Study of Merchant Adoption in Mobile Payment System Based on Ensemble Learning

To explore the influencing factors of the adoption of mobile payment systems from the perspective of merchants, this study builds a data analysis model based on three different ensemble learning algorithms, Adaboost model, random forest and XGBoost model, where static social-economic attributes, dynamic trading behavior and clustering effect variables of merchants are used as independent variables. Moreover, this paper establishes the prediction models and analyzes the prediction accuracy of different models. The results of the study indicate that: 1) Merchants in the housing industry, health hos-pitals and retail industries are more willing to adopt mobile payment systems; 2) The average daily transaction volume and the average amount of each consumer significantly affect the merchant mobile payment adoption behavior; 3) The adoption of mobile payment systems by neighboring merchants significantly positively affected the adoption behavior of merchants; 4) On the basis of the social-economic attributes of merchants, the hit rate and accuracy of the prediction model were greatly improved after adding transaction data.

complete transaction (Zhan & Qiao, 2016). Third-party institutions such as Alipay Wallet and WeChat Pay use the advantages of Internet technology, e-commerce platforms, and social networks to enlarge customer number and business scale. And the business model of the remote payment industry has been relatively formed. Through the continuous integration of online and offline services, various innovative mobile near-field payment methods (mobile phone flash payment, barcode payment) are increasingly accepted by users, which promotes customer loyalty and expands usage scenarios. The offline retail market based on mobile near field payment has become the focus of competition among major payment companies and financial institutions.
As an important participant in the mobile payment industry, merchants' attitude towards mobile payment adoption is particularly important. Exploring the influencing factors of merchants' adoption of mobile payment, formulating efficient strategies to identify and predict the willingness and behavior of merchants to adopt mobile payment, and expanding the acquiring market have become the main means for financial institutions to compete with payment companies. However, there are few studies on the adoption of mobile payment systems by merchants in the existing literature. This article used the mobile payment system as an innovative product and applied the objective behavior data of a merchant bank's acquiring merchants. Based on three integrated ensemble learning algorithms: Random Forest, XGBoost, and Adaboost, analysis models are built to study the influencing factors and decision-making process of merchants' adoption of mobile payment. The research results are compared with the results of Lasso-logistic regression model. The paper is arranged as follows: Section 2 reviews relevant literature; Section 3 introduces the research method and process of the paper. Section 4 analyses and discusses the research results. Finally, summarize the paper and put forward the prospect.

Research on Mobile Payment Adoption Behavior
At present, there are many researches on mobile payment, most of which use empirical research methods, TAM, UTAUT, TTF and other models were used as the theoretical basis to discuss consumer or merchant (anthropomorphic) mobile payment adoption willingness (Li, Sun, & Yan, 2013). Kim et al. (Kim, Mirsobit, & In, 2010) based on the TAM theory, combining the characteristics of individual users (individual innovation, mobile payment knowledge) and mobile payment products characteristics (movability, accessibility, compatibility and convenience), studied influence factors for users' willingness of mobile payment.  (Liu, 2011) proposed in his doctoral dissertation that compared with other influencing factors, the security of mobile payment technology has the strongest impact on user willingness. Zhou believes that in order to improve consumer acceptance and use of mobile payments (Zhou, 2011), it is important to establish initial trust in mobile payments.
Through empirical research in his paper, he found that perceived security, perceived universality, and perceived ease of use have a significant impact on the initial trust of mobile payment consumers.
Dan's research indicated that perceived usefulness and perceived ease of use have a significant impact on consumer acceptance intentions (Dan & Jing, 2011).
At the same time, the externalities of the network are also a significant factor, but consumers show lower perceived risk levels for mobile payments. Oliveira et al. combined UTAUT theory and innovation diffusion theory to explore the influencing factors of user adoption of mobile payment and recommendation of others to use mobile payment (Oliveira, Thomas, Manoj et al., 2016). Liu et al. studied the willingness to adopt mobile payment from the perspectives of health care and incentive based on exchange theory and two-factor theory (Liu, Xia, Li, & Liang, 2017). Guo's research shows that enterprises need to understand consumer payment habits and methods in depth, and achieve enterprise innovation by improving the convenience of operations and enriching mobile operation functions (Guo & Li, 2018). User adoption theory is also applicable to research on merchants' willingness to adopt. Cao studied the willingness of enterprises to adopt mobile payment based on TTF theory with case analysis (Cao, 2008).

