For the past two years I’ve worked for a Japanese startup called iKnow! A service with the sole purpose of enhancing learning and memorization.
This is my little hobby project, Kakuteru.com - a new kind of Lifestreamer.
Just started this January...
...by renting an Apartment with nice purple walls
Which means that we’ll have to eat cupnoodles for the coming weeks
Worked remotely together before starting in January. Started out with a concept & prototype for locating things around you.
This January we decided to take a step back and think things over some more.
What are Recommender systems? What problems are they solving? What’s the buzz?
Let me explain some of our thinking and the fruits of it. Warning: some of this might be food for thought and nothing more.
All right, let’s start with the Internet.
One of the reasons why we’re all in this room together is because of the vast amount of digital information. One of the goals of Recommender systems is to provide more targetted delivery of information. So how did we get here? Let’s do a short recap.
In the early stages the Web was like traditional media, producers delivered content to the consumers.
As you all know this has changed, with now pro-sumers generating vast amounts of often irrelevant information. Social-networks are a huge catalyst to this change. Also, bots and machines are playing active role in the contribution of new data.
Let’s take a closer look at Information itself.
Often information is treated the same way as material objects, even though it’s nature is very different. The most important thing is that material objects can only exist in 1 place and their coherence deteriorates when touched.
Information on the other hand, multiplies itself every time it gets touched. The moment a human has acquired new knowledge, it is copied.
This simple fact seems obvious, but even today entire industries (music, movie) try to contain information in vain. The nature of information is to spread.
Kevin Kelly wrote an awesome essay (which I will refer to in the end) about the Internet being a giant Copy Machine.
With everything being digitized, this has tremendous consequences to business & society.
When looking at simple economics of abundance and scarcity, you can conclude that all that can be Copied will drop in value. This can include music, movies, software and anything that gets digitized. Remember, information will leak whether we like it or not. Containing information is going against the grain.
The things that will get scarce and thus rise in value are the things that cannot be copied by the internet copying machine.
Understanding this is fundamental for post-web 2.0 web innovation & business.
Recommendations and recommender systems are a subset of the broader area of personalization. Personalization can also be described as providing relevance. This is something that can not be copied, since it’s only for one person.
Personalization deals with the attention of people, this makes it a key element in the concept of the Attention Economy: human attention is a scarce commodity.
There’s another change I would like to briefly talk about: the speeding up of the Internet.
And especially the rise of Activity Streams. You’ve probably noticed them on a few of the sites your using. Above here is an example of my Activity Stream on Facebook. But nowadays they can be found on every site with a social component.
If you would see the internet as one big brain, the activity streams would be it’s neurons.
Big players like Facebook, Google, Six-Apart have started drafting the first standards to make activity interchangeable.
The current drafts for those standards specify a simple subject-verb-object structure. Subject ‘peter’ uploaded a picture object.
It’s important to keep in mind, that at the end of the day, ALL recommendations are based on activity.
Also, some of the activity verbs can be used as explicit recommendation input (i.e. weighted stronger) and simple activity like the viewing of a picture or the playing of a Last.fm track can be weighted as implicit.
Some food for thought here. What if you have an activity like say ‘posting a status update’. In this case it’s a Tweet of someone telling about a movie he just watched. With a mashup of technologies, in this case Zemanta and the OpenCog framework we can extract the entity ‘The movie Fight Club’ and also we can measure a certain judgement about it. This would be somewhere in between an explicit and implicit ranking of an object.
So let’s get back to our startup now, Reccoon.
Public attention proﬁle
Public attention proﬁle
Recommendation API calls
1. Provide Attention Proﬁles
2. Facilitate rec enhancing
3. Cross-content hacks and smartness
Massive Activity Aggregator
1. Gather long-term summaries
2. Aggregate all recent activity
3. Semantic analysis of objects
public proﬁle list
from Personalization Widget
Kevin Kelly’s Essays @ http://kk.org
‘Revolutionary Wealth’ by Alvin Tofﬂer
Articles by Alex Iskold on Readwriteweb.om
‘The Black Swan’ by NNT
‘The Numerati’ by Steve Baker
http://activitystrea.ms (AS standardization)
http://developer.zemanta.com/ (Semantic parsing)
http://reccoon.com/ (updates will be posted soon)
http://www.opencog.org/ (NLP & AI)
Most icons used in this presentation are by http://picol.org/