Google · AI/ML Engineer · Mid-Level

Design a Watch Next recommendation system for a video-streaming product that recommends the next video a user is most likely to enjoy when they finish or pause a video. Address: optimization objectives and success metrics, including offline versus online metrics; training examples and user, video, and context features, including how to treat impressions without clicks, clicks without completion, and other implicit feedback; candidate retrieval, ranking, and any re-ranking under a large catalog and low-latency constraints; a reasonable ranking loss, and a multi-task or multi-objective loss if click, watch time, and completion are optimized jointly; the optimizer and its core update equations (for example, SGD or Adam); cold start, popularity bias, feedback loops, content diversity, and duplicate recommendations; and offline evaluation, A/B testing, monitoring, and iteration. Reasonable assumptions are allowed, and the focus is on trade-offs and rationale.