[Retracted] Application of BP Neural Network in Matching Algorithm of Network E-Commerce Platform
Abstract
In order to solve the matching algorithm problem of network e-commerce platform, a method of applying BP neural network in the network e-commerce platform matching algorithm is proposed. First of all, combined with the actual situation of the platform, select 9 factors that are most in line with the company’s actual business model to influence the selection for analysis; secondly, import 60 sets of data into MATLAB software, measure the input and output data uniformly, and divide the sample data matrix into training set and test. Finally, after multiple factor combinations and verifications, it is concluded that in the training model of the five main factors, the prediction results of the model are compared with the real values. The feasibility of establishing the selection model based on BP neural network is proved. Online e-commerce platforms can refer to this model to build a product selection model that meets the needs of the platform, helping enterprises to achieve more efficient product selection work. Since the parameter initialization of the neural network is random, although the output results are different after the program runs for many times, the R2 is still stable between 0.7 and 1.0, which proves that the predicted value made by the system is highly approximate to the real value and can achieve the predicted effect.
1. Introduction
Back propagation neural network (BP network) is one of the most mature and widely used artificial neural networks. Its basic network is three-layer feedforward network, including input layer, hidden layer, and output layer. For the input signal, it is necessary to propagate forward to the hidden node. After the function, the output information of the hidden node is transmitted to the output node, and finally, the output variable result is obtained. The neuron node function is usually taken as the S-type function. BP network can realize any complex nonlinear mapping relationship from input to output and has good generalization ability. It can complete the task of complex pattern recognition, as shown in Figure 1 [1].
With the development of national economy and Internet technology, network e-commerce platform has developed more and more rapidly in recent years. Among many crossborder e-commerce models, the buyer platform reduces the selection cost of consumers, improves consumers’ trust, and is favored by consumers by directly participating in the source organization, commodity selection, logistics, warehousing, and sales process. But at the same time, under this mode, the capital occupation is high, which requires enterprises to accurately perceive the needs of consumers. Platform buyers need to have high selection ability to improve the dynamic sales rate and ensure the healthy operation of the platform. At present, platform buyers rely more on personal experience and traditional data processing methods for product selection. Buyers often fail to make the choice of maximizing benefits due to the deviation of personal subjective judgment.
2. Literature Review
In the face of this problem, most of the existing studies give reasonable suggestions to guide the selection of products on the network e-commerce platform by analyzing many factors affecting the selection of products on the network e-commerce platform, such as using technical selection method, trial and error selection method, and market selection method. However, most of the existing studies give selection suggestions from the perspective of qualitative analysis and lack of selection model after quantitative analysis, and the accuracy of selection is difficult to verify. On the basis of previous studies, this paper studies the selection process and related influencing factors of online e-commerce platform, selects the optimal combination of influencing factors through quantitative analysis, and verifies it. At present, artificial intelligence technology has been applied to all aspects of human life, and the fashion retail industry has also begun to explore the use of artificial intelligence and big data technology to change the traditional operation mode and help enterprises achieve more accurate market prediction and style development. This paper uses the artificial intelligence algorithm (BP neural network) to find the combination of factors that can output the best prediction results, so as to build a selection model that can help platform buyers reduce personal subjective judgment deviation and improve the accuracy of selection, so as to realize the accurate prediction of commodity selection.
