Learning the representation of surrogate safety measures to identify traffic conflict

•We demonstrate the applicability of transformer encoder models with unsupervised pre-training to real-world interaction data for traffic conflict identification.•The method unifies both proximity-based and evasive action-based surrogate safety measures and utilizes the entire time series of the interaction data.•The method eliminates the use of thresholds and interprets the similarities of clusters of traffic conflicts and non-conflicts.•We conduct eight case studies using real-world data that validate the usefulness of the proposed method and found the universal properties of traffic conflict from the identified results.