Raw

I ran across APML the other day, but skimming secondary sources ( Lightweight Attention Preference Markup, Tom Morris on Attention, Thoughts on Tom Morris on APML) the idea didn't click. What this seems to be about isn't so much attention per se, as an approximate, compressed, conceptualisation of it. The spec currently seems very much a first-pass thing (like " good idea - quick, write it down with pointy brackets!"), but I reckon the underlying idea could well be useful.

[PS. It seems Paul checked this out months ago...] 

If I understand correctly, an Attention Profile is a vector space which changes over time, derived from a given data source. Each vector is a real-valued relationship between a resource and a concept, so the profile as a whole describes a kind of topic map ( not a Topic Map). You could say that today this blog is 80% about the Semantic Web, 20% about cats, tomorrow those figures might be reversed. The approximation/compression bit comes in by the concepts being represented by folksy tags, with individual assertions in APML looking like this:

<Concept key="cats" value="0.8" from="dannyayers.com" updated="2007-10-06T01:55:00Z" />

I believe the way in which the data is derived from a given source is left open. Again, if I understand correctly, a person can have an Attention Profile, which may be derived from their activity - browser history, IM etc.

The intended application of this stuff is in filtering, similarity matching between the profile of a given datasource (say a web page) and the profile of a person who might potentially be interested.

It's a shame (but hardly a surprise) the developers decided to invent their own knowledge representation language when there are perfectly good ones already available. SKOS would definitely be in the frame for this. But with a bit of tweaking ( is @from meant to be a named resource? - if so it should be a URI!!) and 3rd-party work on transformation (i.e. GRDDL) interop should be possible.

I do like the way this stuff harks back to old AI - I'm sure there are loads of long-neglected algorithms that could be used to derive the source data & do the filtering. Hmm, I wonder if the (bang up to date) Probabilistic DL Reasoning might be applicable.

Certainly one to watch. I hope they bear in mind other work in the general area, and above all pay attention to WebArch and other good practices.

 

 

 

@en