Lu et al. presented an entity matching method based on distance. The main task of this method is to give appropriate weights to the attributes of each describing entity [2]. Liu stated that based on the idea of reducing the solution space of the genetic algorithm by constructing the feature ellipse of feature point set and the parallel mechanism of genetic algorithm, a feature point matching algorithm based on genetic algorithm search strategy under affine transformation is proposed [3]. Murthy et al. introduced the concept of matching matrix while using the relaxation algorithm. They adopted the idea of two-way matching [4]. Wu et al. proposed a point feature matching algorithm based on the crosscorrelation function. Firstly, crosscorrelation is used to obtain the initial matching between feature point sets, and then epipolar geometric constraints are used to screen the patterns of matching point pairs to obtain accurate feature point matching point pairs [5]. Moholkar and Patil studied the fast point pattern matching algorithm, which uses the clustering method to determine the rotation of two point sets to be matched θ, and the algorithm has high efficiency [6]. Du designed a new feature point matching algorithm. The search strategy of this method adopts the deterministic annealing algorithm, which shows good robustness to false points, high proportion of missing points, and noise [7]. Kung proposed a matching algorithm similar to the iterative matching nearest point (ICP) method. The idea of the algorithm is to select the nearest set of feature points at each step of the iteration to determine the corresponding relationship and then obtain the spatial transformation according to the corresponding relationship [8]. Jin regards platform enterprises as modules based on the business ecosystem, and enterprises can provide users with complementary products or services through the platform [9]. Wang et al. focused on the differences between industrial platforms and product platforms and proposed that the realization of user value by industrial platforms is mainly through the provision of complementary products and services. These enterprises complement each other through the technical platform with common reuse basis provided by industrial platforms [10]. Yiyue et al. pointed out that customs clearance is the biggest barrier restricting the service quality of crossborder e-commerce, and the uncontrollability of its process greatly affects the convenience of transaction [11].
On the basis of current research, a matching algorithm optimization for network e-commerce platform based on BP neural network is proposed. By studying the main factors affecting the selection decision of online crossborder e-commerce buyer platform, 15 influencing factors in five categories are summarized. Combined with practical cases and using BP neural network, a three-layer neural network model is finally constructed, which takes price competitiveness, brand popularity, product popularity, design creativity, and product sales as five inputs and the probability of successful sales of goods as a single output. Since the parameter initialization of the neural network is random, although the output results of the program run many times are different, R2 is still stable between 0.7 and 1.0, which proves that the predicted value made by the system has a high approximation to the real value and can achieve the prediction effect [12].
3. Optimization of Matching Algorithm of Network e-Commerce Platform Based on BP Neural Network
3.1. BP Neural Network
3.1.1. Theoretical Basis
Artificial neural network is an important content in the field of artificial intelligence. Its structure and working principle are designed according to the organizational structure and activity rules of the human brain, but it is not a true restoration of the human brain. It only reflects some characteristics of the human brain and is a simplification, abstraction, or imitation of it. With the prominence of the advantages of neural network and the deepening of its theoretical research, neural network theory has been gradually applied to the automatic matching technology of feature points. BP neural network is one of the most classical and mature networks in the neural network theory system. It was proposed by the team of scientists led by Rumelhart and McCelland in 1986. It is a multilayer feedforward neural network trained according to the error back propagation algorithm. BP neural network can learn and store a large number of input-output mode mapping relationships without knowing the mathematical equations describing this mapping relationship in advance. This self-learning, self-organization, and self-mapping ability of BP neural network makes it more and more widely used [13].
3.1.2. Working Principle of BP Neural Network Algorithm
The learning rule of BP neural network uses the steepest descent method. The basic BP algorithm includes two aspects: forward propagation of signal and back propagation of error. That is, the actual output is calculated in the direction from input to output, while the correction of weight and threshold is carried out in the direction from output to input. The weight and threshold of the network are continuously adjusted through back propagation to minimize the sum of squares of the error of the network. BP neural network works by several neurons through certain connections. Each neuron has a general model, and its general model is shown in Figure 2 [14].
The topological structure of the BP neural network model includes three parts: input layer, hidden layer, and output layer. Theoretically, due to the introduction of hidden layer neurons, a three-layer BP neural network can realize the mapping of multidimensional space RM to RN with arbitrary accuracy.
