Not All RecSys Problems Are Created Equal

The complexity of recommender systems is often distorted by outliers like Netflix, but most problems do not require their level of sophistication. A framework for assessing recommender system challenges is based on two axes: the strength of the baseline, determined by observable outcomes and catalog stability, and the subjectivity of user preferences. Systems with clear conversion signals and stable catalogs can use simpler models like gradient-boosted trees, whereas domains with highly subjective tastes and dense data necessitate advanced deep learning for effective personalization. Understanding where a specific problem falls on this spectrum is more important than simply adopting the most complex, state-of-the-art architecture.