2.1. Business Performance of a Restaurant

Running a successful restaurant business is a complex task that requires consideration of multiple factors. These factors include service quality, competitive pricing, and the presence of similar restaurants nearby, all of which can significantly impact the success or failure of the business (Wang, 2021). Location and geographic diversity also play a crucial role in determining the effectiveness of a restaurant business (Song, 2021). In large metropolitan areas, restaurants must pay close attention to their location and the surrounding commercial facilities and restaurants in the vicinity (Wu, 2021). In today’s digital age, social media has become a powerful tool for consumers to share their opinions on the quality and various services of restaurants. Therefore, innovation activities have become an essential part of increasing the efficiency of the restaurant business (Lee, 2018). Ultimately, the performance of a restaurant business depends on consumer satisfaction, which influences their intention to return and leads to business profits (Susskind, 2019). For social commerce, consumers can rate and review restaurants, reflecting their satisfaction with various factors.

2.2. Social Commerce

Social commerce is a modern approach to online shopping that combines the best of traditional e-commerce with the power of social media to reach and engage consumers within a network (Sohn, 2020). By creating virtual communities, social commerce can effectively promote the sale of products or services online, encouraging participation and sharing equally among users (Liao, 2021). This collaborative approach is precious for consumers who may be hesitant to purchase products they cannot physically touch (Boardman, 2019), as social commerce provides guidance and reduces the inherent uncertainty and risk associated with online shopping (Zhao, 2020).

In today’s world, social media platforms have become an integral part of the consumer decision-making process. Before making any purchase, consumers tend to check out reviews and opinions from previous buyers. This trend is particularly prevalent in the restaurant industry, where customers look for feedback on the quality and service of restaurants before deciding where to dine. According to Gao (2018), consumer feedback plays a vital role in determining whether a product can meet the consumer’s needs to a certain extent. Researchers have adopted data mining techniques from social commerce platforms to better understand restaurant business performance. Chang (2019) conducted research on predicting business performance based on location. The study aimed to identify insights for businesses that coexist and compete in geographic neighborhoods. The research used information from Yelp.com, a social commerce platform that provides member reviews and recommendations for businesses. The study analyzed 10,618 restaurant businesses in 2013 and found that the hybrid model was effective in supporting restaurants with their strategic and operational decisions. Leonard Gunawan (2023) researched the satisfaction levels of restaurant patrons. The study analyzed online reviews on TripAdvisor and used Support Vector Machine (SVM) techniques to classify the sentiment of the reviews. The results indicated that SVM was more accurate, with an accuracy value of 79%, compared to naive Bayes (NB). Overall, social media platforms serve as a data platform that enhances the quality of restaurant service through online reviews by consumers. They are a trusted source of information that helps consumers assess the quality of restaurants. Therefore, consumer reviews on the quality and service of restaurants are crucial.

2.3. Electronic Word of Mouth (eWOM)

Electronic word-of-mouth (eWOM) refers to the exchange of information or opinions, both positive and negative, through the Internet (Lim et al., 2022). This information is sourced from consumers, online social networks, or personal relationships (Chu, 2019), influencing the speed and rate of dissemination that consumers can choose from various platforms, such as online marketplaces or independent websites, and in various formats such as online reviews, blogs, and videos. Consumers can express their opinions through text, images, or ratings (Pyle, 2021). Consumers increasingly rely on electronic word-of-mouth to search for information about service providers and share personal experiences in using services (Golmohammadi, 2020). Consumer emotions are crucial in explaining post-purchase consumer behavior (Yan, 2018). Therefore, electronic word-of-mouth (eWOM) has become a significant reference for making purchase decisions and responding to consumer behavior in the digital age.

2.4. Emotional Intensity

Expressing emotions and feelings through online comments or reviews about products or services on various media platforms can vary significantly. Social networks like Facebook and review websites like wongnai.com are two such platforms that showcase different patterns of consumer behavior. Consumers tend to prioritize their self-image and interactions with other members on social media, often resulting in a positive bias in their comments. However, on review websites, where consumers are usually unfamiliar with each other, comments tend to align more with the genuine feelings of consumers (Liu, 2021). Conflicting online reviews, whether positive or negative, can help consumers decide whether they need more information about a product (Ruiz-Mafe, 2018). Astudy conducted by Li investigated the effect of emotional intensity on perceived usefulness. The study analyzed data from 600,686 reviews of 300 popular U.S. restaurants on Yelp using text mining and econometric analysis. The results showed that positive emotional intensity had a negative impact on the perceived usefulness of the review, whereas negative emotional intensity had a positive impact (Li, 2020). Evidently, consumers’ perceptions and emotional responses play vital roles in their purchase intent, and online reviews can significantly enhance their confidence in decision-making.