3.1.3. BP Neural Network Design
Because the specific application fields of BP neural network are different, it is necessary to design a practical and reliable BP neural network model for specific problems. Usually, the design of BP neural network needs to consider the number of layers of the network, the number of neurons in each layer, the selection of initial weight and threshold, and learning rate, and relevant experts and scholars have summarized some design experience and guiding principles. The details are as follows:
(1) Number of Layers of the Network. It has been proved that the three-layer BP neural network can realize the accurate mapping of multidimensional space RM to RN. Although increasing the number of layers can further improve the accuracy and reduce the error, it also increases the complexity of the network, thus increasing the training time of the network [15]. Therefore, unless special circumstances, priority can be given to reducing network error and improving accuracy by appropriately increasing the number of hidden layer neurons on the basis of three-layer BP neural network.
(2) Number of Neurons in Hidden Layer. In terms of improving the accuracy of network training, although only one hidden layer is used, the method of increasing the number of neurons is much simpler than increasing the number of hidden layers. However, the number of hidden layers is not the more the better. Practice has proved that with the increase of the number of hidden layers, the training error of the network continues to decrease. At the same time, there will be the phenomenon of network training overfitting, which will reduce the generalization ability of the network.
(3) Selection of Initial Weight and Threshold. Because the BP neural network algorithm system is nonlinear, the selection of the initial weight and threshold of the network determines which point of the error surface the network training starts from. This not only directly affects the training time of the network but also relates to whether the training results of the network can get the global optimal solution. For example, if the initial weight and threshold are too large, the input falls into the saturation region of the S-type transfer function after the network weighting calculation, resulting in a very small gradient, which makes the adjustment process difficult to continue. Therefore, it is generally expected that after the initial weighting calculation of the network, the output of neurons will approach zero as much as possible, so that the weight of each neuron can be adjusted at the place where their S-type activation function changes the most [16]. Therefore, it generally takes the random number with the initial weight between [−1, 1]. The initial weight and threshold of traditional BP neural network are selected by random initialization in the [−1, 1] interval. At this point, the algorithm with global optimization characteristics can be used to assist in selecting the initial weight and threshold of the network.
(4) Transmission Function, Learning Rate, and Learning Function of Network. In general, the transmission function of the network hidden layer adopts the S-type function, while the output layer adopts linear function. If the learning rate is too high, it will lead to the global optimal solution, and it is easy to enter the global optimal solution. There are many ways to design the learning function, among which the Levenberg-Marquardt back propagation learning algorithm, which is commonly used and has better performance and training speed, is selected.
To sum up, although the neural network with hidden layer can realize the mapping of multidimensional space RM to RN, if some parameters are selected appropriately in the training process, the training time of neural network can be shortened, and satisfactory training results can be obtained. Among them, the selection of network initial weight and threshold has a great impact on the training process. BP neural network has high nonlinearity and strong generalization ability, but the standard BP algorithm has some disadvantages, such as easy to fall into local minimum solution, slow convergence speed, and oscillation effect. At present, there is no complete theory to guide the design process of BP neural network. At present, its design mainly focuses on the design and optimization of network structure and the adjustment algorithm of network weight and threshold. In this regard, relevant experts have summarized many optimization methods on the basis of a large number of research [17]. In the optimization of network structure design, based on the strategy of traditional optimization theory, algorithms such as waterfall correlation construction method, crosstest method, and neural center surgery optimization are proposed. The gradient descent algorithm is usually used in BP neural network. The gradient descent algorithm has an obvious advantage. It only follows one direction in local optimization; so, the search speed is very fast. However, the gradient descent algorithm is easy to fall into the local optimal solution. Therefore, other algorithms are often needed to optimize it in practical application.
3.2. Proposal of BP Neural Network Matching Method
3.2.1. Starting Point of Method Proposal
Feature point matching is to find the corresponding matching relationship and spatial mapping relationship between two feature point sets. The mapping relationship between image points obtained by the same spatial point on different images is determined to exist, and its mapping relationship is shown in Figure 3.
