessment model is verified in S4, and the buyer evaluation information is inverted. The detailed process is as follows:

[0064] The improved BP neural network algorithm is used to invert the buyer information. It can approximate any function. At the same time, the number of intermediate layers, the number of processing units in each layer, the learning factor of the network, etc. can be set according to the actual situation. It has great flexibility. The BP neural network algorithm is used to invert the buyer information:

[0065] Let the training sample set be X = [X1, X2, ..., X j ], for any sample set X k =[X k1 ,X k2 ,…,X km ], and get the actual output Y k =[Y k1 ,Y k2 ,…,Y kp ] T , its expected output is D k =[d k1 ,d k2 ,…,d kp ] T .

[0066] In the forward propagation of the BP neural network, from the initial input to the I layer, we can have:

[0067]

[0068] in, Represents the weight of logical calculation, and the propagation process from layer I to layer J is:

[0069]

[0070] in, is the output data of the BP neural network when it is forward propagated to layer I, x km is the sample value, w mi is the propagation path from the initial input to the I layer, w ij is the propagation path from layer I to layer J, is the forward propagation speed of the BP neural network from the initial input to the I layer, is the forward propagation speed of the BP neural network from layer I to layer J, f(*) is the inversion result of the BP neural network algorithm for buyer information, and the same applies to the following layers.

[0071] The propagation process from layer J to layer P is:

[0072]

[0073] Then you can get the predicted output:

[0074]

[0075]

[0076] in, is the initial output data of the BP neural network, is the output number of the BP neural network when it is forward propagated to layer I According to, x km is the sample value, w io is the initial input propagation path, w mi is the propagation path from the initial input to the I layer, BP nerve The forward propagation speed of the network from the initial input to the I layer, f (*) is the inversion result of the buyer information by the BP neural network algorithm; after multiple error calculations and network iterations, the reverse error signal propagation of the BP neural network is realized. According to the above principle, the buyer evaluation data set is trained to obtain the inversion buyer information result. If there is a phenomenon of malicious evaluation by the buyer, the corresponding record is removed to ensure relatively objective credit information;

[0077] The error is:

[0078] e kp (n) = d kp (n)-y kp (n)

[0079] where y kp (n) is the actual output of the sample, d kp (n) is the expected output of the sample, and the error energy is defined as

[0080] Then the sum of the errors of the neurons is:

[0081]

[0082] The forward process in the BP neural network is the process of signal propagation forward. In order to minimize the error obtained, it is necessary to back-propagate the error signal and provide feedback to the model. The error back-propagation process is as follows: the correction amount between the hidden layer J and the output layer P in the BP neural network, the correction amount of the weight in the BP algorithm is proportional to the partial differential of the error weight, that is:

[0083]

[0084] Where E(n) is the sum of the errors of the neurons, w IO (n) is the propagation path from layer I to layer O, and λ is a constant.

[0085] And there are:

[0086]

[0087]

[0088] where y kp (n) is the actual output of the sample, e kp (n) is the error, It is the output data of BP neural network when it is forward propagated to O layer.

[0089] You can get:

[0090]

[0091] Its gradient is:

[0092]

[0093] According to the calculation formula of the activation function f(x):

[0094]

[0095] You can get:

[0096]

[0097] So we can get:

[0098]

[0099] Therefore, the correction amount w ko (n) can be expressed as:

[0100]

[0101] From this we can get the iterative formula:

[0102] w io (n+1)=w io (n)+Δw io (n)

[0103] Similarly, the weight correction from hidden layer I to hidden layer J is:

[0104] w ij (n+1)=w ij (n)+Δw ij (n)

[0105] The weight correction from hidden layer M to hidden layer I is:

[0106] w mi (n+1)=w mi (n)+Δw mi (n)

[0107] The reverse error signal propagation of the BP neural network can be realized. According to the above principle, the buyer evaluation data set is trained to obtain the inverse buyer information result. If there is a phenomenon of malicious evaluation by the buyer, the corresponding record is removed to ensure relatively objective credit information.