2.5. Machine Learning

Machine learning (ML) is a powerful technique involving training machines or computers to learn from large datasets, known as training datasets, to find answers and make accurate predictions. This process involves applying various methods and algorithmsto the training data and then using these methodsto predict the outcomes of new datasets (Raschka & Mirjalili, 2019). There are three main types of machine learning:

1) Supervised Learning involves using past data or variables with known answers as training data to predict unknown values or future data. This process involves adjusting parametersto minimize the difference between the target and computed outputs (Jo, 2021). If the variables are discrete, this processis called classification, while if the data or variables are continuous, regression or prediction includes classification with decision trees or neural networks.

2) Unsupervised learning, on the other hand, involves studying data without known answers or relationships between variables. This type of machine learning is used to find relationships or group data with similar characteristics (Kalita, 2022).

3) Reinforcement Learning (RL) is a type of machine learning that enables computers to learn by interacting with their environment through trial and error, similar to human learning. This type of learning consists of two main components: the agent and the environment, where the agent interacts with the environment to learn. RL is based on Markov decision processes (MDPs), which require agentsto make decisions that account for environmental uncertainty and experience (Vamvoudakis et al., 2021). According to a predefined evaluation metric, the agent can receive a good score (Dong et al., 2020). RL aims to find the agent’s policy by maximizing the desired outcome (Palmas et al., 2020).

2.6. Decision Tree

Decision Tree is a popular algorithm used in supervised learning. This algorithm works without parameters, making it a valuable tool for classification and prediction tasks. The decision-making process starts at the root node, which is the starting point for the algorithm’s sequence of decisions. From there, branch nodes connect decision-making conditions, which divide the data into two or more branches based on the child nodes’ conditions. Finally, the algorithm reaches the end nodes, also known as leaf nodes, representing the outcomes obtained after the classification process. Decision Tree is a versatile algorithm used in various applications, including finance, healthcare, and marketing (Sumathi et al., 2022). The process of the Decision Tree is shown in Figure 1.

According to Figure1, the decision tree is a powerful tool known for its rule-based nature. In this approach, if-else rules are created based on the values of independent variables. The model does not rely heavily on equations dictating the relationships between independent and dependent variables. The approach is advantageous because it provides clear and easy-to-understand results, making it easy to apply in various scenarios.

2.7. Artificial Neural Network

Artificial neural networks (ANNs) are complex systems that possess the ability to learn and adapt from experience. They are designed to emulate the functioning of neural networks in the human brain, with operational units interconnected by an extensive network (Desai, 2021). An artificial neural network’s performance largely depends on its learning process, which involves setting the network architecture and relevant parameters (Du, 2019). Typically, the architecture of a multilayer perceptron neural network comprises an input layer, hidden layers, and an output layer. To better understand this architecture, see Figure 2, which provides a visual representation.

Each layer consists of a certain number of neurons, with data flowing continuously forward and the network error values spreading backward. The initial connection weights and criteria of the network start randomly. Then, a training process is employed by comparing the error values of the results with the actual values. The connection weights and criteria of the network are adjusted to reduce the error values until the results are close to the desired outcomes (Graupe, 2019).

2.8. Linear Regression Analysis

Linear Regression is a statistical method that helps to predict a dependent variable using one or more independent variables by establishing a linear relationship between the response variable and the predictor variables in a non-linear format. According to Ciaburro (2018), the primary objective of linear Regression is to determine the coefficients of the predictor variables (X) to showcase how these variables impact the response variable (Y), as depicted in the general equation of regression analysis in equation (1).

Y=a + b₁X₁ + b₂X₂ + b₃ +…+ b_nX_n (1)

When it comes to studying relationships between variables, it’s important to understand the key terms involved. The independent variable, denoted as X, is the variable that is manipulated or changed in an experiment or study. On the other hand, the dependent variable, denoted as Y, is the variable that is being measured or observed. The intercept, denoted as a, representsthe value of the dependent variable when the independent variable is equal to zero. Lastly, the parameter, denoted as b, represents the degree and direction of the relationship between the independent and dependent variables. It’s important to keep these terms in mind when analyzing data and drawing conclusions.

