Several studies have explored transportation mode imputation using GPS data, employing different algorithms. There also exist review papers evaluating the accuracy based on dataset, feature selection, and algorithm type. However, comprehensive comparisons of various categories of algorithms applied to a single dataset -while controlling for dataset differences- are rare. Another critical issue in mode imputation research is the limited evaluation of generalizability and transferability. Common practices, such as splitting data into training and testing sets, often fail to address these aspects effectively due to overlap in trips from the same individuals or geographical regions, undermining claims of generalizability. This study addresses these gaps using a large-scale GPS trip dataset with mode labels, obtained via a smartphone application. The study evaluates accuracy and generalizability of rule-based, statistical, and machine learning methods while accounting for data intensiveness, computational complexity, and implementation feasibility as factors of efficiency. Also, trips are categorized by regions to analyze the performance of models trained in one region and tested in another, offering novel insights into the robustness and transferability of these methods. This comprehensive analysis identifies the trade-offs needed to achieve accurate and efficient mode imputation, enhancing the practical applicability of these techniques in real-time mode detection.