[0108] Beneficial effects of the invention: The invention describes a cross-border e-commerce product credit marking and after-sales system based on a random gradient boosting tree algorithm, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module, to trace the information of manufacturers and agent sellers of cross-border e-commerce products, including the registered trademarks, brand information and operating qualifications of manufacturers and agent sellers of cross-border e-commerce products, preliminarily assess the credit status of manufacturers, mark the sales comparison information of the same product in different international regions, build a system database of cross-border products with MongoDB technology for processing non-relational and relational databases, build a distributed cloud platform and access rights of cross-border e-commerce network agents based on the system database, user interface and cloud service, ensure information security, record product processing problems in the transaction process in the system database, and classify product manufacturer credit from multiple angles based on the improved GBDT algorithm for the transaction evaluation information of the same product in a unit time period. m-1(x) Using the reverse gradient of the cubic loss function, compared with the traditional random gradient boosting tree algorithm, the simulation prediction value iteration is faster, the algorithm efficiency is increased, and based on the results of the product manufacturer's credit classification, the Logistic regression model is used to combine the buyer's evaluation level to build a dynamic credit evaluation model, record the buyer's information in the system database, track the after-sales evaluation and speech, verify the matching degree between the product manufacturer's credit classification and the credit evaluation model, use the BP neural network algorithm to invert the buyer's information, invert the buyer's information, communicate with the buyer, if there is a phenomenon of malicious evaluation by the buyer, then remove the corresponding record, ensure relatively objective credit information, update the system data in real time, update the information within the unit time period to the user interface, provide buyers with a basis for purchase, and provide credit constraints and sales services for cross-border e-commerce. The present invention aims at the problems of complex manual operation, differences in identity, inflexible system distribution and difficulty in distinguishing product credit in the existing international cross-border e-commerce integrated operation and maintenance problems, combines the cloud platform and intelligent algorithm with cross-border e-commerce, and realizes the overall system control of cross-border e-commerce and product credit tracking in different regions. It has a wide application level and low economic cost, can be promoted to social applications, enrich the cross-border free trade system, and bring good social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] The invention is further illustrated by the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0110] Figure 1 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION

[0111] The present invention is further described in conjunction with the following examples.

[0112] See also Figure 1 The present invention aims to provide a cross-border e-commerce product credit marking and after-sales system based on a stochastic gradient boosting tree algorithm to solve the problems raised in the above background technology.

[0113] To achieve the above objectives, a cross-border e-commerce product credit marking and after-sales system based on the stochastic gradient boosting tree algorithm is provided, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module.

[0114] S1. Conduct information tracing of manufacturers and distributors of cross-border e-commerce products, and preliminarily assess the credit status of the manufacturers, including one or more combinations of the registered trademarks, brand information and operating qualifications of the manufacturers and distributors of cross-border e-commerce products;

[0115] S2. Mark the sales comparison information of the same product in different international regions, build a system database for cross-border products using MongoDB technology that processes non-relational and relational databases, and build a distributed cloud platform and access rights for cross-border e-commerce network agents based on the system database, user interface and cloud service to ensure information security;

[0116] S3. Record the product processing problems in the transaction process in the system database, and classify the product manufacturer credit from multiple angles based on the stochastic gradient boosting tree algorithm based on the transaction evaluation information of the same product in a unit time period. The detailed process is as follows:

[0117] The stochastic gradient boosting tree algorithm (Grandient Boosting Decision Tree, GBDT for short) is used. It is a machine learning algorithm used for regression and classification research. It is composed of a series of integrated weak classification models. Each weak classification model gives a prediction value. They are combined according to certain weights to form the final prediction value, and finally a strong classifier model is obtained. Generally speaking, the training goal is to obtain a model so that its predicted value F(x) of the input variable is close to its true value y. The GBDT algorithm adopts a greedy strategy, training only one weak classification model (weak classifier) ​​at a time, that is, let the predicted value h of each base model be m (x) approximates the actual value of the part it needs to predict, and then weightedly combines the predicted values ​​of these base models.

[0118] The structure of each weak classification model is a binary decision tree. In the process of training the weak classification model, the model learns the true value y and the predicted value F after the previous iteration. m-1 (x), that is, the fitting residual, the residual yF m-1 (x) is the reverse gradient of the cubic loss function, as shown below:

[0119]

[0120] Among them, x represents the attribute variable of the product database , after fitting the residual, the previous iteration prediction value F m-1 (x) plus the residual h of this round of fitting m (x) The obtained F m (x), which can reduce the square error loss function and eventually make the predicted value F(x) of the overall model close to the true value y. At this time, each base model fits the reverse gradient of the loss function. After input, training and iteration, the standard model is obtained.

