Recommendation system engineers often feel overwhelmed by the sheer number of models and techniques—FM for feature crossing, DIN for user interest, Transformer for sequences, RankMixer for scaling. Each is understood in isolation, but the big picture remains elusive. This post proposes a unifying framework: every recommendation model can be understood as performing three core functions. First, conditional estimation: predicting the probability of user interaction given context. Second, information organization: structuring user and item data to make estimation tractable. Third, ranking decisions: ordering items to optimize business metrics. By viewing models through this lens, engineers can better understand why certain architectures work, how to combine them, and where new research fits. The author argues that this mental model is more valuable than memorizing individual papers. For practitioners, this means faster debugging, better model selection, and clearer communication with stakeholders. The framework is not tied to any specific implementation, making it a durable reference as the field evolves.
A framework to understand all recommendation models through three core functions, helping practitioners move beyond fragmented knowledge of individual techniques.