Recommender systems need serendipity to stimulate users' active exploration and break predictable consumption cycles. Existing offline beyond‑accuracy metrics usually isolate either historical similarity or global popularity, making it hard to assess both together with actual user relevance. To address this, we propose SPADE (Serendipitous Pareto Distance Evaluation). The key idea is to embed every item into a two‑dimensional space where the x‑axis encodes historical similarity and the y‑axis encodes popularity. For each user, we compute the Pareto frontier that maximally balances popularity and similarity—i.e., items that are not simultaneously dominated on both axes by any other item. Then, for correctly recommended items in the test set, we calculate the minimum Euclidean distance to this frontier and average the distances across users to obtain a serendipity score. Experiments on five public datasets and five baseline algorithms show that SPADE effectively prevents algorithms from gaming offline metrics with irrelevant or non‑personalized recommendations, reliably isolating genuine serendipitous discoveries.
Review: SPADE unifies popularity and similarity into a single evaluation space and uses Pareto frontiers plus Euclidean distance to provide an interpretable, robust serendipity metric, improving offline assessment of recommendation quality.