[0121] In the process of solving the problem by gradient descent method, setting the learning rate v makes it easier to approach the true value. When v decreases, the number of base models will increase accordingly, thereby improving the accuracy of the model. The training algorithm of gradient boosting decision tree is as follows:

[0122] Input the given training set A differentiable loss function L(y, F(x)), the number of iterations M (i.e. the number of decision trees / the number of base models);

[0123] (1) Initialize the model with a constant γ0:

[0124]

[0125] (2) For m = 1 to M:

[0126] Calculate the pseudo residual r min as follows:

[0127]

[0128] According to the training set Construct a weak learner h m (x) is used to fit the pseudo residual;

[0129] The multiplier γ is calculated by one-dimensional optimization as follows m :

[0130]

[0131] v represents the learning rate, and the calculation formula for updating the model is as follows:

[0132] F m =F m-1 (x)+vγ m h m (x)0 <v≤1

[0133] (3) Output F M (x) represents the prediction of a strong classifier composed of a series of weak decision tree models.

[0134] The present invention uses a GBDT cascade classification prediction model for imbalanced data. During the model construction process, multiple learners f(x) are combined to form a hierarchical classifier g(x), and the positive example output results of multiple g(x) are connected in series to form a cascade classifier G(x).

[0135] The cascade classifier sets different thresholds in each layer of classifiers to divide samples and perform classification training. If the test sample of the previous layer classifier meets the threshold standard of the next layer, it can enter the next layer classifier test, and so on.

[0136] The input of the classification model is the original data set s = {x, y}, where x is the buyer's credit rating, x = {d, h, l, c}, and y is the binary label of the buyer's credit performance. The output of each layer of the classification model is the predicted product manufacturer credit z is the comprehensive level output The predicted product manufacturer credit obtained.

[0137] The transaction evaluation information of the same product is divided into negative samples and positive samples according to the hierarchical time standard. The evaluation data shows the problem of label imbalance. The random undersampling method is used to achieve the balance of positive and negative samples. The specific method is: for each level i (i∈[1,m]), the data is randomly undersampled to obtain k independent positive and negative sample balanced data sets, each of which is denoted as s i,j (j is the data set number), k GBDT learners f are trained i,j (x)(i∈[1,m],j∈[1,k]), combine the results of k GBDT learners to obtain the final classification result.

[0138] GBDT is used as the learner f(x). GBDT is an integrated learning model based on CART (Classification and Regression Decision Tree). This model trains a group of weak learners (CART decision tree) in series to gradually fit the predicted delay time to the true value. For the binary classification model, the samples are classified into positive and negative categories, and the sigmoid function is used to calculate the category.

[0139] The input sample set is s i,j =(x,y i )i∈[1,m],j∈[1,k]. Where x is the input feature, x={d,h,l,c}; y i is the actual delay label corresponding to sample x. The learner is trained on a dataset containing n samples at the i-th level. The construction steps of the GBDT model f(x) are as follows:

[0140] Step 1: Initialize the learner f(x), use the logarithmic loss function, and adjust the decision tree parameters so that the loss function L(y i ,f(x)) reaches its minimum.

[0141]

[0142] in, is the model's predicted delay label for sample x; θ1 is the decision tree parameter.

[0143] Step 2: Use the negative gradient γ of the loss function i Fit the residuals, adjust the parameter target of the decision tree to minimize the loss function, and update the model f(x).

[0144]

[0145] Among them, θ2 is the decision tree parameter.

[0146] Step 3: Repeat step 2 to complete L-1 iterations, and complete the probability calculation through the sigmoid function to achieve category discrimination.

[0147]

[0148] Among them, θ1 is the decision tree parameter.

