Machine Learning Fails: Why Microsoft thinks you live in Poltár, Slovakia

Something caught my eye this morning when looking at my Office 365 Outlook calendar. I live in Cambridge, UK. Why would I care about the weather in Poltár, Slovakia?

Apparently, Office 365 thinks I live there:

And I’m not the only one. A quick Google reveals lots of people who have been automatically assigned by the Outlook Calendar to this sleepy Slovakian town (population 5,632):

Is there anyone who can resolve this strange problem? poltar Slovakia looks nice place to visit according to Trip Advisor but it’s not where I live. (Microsoft answers)

Office 365 Calendar is now trolling me with weather alerts for Poltár, Slovakia. No, I’ve never heard of it either. (Twitter)

Weather in calendar is not for my location — It shows for Poltar, Slovakia as opposed to Auckland, New Zealand… wrong hemisphere, wrong season, wrong country.(Outlook uservoice)

What’s so special about Poltár? I decided to check it out on Google Maps. Notice anything strange?

Roads called “13 January” and “9 May”! I cross-referenced with Bing and StreetView, and yes the roads are really called that. Surely this can’t be a coincidence.

I can’t say for sure, but I reckon Outlook tries to guess where you live from the contents of your calendar. It scans all your events for references to nearby landmarks, businesses, towns, and streets.

Your calendar contains, amazingly enough, “13 January” and “9 May” — which could only be references to the streets of your beloved hometown of Poltár.

It seems remarkable that Microsoft would fail to remove dates when preprocessing your calendar data. Perhaps they thought irrelevant clutter like dates wouldn’t impact the algorithm. But, as it turns out, the signal of someone referring to not one but two streets from this small town is strong enough to decide that I’m a Poltárian.

Or maybe there’s another explanation? Let me know.