Frame the Recommendation Problem
0
Start by clarifying the recommendation surface. A social app may recommend friends, accounts to follow, groups, posts, comments, short videos, events, marketplace items, or professional connections. Each surface has different freshness, latency, privacy, and safety constraints.
1
For this interview, define the primary goal as recommending a personalized ranked list of social entities to a user. The system should support multiple surfaces, but we can focus deeply on people-to-follow and content-to-engage recommendations because they exercise social graph traversal, candidate generation, ranking, online features, and feedback loops.
2
A strong answer separates the system into retrieval, ranking, filtering, serving, feedback, and learning. Retrieval finds thousands of plausible candidates from billions of entities. Ranking orders them using rich features. Filtering enforces policy, blocks, privacy, and diversity. Serving returns results within a tight latency budget. Feedback updates features and models.
3
Real systems usually combine graph-based recommendations, collaborative filtering, embedding similarity, content understanding, popularity trends, and business constraints. LinkedIn's People You May Know has historically used graph signals such as common neighbors and professional attributes. Pinterest published work on Pixie and PinSage for graph-based personalized recommendations. Meta has described TAO as a distributed social graph store used for Facebook-scale graph reads.
