A retail company gets very few purchases per customer and relies on many new users. When splitting data for training and testing a custom recommender, how should the data scientist create a test set that reflects real usage?

The correct approach is to use each user's most recent interactions as the test set. This method simulates how a recommender system would be used in practice: predicting future behavior based on past interactions. It ensures the model is evaluated on its ability to recommend new items a user would interact with, rather than items they have already interacted with. Shuffling all records and holding out the last 10% of overall interactions (option 1) could include historical data for users in the test set, which doesn't reflect real-world prediction. Selecting users with the fewest interactions (option 3) or randomly selecting users (option 4) for the test set would not accurately assess the model's ability to make future recommendations for existing users, which is crucial for a recommender system.