Count-based popularity control under-corrects semantic-popularity exposure in text-aware recommender systems

Popularity bias in recommender systems is commonly measured from an item’s own interaction count, leaving the popularity structure encoded by semantically related items largely unaccounted for. Here, we define semantic popularity as the neighbourhood aggregation of train-only item popularity in text embedding space and introduce semantic popularity exposure (SPE) to quantify how strongly recommendation lists favour semantically popular regions. Across five Amazon Reviews 2023 categories, direct semantic control reduced SPE more than matched count-based control at every positive penalty strength, and this ordering remained consistent with independent LightGCN and SASRec recommenders in all 10 dataset-backbone combinations under a baseline-relative 5% NDCG floor. At matched achieved NDCG, semantic control also produced lower mean SPE at every non-baseline utility target, while neighbourhood ablations showed that random-neighbour, category-mean and count-based signals were substantially weaker than semantic-neighbour signals. Across the five dataset-level units, the slope-based count-to-semantic transfer closely tracked the pre-intervention correlation between count popularity and semantic popularity (Pearson \(r=0.873\) ; mean absolute deviation \(=0.029\) ), supporting \(T≈ r\) as a descriptive first-order approximation rather than an exact law. These findings show that interaction-count popularity and semantic popularity represent distinct exposure dimensions in text-aware recommender systems, and suggest that semantic neighbourhood structure should be audited alongside conventional popularity measures when designing popularity-control mechanisms.