Identification and characterization of unusual commuters using a 2-layer transformer model

Source abstract

This study advances activity-pattern recognition by moving beyond broad classifications of homogeneous activity participation to identify commuters whose behavior deviates markedly from prevailing mobility structures within predefined time intervals. An unusual commuter is defined as an individual whose commuting pattern, jointly characterized by commute distance, travel time, commute duration, and the sequencing of pre-work activities, differs substantially from the dominant pattern observed among individuals traveling to work within the same temporal period. To identify individuals that are different, the study adopts a two-layer Transformer framework. The first layer classifies commuting intervals into predominant, moderately predominant, and least predominant periods using average surprisal scores. The Surprisal is the negative log likelihood of the probability of commuting during predefined time intervals. The predominant time intervals are the interval with the lowest average surprisal score, representing the intervals when commuting is most expected. The second layer was integrated to understand why some individuals within the predominant time intervals exhibit high surprisal scores despite the observed low average surprisal scores prevalent within these intervals. The findings reveal a tension between environmentally desirable travel behavior and dominant commuting patterns, while also showing that unusual commuting should not be interpreted uniformly for policy purposes. Low-income unusual commuters may require improved access to employment and affordable travel options because of limited vehicle availability. In contrast, higher-income commuters who undertake longer trips may benefit from flexible or hybrid work schedules and incentives to use public transportation. Household characteristics also shape unusual commuting, as larger households, dependent children, and fewer vehicles increase household coordination constraints. The inconsistent effects of built-environment classifications indicate that spatial context influences unusual commuting selectively rather than uniformly. The cubic relationships further demonstrate that unusualness varies nonlinearly across travel time and distance, rather than increasing uniformly as trips become longer, contrasting with previous studies that assume a linear relationship.