Research

Research statement and completed studies

What I have studied so far, what I want to study next, and how my training in computer science, machine learning and international economics connects the two.

01 — Statement

Research statement

My research interests sit where machine-learning methodology meets questions about how digital technologies spread and what they change. I come to these questions with two kinds of training. A computer-science degree gave me the computational grounding to build and evaluate models directly — my undergraduate thesis implemented TF–IDF weighting and a Naive Bayes classifier from first principles — and several years of software and data engineering taught me to assemble, clean and integrate data from heterogeneous sources.

My master’s degree in international economy and business added the substantive side: how institutions, investment and inequality shape economic outcomes across countries. In my master’s thesis I combined World Bank and V-Dem data into a country-level panel and used fixed-effects models, robustness checks and regional and period-specific analyses to study the correlates of internet adoption. The results were informative but explicitly associational, and the thesis itself identifies reverse causality and limited generalisability across regions as open problems.

Those limitations define what I want to work on next. I am interested in causal machine learning — especially the estimation of heterogeneous effects from observational data — and in reliable machine learning that holds up under distribution shift, for example when evidence from one set of countries, periods or firms is carried over to another. The application I care most about is AI adoption: which firms and countries adopt AI, why, and with what consequences for productivity and inequality. These are developing directions for doctoral research, not completed work.

02 — Themes

Four research themes

Themes are labelled by status. Proposed direction: no completed work yet. Builds on completed work: a new direction that extends finished research. Completed work and ongoing interest: I have already done research in the area.

Proposed direction

Causal and Reliable Machine Learning

A developing direction rather than completed work. My master’s thesis produced associations that could not be read causally and that varied by region. I want to learn and apply methods that estimate heterogeneous effects from observational data and that remain reliable when the population a model is applied to differs from the one it was trained on.

Starting point in completed work

Questions I would like to study

  • How much do the effects of technology-related interventions differ across countries, firms or institutional settings, and can these differences be estimated credibly from observational data?
  • When a model of technology adoption is estimated on one set of countries or periods, how well does it transfer to another?
  • How should robustness and uncertainty be reported when observational evidence informs decisions about digital policy?

Methods of interest

  • Heterogeneous treatment-effect estimation
  • Double/debiased machine learning
  • Causal forests and meta-learners
  • Sensitivity analysis
  • Distribution-shift and out-of-distribution evaluation
  • Uncertainty quantification

Application context

Cross-country and firm-level data on the adoption of digital technologies, including AI.

Builds on completed work

AI Adoption and Technology Diffusion

My completed research concerns ICT diffusion measured by internet use at the country level. Extending this to AI adoption — and especially to firms — is a proposed direction. I am interested both in the determinants of adoption and in its consequences.

Completed work in this area

Questions I would like to study

  • Why does the adoption of AI vary across countries, and do the economic, educational and institutional correlates of earlier ICT diffusion carry over to AI?
  • Which characteristics of firms are associated with adopting AI, and how does firm-level adoption relate to country-level patterns?
  • Is AI adoption associated with changes in productivity or inequality, and can those relationships be identified credibly?

Methods of interest

  • Cross-country panel analysis
  • Firm-level empirical analysis
  • Causal machine learning for heterogeneous effects

Application context

Firm-level AI adoption as the intended research context, complemented by cross-country comparisons.

Completed work and ongoing interest

Computational Social and Economic Analysis

The empirical work I have completed so far combines economic, political, institutional and technological indicators in country-level panels, and studies how relationships change across regions and periods.

Completed work in this area

Questions I would like to study

  • How can large, heterogeneous cross-country datasets be combined so that measurement differences are handled transparently?
  • Which findings from global panel models survive when regions and periods are analysed separately?

Methods of interest

  • Fixed-effects panel models
  • Statistical inference with clustered and robust standard errors
  • Robustness analysis
  • Regional and period-specific analysis

Application context

Economic, political, institutional and technological data at the country level, from sources such as the World Bank and V-Dem.

Completed work and ongoing interest

Natural-Language Processing

My first research project was in natural-language processing, and I have continued to build classical text-classification pipelines. I see text as a useful source of evidence for social and economic questions, including how technologies are discussed and adopted.

Completed work in this area

Questions I would like to study

  • How can text data be used as reliable measurement in empirical social and economic research?

