Artwork Personalization at Netflix for Recommendation

Did you hear about Netflix’s $1M challenge for improving the recommendation engine efficiency back in 2009? I was just reading about a few recommendation related articles in the online space and found some interesting information.

Being a long time user of Amazon, ‘Frequently bought together’ section Or ‘Customers who bought this also bought X’ gave some high level hints on the recommendation usage. Recommendation engines helps in stickiness, helps in product discovery, enhanced customer experience and the perfect option for up-selling/cross-selling. It helps in the movement of non-demand, slow moving items.

Netflix is all about content discovery. I was knowing about some of the facts –

1. If I am seeing a particular actor’s movie, I will be getting recommendations of that actor’s movie
2. Same goes with genre as well (comedy, romance, thriller, sci-fi, drama)
3. If I start seeing some movie and closes it without proceeding much, my recommendation won’t include the likes of the movie which I closed now
4. Naturally new releases will be getting a preference in the recommendations

However, what blew my mind was about a technique called Artwork Personalization at Netflix. Without any doubt, we know that the first thing which attracts a user to click on the next video recommendation is the one single picture (artwork) which is there as a thumbnail. For the same movie, there are multiple artworks being created at Netflix. Consider for 2 people, the next recommended movie is Good Will Hunting starring Matt Damon, Minnie Driver, Robin Williams. For a profile, which is interested in Romance, Image A below will be displayed and for another profile who loves, watching comedy, Image B, where in Robin Williams (a well-known comedian) will be displayed.

Image Source - https://netflixtechblog.com/artwork-personalization-c589f074ad76

Recommendation is personalized at individual user level. This is just brilliant and what makes it fascinating is the few 100 M user base of Netflix and peak handling of 20M+ requests per second with low latency. Artwork Personalization details available here.

In the case of Amazon, system cannot just look for ‘customers who bought this also bought’ just based on the item to customer relationship. The main reason is most customers would have bought some common items (e.g. Masks, Sanitizers, Harry Potter books etc.). If I am trying to buy a TV, they cannot suggest these items to me, even if someone else bought it. Amazon works on a technique called item to item collaborative filtering, wherein more than the similarity between customers, it’s taking into account the correlation between the products. Details available here.

What we started off the topic was with a $1M challenge. Though Netflix gave the prize to a group that improved the efficiency by 10%, they never used that code due to change in business strategies and effort involved. Details available here.

Recommendation based selling is a true example of Steve Job’s quote – "People don't know what they want until you show it to them."