At Home24, Europe’s largest online furniture retailer, we face the challenge of calculating good recommendations for products where we lack enough user behavior data. Although we get fairly good results with recommendations based on product metadata, for some products the metadata is just not granular enough to produce good results. In particular, "material: wood, color: brown" matches a wide variety of products that don’t actually fit together from a human perspective.

To make up for this deficit, we found a way to augment our algorithms with product images - analysing the colors in each image to recommend products that match well visually.