Thanks Anthony for a very comprehensive article, well-explained. For me, one of the most fascinating aspects of recommendation systems is their fallibility. Netflix regularly recommends programs that are hilariously mismatched to my real interests. This disconnect arises out of their fundamental approach to classification: by genre. The idea is that if I like sci-fi show X then I’ll also like sci-fi shows Y and Z (especially if Netflix has just spent hundreds of millions on making Z and is desperate for us all to watch it). But of course my actual decision-making process is far more biased around quality than around genre. I may have enjoyed sci-fi show X because of its superior acting and story-line and conversely have no interest in sci-fi shows Y and Z because of their by-the-numbers approach to story development and the wooden acting of their stars. Eventually, categorization systems will need to become sophisticated enough to move away from simplistic genre-based approaches. And “people like you also bought” approaches may work within a standard deviation or so of the average taste, but will fail across much larger sections of a user base. Undoubtedly we’ll see tremendous improvement over the next couple of decades in recommendation systems, but for today they remain all too often rather clunky, churning out irrelevant or even risible suggestions.