In this work, we study the low-rank MDPs with adversarially changed losses in the full-information feedback setting.

However, this is not always the case, as different tasks can compete with each other and lead to weaker models.

In this work, we introduce WinoGAViL: an online game of vision-andlanguage associations (e.g., betweenwerewolves and a full moon), used as a dynamic evaluation benchmark.

Graph convolution networks (GCNs) for recommendations haveemerged asan important research topic due to their ability to exploit higher-order neighbors. Despite their success, most of them suffer from the popularity bias brought by a small number of active users and popular items.