Procurement Volume Prediction of Cross-Border E-Commerce Platform Based on BP-NN

For logistics management, predictive analysis has never been more crucial than it is now, thanks to the vast number of deal data generated every second by electronic commerce. In an e ﬀ ort to enhance customer service and supply control, e-commerce businesses are progressively utilizing machine learning technologies to enhance projections. The back-propagation neural network (BP-NN) model is used to develop a C-A-BP forecasting model that considers commodity sales characteristics and the data series ’ trend. In order to predict each cluster, a C-BP-NN model is ﬁ rst created, adding sales information as in ﬂ uencing elements into the C-BP-NN model. An A-BP-NN model is used in combination with the ARIMA that is employed for the linear component. These two forecasting models are merged to provide the ﬁ nal results. By comparing the results of the ARIMA, BP-NN, A-BP-NN, and C-BP-NN using the information provided by Jollychic ’ s cross-border platform, the A-C-BPNN was shown to be the best.


Introduction
Customer satisfaction in the supply chain industry is enhanced when goods are prestocked at local warehouses in numerous marketplaces throughout the world, reducing logistics time. Because of the globalized production and sales regions, it takes longer for cross-border e-commerce companies to make arrangements from the acquisition of commodities, shipping, and customs quality control to arrive at the final destination [1]. Data from e-commerce commodity sales form the foundation for logistics management, and hence algorithms and big data analysis methods are frequently used to forecast sales. This data is critical to crossborder e-commerce companies' worldwide supply chain plans [2].
The sophistication of the cross-border e-commerce sector affects sales predictions in a variety of ways, in addition to the volume and variety of transaction data. As a result, e-commerce businesses still face a problem when trying to increase the precision and accuracy of their sales forecasts [3].
ARIMA [4], Procurement Volume Prediction [5], and BP-NN [6] are some of the terms used to describe this research. The following is a brief summary of the main points. Supply chain management necessitates [7] that items be stored in preparation in local warehouses in numerous marketplaces throughout the world, which might significantly cut logistics time. Although the manufacturing and sales sectors of e-commerce items are globalized, it takes cross-border e-commerce firms longer to prepare for the purchase of supplies, transportation, and customs quality inspections, to name a few. Data from e-commerce commodity sales form the foundation for supply chain management, and hence algorithms and big data analysis technologies are frequently used to forecast sales. This data is critical to cross-border e-commerce companies' worldwide supply chain plans [8].
It is not just transaction data that affects sales projections in the cross-border e-commerce business, but many other elements, as well. As a result, e-commerce businesses still face a problem when trying to increase the efficiency of their sales forecasts.
Much research has been done in the field of sales forecasting. Most of these research approaches to predicting sales use either time series (TS) models [9,10] or machine learning (ML) algorithms [11].
In order to extrapolate emerging tendencies from existing data collected, TSs vary from moving averages [12] to the ARIMA groupings. Although useful for sales forecasting, TSs are limited by the expectation of linear development and cannot compensate for other effects like pricing and marketing changes. There are several studies in which univariate forecasting approaches have been used as a benchmark model [13,14].
MLs and TSs have been compared in various current prediction models. ANN outperforms ARIMA in forecasting demand, according to a study cited in [3]. Reference [24] evaluated classic statistical approaches, such as ARIMA and multivariate regression, and found that ANNs performed better than the more statistical approaches. Reference [25] compared SVR to ARIMA, Holt-Winters, and moving average for use in sales prediction with promotion influences.
TS-based MLs have also been used for sales forecasting. Referencing [19], it was shown that combining ARIMA and ANN to simulate both linear and nonlinear aspects of the dataset was more efficient. There was an ARIMA prediction model developed in [20] that was used to train and fit the BP-NN using its residuals. LSTM ensemble prediction was demonstrated in [21] using a unique approach that efficiently aggregates several prediction outputs from a series of independent LSTM networks. As more and more ecommerce data becomes available, several algorithms have been developed to better manage unpredictable sales trends and account for a variety of other aspects. Reference [22] presented a unique method for automatic learning useful features from structured information using CNNs. Training an LSTM model with marketing consumer demands and cross-series information was the goal of reference [23]. ELM was frequently used in predicting, which is more important. Using ELM and distribution optimization, a new data-driven strategy for predicting user behavior was developed in [25]. In [26], a deep learning architecture was used to improve ELM's ability to predict wind speed.
Even though there are a variety of prediction models, the choice of approach is influenced by the features of different items. There is evidence to support the idea that product qualities can influence both search and sales since customers are more interested in the intrinsic properties of items when making a purchase decision. Clustering algorithms have been used to better represent the features of commodities in sales forecasting. When it comes to improving the outcomes of NNs, both fuzzy artificial NNs and aggregating approaches have been applied. Reference [26] used hierarchy self-organizing maps and principal component analysis to create the SVR to address the sales forecasting challenge. In reference [27], an ELM and composite linkage approach were used to develop a sales prediction model. Reference [28] developed an SVR-based clustering sales forecasting system. For forecasting computer retail sales, reference [29] created a clustering-based forecasting model that combines clustering and ML methodologies.
Three-phase BP-NN-based prediction models are created to focus on the two factors (sales attributes and data series trend) indicated above in this study, according to the literature review.
It is necessary to first incorporate numerous salesinfluencing elements using the two-step clustering technique, which is an upgraded version of BIRCH. Each cluster is then modeled using the BP-NN approach, which has been proven to be a successful predictor in numerous data analysis contests, such as Kaggle and many recently published papers.
For the complex sections of sample data, we present an A-BP-NN approach that integrates the characteristics of ARIMA and BP-NN methods, respectively, to improve trend prediction performance. It is possible to build a final mixture strategy that takes into account the various factors that influence sales of products and the time series pattern. The C-A-BP-NN model is shown here.
To sum up, The BPNN and ARIMA have been shown in this paper to be speedy and efficient throughout the initial phase of global search, addressing the aforementioned drawbacks. Based on their advantages and limitations, this paper presents a hybrid computational technique that combines ARIMA with BPNN weight optimization. This process is known as ARIMA and BPNN computation. The ARIMA's global search capability and the BPNN's range search capability are both utilized by the mixed computation technique.
The main contributions of the paper are as follows: (i) We present a unique GA method, dubbed the BPNN and ARIMA, in which a novel ARIMA agent acts as an inventive operator, enhancing BPNN and ARIMA's exploitation and exploration capabilities in the issue search space (ii) We compare the proposed ARIMA performance to well-known and innovative metaheuristic algorithms using well-known benchmark issues   (v) We compare our suggested method to state-of-the-art ANN to measure investment funds' forecasting accuracy The remaining five parts of the document are as follows: Two-step clustering technique, ARIMA parameter estima-tion method, and BP-NN model are only a few of the models and methods briefly discussed in Section 2. The BP-NNbased model is provided in Section 3 to anticipate sales attributes and trends of time series in a three-phase manner. This forecasting model's accuracy is demonstrated numerically in Section 4. Section 5 summarizes the findings and makes a suggestion for further study.

