I'm playing around with an idea for #hackweek (Jan 7-11, 2013) to pull out the "important" things from your Twitter home timeline. Looking for feedback on what you think would work and/or not work.

A few months ago, I built a prototype that texted you whenever 3 or more people mentioned a @username, #hashtag, or URL within a 3 hour window. It worked pretty well but still had a pretty low signal/noise ratio. First step is to learn to calibrate a trend so that popular keywords only "trend" when they spike more than usual.

Looking at just the home timeline, you're limiting yourself, but that kind of focus might be helpful. Handful of ideas off the top:

* 2 or more of your friends replied to another friend's tweet
* 2 or more of your friends retweeted a tweet from someone you don't follow
* Sentiment -- your timeline is happy/outraged/sad
* Word frequency (not just hashtags) -- this word appeared an usually high amount of times within a time window
* Photo slideshow (including Instagram the links to which I am too lazy to click on)

There's also the obvious (and overdone) "here's a news digest of links from my timeline."

Any reason you're focusing on the home timeline versus, say, favorites or social graph or direct messages?

I'm focusing on "timelines" in general. Your home timeline, other people's home timelines, and list timelines.

I think of it as a way to "fuzzy follow" a curated set of people. And to increase the threshold for hearing from these sets so that you only hear about the "best" stuff in them.

It's the difference between following Anils tweets (traditional follow) and following the best of the tweets that Anil reads. I'm as interested in Anil's inputs (what he reads), in other words, as much as his outputs (what he tweets)... there's just more noise in the inputs (from my perspective, since our interests don't overlap 100%).

Does that make sense? Still thinking it through obviously.

I think it makes sense. Using mechanical methods of creating functions that do this is not impossible but has to be given some thought. On the other hand, isnt through algorithms that some platforms do this?

For me this is where Stellar.io solves the problem by looking at the confluence of favorites. I can see the tweets favorited by 3-4 of my friends and they're almost always written by someone I not only don't follow but often haven't ever heard of.

I know favorites are kind of a weird beast not used by most general Twitter users (they prefer the RT), but among my techie friends, everyone favorites and rarely RTs, so it's a good indicator of quality outside your own timeline and onto theirs.

Also, isn't this what the Discover tab on Twitter is supposed to do? Show me notable things from people I don't follow based on various input signals? Normally, it's really general "zomg Beiber got a haircut!" kinds of things for me, but once in a while it's really specific and teases out some small tidbit from the world of cycling that I also follow via Twitter.

Perhaps the goal should be to make the Discover tab much more based on personal network trends instead of global Twitter trends?

Matt, I concur with you. Thanks for letting me know about stellar.io, will look into it.

Yeah, stellar.io and the Discover tab are both trying to solve this problem in various ways. But still favorites won't solve the full problem since they are also used for lots of other things (thanking someone, letting them know you saw/appreciated something, bookmarking for later, etc).

I have this mental image of each of us as a neuron. The people we follow are the dendrites. We are the nucleus and we decide when thresholds of interestingness have been triggered. And our tweets out are the axon, which in turn hook into other people's dendrites.

The question I have is... can we create algorithms that function as nucleus (interestingness filter) augmentation? At least, as "interesting" filter for passing along inputs.

Sounds kinda crazy.

Just to catch you up on the work I did yesterday, here's where I am right now...

twitter.com

I truly believe that RT –for the sake of being a Twitter service– is a lot more powerful than Favorites. My basis for this is really simple. The psychological force behind the stickiness of social networks are Positive Feedback Loops. Taking that into consideration means the source engine for the generation of the data that will later become knowledge (the output on the Personal Trends list) will be a lot richer.

If any of this doesnt make sense let me know and I can further expand. Didn't want to take too much of a post here to get my point across. I've lectured on Social Media, specifically what psychological forces keeps us addicted to them.

I agree that RTs are a big signal. Here's a rough description of my current "personal trend" algorithm that isn't fully implemented yet.

Trend score =
(# people tweeting a particular hashtag, username [non @-reply or RT], or link in last 3 hours) *
(RT count for tweets in trend (max of X)) *
(tweets/hour over last 3 hours - tweets/hour over last 72 hours)

Each input will be weighted too, eventually.

This will show you trends that are recently spiking above their normal activity level amongst the people in a given timeline. And that timeline can be yours, or someone else's, or a list's.

Does each one of those criteria have the same weight in the resulting variable? Or does one weigh more than the other?

They'll be weighted differently. Will have to fiddle with it once I have some real data to play around with.

For what it's worth, I really like some of the recipes you showed in the screenshot that were about links to a site. That's kind of interesting and not something I'd ever thought about, but using Twitter to find mentions of a site is really interesting and would give you not only tips to good posts/pages on those sites but give site owners a new feedback mechanism for their work.

I think one of the toughest, but most valuable, things I'd like to see come from my timeline are opportunities that are uniquely suited to me. By which I mean, if two people who I follow but who don't know each other are converging on a topic, I should be prompted to bring them together.

Case in point, Ben Zimmer helped name "hashtag" as word of the year, and Chris Messina, who invented them, tweeted about it. I follow them both, but had to manually make the connection that I should be introducing/connecting them.

The ultimate level-up for me would be if all my social networks were "rewarding" me by making me feel useful (or look good) by nurturing connections that otherwise wouldn't be made.