Modern recommender systems rely on large-scale ML models that are data-hungry and black-box. Recent advances in LLMs suggest that explicit reasoning can improve learning efficiency, yet it remains unclear how generative LLMs can systematically improve recommendation tasks that are discriminative in nature. Moreover, in personalized settings, LLMs tend to hallucinate. Existing explainable recommender systems either generate explanations independently of predictions or provide post-hoc rationales; in both cases, explanations do not improve accuracy over black-box recommenders. We argue that when properly calibrated to prediction outcomes, natural-language explanations can in fact improve recommendations. We propose RecPIE (Recommendation with Prediction-Informed Explanations), a framework that jointly optimizes prediction-informed explanations and explanation-informed predictions. In RecPIE, the recommendation task guides the learning of consumer representations, which are used by a trainable LLM to generate explanations for why a consumer may or may not like a product; these explanations are then fed back into a neural recommender to improve predictions. The two components are trained alternately, allowing explanations to be progressively refined based on how much they improve recommendation accuracy. Empirically, on next point-of-interest recommendation using Google Maps data, RecPIE improves accuracy by 3-4% over state-of-the-art baselines and matches the best baseline using only 12% of the training data. Human evaluations show that RecPIE's explanations are preferred 61.5% of the time among five competing methods. To our knowledge, this work is among the first to demonstrate that generative explanation and discriminative recommendation tasks can be jointly learned to outperform standalone approaches on either task.

翻译：现代推荐系统依赖于数据需求量大且具有黑盒性质的大规模机器学习模型。大型语言模型的最新进展表明，显式推理能够提高学习效率，然而生成式大型语言模型如何系统性地改进本质上是判别式的推荐任务，目前仍不明确。此外，在个性化场景中，大型语言模型容易产生幻觉。现有的可解释推荐系统要么独立于预测生成解释，要么提供事后归因；在这两种情况下，解释均未能提升黑盒推荐器的准确率。我们认为，当自然语言解释经过适当校准以匹配预测结果时，实际上能够改进推荐质量。我们提出RecPIE（基于预测信息解释的推荐），这是一个联合优化预测信息解释与解释信息预测的框架。在RecPIE中，推荐任务指导消费者表征的学习，这些表征由可训练的大型语言模型用于生成解释，说明消费者可能喜欢或不喜欢某产品的原因；这些解释随后反馈至神经推荐器中以改进预测。两个组件交替训练，使得解释能够根据其对推荐准确率的改进程度逐步优化。在基于谷歌地图数据的下一个兴趣点推荐任务中，RecPIE相比最先进的基线方法将准确率提升了3-4%，且仅使用12%的训练数据即可达到最佳基线的性能。人工评估表明，在五种竞争方法中，RecPIE生成的解释在61.5%的情况下更受青睐。据我们所知，本研究首次证明生成式解释任务与判别式推荐任务可以通过联合学习的方式，在两项任务上均超越独立方法的性能。