Methodologies
2.1. Selecting the Appropriate Features. The volume of data in the e-commerce industry has grown exponentially since the introduction of web technology. Big data is characterized by a range of formats, including Twitter, text, online, music, movie, click-stream, and log files. Techniques such as filterbased or wrapper feature extraction have been explored to minimize the dimension of the data to remove the most unnecessary and superfluous information from diverse sources. When predicting the accuracy of feature subsets in wrapper feature selection, effective training utilizes mathematical resampling (like cross-validation) to predict feature subset accuracy [30]. Different algorithms can use filterbased feature extraction to represent the identical dataset.  In this research, the prediction and clustering techniques use wrapper feature selection to remove uninteresting qualities from multidimensional data based on SD, Pearson correlation coefficient, CV, and feature significance scores, with the following information.
Data dispersion is measured by SD, which may be computed as follows: The coefficient of variation (CV) is a statistical measure of the variability of actual values in the data.
In statistical terms, PCC is a way of measuring the linear correlation between two variables.
where ððY i − YÞ/δ y Þ, δ y , and Y show the standard deviation, standard score, and the mean of Y i , respectively. Each feature is given a score based on how important it is to the model's enhanced decision trees. The more weight is given to a certain trait in decision trees, the more important it becomes. The importance of a single decision tree is derived by calculating the performance metric by each splitting   Clustering techniques are selected based mostly on the size and type of data that was gathered. When working with numerical or categorical data, standard techniques for clustering can be used. Reference [31] presented the BIRCH, a hierarchical technique that is particularly well-suited to big datasets with continuous characteristics. However, in this study, the two-step clustering approach in SPSS Modeler is recommended because of the huge and heterogeneous data. BIRCH's two-step clustering technique uses the scroll distance as a metric to calculate distances between categorical data as well as between continuous data. Preclustering monitoring is similar to BIRCH's two-step clustering procedure in that it keeps the focused parts of acquired information in the form of descriptive analysis. The dense zones are then clustered using a hierarchical clustering technique. Both BIRCH and the two-step clustering method are capable of handling mixed characteristics, but the two-step clustering technique has a novel technique for allocating cluster identification to noisy data.
Since it is a hierarchy approach, it is much more effective at coping with noise and inconsistencies than partitioning approaches. This approach is superior to others in that it automatically calculates the ideal number of clusters. The two-step clustering approach depicted in Figure 1 may be used to cluster items in e-commerce interaction datasets of large and heterogeneous transaction datasets.

