Reader beware, this is a nerd post!
One of the most common types of machine learning is used to detect outliers in a dataset. Today we introduced several “anomaly detection” algorithms into the beta that assess your activities, power thresholds and segment efforts.
Why do we do this? In part, to help us get better predictions, determine when we shouldn’t send you a note alerting you about something that isn’t useful and to help you identify when something’s off with your device data.
We’re currently checking for the following anomalies:
- Ride Activity: Identify a ride that’s inconsistent with your typical rides. For example, when you ride for <5 minutes testing something.
- Watts-kg: Identify either low or high power readings based on your thresholds. This is useful to identify when your power meter may need a battery replacement.
- Segment Efforts: Here we’re looking for outliers in how many watts/kg does it take to complete flat or uphill segments. Long term this will help us alert you to more subtle power meter issues like when the meter isn’t calibrated properly and is reading say 10-15% high or perhaps even when your weight in Strava isn’t quite right. We have other crazy ideas for these outliers we’ll explore once there is more data in the system.
We’ve also added a few new notifications this week:
- FTP Alert: We advise you when there is a material difference in your Strava FTP and our estimate. We do this check when you’ve set a new 20-minute power threshold.
- Power Meter Alert: We advise you when your power meter is reading very low or very high. We’ll improve this soon by comparing efforts across all your bikes with PMs on your starred segments to see if there are more subtle outliers (e.g. bad calibration). This alert is internal only for a few weeks to let us tune the parameters, etc. We’ll send them to you very soon.
You can see the results of these predictions on your activity pages on the site. A red anomaly prediction on an activity page means we didn’t detect one while green indicates we did. For the watts/kg ones we show the offending reading and for segment efforts we just show the count. Don’t be concerned if you see anomalies on your activities. We don’t message you if it’s not a material issue.
I’m excited to see how the Virtual Ride segments hold up as I’m assuming there are fewer factors impacting performance results on those segments. My hope is we can get some great learnings from those and perhaps use them to isolate smaller discrepancies with weight or power calibration. Thankfully they are the most popular segments in the system given athletes in Chicago and the Bay Area both use them a ton!