Application of Machine Learning in Adoption Behavior Research
In recent years, some scholars have applied weak classification algorithms in machine learning, such as Logistic regression and decision tree, to research on product diffusion and adoption behavior. variables based on meta-analysis (MA) and rough sets, which improved the prediction accuracy of the model (Li & Chen, 2014;Li, Peng, & Zhen, 2014). Support vector machine (SVM) and BP neural network, as two widely used machine learning algorithms, can achieve high prediction accuracy, but the disadvantages are also obvious, that is the interpretability of the two models is poor. In the study of mobile payment adoption, Chong extended the theoretical model of TAM, using the output of the structural equation model as the input of the neural network method, and realized feature selection through the advantages of path analysis of the structural equation model (Chong, 2013a(Chong, , 2013b. Chen et al. inherited Chong's model and compared the prediction accuracy of neural networks and multiple regression models during the model detection stage (Chen, Jing, & He, 2015). The sample set for the studies is subjective data collected through questionnaires, which has limitation. Jiao et al. used machine learning models such as Lasso regression (Jiao, Jing, Liu, & Zhang, 2018), Ridge regression, random forest and decision tree to predict the demand for shared bicycles, and obtained the conclusion that the prediction accuracy of the random forest model is higher than that of regression analysis. Compared with the weak classification model, the random forest model integrated with decision tree has a greater improvement in prediction accuracy. However, because there is no fixed channel to collect objective data for mobile payment adoption by merchants, which makes it difficult to obtain research data, there were few researches leverage machine learning for mobile payment adoption by merchants.
Therefore, consider the interpretability of the prediction model to obtain guidance for future, and at the same time try to improve the prediction accuracy of the model, this study chose three different types of ensemble learning algorithms: Random Forest (Breiman, 2001), XGBoost , Adaboost (Freund & Schapire, 1996;Freund & Schapire, 1997), and made a comparative analysis with Lasso-logistic regression model (Tibshirani, 1996). The advantage of Lasso-logistic regression model is that it can get the significance and weight value of each characteristic variable, so that it can make targeted management recommendations, but the prediction accuracy of this model is not as good as ensemble learning.

Research Methods
The paper uses the Lasso-logistic regression model and three ensemble learning algorithms in supervised learning scenarios. First, data collection, variable selection, and data preprocessing are performed. In order to solve the problem of collinearity among variables, this study adopted a stepwise regression method for attribute selection. The filtered variables constitute the input variables of the three models in this paper. Lasso-logistic regression model, Adaboost, random forest, and XGBoost can be used to achieve function of prediction after training on data samples.

Experimental Procedure
The experiment in this study consists of 6 steps: Step 1: Data acquisition. Merchants' static attributes, dynamic transaction data, and clustering effects, etc. are collected in three different ways.
Step 2: Data pre-processing. There are missing values and outliers in the collected data, so only data after data preprocessing can become the data samples of the model in this paper.
Step 3: Attribute selection. This paper uses a stepwise regression method to filter the input variables.
Step 4: For the filtered attributes, as input variables, they are trained by Lasso- logistic regression model, and output the weight value and significance index of each feature variable. Then further filter the independent variables.
Step 5: Input the selected independent variables into the integrated learning model to obtain three different prediction models.
Step 6: Through the classification matrix, evaluate the model from hit rate, coverage rate, accuracy rate and ROC curve.

Model Introduction
The four models in the experiments in this paper are based on the scikit-learn library implemented in Python machine learning tools. The scikit-learn library is built on NumPy, SciPy and matplotlib, and provides efficient data mining and analysis tools. In this paper, a random logistic regression model is used to select variables, and the selected feature variables are added to the four prediction models. In order to avoid the overwhelming influence of variables with small values by large variables on the model, and to simplify the calculation complexity of the model, this paper uses the linear range method to normalize the data after filtering.

1) Lasso-logistic regression model
Logistic regression model is a classification method widely used and studied in industry and academia. The calculation method of this model is simple and the variable interpretation ability is strong. However, if the sample data size is large and there are many covariates, the statistical significance of some variables will The coefficient estimates β in the Lasso-logistic regression model can be written as shown in Equation (2).
In Equation (2), λ represents the harmonic parameter. The key to the variable selection of Lasso-logistic regression model lies in the choice of harmonic parameter λ. The smaller the value of λ, the more parameters the model retains; otherwise, the less it retains.

