Re-ranking recommendation feeds by visual appeal / Eitan Zimmerman (Argmax)

Discover how re-ranking recommendation feeds by visual appeal can improve user satisfaction, using Lab color space, color histograms, and complementary/triadic/quadrant colors to enhance the Artlist footage recommendation engine.

Key takeaways

The RGB color space is not perceptually uniform, meaning that small changes in RGB values do not always correspond to similar changes in perceived color.
Lab color space is a perceptually uniform color space, trying to measure color in a way that is more similar to human perception.
Color histogram can be used to measure color distribution of an image and calculate similarity between images.
Artlist, a high-quality stock content hub, asked to create a recommendation engine for their footage section.
The recommendation engine should be able to rank footage based on its visual appeal and not just similarity.
The team used Lab color space to calculate the histogram of an image and then calculated the complementary, triadic, and quadrant colors.
Complementary colors are those that are opposite each other on the color wheel, triadic colors are those that are equally spaced from each other, and quadrant colors are those that are at 45-degree angles from each other.
The team used a naive approach to color re-ranking, by calculating the color histogram of each image and ranking them based on their similarity.
The team also used user feedback to personalize the recommendation engine and penalize images that are too similar.
The team used a vector database to index images and used similarity search to find similar images.
The team used a collaborative filter to combine user feedback and metadata to generate personalized recommendations.
The team found that using Lab color space and calculating complementary, triadic, and quadrant colors improved the recommendation engine’s performance and user satisfaction.

Re-ranking recommendation feeds by visual appeal / Eitan Zimmerman (Argmax)

More talks