The authors propose a generic model-based re-ranking framework, MultiSlot ReRanker, that optimizes relevance, diversity, and freshness. It uses the Sequential Greedy Algorithm, which is efficient for large-scale recommendation engines. The framework models the mutual influences among items and leverages the second pass ranking scores of multiple objectives. The authors have also generalized the offline replay theory for multi-slot re-ranking scenarios and built a multi-slot re-ranking simulator using OpenAI Gym integrated with the Ray framework. This simulator can quickly benchmark both reinforcement learning and supervised learning algorithms.
Publication date: 11 Jan 2024
Project Page: https://arxiv.org/pdf/2401.06293v1.pdf
Paper: https://arxiv.org/pdf/2401.06293