2.9. Feature Selection

Feature selection is an essential pre-processing step in data analysis that involves reducing data size while maintaining the most appropriate data quality possible. There are several methods for feature selection, including Enter Regression, Forward Selection, and Backward Elimination. Enter Regression is a method of selectively introducing independent variables into the regression equation one at a time, using the criterion of the simple correlation coefficient. The process involves selecting the independent variable with the highest correlation with the dependent variable first and conducting a statistical test to determine if it can significantly predict the dependent variable. The process then continues by selecting the following independent variable in the order of relevance, and the best criteria are applied in each round when combined with previously selected variables. This process continues until no further independent variables can be added to the equation if deemed appropriate. Forward selection is similar to enter Regression, where independent variables are selected for the equation one at a time using the criterion of the simple correlation coefficient. The process involves selecting the independent variable with the highest correlation with the dependent variable and conducting a statistical test to determine if it can significantly predict the dependent variable. The process then continues by selecting the following independent variable in the order of relevance, and the best criteria are applied in each round when combined with previously selected variables. This process continues until no further independent variables can be added to the equation if deemed appropriate (Bhadra & Bandyopadhyay, 2021). Backward Elimination, on the other hand, is a method opposite to variable selection by stepwise addition. It starts with creating a regression equation that includes all independent variables. The p-value of each independent variable in the predictive model is then calculated (Nazarathy, 2021).

Subsequently, the variable with the highest p-value is considered, and its p-value is compared with the predefined significance level, such as Pr=0.05. The variable is removed from the equation if the p-value exceeds the specified significance level. The process is repeated with the remaining variables until no p-value exceeds the designated significance level. The variable selection process is terminated at that point, and the equation is considered appropriate. Besides reducing the data size, feature selection reduces the learning process’s complexity and overfitting. Overfitting occurs when a machine-learning model can make accurate predictions for the training dataset but fails to provide accurate forecasts for new data (Ba et al., 2023). This makes feature selection an essential step in data analysis and machine learning.

2.10. Method Evaluation

The confusion matrix is a valuable tool for evaluating the accuracy of binary classification or variables with only two groups of results. On the other hand, regression analysis uses different measurement methods, such as the R-squared value and mean absolute percentage error (MAPE) or root mean square error (RMSE), to estimate the relationship between a dependent variable and one or more independent variables.

Supervised learning is a type of machine learning that can measure accuracy because it uses a target as the starting point. This target value represents the expected outcome, and the model’s forecast values can be compared against it to measure the expected deviation from the initial data. Therefore, both classification and regression models can be evaluated for their accuracy, which indicates their efficiency.

2.11. Weight by Correlation

The Weight by Correlation operator is a highly useful tool for determining the relevance of attributes by computing the correlation value for each attribute of the input ExampleSet concerning the label attribute. This approach uses Correlation as a weighting scheme and returns the absolute or squared value of Correlation as the attribute weight.

The weight of an attribute is calculated concerning the label attribute using Correlation. The higher the weight of an attribute, the more relevant it is considered. Correlation is a numerical value that measures the degree of association between two attributes, X and Y. Correlation ranges from -1 to +1, where a positive correlation implies a positive association and a negative correlation value suggests a negative or inverse association. In the case of a positive correlation, large values of X tend to be associated with large values of Y, and small values of X tend to be associated with small values of Y. On the other hand, in a negative correlation, large values of X tend to be associated with small values of Y, and small values of X tend to be associated with large values of Y. Suppose two attributes are X and Y, with means X' and Y' and standard deviations S(X) and S(Y), respectively. The Correlation is computed as the summation from 1 to n of the product (X(i) X')∙(Y(i) Y') and then dividing this summation by the product (n 1)∙S(X)∙S(Y), where X(i)-X')∙(Y(i)-Y') and then dividing this summation by the product (n-1)∙S(X)∙S(Y) where n is the total number of examples and i isthe increment variable of the summation. In a positive Correlation, if an X value is above average, the associated Y value is also above average. Then, the product (X(i)-X')∙(Y(i)-Y') would be the product of two positive numbers that would be positive. If the X and Y values were below average, the product above would be of two negative numbers, which would also be positive. Therefore, a positive correlation is evidence of a general tendency that large values of X are associated with large values of Y, and small values of X are associated with small values of Y. In a negative correlation, if an X value was above average, and the associated Y value was below average, the product (X(i) X')∙(Y(i) Y') would be the product of a positive and a negative number, making the product negative. If the X value is below average, and the Y value is above average, the product above would also be negative. Therefore, a negative correlation is evidence of a general tendency that large values of X are associated with small values of Y, and small values of X are associated with large values of Y.

