Analyzing Forest Leisure and Recreation Consumption Patterns Using Deep and Machine Learning

Abstract

Globally, forest leisure and recreation (FLR) activities are widely recognized not only for their environmental and social benefits but also for their economic contributions. To better understand these economic contributions, it is vital to examine how the regional economic levels of customers vary when consuming FLR. This study aimed to empirically examine whether the regional economic level of residents (i.e., gross regional domestic product; GRDP) is classifiable using FLR expenditure data, and to interpret which variables contribute to its classification. We acquired anonymized credit card transaction data on residents of two regions with different GRDP levels. The data were preprocessed by identifying FLR-related industries and extracting key spending features for classification analysis. Five classification models (e.g., deep neural network (DNN), random forest, extreme gradient boosting, support vector machine, and logistic regression) were applied. Among the models, the DNN model presented the best performance (overall accuracy = 0.73; area under the curve (AUC) = 0.82). SHAP analysis showed that the “FLR industry” variable was most influential in differentiating GRDP levels across all the models. These findings demonstrate that FLR consumption patterns may vary and are interpretable by economic levels, providing an empirical framework for designing regional economic policies.

1. Introduction

Globally, forest leisure and recreational (FLR) activities have been extensively studied as a significant component of outdoor leisure [1,2,3]. These activities have been shown to contribute positively to both physical and mental health, while also enhancing individual and social well-being [4,5]. They encompass camping and hiking, in addition to horseback riding, marathon running, fishing, skiing, and other forms of nature-based recreation [6,7]. Beyond their health benefits, these activities also represent a substantial economic contributor, generating considerable value through tourism and related industries.

According to a global report, approximately 3.3% of global GDP was generated by cultural and nature-based recreation [8]. In the United States, outdoor recreation accounted for 2.3 percent (USD 639.5 billion) of GDP in 2023 [9]. Moreover, recent studies highlight not only the industrial scale of FLR but also its role in revitalizing regional economies [7,10]. Forest recreation has been positioned as a means to excavate local history and cultural identity, facilitating the integration of culture and tourism for sustainable development [11,12]. Furthermore, tourism generally has been shown to have both direct and indirect economic impacts on related industries, and industries associated with FLR have the potential to generate similar economic effects [13,14]. These economic and social impacts have been supported by quantitative and empirical studies [12,15,16]. For example, Zhou et al. (2024) [15] demonstrated that nature-based leisure activities enhance regional economic vitality and promote socio-cultural engagement. Similarly, Tyrväinen et al. (2013) [12] showed that forest recreational visits contribute to local economic development. FLR patterns have additionally been analyzed in spatial dimensions to understand spatial consumption behaviors and visitation purposes. While the economic value of FLR is widely confirmed, there is a lack of empirical research investigating the determinants of consumer spending in FLR markets using actual transaction data to interpret their impact on regional economic outcomes.

A recent study on FLR markets reported that there are statistical differences in consuming patterns of forest leisure and sports related to individual incomes [17]. To analyze such markets effectively, examining actual consumer characteristics, including demographic and spending information, is essential [18,19]. Furthermore, statistical approaches are limited in handling complex consumer attributes, and the application of artificial intelligence (AI) techniques has become increasingly widespread across various fields [20,21,22]. Recent studies in South Korea have actively utilized AI to reveal consumer characteristics by analyzing credit card transaction data across various industries and product types [23,24,25,26,27,28,29]. Similarly, global research trends show the continued use of AI in analyzing consumer behavior patterns in a wide range of fields [30,31,32,33]. Although AI has been applied in forestry primarily for remote sensing and policy-related text analysis, quantitative research focusing on consumer behavior remains limited [34,35,36]. Accordingly, further research should clarify how FLR consumption is related to regional economic level (i.e., gross regional domestic product; GRDP), offering empirical insights into regionally differentiated consumption patterns.

