Problem
Netflix personalizes the artwork shown for each title, but the algorithm can only choose from images creators have already made. Creators decide what to produce from experience and from past performance data on thousands of launched titles. Reading through that data by hand is slow and subjective, and it invites confirmation bias. Poster rules may not hold on TV and phone screens, where image size and attention differ. The team needs a way to test creative assumptions and to find unexpected patterns.
Example from the article: for the film Purple Hearts, a romance with a military storyline, the team looks for what worked in images of similar titles on both themes.
Solution in brief
Describe images in two complementary ways. The top-down way labels each image with expert-defined attributes (faces, composition, color) and finds which attributes predict success. The bottom-up way clusters images by pre-trained CNN features, so creators browse groups of similar images, check past performance and take inspiration from them.
Questions this case answers
- How to discover which content features of unstructured items drive engagement when the useful features are not known in advance
- How to surface patterns from historical performance of images so that human creators can act on them without confirmation bias
- How to find which visual properties of an image make it perform well when no hypotheses exist yet?
- How to give designers data-driven guidance without hard-coding a rigid style rulebook?
Tools and methods
- Expert-defined image attributes — color, texture, number of faces, composition, facial expression, objects, modeled with computer vision
- Pre-trained vision models and APIs — face and object detection
- Human-in-the-loop labeling tools — let experts train custom models for subjective attributes
- Pre-trained CNN features plus k-means — cluster images by visual similarity
- Member clustering — group viewers by behavior or by the image they responded to
Google or a one-time email code. No password.