What makes an item interesting?

It's a big question, and an important one, because attention is finite and most people would rather not waste time looking at stuff that is uninteresting (to them). If you have a system that figures out what is interesting then you can filter out the uninteresting stuff.

This is easier said than done.


Sources for interestingness metrics


A number of sources of information may allow us to guesstimate the interestingness of a given item.

1. Context

  • How recent is it? (Especially valuable for news)
  • How authoritative is the source? (assuming some universally agreed-upon way of determining authority)

2. Visitor behavior information


There are a few indicators that can be derived from visitor behavior. Many sites will let you sort their contents by "most viewed", "most discussed", "highest rated" and "most favorited". Number of backlinks is also a popular measure.

These are all basically popularity indicators.

PostRank has a secret formula that combines the following (and more):
  • number of comments
  • number of times item was social boomarked and dugg
  • number of times item was tweeted and liked in FriendFeed
  • number of times item was "clicked" (whatever that means)

See it at work on, e.g., the popular blog BoingBoing. (If you don't mouse over the item scores, you're missing something!)

3. Reader-based information

  • Have I seen it already?
  • Does it need my attention soon?
  • Is it relevant to my interests?

4. Reader-specific information derived from third parties

  • Did somebody direct it specifically to my attention?
  • Did people I trust find it interesting?
  • Did people like me find it interesting?

Based on all this, the ideal item might be something recent that I haven't seen yet, comes from an authoritative source, is relevant to my interests, has been deemed interesting by people I trust and people like me, and was directed by someone specifically to my attention.

The implicit metrics behind common sources of information


Now, it is important to realize that the sources of information we typically rely on will filter information as if there were only one or two metrics that matter. For instance, my incoming twitter stream (well, firehose) assumes nothing is important unless it comes from someone I follow. Here are some more examples:
Source
Metric(s) taken into account to filter content
My twitter stream
  • Reader-based: does it come from someone in my network?
  • Context: is it recent?
Twitter @replies to me
  • Reader-specific from 3rd parties: was it directed specifically to my attention?
  • Context: is it recent?
My aggregator (e.g. Google Reader, Bloglines)
  • Context: is it recent?
  • Reader-based: have I not already seen it?
  • Reader-specific from 3rd parties: does it come from a source I trust?
Google
  • Reader-based personal relevance: does it match my query text?
  • Visitor behavior - inbound links: is the PageRank high?
http://popurls.com
  • Recency: Is it currently popular on the web?
  • Visitor behavior: How many inbound links, visits, duggs (from anyone) did it get?
Trending topics on http://trendistic.com/
  • Recency: Is it currently popular?
  • User behavior: How often has it been repeated (by Twitter users)?
http://twitturly.com/
  • Recency: Is it currently popular?
  • User behavior: How many backlinks (by Twitter users)?
Presumably, a well-balanced info diet will incorporate data sources that favor a diversity of metrics.

Universal, group-specific and user-specific metrics


The PostRank metric is universal, meaning that a given piece of content will have same value regardless of the user. By contrast, a user-specific interestingness metric would incorporate data relating to a specific user (e.g. his personal output or his social graph).

In between the two, a group-specific interestingness metric incorporates data relating to a group or aggregated from its members. Digg and Reddit provide group-specific metrics, with the group being the site's users.

The personal microblogosphere idea consists in computing a metric using only data from a restricted set of people found in the user's social graph, e.g. only friends (radius 1) or friends-of-friends (radius 2).

It must be kept in mind that the more user-specific the metric, the more computation-intensive it gets if you want to satisfy lots of users.

User-specific
Personal microblogosphere
Group-specific
Global
Most relevant to:
one user
one user
Group members
Anyone
Whose input contributes:
Potentially everyone
One user's neighborhood
Group members
Everyone





More about this


Bokardo: Which Movie to Watch? An Overview of Recommendation Systems lists seven "ways to prioritize": Newness, Time-sensitivity, Popularity, Personal Relevance, Social Network Relevance, Authority, and Collaborative.