Auditing Gender Disparities in an Academic Profile Retrieval Pipeline: Indexing Choices and the Evaluation of Preprocessing Mitigation

Academic profile-search and recommendation systems rely on information extracted from historical data to identify and rank potential researchers. However, existing disparities within these data can also influence which profiles become visible to users. This study examines gender representation in a semantic retrieval pipeline built from 60,995 Scopus researcher profiles collected for Centinela, a research collaboration platform developed at Escuela Politécnica Nacional (Ecuador). It also revisits the evaluation of five preprocessing mitigation techniques implemented in the Holistic AI library. The retrieval analysis covers 220 research-area queries and uses all-MiniLM-L6-v2 embeddings with exact cosine similarity. Three versions of the indexed profile text were compared to determine whether including gender, h-index and affiliation information changes the composition of the retrieved results. With the original template, women accounted for 21.0% of the top-10 results, compared with an average of 35.2% in the corresponding research-area pools (a difference of −14.2 percentage points, 95% CI −16.5 to −12.1). This difference decreased substantially when the gender field was excluded, reaching 39.1%, above the pool share but close to the 41.1% share of women in the whole corpus. The retrieved profiles also showed considerably lower h-index values than the corresponding pools. A proportional re-ranking strategy improved gender representation while producing almost no change in the area-based relevance measure. The second analysis evaluates Reweighing, Disparate Impact Remover, Correlation Remover, Learning Fair Representations and Fairlet Clustering using 30 random train-test splits. The results show that the parity previously reported for Reweighing was associated with the use of weighted selection rates rather than an actual change in unweighted predictions. When evaluated on held-out data using unweighted metrics, Correlation Remover and Disparate Impact Remover reduced gender disparities in a non-circular prediction task, although with some loss of utility. The remaining techniques did not provide comparable improvements. These findings highlight how decisions made during data preparation and evaluation can influence conclusions about fairness. Nevertheless, the results are limited to one dataset, one embedding model and a binary gender attribute. The generative component of the platform was not evaluated.