Nipaporn Chutiman

Predicting customer repurchase behavior is critical in e-commerce analytics, yet the contributions of feature engineering and model selection remain insufficiently understood. This study investigates how feature engineering and algorithm choice influence repurchase prediction performance using a large-scale Tmall dataset (228,867 user–merchant records; 6.01% positive rate). Six feature-selection methods, including mutual information (MI), Light Gradient Boosting Machine (LightGBM) gain importance, stepwise logistic regression, least absolute shrinkage and selection operator (Lasso), ridge regression, and Elastic Net, were applied to 37 candidate features. All methods identified merchant historical repurchase rate as the strongest predictor. Seven feature schemes-six selected subsets plus the full set as benchmark-were evaluated via single-split testing and fifty-replication stability analysis. The 28-feature Elastic Net subset matched or exceeded the full set and ranked first in 15 of 50 replications. Seven classifiers were compared under class weighting, which outperformed unweighted training and the Synthetic Minority Oversampling Technique (SMOTE). LightGBM achieved the highest area under the receiver operating characteristic curve (AUC = 0.7222). SHapley Additive exPlanations (SHAP) analysis confirmed the dominant effect of merchant repurchase rate. The findings highlight that feature engineering, particularly merchant-level historical features, contributes more than algorithm selection, informing e-commerce practitioners, data scientists, and marketing researchers.

Heavy rainfall is a major driver of floods, flash floods, and landslides, significantly impacting lives, property, and infrastructure. These risks are intensifying under climate change, which increases both the frequency and intensity of extreme rainfall events. This study analyzed the characteristics and assessed the risk of maximum daily cumulative rainfall in Thailand using the Generalized Extreme Value for the r-largest order statistics (GEVr) model. Daily rainfall data from 1996 to 2025 were obtained from NOAA Physical Sciences Laboratory, covering 690 grid points at a spatial resolution of 0.25°. The optimal order r was determined using the strong-stop criterion. Return levels were estimated for return periods of 2, 5, 10, 25, and 50 years. The results indicated that the GEVr model with r =1 adequately described rainfall extremes in 52.90% of the study areas. However, higher orders (r >1) were required in several regions—particularly the Northeast, North, and Southern—to capture multiple extreme rainfall events occurring within the same period. The Gumbel distribution was identified as the most suitable model in 52.75% of the areas. These findings demonstrate that the GEVr model, combined with spatial analysis, provides an effective framework for flood risk assessment, water resource management, and long-term policy planning.

Particulate matter smaller than 2.5 micrometers (PM₂.₅) is a significant environmental and public health issue in Thailand. The aim of this study was to develop a statistical model to evaluate and predict PM₂.₅ levels in Thailand. The Generalized Extreme Value distribution for r-largest order statistics (GEVr) was used in conjunction with PM₂.₅ data acquired from NASA EARTHDATA to develop the model. The data cover 176 grid points across Thailand during the period 1996-2025. The results showed that 47.73% of the total grid points had an optimal order of the daily maximum PM₂.₅ concentration (r = 1), while the remaining grid points had optimal orders at r = 2, 3, 4, and 5. This reflects that considering other largest daily maximum PM₂.₅ order statistics resulted in a more appropriate model than using only the first order daily maximum PM₂.₅ concentration. The Fréchet distribution was identified as the best-fitting 46.02% of cases, particularly in the Northern, Northeastern, and Central regions. Return level estimation for the 2 to 50-year period shows a long-term risk of PM₂.₅ concentrations reaching levels hazardous to human health. This research indicates that the GEVr model can be an important tool for air-quality risk assessment policy formulation in Thailand.

This research proposes two new estimators which use auxiliary information to derive estimates of population means in cases where populations are considered rare and clustered in line with the General Inverse Adaptive Cluster Sampling (GIACS) framework. A ratio-type estimator is the first to be proposed, while a correlation-type estimator is the second. In this study, theoretical analyses were conducted to derive explicit first-order approximations of the bias and mean squared error for each estimator. The results show that, although the estimators are biased in finite samples, the bias decreases as the sample size increases and converges to zero, implying that the estimators are asymptotically unbiased. Simulation data with symmetric distributions were generated to evaluate their performance. The findings showed that the GIACS estimators using auxiliary information are more efficient than those without. When comparing the performance of estimators that use auxiliary information, it was found that the two proposed estimators, the ratio-type and correlation-type estimators, offered superior efficiency when compared to regression-type estimators. These results confirm that applying auxiliary information can significantly improve the accuracy of estimations within the GIACS framework, making the proposed estimators highly suitable for practical applications involving rare and clustered population.

