Dire need for faster edits

[diversario]

I travel a fair amount, and every time I end up spending 1h+ on dealing with Arc timeline.

Here’s a few screenshots showing tiny(!) portions of the timelines that get generated daily:

There are dozens and dozens of these, they just keep going, and in this case, all of these are me staying in a hotel room - so I have to go through all of that, take 5 taps for every Car or Walking or Airplane segment to mark them as stationary (or bogus, depends on how it looks), and I end up with a single place eventually. This is an insane time sink. There has to be a way to select all these garbage segments and edit them. Whether or not this is the “right” solution doesn’t matter here, because what’s not right is not respecting user’s time. If the classifier eventually becomes so good that there’s no longer any use for bulk edit - great, remove it, but that’s very much not the case currently.

So. Please make this easier somehow

[matt]

Yeah this is an ongoing problem that’s worsened since iOS started becoming a massive liar about location data accuracy. And it’s especially troublesome for frequent travellers (myself included!), due to Trust Factor requiring at least one day’s cleanup work before becoming useful. It’s not much use if you only stay in one place for a few days at a time.

@diversario

Whether or not this is the “right” solution doesn’t matter here, because what’s not right is not respecting user’s time.

Bit of a tangent: The problem with building new UI to fix the symptoms when there’s a way to fix the underlying problem is that it makes it unnecessary to fix the underlying problem. The user is left with new UI and UI steps to perform periodically (albeit more convenient and efficient than the previous) when they really shouldn’t have to do any work at all. The new should-be-unnecessary UI becomes normalised, and the underlying problem gets forgotten.

In terms of improved UI though, I think there’s a case to be made that efficiency improvements to the existing confirm/correct flows could be made that would have general applicability, not just “gloss over this data problem” applicability. So I’m still open to those options too.

@diversario

If the classifier eventually becomes so good that there’s no longer any use for bulk edit - great, remove it, but that’s very much not the case currently.

In this case it’s the underlying recorded location data that’s the problem. iOS tells massive lies about data accuracy (eg reporting 10 metre accuracy when the data is 100+ metres away), which then breaks the filtering and smoothing maths, causing these nonsense drifting items to clutter up the timelines. So the “correct” fix is at Apple’s end, and the next best fix is some way of LocoKit being able to identify the lies immediately, without having to wait for user feedback and calculation of Trust Factors.

Anyway, in terms of possible UI improvements, I think one that could make life a bit easier is to have the long press menu on timeline items offer up a shortlist of type correction options. Like for those airplane items in your screenshot, if you long pressed on them the options menu that appeared could show the top 4-5 type matches, for quick correction. You could long press, select “stationary”, and job done.

I’ve actually already got that task in my “Today” list. Though unfortunately “today” is a misnomer, and instead means “stuff I’d like to start on today, but almost certainly won’t get to”. But anyway, it’s one that I’m hopeful will get into an update in the not too distant future.

For fixing the underlying problem, I still spend time every day cogitating on that. It’s a problem that affects me too, due to travelling to different cities each month then spending the first few days (or worse) cleaning up a new mess, until Trust Factor takes over.

I don’t have any credible ideas yet for that though - it’s a hard problem. The only way that we as humans know the data is wrong is because we know where we really physically are or were. Machine learning / AI can’t do that.

[matt]

Somewhat related: I’ve added a new option to the more/ellipsis menu on the Edit view for path items, to “Mark as bogus” for the entire item.

That gives you a quick way to mark the entire thing as bogus, instead of having to go into the segments and mark each segment bogus. Useful in cases where you’re confident the entire thing is wildly wrong location data.

Not really helpful in the cases in your screenshots - those look like they’re all close enough that they should be marked stationary. But it does speed up the flow and take out a few extra taps when dealing with places where the phone doesn’t just lie by 100 metres or so, but instead drifts over hundreds of metres periodically throughout the night.

[diversario]

Options in long press would be a big help! Please include “mark as bogus” there as well because there’s a lot of those (even when I’m staying at my house I get wild spikes here and there a few times a day )

[diversario]

Oh, in a similar vein - how about a way to clean up segments in bulk? Is there a downside to being able to say “clean up all the locations on this day”? I try to do that for all location (when I remember) and I don’t know what could go wrong if clean up was done automatically even?

[matt]

@diversario

I don’t know what could go wrong if clean up was done automatically even?

Some of the largest support headaches come from people accidentally confirming the wrong place assignment to a visit. It becomes encoded in the Place model as “correct” data, that it will then automatically replicate. Users who don’t frequently check their timelines then end up with potentially tens of visits, or more, automatically assigned to the wrong place, creating a massive manual cleanup job.

It’s less of a problem with accidental/incorrect activity type confirmations - the models have much more activity type data to work with, so are less vulnerable to single incorrect confirmations - but it still ideally shouldn’t happen.

For me personally, I’d like to see improvement on the existing “x unconfirmed items” flow. For example if there’s 10+ items that need confirmation/correction, that flow is often hard to do well, because there’s not enough context information for each item presented. I’ll get caught in situations like, “well yeah, I did go to 7-Eleven yesterday, but I also went to the drug store next door. am I sure this is the correct visit or am I about to create a mess?”

If that flow were “better” somehow, it’d be easier to initiate the flow then quickly tap tap tap through it for a full day, feeling confident that no mistakes were made. That would get the best of both words - ensuring there’s still explicit assessment and confirmation of each item, but without having to expand the map and inspect more closely, tap back to timeline to double check, etc.

I think in the simple cases, where we already feel confident that everything needing confirmation is going to be already correct, it’s likely going to be a situation where the models will also feel the same after only one or two more confirmations, and won’t keep asking in subsequent days.

The models should ideally only be asking for confirmation for either:

1. First time visits to new places, trips along new routes or with new activity types
2. Situations where there’s always going to be low certainty, for example visits to places that are very close to other places that you frequently visit at similar times, or trips along routes that you frequently travel by multiple modes (eg a route that you sometimes cycle, but also sometimes drive or take the bus, etc)

Anyway, I’m waffling again. Simple answer is: I think at this stage probably the biggest win would be in improving the visible context information on the current Confirm view from the “x unconfirmed items” flow. If you really do only need to glance for a second or two at each item before saying “yep, that’s right”, a day with 10 items to confirm could be cleaned up in half a minute.

[matt]

I think the real core of the problem though is the models just don’t quite reach the level of intelligence we want. There’ll be items needing confirmation that we look at and think “that’s obvious - you shouldn’t be bothering me with this, and shouldn’t have made that mistake”.

I’d really like to start making use of LLMs (GPT4 etc) to improve these tasks. Make Arc much smarter, so that it does more of the work, and does more of it correctly. But so far LLMs are still too expensive. I’m really excited to start experimenting with that stuff this year in Arc though. LLMs are getting cheaper, and smarter, so it feels like it’s going to reach a tipping point soon, where Arc’s brains can level up dramatically.