How to implement this recommendation algorithm?

Most recommendation algorithm articles I've read are focused on the Netflix model where users rate items. What I want to do is slightly different (I think).

Let's say that instead, I want to create a site where a user is presented with two pictures of cars. The user can then select which car they like better. The user can repeat this process as many times as s/he likes, but hopefully as they continue, the pictures become more and more refined towards what the user likes.

How would you implement this algorithm? It seems like one possible way would simply be to implement an ELO ranking algorithm and use the order of those results as a "rat开发者_运维百科ing", but that has serious flaws in that multiple items can't be given a maximum rating (which the user may have done if given the ability to rate the items themselves).

Another method, which seems more promising to me, would be to predetermine the general properties of each vehicle (e.g. color, body type, 2 door vs 4 door, etc.) and use those to get a general idea of the properties each user likes and base recommendations off of that.

I'll take a stab at this.

Suppose that each car is given a set of properties. If this set of properties were coded as a vector, one potential method of recommendation would be to use Self Organizing Maps (SOM). The basic gist of a SOM is that is a categorizer of input vectors. If you train a SOM with input vectors representing distinct classes of input, a SOM will start to cluster its storage vectors to be more like each class of input. Note that the original input vector is not retained. To train a SOM with an input vector, the best vector currently in the SOM is picked and then the area around that vector becomes more like the input. Of course, see Wikipedia http://en.wikipedia.org/wiki/Self-organizing_map.

So how does this apply to this situation? Well, one SOM could be used to train on images that the person does like and one could be trained on the ones they do like. Even if there are is no single style that they like, clusters should form around cars they like/don't like. Then seeing if they like a car that has not been picked by them is a matter of finding how well it matches to groups from their likes and dislikes. Note that in this case, it would be best to match up cars that are dissimilar to each other or more likely to not both be liked.

When the person first joins the site, it may be advantageous to allow them to pick a few likes and dislikes right off the bat to seed the SOMs.

Good luck!

Maby it's a bit too late to answer, but you might want to check this out. It's about an MIT professor that argues that 5-star rating, like-rating, etc. don't work, he proposes an algorithm that works with input by pairs, just as you suggest (Car A or Car B). The Algorithm is quite complex but can be found on the link.