This morning I implemented PageRank to sort backlinks in my prototype note system. Mixed results!
＋: Easier to navigate to implicit "neighbor" notes connected via "hub" notes
＋: Surprises me notice how "central" some notes are
－: "Weird" backlinks often end up at the bottom
Conversation
Is the goal of NoteRank specifically to see which Notes rank the highest? What would be your fantasy result here?
1
2
It's fuzzy. Backlinks themselves are one piece of what I think of as "peripheral vision"—serendipitous representations of structure and associations. A high rank tells me that a note is implicitly "near" another note that's important along some axis (not always the axis I want).
3
1
11
You don't know me yet, but I've been experimenting with similar ideas after hearing an inspiring description of your notes system from my housemate . I've been taking it in more of a ML direction: I implemented cosine similarity between notes using BERT representations.
1
5
+: "Weird" backlinks might just be directions in space - I'd LOVE to find an "opposing viewpoint" direction, or train a model with such a direction
+: "Weird" backlinks might also be a certain *distance* in space? Perhaps there's a "Goldilocks zone" for creative connections.
1
3
Pretty sure and have experimented with this too - what have y'all found useful? I think the potential to identify interesting/unusual connections is the potential killer app.
cc - this is what I was describing to you in that waiting room!
1
3
Unfortunately, the content of the text does not also contain the context of the text. Without the context the similarities are often superficial. Makes it very hard to scale as-is. Requires a whole new 'context' system which BERT does not naturally lead to. Still thinking...
1
4
My instinct is unfashionable: rather than worrying so much about ranking and algorithms, worry a lot about info arch and presentation so you can approximate “see everything all the time”. The challenge is to do that usefully—everyone just starts drawing force-directed graphs.
1
3
I agree with your unfashionable instincts! Does that make them fashionable now? 🙃 It's always best to try solving a problem with design first. That said; there are entirely new landscapes of design opening up to us with modern NLP. What if tools-for-thought could also 'think'?
1
1
I buy that. "Unlinked references" in already seem somewhat valuable. It's easy to imagine increasing their value by e.g. using vector embeddings to "fuzz" the edges of that set. Not clear how soon that would come up on an oracular pri-queue relative to design work.
Naive implementations of AI will be less valuable (from a product standpoint) than refined design. A refined implementation of NLP functionality seems to require entirely new forms of UX/UI to solicit the reasonable user inputs or provide a non-frustrating interface to fuzziness.
1
5
Yeah, agreed.