Why people like liking things

When we first started the Preona project one of our most important assumptions was that actively liking things is very cumbersome and few people will want to do it in order to personalise the reading suggestion system.

What we wanted to build was a system that could automagically learn from what the user is doing, be intriguingly implicit and invisible.

This was wrong.

After observing what people are doing online and how they behave, we, or rather I, have come to the conclusion that liking is a very important action that transcends cultural and technological bounds. And that people do it for everything but improving the ratings of what they’re reading.

The reason people like liking things is simple: Feedback.

And I don’t mean giving feedback to developers, or recommendation systems. No, they’re giving feedback on a completely personal level. The act of liking things online has gained a significant cultural and sociological dimension – it’s a lot like giving flowers to people, the meaning depends wholly on the context.

But this makes the whole act very difficult to interpret and base your machine learning systems on. When the context defines the action’s meaning, rather than the action itself that’s a very big problem. Sure, in general you can assume that liking means people would like to read more of that particular content, but what about when they’re just expressing agreement, or mutual dislike of something? Or expressing condolences? Nobody wants to read a lot of sad things …

This sounds like a very difficult problem to solve, but perhaps the solution is incredibly simple.

Implement liking as a purely conversational feature and ignore it completely from a machine learning perspective? I don’t know, we’re going to have to try and see what works best.

An interesting experiment would be to try predicting whether a person will “like” something with Google‘s new prediction API. This probably calls for a weekend project, hopefully I’ll have a chance to play with that in the following days and try it out.

Related articles by Zemanta

• LinkLog: Google Prediction API (dorai.wordpress.com)
• BigQuery and Prediction API: Get more from your data with Google (googlecode.blogspot.com)
• Machines that learn better (physorg.com)
• Google’s search for the perfect learning machine (news.cnet.com)

This entry was written by swizec, posted on May 20, 2010 at 1:42 pm, filed under Real-Time web, Synaptic web, The Web and tagged Application programming interface, Artificial intelligence, Google, Machine learning, Recommender system. Bookmark the permalink.
Follow any comments here with the RSS feed for this post.