Tom Printy’s Blog

Fast Search a company in Norway has just acquired Agent Arts, a recommender system that has been focused on mobile systems. This is great for the Agent Arts guys. I was talking with them when I was looking for my last full time gig. They were based in Australia at the time but they moved to California sometime this year.

While this is well and good what frustrates me the most is that my past fulltime employer Sourcelight Technologies is never mentioned. Sourcelight has been in business for 10 years and has powered the movie recommendations on Hollywoodvideo.com Moviegallery.com comcast.net and digitarian.dk . So why do people ignore Sourcelight? Do they not know about it? The amazing part of this is that there is no VC money involved just some smaller investors. The really interesting part is that Sourcelight started by creating Kiosks that people could use in the store to get recommendations. These kiosks were deployed in Blockbuster stores in Chicago and Texas and other places.

So what does the recommender space look like? It is highly competative. There are a few clients and you always are bidding against the same competitor. I wonder when some of these companies will start cahnging direction or closing. I also think Choicestream will be interesting to watch. They have raised a ton of cash and thier burn rate has to be incredible. I haven’t heard a lot from them lately so I wonder if thier business is slowing.

Recommendation systems are still in thier infancy. The eventaul goal of all these systems in advertising. The company that does that the best first will win this game.

WOW!  Netflix serves nearly a billion movie predictions a day. That is really amazing!  They also collect 2 millions ratings per day.  That is a lot of data.  I would love to see the back end systems that they use to crunch the ratings. I had heard that it takes them 3 days to generate their forecast model now I know why.  You can read more about their technique on forbes.com  I wonder how these numbers relate to what the other recommender systems gather.  I know what we gather at Sourcelight is not nearly as large as this.  I would have thought with all this data they would have released a larger chunk of it for the Netflix prize contest.

I was trolling technorati to see if anyone was talking about Sourclight Technologies one person stated in thier blog that “This is my favorite movie recommendation site. It has been the most accurate so far in my opinion.” I am glad that she has found it useful and hope that she will continue to use it.

Here is an interesting post describing how one person got an RMSE of 0.8937 with the Netflix contest data. Since he didn’t think this  procedure would go any lower he published his method.  What I find really interesting is that his method took a couple of weeks on at 3GHz computer. Clearly this method would not work real time. He also has 256 point in his sytem that is used. I my methods I was only have to have about 130 before I ran into some memory issues.

With recent announced funding that both Choicestream and Aggregate Knowlege are getting there are still some money to be made in this space.  But with the valuation of these two companies I wonder what types of contracts they are going to have to put togther just to stay a float. These guys are also leaving a nice niche where smaller companies can move in and work out deals with some smaller web site/service.  It is interesting to see how much money these guys are going to burn through.