[0149] GBDT multi-classification is to classify samples with multiple classification labels. Compared with GBDT multi-classification, the GBDT cascade classification prediction model performs balanced processing of category data at each level. Each level is independent of each other and can be trained simultaneously. When the predicted product manufacturer credit is greater than the level standard, the hierarchical classifier predicts and outputs a positive example, and enters the next level classifier for prediction; if the hierarchical classifier predicts and outputs a negative example, the calculation is terminated. The prediction results of each level classifier are connected in series through positive examples to obtain the predicted product manufacturer credit;

[0150] S4. Based on the results of the product manufacturer's credit classification, a dynamic credit evaluation model is constructed in combination with the buyer's evaluation level. The buyer's information is recorded in the system database, and the after-sales evaluation and comments are tracked. The matching degree between the product manufacturer's credit classification and the credit evaluation model is verified, and the buyer's information is inverted. Communicate with the buyer. If there is a phenomenon of malicious evaluation by the buyer, the corresponding record is removed to ensure relatively objective credit information. The detailed process is as follows:

[0151] The BP neural network algorithm is used to invert the buyer information. It has most of the advantages of artificial neural networks. The BP algorithm is a supervised adaptive learning algorithm in artificial neural networks, and can also be called an error back propagation algorithm. The basic structure of the BP neural network is composed of nonlinear variable units, which is characterized by good nonlinear mapping performance. In theory, it can approximate any function; at the same time, the number of intermediate layers, the number of processing units in each layer, the learning factor of the network, etc. can be set according to actual conditions, which has greater flexibility and has broad application prospects in optimization, signal processing and pattern recognition, intelligent control, fault diagnosis, etc. Therefore, the present invention uses the BP neural network algorithm to invert the buyer information:

[0152] Let the training sample set be X = [X1, X2, ..., X j ], for any sample set X k =[X k1 ,X k2 ,…,X km ], and get the actual output Y k =[Yk1 ,Y k2 ,…,Y kp ] T , its expected output is D k =[d k1 ,d k2 ,…,d kp ] T .

[0153] In the forward propagation of the BP neural network, from the initial input to the I layer, we can have:

[0154]

[0155] in, Represents the weight of logical calculation, and the propagation process from layer I to layer J is:

[0156]

[0157] The propagation process from layer J to layer P is:

[0158]

[0159] Then you can get the predicted output:

[0160]

[0161]

[0162] in, is the initial output data of the BP neural network, is the output data of the BP neural network when it is forward propagated to layer I, x km is the sample value, w io is the initial input propagation path, w mi is the propagation path from the initial input to the I layer, is the forward propagation speed of the BP neural network from the initial input to the I layer, and f(*) is the inversion result of the BP neural network algorithm for the buyer information; after multiple error calculations and network iterations, the reverse error signal propagation of the BP neural network is realized. According to the above principle, the buyer evaluation data set is trained to obtain the inverted buyer information result. If there is a phenomenon of malicious evaluation by the buyer, the corresponding record is removed to ensure relatively objective credit information;

[0163] The error is:

[0164] e kp (n) = d kp (n)-y kp (n)

[0165] The error energy is defined as

[0166] Then the sum of the errors of the neurons is:

[0167]

[0168] The forward process in the BP neural network is the process of signal propagation forward. In order to minimize the error obtained, it is necessary to back-propagate the error signal and provide feedback to the model. The error back-propagation process is as follows: the correction amount between the hidden layer J and the output layer P in the BP neural network, the correction amount of the weight in the BP algorithm is proportional to the partial differential of the error weight, that is:

[0169]

[0170] And there are:

[0171]

[0172]

[0173] You can get:

[0174]

[0175] Its gradient is:

[0176]

[0177] According to the calculation formula of the activation function f(x):

[0178]

[0179] You can get:

[0180]

[0181] Therefore, we can get:

[0182]

[0183] Therefore, the correction amount w ko (n) can be expressed as:

[0184]

[0185] From this we can get the iterative formula:

[0186] w io (n+1)=w io (n)+Δw io (n)

[0187] Similarly, the weight correction from hidden layer I to hidden layer J is:

[0188] w ij (n+1)=w ij (n)+Δw ij (n)

[0189] The weight correction from hidden layer M to hidden layer I is:

[0190] w mi (n)+1)=w mi (n)+Δw mi (n)

[0191] The reverse error signal propagation of the BP neural network can be realized. According to the above principle, the buyer evaluation data set is trained to obtain the inverse buyer information result;

[0192] S5. Update system data in real time, update information within a unit time period to the user interface, provide buyers with a basis for purchase, and provide credit constraints and sales services for cross-border e-commerce.