3.1. Data Collection Methods

The researcher has conducted a comprehensive study, gathering data from various sources to use as input for analysis. Specifically, restaurant-related data was collected from the wongnai.com website within the Bangkok metropolitan area. This includes detailed information about restaurants, such as their categories, price ranges, parking availability, the number of reviews, credit card acceptance, Wi-Fi service, suitability for groups, provision of alcoholic beverages, website availability, delivery service, and review scores. Furthermore, Facebook information for the restaurants, such as check-ins, page likes, and followers, was also collected. In addition, population density data in the area was obtained from Bangkok’s Geographic Information System (GIS) Technology Center. Google utilized natural language AI to analyze the sentiment of the customers’reviews on wongnai.com. The sentiment analysis will provide valuable insights into the opinions about the restaurants. Lastly, all the gathered information was recorded in a database for the following data preparation steps.

3.2. Data Preprocessing Methods

The methodology employed for data analysis in this research involves utilizing advanced machine learning techniques. A detailed workflow diagram, as illustrated in Figure 3, showcases the step-by-step process of this data analysis approach.

3.3. Data Cleaning

To ensure accurate analysis, it is important to carefully examine and address any outliersin the data set. For instance, if certain restaurants’ values are missing due to incomplete information obtained from websites, these values must be replaced with appropriate substitutes. One common approach is to replace the missing factor values with 0, which allows for a more comprehensive and reliable analysis.

3.4. Sentiment Analysis

This involves analyzing customer feedback from online reviews, where customers who have used the restaurant services write their opinions on the wongnai.com website. Natural Language AI, a cloud service by Google, is used for this analysis. Machine learning can analyze unstructured text by understanding natural language (NUL). This is often used for sentiment analysis to understand customer opinions and obtain in-depth information that can be used to develop business products and services. The results are a score that can be interpreted as follows: 1) -1.0 to -0.25 represents negative sentiment, 2) -0.25 to 0.25 represents neutral sentiment, and 3) 0.25 to 1.0 represents positive sentiment.

3.5. Data Transformation

The researcher has transformed the data into a format suitable for analysis using machine learning techniques, as illustrated in Table 1.

3.6. Clustering

The researcher grouped restaurant types into three categories using the average price, a distinctive characteristic of each type. This was done to understand how each category has factors that influence the restaurant business’s performance differently. This research chose to use the K-means Clustering technique for grouping.

3.7. Feature Selection

The data size reduction process implemented in this study was performed while ensuring that the optimal data quality for analysis was preserved. This was done by utilizing the evolutionary selection technique to calculate the weights of the specified factors. The weight assigned to each factor was determined based on relevance, with the higher-weighted factors being considered more significant. The inclusion or exclusion of a factor was based on its weight calculated during the process. Various methods were used to select the variables, such as forward selection, backward elimination, and enter regression. The weights of the specified factors were calculated to reduce the data size while preserving the optimal data quality for analysis using the evolutionary selection technique. The higher the weight of a factor in the calculation, the more relevant it is considered. If a factor’s calculation results in a higher weight, it is retained. Conversely, if a factor’s calculation resultsin a lower weight, the factor is excluded from the data analysis process. The methods employed in this study include forward selection, backward elimination, and enter regression for variable selection.

3.8. Data Analysis

This study compares the predictive model performance created using machine learning algorithms with RapidMiner software. The models include decision trees, artificial neural networks, and linear regression analysis. The data was divided into two partsfor testing: the training dataset and the remaining data for testing, using the K-fold Cross-Validation method.

3.9. Method Evaluation

The performance of the prediction models was evaluated and compared by considering Accuracy and overall performance measured by the F-measure.

3.10. Weight by Correlation

Select factors influencing review scores reflecting the performance of restaurants by choosing a weight > 0.1.