In AI analysis, consumer characteristics, which are often highly complex, are typically structured in tabular format and are analyzed using classification and regression techniques based on machine and deep learning algorithms [32,33,37,38]. These models, including deep neural networks (DNNs) and tree-based structures, compute the influence of each input variable as it passes through nodes, thereby classifying the target or generating nonlinear regression functions that are difficult to estimate using traditional statistical models [39,40,41,42,43]. Classification model performance is commonly evaluated by comparing metrics such as overall accuracy (OA), F1 score, and area under the receiver operating characteristic curve (ROC-AUC), which enables the selection of the optimal model [44,45,46,47]. To further interpret the results, SHapley Additive exPlanations (SHAP) is often applied to estimate the contribution of each input feature to the prediction outcomes [48,49].

FLR industries hold significant potential to contribute to regional economic development. Consumer behavior in FLR markets can vary substantially depending on individual economic status. Understanding whether and how these consumption patterns in markets differ as per customers’ income levels is essential for designing strategies that support regional economies. However, due to privacy regulations, acquiring actual consumption data remains challenging, and the available data often involves complex, multi-dimensional characteristics that are difficult to analyze using conventional statistical approaches. To address these gaps, we investigated the differences in FLR consumption patterns based on regional economic levels (i.e., GRDP) by applying deep and machine learning models to individual credit card transaction data. Furthermore, SHAP was employed to interpret the importance of each feature and provide transparency in model decision-making. Through this methodological approach, the study may contribute to advancing empirical understanding of income-based consumption behaviors in the forest economic domain by applying interpretable classification techniques to individual-level transaction data, thereby offering a practical framework for classifying regional economic levels.

2. Materials and Methods

2.1. Conceptual Workflow of GRDP-Level Classification

The overall workflow of this study comprised three main stages: data acquisition, GRDP-level classification, and model assessment (Figure 1). In the data acquisition stage, customer credit card consumption data, including demographic and spending information, were collected from a credit card company and preprocessed. FLR industries were identified through expert surveys and subsequently classified according to the Korean Standard Industrial Classification (KSIC) System. These processes yielded a structured dataset with the GRDP level as the dependent variable and a set of independent variables, including sex, age, spending behavior, and expenditure on FLR industries. In the GRDP-level classification stage, one deep learning model (DNN), three machine learning models (random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM)), and one statistical model (logistic regression (LR)) were applied to perform the classification tasks. Finally, in the model assessment stage, model performance was evaluated using OA, F1 score, and AUC, followed by best-fit model selection and feature importance interpretation using SHAP values.

2.2. Data Acquisition and Preparation

As a first step, this study preliminarily examined administrative regions in South Korea that fell within the top 10%, 45%–55%, and bottom 80% quantiles of GRDP before proceeding to GRDP-level classification using actual customer data. These three economic quantiles were based on the assumption that FLR consumption patterns would differ depending on the economic level of residents, thereby enabling comparative analysis across income groups. However, the dataset from the bottom 80% quantile was considerably limited in both size and FLR industry expenditure. Due to these constraints, the bottom 80% quantile data were excluded from the analysis. The final analysis was focused on comparing residents from regions in the top and middle GRDP quantiles.

According to the investigation, 23 cities or administrative districts fell within the top 10% and 45%–55% GRDP levels. To ensure comparability across regions with different economic profiles, while controlling for population size, the Jung-gu district of Seoul (GRDP 10%; top GRDP) was selected as a suitable region for further classification, with a population of 121,312 in 2023, and Naju city of Jeollanam-do (GRDP 45%–55%; mid GRDP) with a population of 117,377 in 2023.

Second, we acquired actual customer credit card consumption information for top GRDP and mid GRDP from a credit card company. Due to strict regulations of customer information privacy, the monthly demographic and spending information of 207,485 customers was collected, including variables such as “period (year-month)”, “administrative region-province”, “administrative region-city/county/district”, “sex”, “age group”, “KSIC middle & small code”, “detailed KSIC name and code,” “transaction amount”, “number of transactions”, and “number of members” for the period from January to November 2023.

