AAAI ICWSM 2027 · Full paper submission
Personalized news platforms are often evaluated as if stated preferences, logged recommendation exposure, and click consumption form a single coherent pipeline. Collapsing these layers can distort audit conclusions: a platform may appear more aligned or diverse than observed click consumption supports, which can misdirect diversity governance or algorithmic intervention. We introduce a reusable Preference–Exposure–Consumption (PEC) audit framework that separates stated preference, observed weighted profile state, logged recommendation exposure, app-surface pathways, and click consumption under explicit observability boundaries. Using six months of logs from a deployed mobile news application, we audit 1,583 user profiles, 95,143 logged recommendation items, and 17,512 click events. Each trace type contributes distinct information; none directly substitutes for another. Preference–consumption alignment exceeds chance-based null baselines but captures only part of users' top-consumed category set. Logged recommendation lists contain clicked articles more often than a date-matched candidate-pool baseline predicts (11.29% vs. 9.13%; top-5 lift 1.47×), but many clicks arrive through other app surfaces. Raw exposure–consumption diversity gaps shrink under count matching, yet concentration mismatch persists in the audit-eligible cohort (HHI gap 0.082). Together, the results show that audit conclusions change depending on whether platforms measure stated preference, logged exposure, or click consumption.
Figure 1. PEC audit framework and relation-specific observability boundaries.
Every output is tied to a relation, a denominator, a validity control, and a bounded interpretation.
Boundary: PEC is a measurement protocol. It is not a ranking objective, a causal estimate, or a production targeting procedure.
The conclusions change with the layer being measured.
RQ1 · Preference
The top consumed category appears in stated interests for 77.22% of eligible users, above chance-based nulls. Mean category Jaccard is only 0.228, showing that dominant interests do not capture the broader consumption set.
RQ2 · Traceability
Only 11.29% of comparable clicks appear in the same-day logged recommendation list, versus a 9.13% date-matched candidate-pool expectation. Top-5 lift reaches 1.47×, while most clicks arrive through other surfaces.
RQ3 · Diversity
Logged exposure is broader than click consumption in the raw logs. Count matching removes much of the entropy gap, yet the audit-eligible cohort retains an HHI gap of 0.082 and JS divergence of 0.471.
RQ4 · Divergence
Same-window diagnostic AUC reaches 0.859 with weighted profile-state features. It falls to 0.731 in the time-split diagnostic and below 0.71 in denser-history cohorts, limiting operational interpretation.
Preferences, logged recommendation exposure, interface pathways, and clicks are informative but non-substitutable traces. The apparent alignment and diversity of the same recommender changes with the layer being measured, so every audit claim should name its observable trace, eligible denominator, and interpretation boundary.
When Exposure Is Not Attention: Auditing the Preference–Exposure–Consumption Gap in Personalized News Recommenders
Public preprint · submitted October 7, 2026 · 16 pages, 5 figures, and 14 tables.
The public artifact exposes metric implementations, aggregate tables, synthetic schemas, figure-generation code, and authorized rerun instructions without redistributing private behavioral traces.
Audit metrics, release-safe aggregate tables, synthetic examples, notebooks, and figure builders.
Raw operational logs, row-level click histories, recommendation lists, user profiles, and model feature rows.
Observational measurement claims only; no causal ranking effect, verified impression, or total-attention claim.
Korea University@misc{park2026exposure,
title = {When Exposure Is Not Attention: Auditing the
Preference--Exposure--Consumption Gap in
Personalized News Recommenders},
author = {Park, Woojin},
year = {2026},
eprint = {2610.10173},
archivePrefix = {arXiv},
primaryClass = {cs.SI},
url = {https://arxiv.org/abs/2610.10173}
}
Ph.D. Candidate in Industrial Artificial Intelligence, Korea University
M.S. in Data Science, Carnegie Mellon University
Co-founder & Chief AI Scientist, insightea
His work spans recommender systems, ranking, user modeling, algorithmic auditing, and reliable AI, connecting experience in operational AI systems with careful empirical measurement.