Researchers have found that using pairwise ranking outperforms single-action reinforcement learning for selecting explanations in offline recommendation systems. They tested six different selection methods, including LambdaRank and teacher-student distillation, on two datasets: a Google Local subset with 2,958 pairs and the MovieLens-1M dataset. The results show that LambdaRank reaches an F1 score of 0.500 on Google Local and 0.329 on MovieLens-1M, outperforming other methods like PPO and GRPO. This study highlights the importance of efficient explanation selection in recommendation systems, which can incur significant serving costs due to latency and cost scaling with traffic. --- Why it matters: This matters because it provides a practical lesson for building efficient explainable-recommendation systems, which can help reduce costs and improve user experience. Engineers can apply these findings to develop more effective offline-pool selectors and KG-path selectors. Source: https://arxiv.org/abs/2608.18531