Third, to identify industries related to FLR, a survey was conducted with ten experts, referencing the Korea Consumer Index and the KSIC [50,51]. A total of 13 mid-level KSIC industries that fell within the top 20% in relevance and scored 40%–90% in impact were selected. These selected mid-level KSIC industries were then reclassified into five FLR industry categories based on thematic similarity and consumption behavior (e.g., tourism and accommodation: hotel services, resort condominiums, other lodging services; cultural shopping: daily goods, arts and crafts, forest products; food and beverages: restaurant services; leisure services: recreational sports, performing arts, operation of tourism facilities; education: general education, cultural education, publishing).

Based on these assessments, 34 detailed sub-industry codes were identified, and FLR-related transactions were extracted from credit card data based on these codes, resulting in a final dataset of 38,433 FLR consumption records. For classification modeling, the GRDP quantile corresponding to each region was used as the dependent variable, and sex, age group, FLR industry category, transaction amount, number of transactions, and number of members were used as independent variables. Nominal variables were transformed using label encoding to make them suitable for machine learning analysis, and numerical variables were standardized to minimize the impact of differences in units and scales across variables.

2.3. Development of GRDP Classification Model

In this study, five models, including machine and deep learning and statistical classification models, were used to analyze regional differences based on consumer transaction data. This analysis focused on a comparative study of top and mid GRDP. Classification was performed by training the models to distinguish between these two regions based on consumer spending patterns using the features extracted by each algorithm. To ensure data homogeneity, 10,000 samples were randomly selected from each region, resulting in a dataset of 20,000 records for training and validation.

In the Python 3.8.0 environment, five classification algorithms (DNN, XGBoost, RF, SVM, and LR) were applied to examine whether cardholders in the top 10% GRDP region and those in the median quantile (45%–55%) could be classified based on consumption behavior. To achieve optimal model performance, hyperparameter tuning was conducted using the Optuna library, with parameter ranges for each model set based on previous studies [52,53,54].

2.3.1. DNN

To classify the structured data using deep learning, a customized DNN architecture was designed based on TensorFlow. DNNs comprise multiple fully connected layers capable of modeling complex nonlinear relationships between inputs and outputs. In this study, the DNN architecture enabled the analysis of intricate interactions between independent variables used in the classification task [55,56,57].

Utilizing Optuna, the best-performing DNN model comprised six hidden layers with 512, 512,256, 256, 128, and 64 neurons. Dropout regularization was applied to the hidden layer between the 4th and 5th layers at a rate of approximately 0.15. The rectified linear unit activation function was used in the hidden layers, and a sigmoid function was applied to the output layer to perform binary classification. The model was trained with the Adam optimizer, using a batch size of 128 for 300 epochs.

2.3.2. RF

RF is an ensemble learning method that constructs multiple decision trees and aggregates their classifications through majority voting. It is renowned for its ability to estimate feature importance in prediction and classification tasks [58]. Each tree was trained on different combinations of samples and features to reduce the risk of overfitting inherent to individual trees.

A model was built using 196 decision trees, each with a maximum depth of 17. The minimum number of samples required to split a node was set to seven, and each leaf node was required to have at least four samples. These settings were selected to balance model complexity, interpretability, and computational efficiency. The Gini impurity criterion was used for node splitting.

2.3.3. XGBoost

XGBoost is an implementation of gradient boosting, designed for performance optimization and computational efficiency. Unlike RF, XGBoost sequentially builds decision trees by correcting the residuals of previous trees. It approximates the loss function using second-order derivatives for both the classification and regression analyses. Moreover, it employs both L1 and L2 regularizations to prevent overfitting.

In this study, the model was trained with 126 boosting rounds, a maximum tree depth of 7, and a learning rate of 0.13. The objective function was set to softmax for multiclass classification, and the evaluation metric was the multiclass log-loss. Features such as the built-in handling of missing values and regularization techniques contribute to the model’s stable performance.