2) Ensemble Learning Model
Integrated learning refers to the method adopted for multiple learner collections, which can be divided into several types: multi-classifier system, mixture of experts, and committee-based learning. The current research focus is still homogeneous ensemble learning. The main idea of integrated learning is to first generate multiple learners through certain rules, and then use a certain integration strategy to combine, and finally comprehensively judge and output the final result. Homogeneous class integrated learning means that multiple learners are homogeneous "weak learners". Based on this weak learner, multiple learnings are generated through sample set disturbance, input feature disturbance, output representation disturbance, algorithm parameter disturbance, etc. After the integration, a "strong learner" with better accuracy is obtained.
Breiman proposed the Random Forest (RF) algorithm in 2001. The classification decision tree without pruning constructed by the CART decision tree algorithm was used as the base classifier. The idea of the random forest algorithm is shown in Figure 1. First, the Bootstrap method is used to extract the training set from the original sample set; then a decision tree model is trained on each training set, and during the growth of each tree, randomly selected some variables from all feature variables, and then select the best attributes among these variables according to the principle of the smallest Gini coefficient; finally gather the prediction results of all base classifiers, and vote to get the final category.

Y. Y. Li, Y. Li
The full name of XGBoost is eXtreme Gradient Boosting, which is an extension of Gradient Boosting Machine. The Gradient Boosting Machine algorithm adopts the idea of gradient descent when generating each tree. Based on all the trees generated in the previous step, it moves towards minimizing the given objective function. Under reasonable parameter settings, a certain number of trees need to be generated to achieve the expected accuracy. When the data set is large and complex, the calculation of Gradient Boosting Machine algorithm is huge.
XGBoost is an implementation of Gradient Boosting Machine, which is beneficial to the parallel calculation of the algorithm and improves the accuracy.
Freund proposed the Adaboost algorithm based on Boosting. The algorithm is shown in Figure 2. The principle is to first assign weights to each sample in the training data, and initialize the sample weights to equal values, and conduct training to get the first base classifier; after the first base classifier weights

Analysis of Influencing Factors Based on Lasso-Logistic Regression
The sample set in this study includes three aspects: merchant static attribute variables, merchant dynamic transaction variables, and merchant clustering effect variables. The Lasso-logistic regression model was used to analyze the experimental data in order to obtain the influencing factors that influence the merchants' adoption of mobile payment, so as to obtain the enlightening significance of management.
After selecting variables by random regression model, the coefficients of each variable calculated by Lasso-logistic regression model are shown in Table 2. It can be seen from

2) Dynamic merchant transaction variables
The average daily transaction volume of the merchant and the average number of each consumption amount significantly affect the merchant's mobile payment adoption behavior. The transaction volume is significantly positively correlated with the adoption behavior; the average amount of each consumption amount is significantly negatively correlated with the adoption behavior. That is, merchants with large transaction volume and low single transaction amount are more willing to adopt mobile payment. Because mobile payment can save transaction payment time, the greater the transaction volume, the more obvious the advantage of saving time; at the same time, the lower the amount of each transaction, the lower the degree of concern for consumers and merchants on payment security, and they are more willing to accept Mobile payment methods.

3) Merchant cluster effect variable
From the analysis of the influence variables of the surrounding merchants' adoption behavior, the proportion of the surrounding adopted merchants significantly affects the other merchants' adoption behavior. This conclusion shows that the willingness of merchants to adopt will be affected by neighboring merchants.

Predictive Analysis of Adoption Behavior
The three indicators (Equations (4)-(6)) of model hit rate, coverage rate and accuracy rate are used to evaluate the effectiveness of the model. Among them, TP indicates the number of forecasts adopted and actually adopted; FP indicates the number of forecasts adopted and actually not adopted; FN indicates the number of forecasts not adopted and actually adopted; TN indicates the number of forecasts not adopted and actually not adopted. The comparison results of different algorithm performance indicators are shown in Table 3. In order to more accurately measure the accuracy of each model, the ROC curves of the four models are drawn, as shown in Figure 3.
It can be seen from Table 3 and Figure 3 that the accuracy, precision, and coverage of the three integrated learning models of random forest, XGBoost model, and AdaBoost are significantly higher than the Lasso-logistic model, which reflects the advantage of predictive accuracy of ensemble learning; the prediction accuracy of the AdaBoost model is lower than that of the random forest and XGBoost models, indicating that the accuracy of the optimized model based on sample perturbations is not suitable for the data set of this study.

Conclusion and Further Work
This study builds analysis model based on three Ensemble learning algorithms: Adaboost, Random Forest, and XGBoost, using social-economic attribute variables, dynamic transaction variables, and clustering effect variables of the merchants as independent variables, researched influencing factors of merchants' adoption of mobile payment services. Comparative study was performed