Do Calibrated Recommendations Affect Explanations? A Study on Post-Hoc Adjustments

Recommender systems generate suggestions by identifying relationships among past interactions, user similarities, and item metadata. Recently, there has been an increased focus on evaluating recommendations based not only on accuracy but also on aspects like transparency and calibration. Transparency is important, as explanations can enhance user trust and persuasion, while calibration aligns users’ interests with recommendation lists, improving fairness and reducing popularity bias. Traditionally, calibration and explanation are applied in post-processing. Our study investigates two key research gaps: (1) the impact of graph embeddings in model-agnostic knowledge graph explanations, exploring their under-researched potential compared to syntactic approaches to produce meaningful explanations; and (2) the effect of calibration on recommendation explanations, assessing whether calibrated recommendation reordering influences the outcomes of explanation algorithms. We evaluate the quality of explanations using a set of metrics, such as diversity, which measures how well different interests of the user are covered; popularity, which assesses how well explanations avoid favoring already popular items; and recency, which examines the inclusion of recently interacted items. Our findings demonstrate that graph embedding methods are effective in generating high-quality explanations using these offline explanation metrics, and that post-hoc knowledge graph explanation algorithms are robust to calibration changes.