2.3.4. SVM

An SVM is a margin-based classifier that aims to find a hyperplane that maximizes the margin between different classes. Strategies such as one-vs-rest or one-vs-one are typically applied. In this study, a radial basis function kernel was used to capture nonlinear boundaries for multiclass classification. The penalty parameter (C) was set to 9.96, and the gamma value was set to 0.98. Probability estimates for ROC analysis were obtained using Platt scaling. All features were standardized using z-score normalization before training.

2.3.5. LR

LR is a linear classification model that estimates the log odds of an outcome based on a linear combination of input features. It is widely used as a baseline model in multiclass classification due to its simplicity, speed, and interpretability. In this study, the regularization parameter (C) was set to 0.01, and the solver was specified as “liblinear.” The number of iterations was adjusted accordingly. To address class imbalance, the class weight = “balanced” option was used to optimize the model.

2.4. GRDP Classification Models Assessment

The performance of the classification models was evaluated using three metrics—OA, F1 score, and AUC. OA measures the proportion of correct classifications out of the total classifications and reflects the overall performance of a classifier. However, when class imbalance is present, accuracy alone may overestimate model performance, necessitating the consideration of additional evaluation metrics. The F1 score is the harmonic mean of precision and recall and provides a balanced assessment of model performance, particularly for the positive class. Precision refers to the proportion of correctly classified positive instances out of all classified positives, whereas recall indicates the proportion of correctly classified positive instances out of all actual positives. The F1 score is particularly useful when precision and recall are equally important. AUC represents the area under the ROC curve and reflects the probability that the classifier will correctly distinguish between randomly selected positive and negative samples. AUC values close to 1 indicate high classification performance. This metric is particularly suitable for evaluating the tradeoff between sensitivity and specificity in binary classification problems. Theoretically, the AUC is derived by plotting the ROC curve based on sensitivity (recall) and specificity (1—false-positive rate) over various threshold values and calculating the area under the resulting curve. This quantifies the ability of the model to correctly distinguish between positive and negative instances. OA, precision, recall, and F1 score are defined as follows:

$$\mathrm{O}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{l} \mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(1)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{TP}{TP+FP}}$$

(2)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{TP}{TP+FN}}$$

(3)

$$\mathrm{F}1={\displaystyle \frac{2(Precision\times Recall)}{Precision+Recall}}$$

(4)

2.5. SHAP-Based Feature Importance Analysis

To identify the most influential features in the optimal models, this study employed the SHapley Additive exPlanations (SHAP) technique, which applies to three classification models of excellent performance. SHAP allows the interpretation of complex machine and deep learning models by quantifying feature importance even when the models themselves are not inherently interpretable [48,49]. Unlike local explanation methods such as local interpretable model-agnostic explanations, which compute feature weights based on localized perturbations, SHAP is grounded in cooperative game theory and calculates Shapley values. These values represent the average marginal contributions of each feature across all possible feature combinations. By comparing classifications made using a specific feature with average classifications, SHAP computes the contribution of the feature to the model output. Through this mechanism, SHAP decomposes the classification into additive feature contributions, enabling a consistent and unified measure of feature importance across different types of models. In this study, SHAP was used to evaluate the contributions of the independent variables for both interpretation and model refinement.

In the SHAP summary plots, each dot represents a single observation, with color gradients indicating the value of the corresponding variable. For continuous variables such as transaction amount, transaction frequency, and number of users, red tones represent higher values while blue tones indicate lower values. For categorical variables encoded numerically, color interpretation follows the order of encoding. In terms of case of gender, female participants (encoded as 1) are presented in red, and male participants (encoded as 0) appear in blue. For age groups, values range from teenagers to individuals in their 60s, with redder colors indicating older age groups and bluer tones representing younger groups.

