Enhancing Diaspora Scientist Identification Through SMOTE ‐Driven Machine Learning and Transformer‐Based Hybrid Models on Multi‐Institution Dataset

This paper explores machine learning‐based discovery of Indian‐origin scientists working overseas with a fused multi‐institution dataset of 4280 records retrieved in seven research institutions: Ames Laboratory, AMNH, Broad Institute, DRI, NIDDK‐NIH, Salk Institute, and Van Andel Institute. The primary task is binary classification of scientist profiles into Indian‐origin and non‐Indian‐origin categories under a class‐imbalanced environment, where 79.9% of the records belong to the non‐Indian class and 20.1% belong to the Indian‐origin class. Random Forest, Support Vector Machine, and K‐Nearest Neighbors were compared with and without the Synthetic Minority Oversampling Technique (SMOTE), followed by hyperparameter tuning, ensemble learning, and a transformer‐enhanced SVM hybrid model. The findings indicate that the impact of SMOTE was dependent on the classifier. Although SMOTE did not have a uniformly good impact on all classifiers, it did increase minority‐class recall of SVM by 0.0896 to 0.4924 before tuning. Following the hyperparameter optimization using SMOTE, SVM demonstrated the best overall performance on minority class, with an accuracy of 0.9117, precision of 0.9109, recall of 0.6205, and F1‐score of 0.7361. Ensemble techniques and in particular hard voting gave consistent results but failed to beat the tuned SVM. The hybrid model that was improved with the transformer created similar performance to the tuned SVM and with the addition of the semantic representation ability of profile text which included name, qualification, designation, and institutional information. In general, the research shows that to identify diaspora scientists reliably, care should be taken when handling imbalance, tuning using cross‐validation, and component‐wise evaluation. The results indicate that classical classifiers, and in particular tuned SVM, are still highly competitive in structured and semi‐structured profile‐identification environments in the future.