Skip to content
View MelikaBaghi's full-sized avatar

Block or report MelikaBaghi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
MelikaBaghi/README.md

Melika Baghi

Ph.D. Candidate in Industrial & Systems Engineering at Georgia Tech, working at the intersection of machine learning, statistics, and data-driven decision making. I develop methods for high-dimensional, structured and incomplete data, with applications spanning healthcare, scientific computing, and sequential decision problems.

Python · PyTorch · scikit-learn · XGBoost · SQL · Statistical ML

I also hold an M.S. in Statistics (Georgia Tech) and an M.S. in Industrial Engineering with a focus on Healthcare Systems. Advised by Kamran Paynabar and Xiao Liu.

melikabaghi.github.io · Google Scholar · LinkedIn · mbaghi3@gatech.edu


Project What it does Links
Medical tourism hospital simulation Hybrid agent-based + discrete-event model of a multi-specialty international-patient department, with channel choice, physician switching, abandonment and deterioration modeled explicitly. A 256-run factorial screen finds bed capacity the most consistent driver of waiting; against an otherwise identical discrete-event-only model, the hybrid model's mean inpatient admission wait is 42% lower (14.8 vs 25.8 days, p < 0.001). code · paper · DOI
RouteCert Certifies a prediction when the evidence already collected determines the decision, and otherwise abstains and says which modality to acquire next, with finite-sample guarantees conditional on the modality pattern. Evaluated on four datasets spanning ECG, speech and wearable sensing. code · paper
cXGBoost Constrained gradient boosting that enforces the manifold-chart bound inside every leaf, so predicted reduced-order bases are valid by construction. On four parametric problems (cylinder wake, Kolmogorov flow, the Kuramoto–Sivashinsky equation, a forced beam) it has lower mean reconstruction error than the projected Gaussian process and Grassmann interpolation. code · paper
State-EXP3 Delayed bandits with drifting state losses: one delayed outcome is charged to every action that could have produced that state, at a cost set by an effective dimension. Reduces regret up to 79% against action-level weighting and 32–68% against a rate-optimal delayed-bandit baseline. code · paper
Tensor-structured surrogates Keeps the multiway structure of simulation output instead of flattening it: every mode gets its own subspace, interpolated on its own manifold, with cores aligned before one low-rank surrogate is fitted. Across five PDE benchmarks in 2D and 3D, the interpolation size is 13× to 482× smaller and cached queries run 3.3× to 8.5× faster than a single-subspace surrogate, with mean errors within 0.22 percentage points. paper
LLM agents as simulated users Three dated OpenAI model snapshots, with a fixed prompt, strict JSON schema and cached replies, acted as 36 synthetic patient profiles inside a queueing simulation of eight capacity policies (120 paired seeds). Matched on baseline rates, they disagreed on the best policy under a demand surge; simple behavioral rules picked the same policy, so the study stopped at its go/no-go gate. details
Retrieval-augmented claim verification Local RAG over 5,183 SciFact abstracts: BM25, two dense retrievers, rank fusion and reranking, with an open 1.5B LLM returning JSON verdicts with citations. Hybrid retrieval raised nDCG@10 from 0.678 to 0.724 over BM25 (p = 0.0008); one distractor abstract cut faithful rationales from 0.251 to 0.142. details
Status
A Product-Grassmann Tucker Surrogate for Parametric Field Prediction Submitted, Journal of Computational Physics
Constrained Extreme Gradient Boosting for Parametric Proper Orthogonal Decomposition Subspace Prediction Under review
Managing Shared Hospital Capacity for Medical Tourism: Patient Behavior and Hybrid Simulation Under review, Health Care Management Science
Conditional Validity for Adaptive Modality Acquisition: When the Policy Chooses Its Own Calibration Group Under review
Pooling and Drift in Delayed Bandits Under review
Supervised Active-Subspace Input Reduction and Risk-Controlled Field Uncertainty Completed manuscript

Full list on Google Scholar · CV · Resume

Interests

High-dimensional statistical learning · Uncertainty quantification and conformal prediction · Multimodal learning and missing data · Sequential decision making and bandits · Reduced-order and surrogate modeling · Tensor and geometry-aware learning · Scientific machine learning · Evaluation of LLMs and AI agents

Pinned Loading

  1. cXGBoost cXGBoost Public

    Constrained Extreme Gradient Boosting for adapting reduced-order models: the estimator and the manuscript.

    Python

  2. IPD-medical-tourism-simulation IPD-medical-tourism-simulation Public

    Hybrid agent-based and discrete-event simulation of a hospital serving medical tourists and local patients: model, data and paper.

  3. RouteCert RouteCert Public

    RouteCert: finite-sample risk control conditional on the terminal modality pattern created by adaptive acquisition

    Python