Right, but this solution tries to fit what is otherwise continuous into discrete so his calculations are easier. It is naive and unrealistic. I've never seen

Right, but this solution tries to fit what is otherwise continuous into discrete so his calculations are easier. It is naive and unrealistic. I've never seen that succeed, and the 'approximations' people do with those kind of numbers are scary. Are all likes equal in weight? Does my affinity remain constant over time? Am I capable of liking a countable number of concepts? Do I naturally revisit all items I've liked in my past life and re-evaluate if they make my cut? The solution isn't in limiting likes, it's from finding a way to measure 'likes' using actual human behavior passively. For example, on newegg, sorting by 'most popular harddrive' is usually a better proxy than 'highest rated' for the quality I'm looking for. People vote with their wallets, and newegg was among the first to offer that data to inform recommendations. The tradeoff is this behavior helps entrench the winners where markets follow power laws, and that is why Amazon, for example, doesn't offer a similar feature (Amazon's is 'popular and new'). There will be a startup (if there isn't already), who will be acquired by a bank, designed to mine mint data with a POS receipt log and turn that into marketing data (customers who buy these bananas also buy gym membership at ...). JPMorgan (I think?) bought a groupon competitor for this reason. Facebook is tackling this from the other direction by combining browsing history with experian.