Preclustering.
Using the BIRCH algorithm's clustering feature (CF) tree growth, it is possible to execute outlier treatment while simultaneously reading each record in a data collection. The datasets in dense regions are then used to generate a CF tree, which yields subclusters.

The Assignment of Cluster Membership.
It is determined which data records belong in which clusters by determining how far apart data records are from each other in terms of log-likelihood.

Results' Validation.
In order to evaluate the performance of the clustering findings, the coefficient φ is used. Better clustering results can be achieved with a value of 1 greater than zero: where α 1 is the average distance among the sampling and its groupings and α 2 is the average distance among the sampling and its separate cluster.

The Optimization of the ARIMA model's Parameters.
Moving averages and autoregressive models are combined to form the ARIMA model. Reference [31] explains how to use the Box-Jenkins technique in the time series concept to build an ARIMA (p, d, and q) model. Since its parameters are often established by plotting ACF and PACF data, the ARIMA is susceptible to judgment variation when it comes to defining parameters. As an alternative, the function called ARIM () in the R package is often utilized to autonomously create an ideal ARIMA framework, which composes of the shortcoming of ARIMA during parameter judgment. The ARIM () function is coupled with the findings of ACF and PACF plots in order to increase ARIMA's fitting

The Three-Phase Prediction Model
Research in this study proposes a three-phase C-A-BP-NN prediction system based on both the sales attributes and the data's tendency.
New NN models are developed for the first phase of forecasting in Phase 1, which combine distinct clustering characteristics into forecasts. BP-NNs are used to model each of the generated clusters, after which the two-step clustering approach is used to divide the commodities into distinct groups depending on characteristics.
Two NN models are described in Phase 2: the ARIMA-BP-NN architecture, which uses ARIMA's linear fitting capability and BP's nonlinear mapping ability to forecast trends in time series. The linear component of the data series is predicted using ARIMA, the nonlinear part is revised using ARIMA residuals, and a BP-NN is established using the rolling prediction approach.
The C-A-BP-NN is created by merging the C-BP-NN with the A-BP-NN at this stage. A-BP-NN and C-A-BP-NN findings are given a weighted average in the C-A-BP-NN in order to reduce some errors of squares. This reflects the reliability and reliability of sales characteristics and data series. As seen in Figure 3, the proposed three-phase strategy is shown, and the details are provided below. Eight data series are included in the source data series, as indicated in Table 2, between the dates of January 3rd, 2017, and December 29th, 2018.

Data Description
(iii) Clustering series  Wireless Communications and Mobile Computing Ten continuous variables and six categorical features are derived from the reconstruction of the source data series in the collection of clustering characteristics. Table 3 depicts the clustering series feature descriptions.

Experiments.
The following consistent experimental settings are developed to verify the theoretical model's performance in accordance with performance assessment indices.

Uniform Dataset.
Training, validation, and test datasets are created, as shown in Table 4, to suit the requirements of different models. Here is a breakdown of how the data application works: A sample period of 381 days is covered by the clustering series. Training set 1, which includes data from the first 347 days of the clustering series, serves as the basis for two-step clustering models in the C-BP-NN model. After a two-step clustering process, the obtained data is utilized to build BP-NN models. The C-BP-NN model is tested using the remaining 34 days of data. As part of the validation process, two-step clustering is performed to partition and deploy the test set. (1) The ARIMA predicting is performed using the second training dataset, which contains data from the first to the 277th day, and the validation data is used to estimate ARIMA multicollinearity, which is then utilized to build the A-BP-NN network. By using a specified set of data, it is then evaluated. C-BP and A-BP-NN designs were fitted with quick and effective weights using the latest 34 data samples.

