Image Optimization at Netflix
At Netflix, we figured out that optimizing which images to show users when suggesting videos makes a big difference. For example in the image above we see 6 images describing the new Netflix original "The Unbreakable Kimmy Schmidt". The image at the bottom right attracts more users than the default image at the top left. Using machine learning to learn what images users are more attracted to is now a big effort that significantly increases engagement. This is a broad new initiative at Netflix that includes multiple teams (including myself).
In retrospect, it is not very surprising: users scan the page of videos and look for something appealing. The human brain processes images more efficiently than text, and a visually appealing image can grab the attention effectively.
More details on this effort are available in a non-technical blog post, and a technical blog post where more examples are shown. The non-technical blog post discusses issues like what are the characteristics of appealing images - more or less human subjects (less)? Heroes vs. villains (villains)? Is there a difference in preferences based on regions (yes)? These are all questions that are quite interesting in their own right from a psychological perspective. The technical blog post discusses the systems that power the new feature and the AB testing that was used to develop it.
I feel like this is the streaming version of having a good book cover to attract readers.
Hi Guy, It’s a nice piece. Image optimization plays a key role even in brick and mortar businesses such as consumer packaged goods where brands fight for both shelf space and consumer attention. Closer home, in the online world, image optimization plays a key role in home page and landing pages design. I did not understand how Netflix’ secret sauce tastes different from others? Is it a better algorithm compared to peers? Is it about serving an optimized video page to each of 75 mil subs based on their previous viewings?
Good one. How about aspects of gender- female vs male and then segregated by region? That would add to complexity and that's where the user engagement becomes tricky to crack. Analysis over a period of time then helps to understand "one pattern" of the behavior. Pattern changes with situations like - for offers/sale on some episodes users engage differently than their perceived behavior. Color too plays an interesting role. Lots to consider when addressing user engagement. This is an awesome and exploratory field. Keep it coming..:-)
I guess you'll soon post articles on deep learning :) but before that, why not use the stock cover provided by film producers? are those pictures not optimum?