Building a recommendation system is hard. Not just the engineering, but the decision-making process before you write a single line of model code. Will you use collaborative filtering? Will you have sufficient user behavior signals to apply a deep learning approach? What if you’re launching a new product with zero data, a classic cold start problem?
Most data teams either rely on institutional knowledge (“we always use matrix factorization here”), spend weeks benchmarking approaches, or simply apply whatever the latest conference paper recommends. None of these approaches is systematic and none of these approaches scales.
Top comments (0)