feat: started working on force-directed graph for visualizing category-transitions

[ErikBjare]

Generates a graph where:

• node radius is based on the time spent in the category.
• edge force & thickness is based on the number of transitions (not the best metric)

Uses category colors.

TODO

• What should count as a transition/context-switch?
  • Currently, even the briefest of transitions are counted as just as significant (could reduce this by merging/flooding category-events and removing short category-events)
  • Also shouldn't count transitions wider than x seconds apart (or fill such gaps with events that have an "afk" category, since that might also be interesting to see)
  • @casaout might have some ideas
• Is there a way to get directed-graph behavior, with animated edges to see direction things are flowing?
  • Seems super overkill and hard, but maybe awesome?

Screenshot (WIP)

[codecov]

Codecov Report

Merging #365 (6f9f20f) into master (f9481d0) will decrease coverage by 0.28%.
The diff coverage is 7.14%.

@@            Coverage Diff             @@
##           master     #365      +/-   ##
==========================================
- Coverage   26.00%   25.71%   -0.29%     
==========================================
  Files          26       26              
  Lines        1419     1431      +12     
  Branches      219      219              
==========================================
- Hits          369      368       -1     
- Misses        998     1011      +13     
  Partials       52       52              

Impacted Files	Coverage Δ
src/route.js	0.00% <0.00%> (ø)
src/stores/activity.ts	33.20% <0.00%> (-0.39%)	⬇️
src/stores/categories.ts	63.15% <9.09%> (-7.09%)	⬇️
src/util/classes.ts	82.82% <10.00%> (-2.59%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[casaout]

Thanks, Erik, for pinging me!

It might be helpful to get a little bit more context on the feature aims and what you're trying to achieve/visualize, to be able to answer your question better.

Am I assuming correctly that you create categories based on the app/website usage that ActivityWatch detects?

Generally, I think it might be really challenging to come up with good, universal rules/heuristics that define which category switches (or combinations) account for a context switch. It's also highly dependent on the definition of 'context switch'; I'd probably define it to something that is similar to a task (i.e., task switch)... In some preliminary work, we've tried to train models to identify task switches and types based on users' self-reports; but these are highly individual.

What we're doing in PersonalAnalytics is to treat 'afk' as it's own 'IDLE' category. This is set if there is no user input for 2 (or 3) minutes, I think. The thing is, however, that the user might be taking some notes (but is still within the same context/task) or doing a biobreak (which would probably be a context/task switch).

[ErikBjare]

It might be helpful to get a little bit more context on the feature aims and what you're trying to achieve/visualize, to be able to answer your question better.

I'm not quite sure yet, this is quite experimental. I wanted to try and visualize how I context switch between categories in some neat way.

Asked you because I figured you had a good view of the prior art and could point me in the right direction (since I noticed I was quite lost). Didn't want to reinvent the wheel.

Am I assuming correctly that you create categories based on the app/website usage that ActivityWatch detects?

Yes, the categorization happens using regular expressions, and form a hierarchy.

This is set if there is no user input for 2 (or 3) minutes, I think.

We use 3min :)

The thing is, however, that the user might be taking some notes (but is still within the same context/task) or doing a biobreak (which would probably be a context/task switch).

Yeah, I've figured that some context switches are larger/harder than others. Perhaps one could let the user set a threshold for how "large" switches have to be to count, somehow (maybe depending on gap size, low-likelihood transition, ...).

Will read that paper, thanks!

[casaout]

A few things that came to mind:

• The interesting thing is that I haven't been able to find a proper solution in related work and myself yet. So it's definitely awesome to have another great mind thinking about it ;)
• You could also consider the time spent in an activity (category) before switching to the next one. I could imagine that you, e.g., switch less when working in the IDE.
• Maybe the transition itself is the most interesting. Are there any patterns? (e.g. after social media, a person always switches to email briefly; or, more switches towards the evening when tired) Could some patterns (e.g. constant switching to email/IM when coding) indicate that you are not very focused/productive/...?
• Regarding transitions between activities, one thing, we also looked at briefly was to identify patterns of switches that frequently come together (code -> stackoverflow -> code, email -> IM -> task list). We also did some preliminary animated visualizations, but I cannot find them anymore :( Somewhat related, maybe Chapter 5.2 in our paper on developers' work lives could be interesting too.
• Generally, I believe such a visualization will benefit from user input, be it manual setting of thresholds/creating simple queries OR self-reporting (productivity/tiredness/focus/progress).