Pavel Metrikov, Chloe Gu, Xiaoyu Li, Byungki Byun, Swarnangsu Acharyya · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3841122
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Learning-to-rank systems often use a global scorer despite systematic variation in user preferences. We propose a model-agnostic latent-trait framework for probabilistic ranking objectives. A low-dimensional profile activates shared ranking directions linearly, enabling tractable profile inference with nonlinear document-feature effects. Our iterative route encodes all traits as G profile features in one ranker rather than learning or serving G factor models. Ranker fitting reuses an unmodified standard LTR package, and serving uses one model. An offline module probes the ranker to construct an additive surrogate for profile updates across neural and tree rankers. To our knowledge, this is the first continuous user-specific latent personalization under a ranking loss for standard LambdaMART without architecture changes. In simulation, the method outperforms non-personalized and history-encoder baselines, incurs a modest penalty when G is overestimated, and yields probe-surrogate updates that transfer to the trained ranker. In a retrospective offline Message Search evaluation, personalized LambdaMART outperforms its non-personalized baseline. TreeSHAP provides model-relative trait interpretations.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.