I’ve been using Arc for very many years (in the paid version since available). In recent months, the classification feels like it’s getting worse and worse. It’s almost useless at this point.
Sometimes I try to teach it and go through a weeks visits and correct each movement type and location, but it does not seem to help.
I’m attaching a screenshot of a clearly implausible classification (although I would love to own that car) along with some debug info that looks suspicious to me—why are there so many updates pending?
I am trying to add the screenshots that I mentioned in the original post, which discourse’s spam protection disallowed me from posting. But it’s telling me now that I’m responding too quickly.
I never swipe-up-kill apps, and arc only occasionally (once every couple of weeks) complains that iOS terminated it. Phone is always on (flight mode at night), background app refresh off for most apps, but on for Arc.
In your first screenshot, the significant detail is the “Most common speed” line, which shows 38 km/h. That’s a very plausible speed for a car.
The difference between “most common” and “average” speed is important - the former means the speed at which most of the data was recorded. So for example if you’re travelling at walking speed for 10 minutes, then for a brief moment the phone messes up and records some data rapidly moving away, at 1000 km/h, the average speed for that trip will shoot up absurdly high, but the most common speed will still be walking speed.
The classifier is classifying the individual recorded samples within the trip, so if most of those samples are at walking (or in this case car) speed, the overall conclusion will be walking or car, even though the average speed might show something quite different, if the phone blipped out and produced some nonsense data briefly.
Good to hear! Sounds all correct settings then.
From your screenshots, it looks like it’s still updating the models, but has fallen quite far behind. It shows over 2000 models in the queue! Which usually means you’ve done some air travel at some stage recently.
Flying means passing through hundreds of model regions, which then queues up a model update for each of those. The good news is those models update extremely quickly (less than a second for each), because there’s very little data in them.
The model updates background task also always starts by doing the most important models first. Those being the models for where you are right now. That way you should be getting best possible classifier results for the current day’s data most of the time.
Though with over 2000 models in the queue I’m guessing you might’ve restored from backup recently? The models aren’t stored in the backups, so they need to be rebuilt after restore, which can also create quite a backlog.
In my testing of many restores from backup, I typically see the queue backlog clear out within about a week or two. Although it’s usually only the first 2-3 days after restore that I notice, because it will be first updating the most important models, so once that’s done I don’t notice any degradation in classifier results.
Do you have any other examples of classifier results that are particularly bad? That first screenshot looks to be more likely a problem of nonsense/bogus data, and the classification of car may still be the most correct one, given that the most common speed is sensible for car.
If you have done a restore from backup recently, then you’ll probably see worse classifier results for old data, in locations you haven’t been to recently. Those model updates won’t have been pushed to the top of the queue, due to not being near your current location.
