First Supply, Now Demand; Will Amazon, Netflix, or Google Figure it Out First?
Authors love Amazon rankings. Consumers love Amazon recommendations. Both are bunk. How do I know? Because no one — not the smartest minds in mathematics — has figured out how to pattern demand effectively. At Amazon it is all smoke and mirrors. See Chris Anderson’s book The Long Tail for the (lack of) details of Amazon’s magic formula. They won’t share it — they can’t share it — because it is too wishy-washy. Grocery store dressing is their secret sauce.
In October 2006, Netflix pointed out that predictive ranking systems wore no cloths. They gave away a data set and challenged anyone and everyone to improve their recommendation system. Better their proprietary system by 10%, get 1 million dollars. Free R&D for Netflix — sure, but brilliant none the less. This month’s Wired magazine provides an update on the contest.
The Harry Potter Problem
During the time I worked at Indigo.ca, I learned about a problem from datamining-101 — the Harry Potter problem. When you do a simple customers-who-bought-this-also-bought-this recommendation you run into Harry right away. Everyone bought Harry Potter so your result set becomes useless almost immediately. The compromises begin. To build a better algorithm according to Wired’s Jordan Ellenberg, you apparently start with the rank of the item’s nearest neighbour, you model-in ‘human generated dimensions like highbrow versus lowbrow‘, then you back out as much noise as you can by ‘overfitting’. All fascinating stuff.
The Human Factor
The Wired piece focuses on Gavin Potter — a contestant with a really attractive story. Potter is an amateur and he is using his training in psychology to bring value to his algorithm. Potter says in the article “the 20th century was about sorting out supply. The 21st century is going to be about sorting out demand.” Abolsultely GD right! Potter includes things like inertia — people’s tendency to rank quality relative to recent experience rather than in absolute terms — in his calculations. Queue Mark Hurst et al that champion putting people first — and modeling actual behaviour into business solutions. The idea is compelling for sure, but I wonder if this problem needs fewer dimensions rather than more.
The Promise of Google Book Search
With the net bubbling with talk — by people way smarter than me — of the semantic and social web, I hesitate even bringing it up, but I think an interesting tact to take with recommendation engines would be to strip out human opinion and solely look at the connections between works — both contextually and socially. Aim for some kind of measure of relative cultural currency. Is The Gradute better than The Matrix? Are either better than the next Pixar movie that hasn’t even been imagined yet? Who cares. Tell me the relative ‘importance’ of a work within a certain social/cultural cluster or context (think meme tracking) and I will decide for myself. Deep linking between movies is a ways off, so sorry Netflix, but lookout Amazon. Google has the processing power, the know how (open social) and the database (Google Book Search) to start experimenting with a semantic based book reco engine before anyone else. Of course there are more lucrative areas for Google to mine data but here’s to hoping.
No Comments Yet