Uniform Evaluation Indexes.
Validating forecasting models have traditionally relied on a variety of performance criteria. In order to differentiate the best forecasting model, standard assessment metrics are used. A model's accuracy improves as its size decreases.    Continuous characteristics are standardized; outliers are handled at a noise level of 25%; log-likelihood proximity is used to assess distance; BIC is used to cluster the data. Figure 4 illustrates the 2.64 ratios of sizes for each cluster, and the proportion is just right. As a result, the quality of the clusters is satisfactory.
(2) The C-BP-NN models are built in this step. The initial features are chosen from each of the twelve clusters according to the relevance of the features. C-BP-NN models for each cluster are then built using the chosen attributes of each cluster and the SKU marketing in Table 3 as the input and output variations, respectively As an illustration of the modeling methods for the BP-NN, consider one of the 12 clusters. Table 3 lists the characteristics, and Figure 5 shows the seven that have been narrowed down from that list. Among the most important elements are the following: F1, F3, F5, F6, and F7. Although F2 and F4 make less of an impact on the forecast, they are still important factors.
The cluster is pretrained by using the default input and output, which are the 11 cluster characteristics in Step 1 and the associated SKU sales in Table 3.
(3) Optimization of parameter values: When optimizing BP-NNs, the most important step is determining which input and output parameters are optimal. Parameter optimization, on the other hand, has less of an effect on the algorithm's accuracy. To improve the BP-NN, just the fundamental parameters, such as the number of layers, momentum, and the learning rate, are changed. The model can be balanced since increasing the number of layers will make the model more complicated and more likely to overfit, while raising the values of momentum and learning rate will make the model less conservative. Figures 6-8 demonstrate the average variations in the convergence curve based on the number of layers, momentum, and learning rate, respectively, of the BP-NN and other benchmark prediction models 4.5. Models for Comparison. As a comparison between our suggested models and other models, the following methods are selected: ARIMA is a popular time series model for estimating future sales. By establishing the specified characteristics and the associated input and output, the BP-NN model is formed and optimized. In the NN of the C-BP system, to predict sales based on the two-step clustering model's results, the BP-NN considers commodity sales characteristics. In the A-BP-NN, the ARIMA residuals are revised using the A-BP-NN. When it comes to time series modeling, BP-NN is employed for the nonlinear component and ARIMA for the linear part. In the C-A-BP-NN, the C-BP and A-BP-NNs are combined in this model to maximize their respective strengths and minimize their respective weaknesses. The simulation results are shown in Figure 9 and Table 5. Also, the Tylor diagram for various utilized methods is shown in Figure 10.

Conclusions
C-A-BP-NN, a BP-NN prediction technique that takes into account the sales features and patterns of the sample data, was recommended in this study. First, the C-BP-NN is introduced, which combines the clustering and BP-NNs in order to reflect the sales characteristics of commodities in the forecasting process. Based on certain criteria, the prediction may be made using the two-step clustering approach, which divides data into numerous groups. It is then time to build unique clusters' C-PB-NNs with the help of a BP-NN model. The proposed A-BP-NN utilizes ARIMA's strengths in forecasting the trend of data series while also addressing the

10
Wireless Communications and Mobile Computing nonlinear component of the data series, which ARIMA's drawbacks cannot. By comparing AICs for various parameters, we can find the best ARIMA model, and then we can use it to forecast the linear component of the data series.
The ARIMA residuals are used as input and output for the rolling prediction of nonlinear data series by the trained BP-NN. By combining the anticipated residuals from the BP-NN with ARIMA forecast values, we arrive at the final A-BP-NN results.
To summarize, the C-A-BP-NN is constructed by combining the strengths of the C-BP and A-BP-NNs and weighting them appropriately. This means that a concatenation of the findings of the two models is used to arrive at the final prediction values. It is necessary to make comparisons with other existing models in order to determine the efficacy of the C-A-BP-NN. The C-A-BP-NN's predicting ability is evaluated using four popular evaluation indices, including ME, RRMSE, RMSE, and MAE. According to the results of this experiment, the C-A-BP-NN surpasses all other methods, demonstrating that this model may be used as a guide when choosing forecasting models for an ecommerce firm.
As an example of future research areas, instead of using only one model for all commodities, the e-commerce corporation should use various models for each commodity. The two possible expansions have been proposed for further study. On the one hand, because there could be no model with minimal values for all assessment indicators, selecting the ideal model becomes challenging. In order to tackle this challenge, a thorough assessment index of predicting ability will be developed. When it comes to sales forecasting, various significant aspects like inventory costs, order lead times, delivery times, and transportation times are really taken into account in an effort to improve inventory management.

Data Availability
No data were used to support this study.