MAPS: Netflix’s Multimodal Asset Personalization at Scale

Netflix's artwork personalization system can use what a new image contains before that image has accumulated much viewing history. It converts artwork into a numerical description of its visual content and combines that with interaction data. A shared model then learns across five artwork formats, allowing preferences learned from busier formats to inform sparsely used ones. Both changes mattered: Netflix reports that adding image descriptions alone or consolidating models alone produced no significant improvement in its core online metrics, while combining them did. The result supports using content and shared interaction history together to personalize newly introduced artwork.