Imputing trip purposes for long-distance travel

Planning and policy analysis at the national, state and inter-regional corridor levels depends on reliable information and forecasts about long-distance travel. Emerging passive data collection technologies such as GPS, smartphones, and social media provide the opportunity for researchers and practitioners to potentially supplement or replace traditional long-distance travel surveys. However, certain important trip information, such as trip purpose, travel mode, and travelers' socio-demographic characteristics, is missing from passively collected travel data. One promising solution to this data issue is to impute the missing information based on supplementary data (e.g., land use) and advanced statistical or data mining algorithms. This paper develops machine learning methods, including decision tree and meta-learning, to estimate trip purposes for long-distance passenger travel. A passively collected long-distance trip dataset is simulated from the 1995 American Travel Survey for the development and validation of the machine learning methods. The predictive accuracy of the proposed methods is evaluated for several scenarios varying with trip purposes and the extent of data availability as inputs. This research design will provide not only a practically useful approach for long-distance trip purpose imputation, but also generate valuable insights for future long-distance travel surveys. Results show that the accuracy of the trip purpose imputation methods based on all available data decreases from 95 % with two purposes (business and non-business) to 77 % with four purposes (business, personal business, social visit, and leisure). Based on a two-purpose scheme, the predictive accuracy of the imputation algorithms decreases from 95 % when all input data is used (a full-information model), to 72 % with a minimum information model that only utilizes the passively collected data. If traveler's socio-demographic characteristics are available (possibly through other imputation models), the predictive accuracy only decreases from 95 to 91 %.