herec

We were halfway through a code review when a senior engineer pointed at a variable named herec and asked, "Is that a typo,? Or are you naming things after Czech actors? " It was neither. The lowercase herec appears across GitHub repositories, research discussions, and production codebases as shorthand for HERec, a recommendation algorithm that learns from heterogeneous information networks. The name collision isn't a bug. It's a reminder that recommendation problems are about relationships between different kinds of entities, not just users and items.

Most engineers miss the real lesson of herec: recommendation quality depends less on the model than on how you represent relationships between different entity types.

I've spent three years building recommendation pipelines for marketplaces and content platforms. Along the way, I inherited a codebase where herec served as the baseline for every experiment involving user-item interactions plus side metadata. This article breaks down what the algorithm actually does, why meta-paths matter, where it fails in production, and how it compares to modern graph neural networks. No fluff, no hand-waving.