DivNet: Diversity-Aware Self-Correcting Sequential Recommendation Networks

As the last stage of a typical $\mathit{\text{recommendation}}\mathit{\text{system}}$, $\mathit{\text{collective}}\mathit{\text{recommendation}}$ aims to give the final touches to the recommended items and their layout so as to optimize overall objectives such as diversity and whole-page relevance. In practice, however, the interaction dynamics among the recommended items, their visual appearances and meta-data such as specifications are often too complex to be captured by experts’ heuristics or simple models. To address this issue, we propose a $\underset{‾}{\mathit{\text{div}}}\mathit{\text{ersity-aware}}\mathit{\text{self-correcting}}\mathit{\text{sequential}}\mathit{\text{recommendation}}\underset{‾}{\mathit{\text{net}}}\mathit{\text{works}}$ ($\mathit{\text{DivNet}}$) that is able to estimate utility by capturing the complex interactions among sequential items and diversify recommendations simultaneously. Experiments on both offline and online settings demonstrate that $\mathit{\text{DivNet}}$ can achieve better results compared to baselines with or without collective recommendations.