[0193] Beneficial effects of the invention: The invention describes a cross-border e-commerce product credit marking and after-sales system based on a random gradient boosting tree algorithm, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module, to trace the information of manufacturers and agent sellers of cross-border e-commerce products, including the registered trademarks, brand information and operating qualifications of manufacturers and agent sellers of cross-border e-commerce products, preliminarily assess the credit status of manufacturers, mark the sales comparison information of the same product in different international regions, build a system database of cross-border products with MongoDB technology for processing non-relational and relational databases, build a distributed cloud platform and access rights of cross-border e-commerce network agents based on the system database, user interface and cloud service, ensure information security, record product processing problems in the transaction process in the system database, and classify product manufacturer credit from multiple angles based on the improved GBDT algorithm for the transaction evaluation information of the same product in a unit time period. m-1(x) Using the reverse gradient of the cubic loss function, compared with the traditional random gradient boosting tree algorithm, the simulation prediction value iteration is faster, the algorithm efficiency is increased, and based on the results of the product manufacturer's credit classification, the Logistic regression model is used to combine the buyer's evaluation level to build a dynamic credit evaluation model, record the buyer's information in the system database, track the after-sales evaluation and speech, verify the matching degree between the product manufacturer's credit classification and the credit evaluation model, use the BP neural network algorithm to invert the buyer's information, invert the buyer's information, communicate with the buyer, if there is a phenomenon of malicious evaluation by the buyer, then remove the corresponding record, ensure relatively objective credit information, update the system data in real time, update the information within the unit time period to the user interface, provide buyers with a basis for purchase, and provide credit constraints and sales services for cross-border e-commerce. The present invention aims at the problems of complex manual operation, differences in identity, inflexible system distribution and difficulty in distinguishing product credit in the existing international cross-border e-commerce integrated operation and maintenance problems, combines the cloud platform and intelligent algorithm with cross-border e-commerce, and realizes the overall system control of cross-border e-commerce and product credit tracking in different regions. It has a wide application level and low economic cost, can be promoted to social applications, enrich the cross-border free trade system, and bring good social and economic benefits.

[0194] The present invention also provides a computer-readable storage medium, in which at least one instruction is stored, and the instruction is loaded and executed by a processor to implement the above method. The computer-readable storage medium may be a ROM, a random access memory, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc. The instructions stored therein may be loaded by a processor in a terminal to execute the above method.

[0195] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0196] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable terminal device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0197] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cross-border e-commerce product credit marking and after-sales system based on the stochastic gradient boosting tree algorithm, including a cross-border e-commerce product cloud platform module, a cross-border e-commerce product credit rating classification module and a cross-border e-commerce product after-sales module; S1. Conduct information tracing on manufacturers and distributors of cross-border e-commerce products, and preliminarily assess the credit status of the manufacturers; S2. Mark the sales comparison information of the same product in different international regions, build a cross-border e-commerce product database, and build a distributed cloud platform and access rights for cross-border e-commerce network agents based on the system database, user interface and cloud service to ensure information security; S3. Record the product processing problems in the transaction process in the system database, and classify the product manufacturer credit from multiple angles based on the random gradient boosting tree algorithm based on the transaction evaluation information of the same product in a unit time period; S4. Based on the results of the product manufacturer's credit classification and combined with the buyer's evaluation level, a dynamic credit evaluation model is constructed, buyer information is recorded in the system database, after-sales evaluations and comments are tracked, the matching degree between the product manufacturer's credit classification and the credit evaluation model is verified, buyer information is inverted, and buyers are communicated with. If there is a phenomenon of malicious evaluation by buyers, the corresponding records are removed to ensure relatively objective credit information; S5. Update system data in real time, update the information within a unit time period to the user interface, provide buyers with a basis for purchase, and provide credit constraints and sales services for cross-border e-commerce; The transaction evaluation information of the same product in a unit time period in S3 is used to classify the product manufacturer credit from multiple angles based on the improved GBDT algorithm. The detailed process is as follows: Adopting a greedy strategy, we train a weak classification model each time, and let the prediction value h of each base model m (x) approximates the actual value that it needs to predict, and then weights the predicted values ​​of these base models; the structure of each weak classification model is a binary decision tree. In the process of training the weak classification model, the model learns the actual value y and the predicted value F after the previous iteration. m-1 (x), that is, the fitting residual, the residual yF m-1 (x) is the reverse gradient of the cubic loss function, and x is the input manufacturer credit independent variable, as shown in the following formula: Among them, x represents the attribute variable of the product database. After fitting the residual, the previous iterative prediction value F m-1 (x) plus the predicted value h obtained in this round m (x) The obtained F m (x), which reduces the square error loss function, and finally makes the predicted value F(x) of the overall model close to the true value y. At this time, each base model fits the reverse gradient of the loss function. After input, training and iteration, the standard model is obtained. Using the improved GBDT algorithm, the residual yF m-1 (x) Backward gradient using cubic loss function.

