Author Topic: Behind Amazon's ranking system from a machine learning perspective  (Read 317 times)  

Offline Jonathan C. Gillespie

  • Status: Arthur Conan Doyle
  • ****
  • Posts: 995
  • Gender: Male
  • Atlanta, GA
  • Relentlessly Patient
    • View Profile
    • Fiction For Every Reader
I was at MLCONF, the Machine Learning Conference, earlier this year in Atlanta. I learned a variety of lovely facts of interest, key among these being that all our day jobs are doomed. Seriously.

One of the presenters was from Amazon, and we got a rare peek behind Amazon's ranking system. What I learned is it isn't just actual purchases that go into determining a product's rank. Amazon actually builds rank by weighing a variety of actions a consumer takes against a product. I have no idea if this applies to books, because the examples given were clothing and other products, but I 'd imagine there isn't much difference in the engine's treatment in the bookstore.

Each action that represents greater interest/commitment from the consumer is treated with more and more weight by the algos. So, for example, having a search query just display your product nets you a few "points". When a customer actually clicks on your product, that's more weight. A purchase is, of course, the action carrying the greatest weight. All these factors did--the machine learning expert stated--contribute to a product's actual rank.

If I had thought of it at the time, I'd have asked about reviews and clicking "more" in descriptions, and such, but for some reason that escaped me.

Any way, there you go. Something to chew on.


I write worlds on paper, then destroy them.
Jonathan C. Gillespie | Official Site | Twitter | facebook | Newsletter

KBoards.com

  • Advertisement
  • ***

    Offline ilamont

    • Status: Arthur Conan Doyle
    • ****
    • Posts: 636
    • Gender: Male
    • Mass.
      • View Profile
    Re: Behind Amazon's ranking system from a machine learning perspective
    « Reply #1 on: December 09, 2016, 12:21:34 pm »
    Thanks for sharing. If it indeed applies to books as well, it points to the importance of strong covers, titles, descriptions, and metadata, all of which impact the likelihood of someone taking action relative to other titles. It also suggests that reinforcing feedback loops are at work, which is great for titles that the Amazon algorithm promotes but really harmful for the many more that aren't getting much love.

    On the other hand, now that AMS is open to everyone, maybe that's an opportunity to level the algorithmic playing field ... at least while clicks are still cheap.

    KBoards.com

    • Advertisement
    • ***