Answers are grounded in this paper and cite their sources.
Serendipity, the discovery of unexpected yet valuable items, has become an important objective for recommender systems (RS) aiming to reduce filter bubbles and support long-term user satisfaction. Although accuracy-driven models predict known preferences effectively, they can restrict exploration and reinforce narrow consumption patterns. Despite increasing attention to serendipity as a beyond-accuracy objective, the field remains fragmented because definitions, evaluation protocols, and ethical considerations are not consistently aligned. This systematic review examines serendipity in RS from 2020 to 2025, covering 147 studies and 98 algorithmic contributions. We synthesize diverse definitions into a unified conceptual framework centred on relevance, novelty, and unexpectedness. We also present a methodological taxonomy that traces the field’s evolution from heuristic and traditional approaches to deep learning, knowledge graph, generative, and LLM-based methods. Reported results show that recent LLM-based approaches can achieve strong performance on serendipity-oriented tasks, with \:NDCGₛₑᵣₑₙ scores of up to 0.398. We further review evaluation practices and highlight the gap between computational proxies and subjective user experience. Through quantitative ethical analysis, we show that 74.1% of studies acknowledge ethical concerns, but only 6.8% include dedicated ethics sections, 19.7% operationalize ethical metrics, and 5.4% address content safety. By integrating conceptual, methodological, evaluative, and ethical perspectives, this review provides guidance for developing serendipity-oriented RS that balance relevance with beneficial discovery.
No takes yet. Share an insight, caveat, or question.
Najafabadi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: