Can cross-user behavior reveal news relations that individual histories miss?

News recommendation faces a sparsity problem: each user's reading history is short relative to the news corpus, and articles are short-lived. Most content-based methods extract user interest representations from the user's individual reading history alone — meaning they can only recommend articles similar to ones the user has already read. If a user's history contains only an "orange" news article and the candidate set has nothing similar, the system cannot make a meaningful recommendation.

GLORY's insight is that other users' click sequences reveal global relations between articles that any individual user's history cannot. If users in aggregate tend to read article X and article Y in the same session, those articles are related — even if no individual user's history contains both. A global news graph constructed from all users' clicks captures these relations.

Two encoders use this graph. A global-aware historical news encoder gives global perspective to the user's history. A global-aware candidate news encoder uses a global entity graph to find effective associations for candidate articles. When user U1 has read only an orange article, the global graph reveals that orange articles connect to blue articles and green articles in cross-user co-clicking patterns; the system can then recommend blue or green candidates that don't textually match the orange article but match it in the cross-user behavioral structure.

The general lesson: aggregating across users surfaces structure that no individual user can see. When per-user data is sparse, the population's collective behavior is the source of additional signal — and explicit graph construction from cross-user co-occurrence is one effective way to extract it.

Source: Recommenders Personalized