The FLR industry type variable is also numerically encoded according to increasing economic relevance in leisure activities. Here, redder shades correspond to industries such as food and beverages or leisure services, which are located at the higher end of the encoding scale. Bluer colors typically represent tourism and accommodation or education services, which occupy the lower end of the spectrum. This color mapping helps to intuitively identify which types of consumers and industries contribute more strongly to the prediction of regional economic levels.

3. Results

3.1. Descriptive Statistics for Top and Mid GRDP

Before the main analysis, descriptive statistics were computed for the datasets corresponding to each GRDP quantile region (

Table 1. Descriptive statistics of monthly card usage.

		Total Dataset		ForestLeisure and Recreation-Related Expense
Division		Top GRDP (n)	Mid GRDP (n)	Top GRDP (n)	Mid GRDP (n)
Gender	Male	68,117	40,920	11,508	5900
	Female	65,454	32,994	12,041	5384
Age	10s	5187	2161	937	329
	20s	22,528	14,500	4023	2166
	30s	25,374	13,193	4942	2239
	40s	25,129	13,834	4808	2196
	50s	23,586	14,341	3710	2023
	Over 60s	31,767	15,885	5129	2331
Forest leisure and recreation industries	Tourism and accommodation	-	-	2518	608
	Education	-	-	3009	1164
	Cultural shopping	-	-	9859	3850
	Food and beverages	-	-	6383	5092
	Leisure service	-	-	1780	570
Monthly spending amount	Mean	2,632,996	2,324,204	1,799,540	1,788,422
	Std	9,640,675	14,592,324	4,139,138	5,709,750
	Min	600	550	4600	1800
	25%	185,505	152,853	168,660	176,253
	50%	617,200	432,265	460,500	470,000
	75%	2,038,845	1,331,211	1,372,000	1,332,078
	Max	338,627,655	591,602,094	50,583,050	104,947,940
Number of transactions	Mean	81	63	54	47
	Std	249	321	113	145
	Min	3	3	3	3
	25%	7	6	6	6
	50%	16	13	15	13
	75%	53	35	44	33
	Max	5958	11,918	1188	2549
Number of members	Mean	25	21	22	20
	Std	38	56	32	49
	Min	3	3	3	3
	25%	5	4	4	4
	50%	10	7	9	7
	75%	27	15	23	16
	Max	441	1172	270	698

). In the original dataset, the top GRDP (i.e., Jung-gu) and the mid GRDP (i.e., Naju) presented a sex imbalance, with approximately 5300 more male than female consumers. However, in the filtered dataset containing only transactions related to FLR industries, the number of card transactions by women in the top GRDP exceeded that by men. In terms of age, consumers in their teens showed the lowest number of transactions in both the original and filtered datasets, whereas those aged 60 and older accounted for the highest number of transactions in both the top GRDP and mid GRDP. After filtering for FLR industries, the transaction count in top GRDP decreased from 133,571 to 23,621 and in mid GRDP from 72,284 to 11,284. Despite this reduction, the ratio of transaction counts between top GRDP and mid GRDP remained consistent at approximately 2:1.

In the original dataset, the average expenditure of the top GRDP was 308,792 KRW higher than that of the mid GRDP. However, after filtering, the average spending of the top GRDP slightly exceeded that of the mid GRDP by 11,118 KRW. Although the standard deviation decreased in both regions after filtering, the variation in mid GRDP remained greater than that in the top GRDP. Before filtering, the minimum values were approximately 600 and 550 for the top GRDP and mid GRDP, respectively, increasing to 4800 and 1800 after filtering. In contrast, the maximum value for the mid GRDP was consistently approximately twice that of the top GRDP, both before and after filtering. The median values were 184,935 KRW higher in the top GRDP before filtering. However, after filtering, the mid GRDP exceeded the top GRDP by 9500 KRW. Both the mean and median transaction counts decreased after filtering, although the top GRDP maintained a higher median usage. However, the maximum transaction count in the mid GRDP was more than double that in the top GRDP. The number of unique members also declined after filtering, and the top GRDP had higher average and median numbers of users. Similar to the transaction count, the mid GRDP showed more than twice the maximum number of users compared with that in the top GRDP, indicating regional differences in the feature characteristics between the top and mid GRDP.

