Conventional popularity measures typically use an item’s own interaction count. The study instead defines semantic popularity as a neighbourhood aggregation of train-only item popularity in text-embedding space. It then introduces semantic popularity exposure, or SPE, to quantify how strongly recommendation lists favor regions of an item catalogue that are popular among semantically related products. This distinction allows researchers to assess exposure inherited from an item’s semantic neighbours as well as popularity associated with the item itself.

The analysis combined the five Amazon categories with two independent recommender architectures, LightGCN and SASRec, creating 10 dataset-backbone combinations. Under the baseline-relative NDCG floor, the advantage of semantic control over matched count-based control was consistent across all 10 combinations. A separate comparison at matched achieved NDCG also found lower mean SPE with semantic control at every non-baseline utility target. These results concern the study’s SPE metric, recommendation-utility measure and tested product categories; they do not establish improvements in clicks, purchases or other real-world outcomes.

Ablation experiments tested whether the result depended on how semantic popularity was represented. Random-neighbour, category-mean and count-based signals were substantially weaker than semantic-neighbour signals. The comparison indicates that the text-embedding neighbourhood supplied information about popularity exposure that the alternative signals did not capture as effectively. It also suggests that the observed advantage cannot be interpreted simply as a new label for conventional count-based popularity control.

The researchers additionally examined how count-based control transferred to semantic control across the five dataset-level units. A slope-based transfer measure closely tracked the pre-intervention correlation between count popularity and semantic popularity, with a Pearson correlation coefficient of 0.873 and a mean absolute deviation of 0.029. The authors present T ≈ r as a descriptive first-order approximation rather than an exact law, and the correspondence does not guarantee identical behavior in other catalogues, domains or embedding models.

Overall, the study identifies interaction-count popularity and semantic popularity as distinct dimensions of exposure in text-aware recommendation. Its reported results support auditing semantic neighbourhood structure alongside conventional popularity measures when evaluating mechanisms designed to limit the prominence of broadly popular content. The research appears as an accepted, open-access article in Scientific Reports, with the supplied early version identified as subject to further edits before replacement by the final Version of Record.

Subject of Research: Evaluation of semantic popularity control in text-aware recommender systems

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

Article References: Esmeli, R., & Abdullahi, H. (2026). Count-based popularity control under-corrects semantic-popularity exposure in text-aware recommender systems. Scientific Reports. https://doi.org/10.1038/s41598-026-76829-7

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76829-7

Keywords: recommender systems, popularity bias, semantic popularity, text embeddings, SPE metric, LightGCN, SASRec, Amazon Reviews 2023, NDCG, exposure control, Count-based, popularity