Akademik Veri Yönetim Sistemi
TRAVEL BEHAVIOUR AND SOCIETY, cilt.43, 2026 (SSCI, Scopus)
Selecting the right variables is essential in travel behavior modeling for transportation planning. Traditional methods, like choosing from highly correlated predictors or relying on past studies, can reduce the effectiveness of models. Using robust methods to identify relevant variables helps minimize errors, enhances model understanding, and simplifies future predictions by focusing on key factors, making applications more reliable and efficient. In this study, the data from the household travel survey within the Eskisehir Transportation Master Plan (conducted in 2001 and 2015) were used for the theoretical modeling. The objective of the study is to develop models for non-home-based travel purposes (e.g., banking, shopping, socializing, visiting, entertainment, recreation, sports, etc.) by incorporating socio-economic demographic parameters and the land-use data to understand the relationships between socio-demographic variables and Other-Purpose Trips (OPT) behavior. Various theoretical methodologies, including classical Multiple Linear Regression (MLR) in travel models, Ridge Regression, advanced variable selection and machine learning techniques such as Least Absolute Shrinkage and Selection Operator (Lasso), Elastic Net (ENet), Adaptive Lasso (ALasso), and Adaptive Elastic Net (AEnet) are applied in the study. Ridge Regression and machine learning techniques are implemented to address multicollinearity problem that cannot be handled with the traditional MLR models. Among the 2001 production models, ENet is approximately 29% more successful than MLR in terms of Cross Validated Root Mean Square Error (CVRMSE). Similarly, ENet demonstrates a 17% higher success rate in predicting the target year (2015) based on Root Mean Squared Error (RMSE). In the 2015 production models, the most successful predictions according to CVRMSE are obtained from AEnet, with a prediction power approximately 45% higher than MLR. Among the 2015 attraction models, AEnet and ALasso, approximately 37% more successful than MLR according to CVRMSE, are found to be the most successful models.