3.2. Machine Learning and Statistical Analysis of Customer Data Derived from Credit Card Sales Data

Among the five models, DNN showed the best overall performance, with top scores in both F1 and AUC, at approximately 0.73 and 0.82, respectively (

Table 2. Comparison of the performance of each model.

Model	OA	F1 Score	AUC
DNN	0.73	0.73	0.82
RF	0.73	0.73	0.81
XGBoost	0.72	0.72	0.81
SVM	0.72	0.70	0.80
LR	0.58	0.61	0.62

DNN, deep neural network; RF, random forest; XGBoost, extreme gradient boosting; SVM, support vector machine; LR, logistic regression; OA, overall accuracy; AUC, area under the curve.

). On the other hand, RF showed the highest OA, exceeding that of DNN by approximately 0.01. XGBoost showed minimal difference in overall performance compared to RF, with a slightly lower accuracy (by approximately 0.01). Additionally, the SVM model achieved an AUC of approximately 0.7981 (rounded to 0.80), which is close to the commonly accepted threshold for an excellent model. Its F1 score was approximately 0.70, showing only a slight difference compared to the top three models. LR showed the weakest overall performance, even after parameter optimization, with an AUC of only 0.66 and the lowest accuracy at 0.6.

Further analysis using confusion matrices revealed that the DNN model accurately classified both true positives (TPs) and true negatives (TNs). It showed slightly better performance than RF, XGBoost, and SVM in classifying TNs. The RF and XGBoost models were slightly better than DNN in classifying TPs and tended to classify more instances as belonging to the top GRDP. In contrast, LR showed a tendency to classify most cases as mid GRDP and had significant difficulty in accurately identifying TPs compared with the other models. Across all five models, the AUC scores were higher than the accuracy scores, and a visual analysis confirmed that the AI-based models outperformed the statistical model (Figure 2).

3.3. Feature Importance of GRDP Models

Based on overall model performance, a final SHAP analysis was conducted for the top three models—DNN, XGBoost, and RF. In the DNN model, the “FLR industries” was identified as the most influential variable (Figure 3). Moreover, a higher transaction count had a negative SHAP value, decreasing the classification output, whereas a lower count increased the output but had less impact. Other key features included the number of members and the monthly spending amount. The model responded positively to a higher number of users and spending amount, contributing to an increase in the classification values. In contrast, sex generally had a low impact, as indicated by the narrow SHAP value distributions, suggesting low feature importance.

In the RF model, the most important features included the FLR industries, the number of users, and the spending amount (Figure 4). Notably, certain industry types strongly increased classification values, presenting a broad distribution. This distribution indicates a significant influence due to the unique characteristics of specific industry groups. Additionally, the higher numbers of users tended to decrease classifications. However, sex and age groups had relatively low importance, with SHAP values concentrated near zero, implying limited influence on model classifications.

On the other hand, FLR industries, the spending amounts, and the number of transactions were the most important features in the XGBoost model (Figure 5). The FLR industries also had SHAP value distribution, indicating a significant influence due to the unique characteristics of specific industry groups. Unlike in the other models, the spending amount had relatively high importance in XGBoost, suggesting this model was more sensitive to credit card expenses than to the size of the user base.

4. Discussion

4.1. Comparison and Selection of Best-Performing GRDP Classification Models

This section evaluates and compares the performance of the five classification models using the OA, F1 score, and AUC metrics to identify the most suitable model for classifying GRDP levels. We emphasize the interpretation of performance differences in terms of the threshold sensitivity, class balance, and algorithmic characteristics.

In the present study, the average difference between the OA and F1 scores across the models was approximately zero, indicating minimal discrepancy. Although most models, except for LR, showed slightly lower F1 scores by approximately −0.1, this range is considered acceptable based on findings from similar studies, which have reported differences of up to −0.15 between OA and F1 scores in machine learning models [58,59]. In addition, the RF and XGBoost models were evaluated as having satisfactory AUC performances in [58], suggesting that their strong performance in the current study is unlikely to result from overfitting.

