Recommendation systems operate in two main stages: candidate generation and ranking. Candidate generation retrieves a broad set of possible items from the platform's content inventory, the pool of available content the system can choose from at any moment. Ranking then scores and orders those candidates to choose what the user actually sees. Both stages rely on signals, which are measurable user or content attributes that help predict outcomes like watch time, satisfaction, or shares.
Signals come in two flavors. Explicit signals include follows, likes, "not interested," and reports — actions the user deliberately takes. Implicit signals include watch time, pauses, rewatches, and even scrolling speed — behaviors that reveal preference without a conscious button press. Negative signals, such as quick skips, "not interested" taps, hides, and reports, reduce future recommendations, while strong positives like replays, shares, and saves push a piece of content further. Watch time and completion rate are especially valuable for video because they are easy to measure at scale and correlate with intent matched. Replays suggest the video was compelling or loopable, shares indicate strong emotion or usefulness because they cost more than a like, saves suggest long-term reference value, and comments signal engagement, though many platforms look at quality, not just raw counts.
Several dynamics shape how these signals play out over time. Freshness ensures newer content gets a chance even when it is not yet proven, and old items may decay unless they keep performing. Embeddings, or learned numeric representations of users and content, capture similarity — what topics, formats, and audiences tend to go together — and similarity is what drives most recommendations. The cold start problem appears in two places: a new account has little data, so the system relies on defaults and early interactions, and a new post has no performance history, so it is tested on small audiences first, where strong early signals can dramatically expand distribution. These dynamics create feedback loops, since what you see affects what you do, which trains what you see next, and can reinforce narrow interests or amplify popularity bias, where already-popular content gets even more exposure. To counter these tendencies, platforms add diversity and serendipity constraints, balance exploration (testing new content) against exploitation (showing known favorites), enforce safety and policy constraints through downranking, and increasingly incorporate long-term satisfaction metrics like repeat visits, reduced reports, and survey responses to balance raw engagement.