Reviews: DropoutNet: Addressing Cold Start in Recommender Systems

Well written and with a good survey of related work. The paper addresses the following problem: when mixing id level information (user id of item id) with coarser user and item features, any model has the tendency to explain most of the training data with the fine grained id features, using coarse features only to learn the rarest ids. As a result, the generated model is not good at inference when only coarser features are available (cold-start cases, of a new user or a new item). The paper proposes a dial to control how much the model balances out the fine grained and coarser information via drop-out of fine grained info while training. The loss function in Equation (2) has the merit of having a label value for all user-item pairs, by taking the output of the low rank model U_u V_v T as labels.