Meanwhile, the AUC exceeded OA by up to 0.09. This may indicate that, despite the models’ reasonable performance, the fixed threshold used in classification may have led to ambiguity in defining the decision boundary for the positive class. Supporting this, previous studies have reported that AUC can be up to 0.06 higher than OA, particularly in cases involving imbalanced data [58,60,61,62,63]. These studies noted that data imbalance may distort model classifications, and overfitting could also be a concern. In contrast, our study used balanced data with an equal number of samples (10,000) from each GRDP class, eliminating potential bias due to class imbalance. Therefore, the observed gap between AUC and OA may be influenced by both the characteristics of the data and the inherent characteristics of the models, even after mitigating class imbalance. To enhance the generalizability of the results, we suggest applying the proposed framework to datasets from other credit card companies to determine whether the observed classification patterns remain consistent across different data sources.

The LR model exhibited relatively poor performance despite parameter optimization using the Optuna framework. Although its AUC slightly exceeded 0.6, which is commonly considered the minimum validation score for good models, its OA remained slightly over 60%, indicating limitations in handling large-scale data (20,000 records) for classification tasks. This finding was also reflected in the ROC-AUC curve, in which LR showed a conservatively biased classification of the positive class, resulting in a low TP rate in certain segments. This aligned with the confusion matrix, which suggested that the high-threshold setting of the model prevented it from responding sensitively to positive cases.

In terms of model performance based on the AUC scores, the DNN, XGBoost, and RF models all exceeded the standard of an excellent model (=0.8), signifying a strong classification capability. The ROC-AUC of these models showed smooth transitions without sharp inflection points, suggesting a balanced tradeoff between TP and false-positive rates. Among all models, DNN achieved the highest overall performance and effectively captured complex consumer behavior patterns, making it a valuable tool for GRDP classification. However, the XGBoost and RF models also demonstrated comparable performance despite being computationally lighter than DNN. This could be attributed to the relatively small number of variables, which included only six demographic and consumption-related features despite the complex data structure observed in the descriptive analysis [64,65].

Furthermore, this study investigated whether deep and machine learning techniques could effectively distinguish consumption patterns of FLR participants between top and mid-level GRDP regions. Given this research objective, the comparative analysis based solely on the top 10% and 45%–55% quantile groups may reasonably be considered adequate, despite the exclusion of the bottom quantile due to data limitations. Moreover, potential data biases should be acknowledged, particularly those caused by differences in credit card usage across demographic groups such as age or income level. These biases may influence the observed spending patterns and should be addressed in future data collection and model refinement.

4.2. Importance Features for Classifying GRDP Level

This section analyzes the input variables that most significantly influenced GRDP classification across different models, with a particular focus on model interpretability through SHAP values.

Interpreting the variables classification models rely on to classify GRDP levels is essential. Despite the high classification performance of deep learning models, their complex and nonlinear internal structures pose challenges in understanding the contributions of individual features. The SHAP method is typically employed to address this [48,65,66,67]. However, tree-based machine learning models such as RF and XGBoost provide more interpretable variable importance measures due to their branching structure, and SHAP reportedly works effectively with these models [58,68,69,70]. Given these differences in model interpretability, a unified and quantitative approach for assessing feature importance across models is necessary.

Therefore, the SHAP technique was used in the current study. Considering SHAP is model-agnostic, it enables the quantitative assessment of feature contributions in both machine learning models (e.g., XGBoost and RF) and complex deep learning models [71,72]. Through this approach, we were able to identify the relative importance of key features in DNN models, which were previously considered “black boxes,” and visually compare how each model utilized input variables.