4.1. Data Collection

We have conducted a comprehensive study and collected data from various sources using a program developed with C# .NET to retrieve web page content. The program has been designed to extract data from websites based on multiple factors observed on the web pages for a thorough analysis. The study has encompassed the following factors:

1) Restaurant data within the Bangkok area, selected from restaurants with a Facebook fan page and information on the Wongnai website, totaling 1,750 establishments. After a thorough review of errors and missing data, 113 entries were identified, resulting in a final count of 1,637 restaurants. The study has included various aspects such as restaurant categories, price ranges, parking availability, number of reviews, credit card acceptance, Wi-Fi service, suitability for groups, availability of alcoholic beverages, website functionality, delivery service, and review scores.

2) Facebook fan page data for restaurants, including check-ins, page likes, and follower counts.

3) Population density data in the area obtained from Bangkok’s Geographic Information System (GIS) Technology Center.

4) Sentiment level analyzed using Natural Language AI developed by Google, based on customer reviews for each restaurant on Wongnai.com.

The data will now be stored in a database for the data-cleaning process.

4.2. Clustering

To analyze the varioustypes of restaurants, we classified them into three distinct groups. To achieve this, the researcher utilized the K-means Clustering technique, which involved grouping the restaurants based on similar characteristics and features. The results of this analysis were then presented in Table 2, which provided a clear and concise illustration of the different restaurant types and their respective features.

4.3. Forecasting Model

The research employs RapidMiner software to develop three forecasting models: Decision Tree, Logistic Regression, and Artificial Neural Networks. Cross-validation is used to evaluate the performance of each type, where the data is split into ten equal parts using the 10-fold cross-validation technique. Once the data is divided, variable selection is performed through Enter Regression, Forward Selection, and Backward Elimination. Finally, the performance of the best forecasting models from each type is compared. The findings of this research are asfollows: the Decision Tree model performed the best, followed by the Artificial Neural Network and Logistic Regression models. However, all three models showed similar accuracy levels, and the difference in performance was not statistically significant.

Based on the data presented in Table 3, it is evident that the artificial neural network is the most efficient algorithm for forecasting. This algorithm has considered all variables (Enter Regression) and achieved an impressive accuracy rate of 85.22%. Furthermore, the artificial neural network has emerged as the most efficient algorithm when considering the overall prediction performance based on the F-measure. It has displayed an overall prediction performance of 87.47%, which is a remarkable achievement in forecasting. These findings suggest that the artificial neural network is highly dependable and trustworthy for making accurate predictions.

Table 4 displays the outcomes of the enter regression method used to select the factors that affect the review scores of low-priced restaurants on social commerce based on the artificial neural networks forecasting model. The analysis revealed that Sentiment, Seats, Group, Wi-Fi, and Population Density are the crucial factors that influence the performance of low-priced restaurants. These factors were chosen based on their weight, which was greater than 0.1 (> 0.1).

According to the experimental results, the best algorithm for forecasting is the artificial neural network. This algorithm uses the "Enter Regression" method, where all variables are considered, resulting in an accuracy of 91.08%. Additionally, when considering the overall prediction performance (F-measure), the artificial neural network outperforms other algorithms with an overall forecasting performance of 88.06%.

According to the data presented in Table 6, the artificial neural networksforecasting model hasidentified several key factors influencing the review scores of low-priced restaurants on social commerce platforms. These factors were selected using the enter regression method and were assigned a weight >0.1. The most significant factors were Sentiment, Price, Alcohol, CreditCard, and Like. These factors play a crucial role in determining the performance of low-priced restaurants on social commerce platforms. On the other hand, population density was identified as the least significant factor, which has a negligible impact on the review scores.

According to the data presented in Table 7, the artificial neural network algorithm has proven to be the most efficient in forecasting, outperforming all other algorithms by a considerable margin. The algorithm employs the backward elimination method for variable selection and has achieved a high accuracy rate of 85.46%. When assessing the overall forecasting performance (F-measure), the artificial neural network algorithm remains the most efficient, with an impressive overall prediction performance rate of 85.89%. In analyzing moderate-priced restaurants’ performance in social commerce, the backward elimination method was used with the artificial neural networks forecasting model. The result is presented in Table 8, which shows the factors influencing review scores. The factors were ranked based on their weight by correlation, with a weight > 0.1 being considered significant. The three factors most significantly impacting review scores were Sentiment, Seats, and Price. On the other hand, the factor with the most negligible effect was CreditCard.