When new users join a platform without any interaction history, recommendation systems can’t predict their preferences. We address this cold start problem by transferring knowledge from movie ratings to book recommendations while keeping demographic insights intact. Our approach uses three neural networks: one for user demographics, one for item features, and one for combining them. The key idea is selective backpropagation during adaptation—we freeze the demographic network to preserve what we learned about users from movies, while the item and combination networks adapt to books. We test this on two settings: movies to books and movies to music. Results show 28.5% better accuracy than standard transfer learning and 41.2% improvement over baseline methods. Our selective freezing keeps demographic knowledge while enabling domain-specific learning, giving new users personalized recommendations based only on their demographic information.