SHAP analysis revealed that the “FLR industries” variable consistently showed the highest importance across all models, suggesting that FLR industrial structure is a key determinant of GRDP levels. In contrast, the importance of “transaction amount” varied across models, being highly influential in the XGBoost model but contributing relatively little in the deep learning model. This suggests that deep learning models may dilute or redistribute the influence of specific variables through nonlinear interactions and multilayered structures [66,67,73].

Although expenditure is often assumed to be the most decisive predictor, our SHAP analysis shows that the FLR industry type variable plays a considerably larger role in the model’s decisions. In the total performances of the best models, higher-value observations associated with food-and-beverage and leisure-service sectors consistently receive positive SHAP values, indicating that heavier spending in these categories increases the likelihood that a region is classified as high-GRDP. In contrast, observations linked to tourism and accommodation tended to cluster on the negative side of the SHAP axis, suggesting that short-stay or irregular tourism demand does not necessarily translate into sustained local economic circulation. Cultural shopping activity also shows a weak negative contribution, implying limited economic leverage under the present data conditions.

These findings offer actionable implications for regional policy development. In particular, food-and-beverage and leisure-service industries, which exhibited strong positive contributions in the SHAP analysis, may serve as strategic leverage points for enhancing regional economic vitality. Policy interventions could focus on developing region-specific culinary products and experiential programs that promote repeat visitation and extended stay. Furthermore, integrating consumption-oriented elements, such as localized food or seasonal leisure events, into existing festivals and tourism initiatives may increase per-visitor spending and foster sustained economic circulation within local communities. These strategies align with the industry categories identified by the model as having the most substantial influence, suggesting a data-informed foundation for targeted and scalable regional development planning. By incorporating local historical and cultural contexts, such strategies can further be refined to support FLR programs tailored to the economic profiles of individual regions.

However, this study inferred economic levels indirectly using GRDP and focused on a limited set of consumption categories. Therefore, further research is needed to apply this methodology to diverse regions, particularly those targeted for regional revitalization, and to refine the selection and inclusion of input variables to derive more generalizable policy implications.

5. Conclusions

Based on the results of this study, actual card expenditures related to cultural activities varied in terms of total usage frequency and spending amount according to the GRDP quantiles. Overall, the top GRDP exhibited a greater consumption volume than mid GRDP. However, when focusing on the five FLR industry categories, mid GRDP demonstrated higher total and average expenditures per transaction relative to usage frequency, indicating that descriptive statistics alone are insufficient to fully interpret the underlying consumption characteristics.

To classify GRDP quantiles of residents based on such complex FLR consumption behaviors, deep and machine learning algorithms (DNN, RF, and XGBoost) were applied. These models successfully revealed the multivariate spending patterns of consumers with high accuracy and explanatory standards (AUC $\ge$ 0.8). Moreover, AI-based models provide insights into how actual spending patterns reflect the underlying regional economic conditions. These results suggest that the proposed analytical framework can serve as a practical basis for designing region-specific FLR policies.

Certain models presented limitations in classification sensitivity, which was reflected in the lower OA compared with the AUC. Moreover, due to strict customer policy, we were able to acquire a few variables, comprising only demographic and expenditure information. These limitations may be addressed by including additional variables, such as educational attainment, consumer movement routes, or regional transaction locations, in future research.

In this study, we presented an empirically analytical framework for classifying regional economic levels (i.e., GRDP) based on FLR consumption data. This framework may offer methodological insights into how FLR consumption reflects income-related economic patterns across regions. Using deep and machine learning methods, we were able to classify consumer transaction data patterns that are often difficult to identify using traditional statistical techniques. Moreover, it was confirmed that forest-related industries may demonstrate economic significance in terms of consumer income levels. Notably, this study not only evaluated model performance but also explored the relative importance of input variables, which may provide policy-relevant implications for region-specific FLR planning and sustainable economic development. Future studies could incorporate more diverse datasets (e.g., qualitative surveys or location-based services) to enrich the interpretation of FLR consumption patterns and validate the applicability of models in other regional or international contexts.