Although the mapping relationship between spatial points and image points and between image points on different images becomes complex due to the influence of shooting angle and environment, if the influence of external noise is not considered, the mapping relationship of all feature point sets of the image is basically the same. Therefore, if the corresponding relationship of a certain number of feature points is determined in advance by some method, the matching of the remaining feature points can be completed in theory. Generally, the matching method based on the position of platform feature points often needs to establish its accurate mathematical model. If the mathematical model is not accurate enough, it will directly reduce the correctness of the matching results and even lead to matching failure. BP neural network has the ability of self-learning and self-mapping. Its input and output relationship itself can be regarded as a mapping relationship. Users only need to provide some a priori samples for network learning without knowing the specific mapping relationship between sample input and output [18]. Therefore, when the matching relationship of some feature points in the feature point set of the platform is known, the matching of the remaining feature points can be completed by establishing the BP neural network mapping model. The feature point matching method based on BP neural network is actually equivalent to using BP neural network to fit an inherent and exact mapping relationship. Generally, BP neural network has two preconditions in fitting application:
(1) There are enough known samples for network training and self-learning. (2) There is an exact but unknown or functional relationship between the input and output of samples. As mentioned earlier, there is a certain mapping relationship between images taken from different angles of the same space object, and when some matching points are obtained by some method, it meets the premise of using BP neural network.
3.2.2. Idea of BP Neural Network Matching Method
Taking the matching of the platform as an example, this paper takes the characteristic points of the known matching relationship on the platform as the training samples of the network, trains the network, obtains the mapping relationship of the platform simulated by the network, and uses the mapping model to predict the corresponding relationship between the point sets to be matched. That is, the feature points in the left figure are successively input into the trained network model, and the predicted position of the matching points of these feature points in the right figure can be obtained through the calculation of the network model. Due to certain errors in network prediction, it is impossible for the predicted position to coincide with the actual position; that is, there is a certain distance. However, as long as the error is within the allowable range, it can be considered that there are corresponding matching points in these feature points in the figure. Otherwise, it is considered that there are no matching points in the graph. If the corresponding matching point exists, the feature point closest to the predicted position is selected as the correct matching point within the allowable error range according to the principle of maximum similarity. The flow chart of the whole process is shown in Figure 4 [19].
In fact, in the feature point matching method based on BP neural network, the most important work in the early stage is to establish the BP neural network model. It can be said that after the network model is established, the mapping relationship between feature points will be known, and then the feature point matching will be completed.
3.3. Establishment of BP Neural Network Matching Method Model
As mentioned above, the most important and core task of using BP neural network for feature point matching is to establish an ideal network model that can correctly reflect the mapping relationship between platform feature point sets [20]. The matching method based on BP neural network includes three parts: sample classification and processing, the design of BP neural network topology, and the design of matching constraint criteria.
3.3.1. Sample Classification and Processing
To complete the matching process of feature points to be matched, we must first analyze and process the sample data. The analysis and processing of sample data in this subject mainly include the normalization of samples and the division of samples.
(1) Sample Normalization. Normalization is to limit the data to be processed to a certain range after being processed by some algorithm. Firstly, normalization is for the convenience of later data processing. Secondly, it speeds up the convergence when the correction program is running and can convert the dimensional data into dimensionless data, so as to eliminate the phenomenon that some attributes of the sample cannot play a practical role due to the large difference between the data [21]. Although normalization cannot be carried out, because the transfer function used by BP neural network is the Sigmoid function, its value is between 0 and 1, it is often sensitive to the number between [0, 1], and the output function of neuron is the most sensitive between [0, 1] ([-1, 1]). Therefore, in order to improve the efficiency of training, the sample data should be normalized before the network training.
