Investigating the impact of inferring trip purposes in a daily trip generation model

Smartphones are increasingly being used to record travel behaviour data semi-passively, but low engagement rates at the verification stage leads to large amounts of untagged trip diaries (i.e. trip purposes and modes). Researchers either discard the untagged observations in the modelling stage or assume the labels assigned by the inference algorithms are error-free, i.e. have a deterministic outcome rather than a probabilistic one. In this study, we check the impact of inferring trip purposes probabilistically vs deterministically in a daily trip generation model, and if it is beneficial to utilise inferred untagged datasets as opposed to working with tagged datasets only. We use travel diaries collected in Trondheim, Norway, where a third of the trip purposes are untagged. We observe a significant loss in the predictive performance of the daily trip generation model when the trip purposes are inferred deterministically rather than probabilistically. Therefore, it is recommended that researchers working with passive data sources consider the uncertainty in the inference process. We also find that the daily trip generation model developed using both tagged and inferred untagged datasets is more efficient but has a slightly lower predictive performance than the model that uses the tagged dataset only, indicating some potential benefits of utilising inferred untagged datasets. However, we conclude that data quality is far more important than the number of observations.