Recommendation Engines

Photo by http://www.flickr.com/photos/good_day/41798369/

Photo by good_day

I started using Netflix and I’m both impressed with the recommendation engine and curious. Amazon does a great job using a combination of items and behaviour to present thing I might like to purchase. I find the Netflix Canada content recommendations good but not quite as good. This could be the difference that my Amazon history dates back to June 24, 1999 with my @davidcrow.ca email address. I started thinking about how simple these recommendations seem, i.e., basic inputs and outputs but potentially very powerful business implications. You can see recommendations at Amazon, TiVO, Netflix, Pandora, etc. It’s a great tool for discover and can potentially increase sales and business metrics (back in 2006 Amazon was reporting 35% of sales come from recommendations). Recommendations can be incredibly valuable business drivers (see Digg Recommendation Engine to drive traffic) and I was very curious at how others had  approached this problem, specifically if there were methods or techniques that I could leverage. (I’ve been following the creation of a taste graph or interest graphs by companies like Hunch and Quora and Gravity). The other interesting spot where I see a lot of people discovery is on dating sites, however, I think the success metrics of profile matching for certain dating sites is very different (think eHarmony vs Ashley Madison).

There is a lot of really great tools. I built a Hadoop instance and had Mahout up and running on my MacBook Air in about an hour. It’s not as fancy as the backend at Backtype (check out all of their tech). I’m curious at what others like EventBrite are using to power their social discovery of events and Chomp for mobile apps.

  • http://bmannconsulting.com boris

    I’m not that impressed by Chomp. I’m fascinated by the Apple “Genius” part in the App Store. It works very well.

    I’m also wondering when Chomp, AppShopper, and others add wishlists and purchased info for eBooks and music. Like Hunch, I bet you can make some inferences / correlations on purchases of different kinds of content that actually enforce recommendations cross-content.

  • http://blog.hiremebecauseimsmart.com stat arb

    OK but some part of that 35% of sales via recommendations were just people in the mood to buy something — you can’t just subtract 35% off their revenue to project the no-recommendation-engine case.

    • http://davidcrow.ca/ davidcrow

      Hoping I did not infer or imply this, my point was that by augmenting the buying behaviour by presenting “recommended items” that you can extend that behaviour. In the case of Amazon that using their recommendation engine they increased their sales by 35%. For others and other activities the impact/outcome will be very different. The discussion at http://news.ycombinator.com/item?id=537371 sums this up very nicely.

      Robert Grossman provides a great analysis of the Netflix strategy lessons at http://rgrossman.com/2009/07/05/three-lessons-in-analytic-strategy-from-the-netflix-prize/

      For Netflix happier users means more long term revenue. For Amazon, better recommendations means more spend. What does it mean for Twitter, or Facebook, or TripAdvisor?

    • http://davidcrow.ca/ davidcrow

      Hoping I did not infer or imply this, my point was that by augmenting the buying behaviour by presenting “recommended items” that you can extend that behaviour. In the case of Amazon that using their recommendation engine they increased their sales by 35%. For others and other activities the impact/outcome will be very different. The discussion at http://news.ycombinator.com/item?id=537371 sums this up very nicely.

      Robert Grossman provides a great analysis of the Netflix strategy lessons at http://rgrossman.com/2009/07/05/three-lessons-in-analytic-strategy-from-the-netflix-prize/

      For Netflix happier users means more long term revenue. For Amazon, better recommendations means more spend. What does it mean for Twitter, or Facebook, or TripAdvisor?

    • http://twitter.com/ravipathak ravi pathak

      I think another fact worth mentioning here is that only 16% of visitors came to website with the intent of shopping. If you consider others as browsers, you need to up them to buyers and recommendation play a significant role to do exactly that, however not to counter that there are people who come with strong intent to buy a specific product, mostly recommendations doesn’t affect them. Its the other set of visitors (browsers)  to whom it affects without they knowing it , I believe.

      • http://isomorphismes.tumblr.com isomorphisms

        Ravi, Why do you think recommendation plays a more significant role in turning non-buyers into buyers rather than better satisfying (and retaining) customers who were already intending to buy something? Your theory sounds logically possible, but why do you believe it.

        • http://twitter.com/ravipathak ravi pathak

          My hypothesis is the customer who has intention to buy, have decided on a product. They have more or less clear goal to be achieved via an e-commerce store. Therefore, they are less likely to be distracted by recommendations for up-sell/cross-sell. 

          Even if my assumption is not correct, I am fairly certain that you have less than 10% of your visitors who come with a defined objective on your website. I call rest 90% as browsers. This group of  90% are either researching, trying to see what could fit their bill, or doing some random things that one can never imagine. (for e.g.one of my friend who runs a web development shop, tries to buy from an e-commerce store , takes a video of how slow the experience was to sell his services :))Recommendation helps this group ! Think if you are on a offline store with nobody(to tell you whats interesting) except a cashier only looking to collect your money. Would you buy something from that store. May be yes, but only if you have a dire need.

          • http://isomorphismes.tumblr.com isomorphisms

            Ravi, I don’t think that can be correct with Netflix and Amazon. There are so many repeat users who are kept on as repeat users through effective recommendations. Not only that but people who come to buy one book/movie and end up buying two books/adding multiples to the queue.

          • http://twitter.com/ravipathak ravi pathak

            If we somehow do happen to cross the road , I may be able to share may be an A/B test which partially butress what I was saying, but I do agree, it was not for Amazon so can’t claim that for amazon !