Given a user *u *and an item *i*, we calculate a score that serves as a *proxy* for preference. This means that high scores mirror high preference and vice versa. This score allows us to rank a set of items according to their relevance for a given user. The score can be a user’s probability to interact with an item. Thus, we aim to predict user-item interaction probabilities. To compute this probability, we use a deep neural network with a single output unit. This unit uses the sigmoid function for activation, which yields output values in the interval (0, 1). Thus, we can interpret the network output as probability and use it for ranking. The network is then trained on distinguishing between preference and disregard. Therefore, we label all positive user-item combinations with 1 and negatives with 0. As a result, the learning task presents itself as a binary classification task. Excelling at this task can enhance the relevance of our recommendations.