This is so very city-specific that it would only make sense for a small segment of users. On top of that colour coding train lines collides massively with colour coding transport types and messes up visual semantics. My vote hence strongly against it.

Sascha, the majority of Arc’s users live in large cities with extensive train networks. This functionality will be relevant to the overwhelming majority of Arc’s users.

This would be cool!

Are you thinking about getting this data from published GTFS feeds (e.g., https://transitfeeds.com/)? It could even support specific bus lines.

How would this work when a single trip contains a transfer to a different line?

Yinon, I wasn’t aware of that data source! But now that I am, I’ll definitely be looking for ways to build that into the ML models 😄

My plan was to simply use the existing ML classifier architecture, because it is already capable of distinguishing different above ground lines based on coordinates, direction, speed, time of day, and more subtle factors such as differences in accelerometer data (eg different carriages on different train lines).

But bootstrapping the models from existing route information would be a big step up. And if OMD can also provide real time data, then I’d love to build that into the app too, to give the classifiers an even bigger step up 😄

You might also be interested in https://transit.land/

I was not aware of these public transportation feeds. Still remains a challenge: for example here in Berlin there are many overlapping transport lines, therefore location data alone is definitely insufficient for recognising the actual transport taken. Even if your machine learning model is able to solve this, the second aspect of my argument against colour coding train lines prevails. You can either colour code transport lines or means of transport but not both at once.

Sascha, it will be up to you whether your install of Arc learns to identify the train lines or not. If you correct them to simply “train” or “metro”, that is what Arc will learn that you prefer, and it won’t try to classify them by their individual names.

Arc’s activity type ML is unique to each individual, in that it learns specifically how you like things to be classified. The shared ML models are only used as an initial stepping stone, until Arc has learnt enough about you personally. Once Arc knows your personal habits and preferences well enough it almost completely ignores the shared ML models.