So Arc after the most recent update is labeling what it used to correctly interpret as car transportation as skiing or hiking for some reason. I’ve been correcting this as I go along so I assume this should take care of the issue?
Btw, for a one man show you’re doing an incredible job. I can’t recall what I’m paying for the Arc subscription but whatever it is it’s a steal for sure.
Hi @Alaheron! Thanks for your kind words
Just today and yesterday @Astra and I have been discussing this exact issue. I’ll probably ship an update in the next few days that improves it a bit. Though we’re still brainstorming better ideas.
Basically what’s happening is this Arc no longer relying on server-provided activity type models for classification when your own personalised models are insufficient - that small bit of server-side dependence is now fully gone.
But for users who don’t yet have enough of their own data to flesh out their own reliable activity type models, I’ve added in a “bundled” model as a fallback for the classifier to make use of in those cases. That bundled model is literally just my own personal “global” model, trained on my own personal travel data, which won’t be a universally great fit for everyone.
That bundled model should be receiving only a very small amount of “weight” in the calculations, when you’re travelling in areas where there’s already enough of your own data available. But it looks like that maths isn’t working great in some cases, and the bundled model is still getting given too much weight, throwing things off.
So the next update this week will probably be a quick fix to simply halve that calculated weight, until we think up a smarter idea. Though if we’re lucky, that might be enough to fix it! Fingers crossed
Yep, definitely! If all goes well, those corrections you make should also fix the problem within only a few days.
But because this new architecture is very new, I don’t yet have a good instinct for whether it really is “only a few days” or maybe longer. Either way, the upcoming update with that different weighting maths should hopefully dramatically shorten that needed adjustment time.
Just want to echo two things!
1: you are amazing for creating and supporting this fantastic app that I’ve now been using for years; thank you!!
2: I’m also experiencing this skiing/hiking glitch, and it’s coincided with a Christmas road trip. I’ve found that even in segments where I’m travelling by car, 120 km/h without stopping for several hours, it’s been guessing skiing with many many stationary segments interspersed. My previous experience has been that when I correct segments like these, most (or all) of the little stationary segments are smoothed out – or I can also use the new “clean up” function to do the same thing.
However, I’ve found that when I change these skiing segments to car (for example), it’s not smoothing out the stationary segments. For a long span of time in the car, this means dozens and dozens of stationary segments that need to be individually corrected. The overall category for the larger segment changes from skiing to transport (rather than car). See attached screenshot!
I have no doubt that this will be corrected by an upcoming update going forward! Just wondering if the retro fix that you mentioned in your response above will also apply to segments like these, before I settle in for a major correction spree
Thanks again! I love Arc!
Eek, yeah so sounds like it’s much more common than I’d hoped I’ll try to get an update out this weekend or early next week.
Aside: It’s really weird that it’s seemingly picking skiing and hiking for almost everyone experiencing the problem, just from a mathematical perspective. I can’t see how the model would come to conclusions like that. It doesn’t add up
Anyway, one way or another I’ll get it fixed. I’m just surprised it’s doing something that doesn’t seem to make mathematical sense. But luckily I think the solutions that will work won’t require actually understanding why it’s doing that weird thing. Phew.
Thanks @matt !
You might be amused that yesterday, the app was thoroughly convinced that I skied down the West Side Highway in Manhattan
Hah, that’s a weird one
Good news: I submitted an update a few hours ago, and it went into review and has gone live on the App Store already!
Usually it takes 24 hours or so, but this time it went into review really fast. Anyway, the update hopefully fixes the problem (or at least greatly reduces it). Let me know how much of a difference it makes. Thanks!
Worked like a charm! Thank you again!!