Exercise sharing app Strava has removed more than two million rides suspected to have been done on e-bikes, but logged as ‘normal’, from its leaderboards.
The move comes as part of a wider initiative to clamp down on cheats and “unrealistic” performances. Strava announced last year that it was using AI and machine learning tools to analyse activities and ensure integrity on its platform, which awards KOM and QOM titles to users with the fastest segment times.
“Over the past few weeks, we reprocessed the top 100 activities on each of all Ride Segment Leaderboards to address long-standing issues with anomalous activities showing up in results,” a Strava engineer named James wrote.
Within the work, the platform carried out what it calls “enhanced e-bike detection”, which involves a machine learning model “trained to catch activities recorded on an e-bike but uploaded as normal rides”.
Strava says it has since restored 293,000 users to their “rightful spot” in leaderboard top 10s.
In May last year, Strava announced that it had wiped 4.45 million activities from its leaderboards in its hunt against cheats. The company said at the time the deleted activities had been uploaded with the “wrong sport type” or “recorded in vehicles”.
Strava’s rules specifically state that users should not upload public ‘ride’ activities, which appear on leaderboards, if they include data recorded in a car or motorcycle, using an e-bike, or pacing a vehicle.
The company’s machine learning model, which it has been using for over a year, takes into account “57 factors” when analysing activities. In a previous Reddit post, Strava revealed more details about the system, saying it “looks at every activity holistically and uses dozens of different features like acceleration, variance of speed, uphill average speed, and others.”
Strava has stressed that its leaderboard analysis is continuous and “never really ‘done’” – “we know there’s still more to improve when it comes to removing anomalous activities,” engineer James wrote on Reddit.
Read the full article here
