Modern social feeds are powered by recommendation systems that decide what appears next as you scroll. The process typically begins with a candidate set, which is the shortlist of posts or videos retrieved for you before the final ranking decision. To build this shortlist, platforms use a retrieval step that quickly scans a massive catalog of content and pulls out potentially relevant items. Because the full catalog may contain billions of posts, retrieval is designed to be fast and approximate, favoring recall over precision.
Once retrieval narrows the pool to a manageable shortlist, a more expensive ranking model scores each item in detail. This two-stage design lets platforms scale: retrieval handles the computational heavy lifting of finding candidates, while ranking brings accuracy to the final order you actually see. Each ranking model relies on features, which are input variables such as your recent watches, the video's length, or your relationship with the creator. These features are used to predict outcomes like the probability you will watch, like, or share.
A subtle hazard in ranking experiments is label leakage, where a feature indirectly includes future information or the target itself. Offline results may look impressive, but production performance can collapse once the leak is removed. Practitioners audit features carefully so the model learns from genuinely predictive signals rather than artifacts of how the data was collected.