[glutamate]

That is not at all dumb. Doing the preference learning for a huge number of users lets you set up a prior over the preferences that includes correlations between different attributes. Then you use the population prior together with the data for the individual to get the posterior for their personal preferences, which will then take into account the correlations revealed in the population data. This is what we do with http://findmelike.com where we perform the Bayesian update over the user's preferences after every user interaction.

[bru_]

Ha ha... I'm not against learning preferences at all... I think you misunderstood my comment