Data 607 - Discussion 11 - Netflix Recommendation Engine

November 7, 2018

Cinematch: Netflix (Old-School DVD) Recommendation Engine

Netflix's (pre-streaming) DVD recommendation engine, Cinematch, attempted to guess what your opinion on unnrated movies would be based on your previous ratings
The personalized recommendations projected your rating of the movie on a five-star scale, allowing for partial stars rounded to the tenth on the scale (projected 3.7 of 5) and driving suggestions on the site's user interface
The Netflix prize - which in Oct. 2006 offered $1 million to any person or team that could improve their recommendations by more than 10 percent - was arguably the first publicized citizen data science challenge and foretold of Kaggle
As a movie fan with eccentric tastes, the Netflix projected rating was by far the best indicator of what I would think of a movie and a great guide to picking selections

Target users: Active Netflix customers - movie fans
Key Goals: Find and watch movies they love so they keep their subscription and tell their friends
How the Recommendation System Helps Them Accomplish Goals: Make it easy to find movies they'll love

Nearly 3 years after announcing Netflix prize competition, a team finally crossed the 10 percent mark for recommendation improvement
Here's a book on Netflix's growth that covers the contest from a relative non-technical standpoint and a great NYT article about the competition
Here's an online exercise in trying to create a recommendation engine for Netflix using Excel that can be helpful to reverse engineer the process
Despite the team meeting the prize's goal, the entire solution was never implemented
This was a function of cost and, more importantly, Netlifx's shift to streaming

Netflix Instant (streaming service) launched in 2007, and the company realized that while people may love some French independent movie and rate it 5 stars, they frequently aren't in the mood to watch it
High-brow movies would stay in a user's queue for months or years; once shipped, many users would end up keeping those movies at home for weeks or month
With the treasure trove of data they got from Instant, they realized users would love Moonlight but more frequently wanted to watch Die Hard
Their recommendations shifted from being based on ratings to what a user actually watched

Target users: Active Netflix customers - movie fans
Key Goals: Find and watch movies they love so they keep their subscription and tell their friends
How the Recommendation System Helps Them Accomplish Goals: Uses the user's on-demand viewing choices to recommend similar movies they will watch

The shift from recommending based on ratings to behavior is a very interesting move that likely improves Netflix's customer retention and sign-up rates
Netflix Instant includes a simplistic thumps-up and thumbs-down rating system, but it's updated recommendation system is still primary behavior-driven
The five-star rating system persists on the DVD side of Netflix, but there's less info and recommendations do not seem to be driven by this
I miss the old system, which, anecdotally, seemed to accurately project my feelings about a movie at least 90% of the time