Serendipity: Accuracy’s Unpopular Best Friend in Recommenders

Eugene Yan · 6 years ago · 17

Recommendation systems typically optimize for accuracy metrics like NDCG and recall, but serendipity—measuring surprising and novel recommendations—matters more for long-term user engagement and business health. The article reviews metrics for measuring diversity, novelty, and unexpectedness across 10+ research papers, with formulas like cosine similarity and point-wise mutual information used to quantify how different recommendations are from user history and population trends. Balancing serendipity with relevance prevents user boredom, improves long-tail product sales, and can outperform accuracy-focused approaches in A/B tests despite worse offline evaluation scores.