(2) Sample Division. After the network model is established, it is necessary to analyze its approximation ability to the laws contained in the samples. Its approximation ability, that is, generalization ability, should be evaluated by the error size of nontraining samples. The most direct and objective criterion to judge whether the established network model can effectively reflect the real law is that the error of other samples other than training samples is as small as possible and close to the error of training samples [22]. Through repeated experiments, it is found that the trend law of training sample error and sample error to be tested is inconsistent, and sometimes, there is an opposite trend, while the trend of verification sample training error and sample error to be tested is basically the same. This shows that the network training may enter the state of over fitting, resulting in a small error of the training samples and a large error of the samples to be tested. Therefore, we cannot use all the known samples for network training but reserve a small number of samples to evaluate the network performance and finally determine the optimal network result. Therefore, before network training, all training samples are divided into training samples and nontraining samples. In this paper, the result of the minimum training error of the verification sample is retained as the final network result, and the network is used to match the feature points to be matched. In this paper, this method is called the method of using validation samples to determine the final network results [23].
3.4. Development History of Online Crossborder e-Commerce Platform Industry
- (1)
Crossborder e-commerce phase 1.0: at this stage, China’s crossborder e-commerce is fledgling. As a tentative complementary channel with traditional international trade, the focus is on channel development and marketing promotion. Using the Internet as a channel to open up the market reduces the operating costs of small and medium-sized enterprises participating in international trade and makes them have more advantages in international competition. At the same time, the effectiveness of marketing promotion is also particularly critical. Most of the companies involved in providing such services are large-scale, represented by Alibaba International Station and Global Tesco
- (2)
Crossborder e-commerce phase 2.0: after creating the crossborder e-commerce trade mode, foreign trade enterprises at this stage began to use e-commerce platforms providing various services more skillfully and began to integrate resources and establish their own crossborder e-commerce companies with the help of seller’s platform, payment platform, and logistics platform, and the proportion of online transactions gradually increased
- (3)
Cross border e-commerce phase 3.0: at this stage, foreign trade enterprises use the platform of crossborder e-commerce more deeply and make full use of various tools. Some foreign trade enterprises began to establish their own overseas warehouses and overseas teams and received more feedback from overseas consumers through the Internet, producing corresponding products for sale [24]
3.5. Development Status of Online Crossborder e-Commerce Platform Industry
Online crossborder e-commerce has developed against the background of the shrinkage of the traditional international trade market. The global financial crisis in 2008 pushed the international trade competition to a white hot stage. With the deepening of trade friction, the appreciation of RMB, and the increasing cost of means of production such as labor and land, the traditional import and export trade has been hit hard. In this context, the development of crossborder e-commerce has sprung up. The total amount of crossborder e-commerce transactions in 2010 was 1.3 trillion yuan, which had doubled to 290 million yuan by 2013.
4. Design and Application of Selection Model Based on BP Neural Network
4.1. Data Set Establishment
Select an online crossborder e-commerce buyer platform for empirical research. The enterprise is a crossborder e-commerce mainly engaged in apparel goods of overseas luxury brands, and its goods suppliers are mainly Italian luxury dealers. Based on the factors affecting the selection in Table 1, the person in charge of the buyer of the platform and several professional buyers engaged in the selection work in the enterprise were interviewed. Combined with the actual situation of the platform, 9 factors affecting the selection that most accord with the actual business model of the company were selected, which are price competitiveness, brand popularity, product popularity, design creativity, fabrics and materials, compatibility, wearability, natural environment, and past sales of products (similar or similar). At the same time, according to the degree or subjective evaluation, from low to high, a 5-level scale is adopted to formulate quantitative scoring standards for each factor. For example, among the price competitiveness factors, “1” means that the product price competitiveness is very low, “5” means that the product price competitiveness is very high; In terms of fabric and material factors, “1” means that the product fabric and material are very poor, and “5” means that the product fabric and material are very good. Then, with the permission of the crossborder e-commerce buyer platform, the commodities with a selling out rate of more than 80% and a selling out rate of less than 20% in 2019 were randomly selected, and a total of 60 commodities were collected, of which the commodities with good sales and poor sales accounted for half, respectively. The distribution of goods by category is shown in Table 2. Commodity data includes brand item number, commodity name, purchase price, category, color, commodity picture, and historical sales data.
| Factors affecting selection | Factor segmentation | Examples of evaluation methods |
|---|---|---|
| Product attributes | Color/pattern | Does it attract buyers’ attention |
| Fabric/material | Whether there is good performance | |
| Matching degree of fashion trend | Product popularity | Matching degree between attributes and popular trends |
| Environmental factor | Natural factors | Is it suitable for this season |
| Sales and inventory | Procurement budget | Whether the purchase budget is met |
| Supplier situation | Supplier contract conditions | Are the contract details reasonable |
| Product category | Sold out rate is higher than 80% | Sold out rate is less than 20% |
|---|---|---|
| Blouse | 15 | 6 |
| Women’s skirt | 4 | 4 |
| Men’s coat | 12 | 16 |
Finally, the commodity data of the 60 pieces (sets) of commodities are made into a table, and the person in charge of the buyer of the platform evaluates each commodity according to the 9 factors affecting the selection. The scores of each influencing factor of each commodity are summarized into the selection model input vector x = (x1, X2, ⋯, X60) t. In the output vector o, the goods with a sold out rate of more than 80% are assigned as “1,” and the goods with a sold out rate of less than 20% are assigned as “0.”
4.2. Establishment of Neural Network
The learning rate is set as 0.001, and the maximum number of training steps is 5000. After the training, the trained neural network is tested with 20 groups of data from the test set, the prediction results are output, and the error is compared with the real value.
4.3. Result Collection and Analysis
After many factor combinations and verification, it is finally concluded that the accuracy of the output prediction result is the best in the training model that retains the five main factors of price competitiveness, brand popularity, product popularity, design creativity, and product sales. The number of input layer nodes is 5, the number of hidden layer nodes is 11, the number of output layer nodes is 1, the learning rate is 0.001, and the maximum number of training steps is 5000 [25]. The comparison between the predicted results of the model and the real value is shown in Figure 5. The ordinate in the figure is the probability of successful sales of goods, “0” represents poor sales of goods (the selling out rate is lower than 20%), and “1” represents good sales (the selling out rate is higher than 80%). Correspondingly, in the selection forecast, “0” means that the buyer platform should not select this commodity in the selection process, and “1” means that this commodity should be selected in the selection process. The circle points in the figure represent the real sales of goods, and the solid points represent the predicted values given by the selection model. It can be seen from the figure that the predicted value of the model is basically consistent with the real value. R2 is the tracking of the predicted value to the real value during the training process. The closer R2 is to 1, the stronger the explanatory ability of the variables in the model to the output variables and the better the prediction accuracy. R2 shown in the figure is 0.95, indicating high prediction accuracy. At the same time, because the parameter initialization of the neural network is random, although the output results of running the program for many times are different, R2 is still stable between 0.7 and 1.0, which proves that the predicted value made by the system is highly approximate to the real value and can achieve the prediction effect.
5. Conclusion
By studying the main factors affecting the selection decision of online crossborder e-commerce buyer platform, this paper summarizes 15 influencing factors in five categories. At the same time, combined with practical cases and using BP neural network, a three-layer neural network model is finally constructed, which takes price competitiveness, brand popularity, product popularity, design creativity, and product sales as five inputs and the probability of successful sales of goods as a single output. The prediction results of the model are relatively stable and accurate. The model verifies the feasibility of constructing the selection model based on BP neural network. The crossborder e-commerce buyer platform can refer to this model and build a selection model that meets the needs of the platform according to its own business model, so as to help enterprises achieve more efficient selection. Because the parameter initialization of the neural network is random, although the output results of the program are different after running for many times, R2 is still stable between 0.7 and 1.0, which proves that the predicted value made by the system is highly approximate to the real value and can achieve the prediction effect.
Conflicts of Interest
The author declares no conflicts of interest.
Open Research
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