Adaptive cluster sampling (ACS) is a sampling technique commonly used for rare populations that exhibit spatial clustering. However, the initially selected sample units may not always satisfy the specified inclusion condition. To address these limitations, general inverse sampling has been incorporated into ACS, in which the initial units are sequentially selected, and a termination criterion is applied to control the number of rare elements drawn from the population. The objective of this study is to develop an estimator of the population mean that incorporates auxiliary information within the framework of general inverse adaptive cluster sampling (GI-ACS). The proposed estimator is constructed based on a regression-type estimator and analytically examined. A simulation study was conducted to validate the theoretical findings. Three scenarios were considered, representing low, moderate, and high correlations between the variable of interest and the auxiliary variable. The simulation results indicate that the proposed estimator achieves lower variance than the GI-ACS estimator that does not utilize auxiliary information across all examined correlation scenarios. Therefore, the proposed estimator is more efficient and preferable when auxiliary variables are available.

This paper presents new concepts of continuous multifunctions defined from a generalized topological space into a bitopological space, called upper $\mu(\sigma_1,\sigma_2)$-continuous multifunctions and lower $\mu(\sigma_1,\sigma_2)$-continuous multifunctions. Furthermore, several characterizations and some properties concerning upper $\mu(\sigma_1,\sigma_2)$-continuous multifunctions and lower $\mu(\sigma_1,\sigma_2)$-continuous multifunctions are discussed.

Adaptive cluster sampling (ACS) is a sampling method commonly employed when the population is rare and exhibits clustering. However, the initial sample selection may include units that do not satisfy the specified condition. To address this, general inverse sampling is incorporated into ACS, where the initial units are selected sequentially and termination criteria are applied to regulate the number of rare elements drawn from the population. The objective of this study is to develop an estimator of the population mean by utilizing auxiliary information within the framework of general inverse adaptive cluster sampling. The proposed estimator, constructed on the basis of a regression-type estimator, is analytically examined. A simulation study was conducted to validate the theoretical results. In this study, the region of interest was divided into 400 square units (20 rows by 20 columns). The results demonstrate that the proposed estimator, which incorporates auxiliary variables, consistently yields a lower variance than the conventional mean estimator without auxiliary information. This superiority holds across all scenarios considered, specifically when the predetermined number of rare units r ranges from two to ten. Therefore, the proposed estimator is shown to be more efficient than the estimator that does not employ auxiliary information.

The objectives of this research were to develop a model to forecast and estimate the return levels for daily maximum PM2.5 concentrations in Thailand, applying Extreme Value Theory (EVT) with the Generalized Extreme Value (GEV) distribution under eight models for stationary and non-stationary process. This research utilized reanalysis data from the NASA EARTHDATA satellite, represented as grid points with a spatial resolution of 50 × 62.5 km, enabling the analysis of daily maximum PM2.5 concentrations across 176 grid points from January 1, 2009 to October 31, 2024. The analysis revealed that Model 2 (μ(t)=β₀+β₁t where σ and ξ are constants) is the most suitable model for five grid points, namely Sa Kaeo Province, Uthai Thani Province, Nakhon Ratchasima Province, Bueng Kan Province and Mae Hong Son Province, whereas Model 1 (μ, σ and ξ are constants) is suitable for the remaining 171 grid points. Estimating the return levels for return periods of 5, 10, 25, and 50 years showed that Northern Thailand had the most extreme daily PM2.5 concentrations, for all return periods especially Mae Hong Son Province. The results of this analysis can serve as valuable information to support decision-making for response planning in high-risk areas, aiding in efficient resource allocation and preventive measures.

This research aimed to develop a scientific and accurate credit risk assessment model for small and medium-sized enterprises (SMEs) to support banks in credit decision-making. An improved decision tree model is proposed by integrating regularization to control complexity and employing an ensemble learning approach to enhance prediction accuracy. Multiple weak classifiers are iteratively refined using gradient descent optimization to form a robust, strong classifier. The model is trained through supervised learning, with the default probability of SMEs as the objective function, enabling a quantitative assessment of credit risk. The findings show that the proposed gradient descent decision tree algorithm achieves an AUC of 0.99 under 70% and 80% training set ratios, outperforming Adaptive Boosting (AUC = 0.97), Random Forest (AUC = 0.91), and traditional decision trees (AUC = 0.82). To further optimize bank loan strategies, this paper constructs a nonlinear multi-objective programming model that maximizes expected loan returns while considering risk constraints. The proposed approach not only improves credit risk prediction but also assists banks in formulating optimal lending strategies. This study advances credit risk modelling by integrating regularization and ensemble learning, offering a novel and practical solution for SME credit assessment.

This paper presents new classes of open and closed functions defined between bitopological spaces, called weakly $\theta(\tau_1,\tau_2)b$-open functions and weakly $\theta(\tau_1,\tau_2)b$-closed functions. Moreover, several characterizations and some properties concerning weakly $\theta(\tau_1,\tau_2)b$-open functions and weakly $\theta(\tau_1,\tau_2)b$-closed functions are established.