# Summary ![[Pasted image 20241010010731.png]] This paper details YouTube's content recommendation algorithm (shown in Figure 2). The approach consists of two neural network architectures: a candidate generation network and a ranking network. The candidate generation network receives the user's YouTube activity history as input and outputs a small subset of videos from the large video corpus. The candidate generation model performs a coarse-grained ranking using collaborative filtering. The ranking model gives much finer-grained ranking by ingesting a rich set of features describing each video and the user, and outputting a score for each video based on that data. # Key Points - Main challenges for recommending content on YouTube include: 1. **Scale** - YouTube operates on an unprecedented scale, servicing billions of daily users recommendations from a corpus of billions of videos. Any recommendation algorithm for YouTube needs to scale to an immensely-large user base and video corpus, disqualifying a number of effective, poorly-scaling recommendation algorithms 2. **Freshness** - new videos are constantly being uploaded to YouTube, and new user actions are constantly being taken that should inform the user's future recommendations. Any recommendation algorithm needs to be responsive to these new actions and content uploads, balancing the tradeoff between recommending new content and existing content. This can be framed as an exploration-exploitation trade off as in reinforcement learning 3. **Noise** - YouTube doesn't have a true measure of user satisfaction, and thus has to predict noisy proxies (such as user watch time). This makes the prediction task more difficult because we cannot predict our true target - Model evaluation during training used traditional metrics like precision, recall, and ranking loss, but the true test of the effectiveness of the model was an A/B test examining metrics that correlate with our target (such as click-through rate or watch time) -