Recommendation is really hard. It's not a matter of execution, it's a matter of research -- and that's a risky proposition, because the research may not pan out. I know because I tried it (recommending news, but the problem is similar).
I had two or three ideas for recommendation algorithms. My cofounder had quite a few more than that. Over the course of a year with very little income, we iterated intensively on recommendation engines -- we tried four or five completely different ideas, and refined each one until we got stuck. We had a small, but active, community of people who would try any of our recommendation engines and give us feedback. And the feedback was always "this isn't good enough."
I don't know how to solve the problem. The Netflix Challenge results were not especially algorithmically impressive; most of the benefits seemed to come from iterative tuning to the dataset. I admit that I haven't followed much recent research. I think we need an algorithmic breakthrough (on the order of PageRank) before recommendation startups will succeed.
I don't think it's a recommendation problem. He doesn't want to find products that he naturally likes. He wants someone to take a product that isn't exciting in itself and weave a compelling story around it: what the product means and what the product says about a person who appreciates it. He wants to feel like a connoisseur without being able to recognize quality himself. He wants a salesman to teach him how to appreciate the product and, at the same time, convince him that his appreciation makes him a more sophisticated person.
I had two or three ideas for recommendation algorithms. My cofounder had quite a few more than that. Over the course of a year with very little income, we iterated intensively on recommendation engines -- we tried four or five completely different ideas, and refined each one until we got stuck. We had a small, but active, community of people who would try any of our recommendation engines and give us feedback. And the feedback was always "this isn't good enough."
I don't know how to solve the problem. The Netflix Challenge results were not especially algorithmically impressive; most of the benefits seemed to come from iterative tuning to the dataset. I admit that I haven't followed much recent research. I think we need an algorithmic breakthrough (on the order of PageRank) before recommendation startups will succeed.