Calculating recommendations based on product images #bbuzz
summary
A one-day prototype proves image colors boost recommendations: extract ten centroids per product by k-means on pixels (background removed by border variance), compare centroids with Euclidean distance in Lab color space, and combine that score with attribute-based results using an ensemble; scale by bucketizing with locality-sensitive hashing. Results show a 10% uplift in click-through rate, 12% CTR uplift for low-behavior products, and a 4% overall conversion uplift, improving recommendations where metadata is too coarse.
AI-written summary · about
chapters (3)
related talks
21mShopping Recommendations
27mTransform online shopping experiences with personalized recommendations
32mLeveraging recommender systems to personalise search results
22mExploiting Auxiliary Information for Better Recommendation
38mThe Power of the Bonus Card: Road to Personalised Search
28mThe Power of Personalized Customer Experiences Through AI Innovations
33mPeople who liked this talk also liked ... Building Recommendation...
25mUsing Images in Recommender Systems