2. According to claim 1, a cross-border e-commerce product credit labeling and after-sales system based on stochastic gradient boosting tree algorithm is characterized in that: The information traceability in S1 includes a combination of registered trademarks, brand information and operating qualifications of manufacturers and agent sellers of cross-border e-commerce products.

3. According to claim 1, a cross-border e-commerce product credit labeling and after-sales system based on stochastic gradient boosting tree algorithm is characterized in that: The system database in S2 is mainly built using MongoDB technology that processes non-relational and relational databases.

4. According to claim 1, a cross-border e-commerce product credit labeling and after-sales system based on stochastic gradient boosting tree algorithm is characterized in that: In S4, a dynamic credit evaluation model is constructed in combination with the buyer's evaluation level. The detailed process is as follows: The dynamic credit assessment model is constructed using the Logistic regression model. In the practice of corporate credit assessment, the linear structure contained in the model makes it have good stability and interpretability. Assume that in a cross-border e-commerce product transaction, there are n buyers, X = (x1, x2, ..., x m ) is the characteristic variable that affects the buyer’s credit performance, m is the number of variables, and each buyer is given a label v according to his credit performance, where v is a binary variable indicating whether the buyer’s credit performance is normal. If we want to assess the probability of a buyer’s credit performance defaulting, we need to calculate the probability p of the model’s prediction result v = 1, where p = f(v = 1|x1,x2,…,x m ), the specific expression of logistic regression logit(p) is: Among them, (β0,β1,β2,…,β m ) is the unknown coefficient of the model, exp(β0+β1x1+β2x2+…+β m x m ) represents the desired calculation; The maximum likelihood estimation method is used to solve the unknown coefficients in the formula, from which the probability of buyer credit performance p(v=1|X) can be obtained: p(v=0|X)=1-θ Combine the probability functions of v into: According to the Bernoulli distribution, the maximum likelihood function l(β,X) is written as: Then the log-likelihood function ln(l(β,X)) is: Among them, θ i Represents the unknown parameter to be calculated. For Logistic regression, we get ln(l(β,X)): Among them, x1, x2, x i is the parameter of Bernoulli distribution.

5. According to claim 1, a cross-border e-commerce product credit labeling and after-sales system based on stochastic gradient boosting tree algorithm is characterized in that: In S4, the matching degree between the credit classification of the product manufacturer and the credit assessment model is verified, and the buyer evaluation information is inverted. The detailed process is as follows: Use BP neural network algorithm to invert buyer information, let the training sample set be X = [X1, X2, ..., X j ], for any sample set X k =[X k1 ,X k2 ,…,X km ], and get the actual output Y k =[Y k1 ,Y k2 ,…,Y kp ] T , in the forward propagation of the BP neural network, from the initial input to the I layer: in, Represents the weight of the logical calculation, and the predicted output is: in, is the initial output data of the BP neural network, is the output data of the BP neural network when it is forward propagated to layer I, x km is the sample value, w io is the initial input propagation path, w mi is the propagation path from the initial input to the I layer, is the forward propagation speed of the BP neural network from the initial input to the I layer, and f(*) is the inversion result of the BP neural network algorithm for the buyer information; after multiple error calculations and network iterations, the reverse error signal propagation of the BP neural network is realized. According to the above principles, the buyer evaluation data set is trained to obtain the inverted buyer information result. If there is a phenomenon of malicious evaluation by the buyer, the corresponding record will be eliminated to ensure relatively objective credit information.