Sunil Gupta

Research profile

Research summary

Sunil Gupta is a machine learning researcher at Deakin University (AU) whose recent work spans two complementary tracks: personalized federated learning and statistical methodology for clinical prediction. An h-index of 18 with an i10-index of 30 reflects steady, mid-sized contributions, and current output is strong — 19 works in the last three years, with annual counts rising from 4 (2021) to 8 (2025) and 5 already logged for 2026. Topic profiles concentrate on artificial intelligence in healthcare (19) and machine learning in healthcare (15), alongside domain adaptation and few-shot learning (11), Bayesian methods and mixture models (11), and topic modeling (7). The four 2026 arXiv papers in Evolve's catalog — similarity-aware personalized federated learning (SAPE-FL), a ReLU-based sample-size framework, and two oversampling studies predicting depression symptoms and remission in university students — position the researcher as a middle or last author rather than corresponding author. The field split is even (50/50) between federated learning and ML methodology/biostatistics, indicating a deliberately broad method-application portfolio. Collaboration is active: 19 distinct co-authors in three years, anchored by Svetha Venkatesh and including Jill Newby and Alexis Whitton on digital mental health. Reproducibility, assessed over 2 papers (mean 38), is the weakest radar axis, with internal-only validation flagged.

Research keywords

Highlights

• SAPE-FL: similarity-aware personalized federated learning

  In this 2026 arXiv paper (middle author) Gupta and colleagues propose SAPE-FL, which anchors each client to global and peer models via hybrid similarity to handle statistical heterogeneity and non-IID data; Evolve finds broad empirical gains but idealized convergence theory and rates reproducibility medium.

• A statistical framework for ML sample-size determination

  This 2026 preprint decomposes ReLU networks into piecewise linear regions to derive power-based sample sizes using Cohen effect sizes and the noncentral F-distribution. Evolve calls it clever but flags that effect sizes come from noiseless proxy surfaces and validation remains internal; reproducibility rated low.

• Oversampling methods for university student mental-health prediction

  Two 2026 papers with Gupta as last author: a novel oversampling method predicting amotivation and anhedonia symptoms among university students, and SMOTE-VAR, an uncertainty-aware oversampling approach for predicting depression remission. Neither has been reviewed by Evolve yet.

Timeline

1. 2026Publishes SAPE-FL on arXiv: similarity-aware personalized federated learning for heterogeneous, non-IID environments (middle author).
2. 2026Publishes a statistical approach to estimating sample size of machine learning models, built on ReLU piecewise-linear decomposition (middle author).
3. 2026Last-author paper predicts amotivation and anhedonia symptoms among university students using a novel oversampling method.
4. 2026Last-author paper introduces SMOTE-VAR, an uncertainty-aware oversampling method for predicting depression remission in university students.

Papers in the Evolve catalog

• 2026-09-09

  A Statistical Approach to Estimating Sample Size of Machine Learning Models

• 2026-09-02

  Similarity-Aware Personalized Federated Learning in Heterogeneous Environments

Frequent co-authors