We follow these steps to arrive at a mode to make content-based recommendations:
First we take our items dataset and identify the features we want to encode for each item. Next we generate a pretend item for a user, based on a user's interaction with items. We can use a user's activity with items such as clicks, likes, purchases, and reviews. So essentially each user now encoded has the same features, and is represented the same as other items in the dataset. Therefore we have a set of feature vectors for all items, and also a pretend feature vector for the target user:
User -> Likes -> Item profile
Now the ...