Learning Vector Representations of Bicycle Itineraries

Over the past ten years, bicycles have become an increasingly attractive transport and leisure activity mode. As the number of cyclists tends to increase, there is a need to improve their experience. Multiple techniques have already been created for solving trip destination, travel time prediction, and other tasks. Most of the created models were detached from the geographical semantic characteristics of the trips, making these approaches bounded to city local patterns, they were trained. This work proposes the bicycle itineraries vector representation technique based on the state-of-the-art Natural Language Processing embedding method, using OpenStreetMap functional data and Uber H3 map grid representation approach. There were used the Hex2vec and Document embeddings techniques, to obtain the bike trajectory embeddings. In order to determine the quality and accuracy of the new method, trained vectors were applied to solve trip type classification and travel time prediction tasks. After numerous tests, the best models were presented and analyzed, which showed satisfactory results.