Naman Agrawal

About

Research Interests

My research weaves together several intellectual threads that share a common ambition: to bring high-resolution data and rigorous quantitative methods to bear on questions about how societies organize, communicate, and make decisions.

Publications & Working Papers

Peer-Reviewed Publications

• Phase-Aware Deep Learning with Complex-Valued CNNs for Audio Signal Applications Forthcoming

  Agrawal, N.

  To be presented at the IEEE Conference on Artificial Intelligence (CAI 2026), Granada, Spain.

  Proposes complex-valued CNN architectures, with complex batch normalization, principled initialization schemes, and a comparative study of complex activation functions, to exploit phase information typically discarded by real-valued networks. Evaluated across three stages: image classification baselines, MFCC-based audio classification, and phase-aware graph neural networks. Results demonstrate that complex architectures can match real-valued baselines while offering improved handling of phase structure in audio pipelines, with a careful analysis of stability–expressivity trade-offs across activation choices.

• Analysis of Student–LLM Interaction in a Software Engineering Project Published

  Agrawal, N., Shariffdeen, R., Wang, G., Rasnayaka, S., & Iyer, G. N.

  IEEE/ACM International Workshop on Large Language Models for Code (LLM4Code), ICSE 2025, Ottawa, Canada.

  Tracked 126 undergraduate students over a 13-week semester, analyzing conversational logs, 730 LLM-generated code snippets, and code complexity metrics including cyclomatic complexity, control flow graph depth, and Halstead effort. Demonstrates that dialogue-based ChatGPT interactions produce shorter, structurally simpler code compared to GitHub Copilot, highlighting how conversational prompting reduces code complexity while satisfying project requirements.

Software

• covalchemy: R Package for Constructing Joint Distributions with Control over Statistical Properties R Package

  Agrawal, N. (2025)

  Provides tools for constructing multivariate distributions that preserve empirical marginals while allowing precise control over correlation structure, mutual information, and higher-order dependence via copula-based methods. Benchmark experiments demonstrate marginal fidelity with less than 1% error. Designed for research applications and statistical pedagogy, including as an engine for reproducible Simpson's paradox benchmarking.

• autoharp: R Package for Semi-Automatic Grading of R/Rmd Scripts R Package

  Agrawal, N., Gopal, V. et al. (2026)

  A customizable toolkit for assessing student R and R Markdown submissions, offering automated correctness checks, run-time profiling, and static code analysis. Supports educator workflows with functions to render and evaluate scripts, extract code structure, and generate comprehensive diagnostic reports.

Working Papers

• Cultural Flows in the Digital Age: Evidence from YouTube's Global Music Charts Working Paper

  Agrawal, N.

  Analyzes 1.26 million weekly entries from YouTube Music's Top-100 charts across 60 countries (2021–2025) to map global music circulation. Graph-based clustering reveals three stable listening blocs: a cohesive Latin American community, a broad global mainstream cluster linking Europe and Asian markets, and a distinct East African group. Bilateral similarity regressions confirm that shared language is the strongest predictor of musical affinity, with asymmetric diffusion networks positioning the United States as the central hub and Mexico as the pivotal conduit for Spanish-language content.

• Seventy-Five Years of the Supreme Court of India: An Empirical Analysis of Judicial Decision-Making, Citation Networks, and Institutional Evolution Working Paper

  Agrawal, N.

  Constructs a novel dataset of 42,833 judgments with structured metadata extracted using large language models, covering party characteristics, legal doctrines, constitutional provisions, precedent citations, bench composition, and case outcomes. Estimates determinants of appellant success using OLS, logit, probit, and high-dimensional fixed effects, with causal identification through IV, RDD, and DiD designs. Maps citation networks to identify the 50 most influential precedents and detect doctrinal communities via Louvain clustering.

• Crisis, Compassion, and Capital: The Economics of Global Charitable Giving Working Paper

  Agrawal, N.

  Examines nearly 50,000 projects across 201 countries representing over $575 million in charitable donations (2002–2025). Using the Russian invasion of Ukraine as a natural experiment through difference-in-differences and event study designs, finds 300%+ funding surges for Ukraine-linked projects with evidence of both additionality and substitution effects. Text analysis reveals that urgency keywords increase funding by 35–40% and life-saving language by 89%, operating primarily through new donor acquisition. Documents substantial geographic inequality, with North American projects receiving 2–3× more funding than African counterparts.

• Copula-Based Approaches to Engineering Simpson's Paradox and Applications to Fairness Evaluation Working Paper

  Agrawal, N. & Gopal, V.

  Presents a reproducible copula-based pipeline, using quantile inverse-CDF mapping, targeted copula dependence, assignment optimization, and simulated annealing, to generate tunable Simpson's paradox datasets for algorithmic fairness benchmarking. Demonstrates that pooled estimators can mask severe subgroup failures, providing a practical testbed for fairness diagnostics. Implementation is available in the covalchemy R package.

• Equilibrium Analysis for Strategic Information Revelation in Data Markets Working Paper

  Agrawal, N.

  Develops a tractable signaling model in which data sellers choose how much of a dataset to disclose and buyers update beliefs accordingly. Derives conditions for pooling versus separating equilibria and uses simulations to show how disclosure costs and prior beliefs shape market vulnerability to adverse selection. Highlights policy implications including standardized disclosure requirements, certification mechanisms, and third-party audits.

• Rigidity as a Signal: A Game-Theoretic Explanation for Persistent Vacancies in Urban Real Estate Markets Working Paper

  Agrawal, N.

  Models vacancy as a deliberate, costly signal in a two-period game with heterogeneous sellers, deriving multiple equilibrium types: pooling-holdout, pooling-discount, and two separating forms. Analyzes comparative statics and demonstrates how targeted policy levers, including vacancy taxes and price-ratio mandates, can shift equilibria toward more informative outcomes and improve allocative efficiency in urban housing markets.

Research & Professional Experience

Teaching

I served as an Undergraduate Teaching Assistant across the Departments of Computer Science and Economics at NUS from July 2022 to July 2025, and received multiple teaching excellence awards during that period.

Workshops organized: Generative AI for Business · Computer Vision with OpenCV · Network Analysis with tidygraph (R User Conference) · Introductory & Advanced LaTeX · NUS SoC Summer Workshops

Selected Projects

Honors & Awards

Contact