A social feed can use models to estimate which posts someone may like or reply to. A model is a learned computational representation that uses patterns in data to produce predictions; here, those predictions help rank posts. section 2.2.3
Two predictions, one feed
A like model and a reply model can assign different scores to the same post. The platform can weight those scores to create one ranked list, changing which posts appear higher. section 2.2.3
Models can also learn from labeled examples to identify content relevant to a policy, or use historical spam data to distinguish spam from legitimate messages. Deep learning can discover useful features and their weights, but its decisions may be difficult to explain. Labels and predictions also have limits: deciding whether policy-related content supports or opposes a proposal can require context that even a human may lack. section 4.7.1 section 2.4
Figure 2.2 Two different kinds of engagement are predicted: A first model predicts the probability of a Like, and another one, the probability of Reply. Each item is thus assigned two different scores.
The figure shows how separate models give each item a like score and a reply score. section 2.2.3