Methods of interest

  • Sentiment analysis and text classification
  • TF–IDF representations
  • Naive Bayes, logistic regression and linear SVM
  • Classical NLP preprocessing pipelines

Application context

Reviews and other user-generated text; potential measures of technology adoption drawn from text.

03 — Completed research

Completed studies

Each study follows the same structure. The theses are individually authored; the political-stability study is a co-authored course project and is not a peer-reviewed publication.

Master's Thesis2024Sole author

Investigating the Multifaceted Factors Influencing Information and Communications Technology (ICT) Growth

Corvinus University of Budapest · Department of World Economy

Research question

Besides the level of economic development (GDP per capita), which factors are most strongly associated with differences in internet usage across countries and over time?

Motivation
Much of the literature treats ICT as a driver of economic growth. This thesis reverses the lens and treats ICT growth as an outcome shaped by social, economic and political conditions, bringing educational inequality, political competition and foreign direct investment into the same model as income.
Data
Individuals using the internet (% of population) and FDI net inflows from the World Bank’s World Development Indicators; educational inequality (a Gini coefficient originally from Clio Infra), political competition (originally from Polity 5) and GDP per capita, all obtained through the V-Dem dataset.
Sample
A combined panel covering 218 countries. The main fixed-effects model is estimated on 2,592 country-year observations from 124 countries; the educational-inequality series ends in 2010, and the period-specific models cover 1991–2000 and 2001–2010.
Methods
  • Panel regression with country and year fixed effects, estimated in Stata
  • Robust standard errors clustered by country
  • Alternative specifications: Liberal Democracy Index in place of political competition; total GDP in place of GDP per capita
  • Dynamic specifications: a lagged dependent variable, and first and second differences of internet usage
  • Separate models for 1991–2000 and 2001–2010
  • Separate models for nine V-Dem world regions (the Pacific had too few observations)
Main findings
  • GDP per capita was positively associated with internet usage in the main model (coefficient 3.245 percentage points per US$1,000; p < 0.01) and stayed positive and significant in every other specification that included it. It was the most consistent correlate across regions, although it was not significant in every region.
  • FDI net inflows were positively associated with internet usage (coefficient 0.0544 percentage points per US$1 billion; p < 0.01). The coefficient was larger for 1991–2000 than for 2001–2010 and was not significant once lagged internet usage was included.
  • Educational inequality had a positive, statistically significant coefficient in the main model (0.773; p < 0.01), which the thesis interprets as internet use being concentrated among more educated groups in unequal societies. The sign and significance varied by region — for example, negative (at the 10% level) in Eastern Europe and Central Asia.
  • Political competition was negatively associated with internet usage in the main model (−0.031; p < 0.01), but the coefficient was not significant with a lagged dependent variable or in either period-specific model, and it was positive (at the 10% level) in East Asia.
  • Model fit the main model reported an R² of 0.791.
My contribution
Sole author: formulated the research question, assembled and merged the World Bank and V-Dem data, specified and estimated all models in Stata, and interpreted the results.
Limitations
  • Results are associations; possible reverse causality (for example, between income and ICT adoption) means they should not be read as causal effects.
  • Reliance on secondary data brings missing observations and cross-country differences in measurement.
  • Relevant factors such as culture and informal governance are not measured.
  • Heterogeneity across regions limits how far global estimates generalise to individual countries.
  • Fixed effects remove only time-invariant differences between countries.
Status
Completed — MSc thesis (degree awarded 2025)
Keywords
  • ICT
  • digital inclusion
  • educational inequality
  • political competition
  • FDI
  • economic development
  • internet penetration
View Code
Bachelor's Thesis2019Sole author

Opinion Mining Using Naive Bayes Algorithm on Online Accommodation Customer Reviews

Universitas Gadjah Mada · Department of Computer Science and Electronics

Research question

Can an opinion-mining system based on TF–IDF weighting and a Naive Bayes classifier sort accommodation reviews into positive, negative and neutral classes — and how does the handling of stop words affect its performance?

Motivation
Popular listings on accommodation platforms such as Airbnb accumulate more reviews than a guest can reasonably read, which is especially costly for readers who are not fluent in English. Automatically summarising review polarity could help both guests and hosts.
Data
English-language reviews collected manually from Airbnb’s “homes” category for listings in Indonesia, keeping only reviews longer than ten words.
Sample
750 reviews, balanced across classes: 250 positive, 250 negative and 250 neutral. Labels were assigned manually by counting positive and negative opinion words, taking negations into account; reviews with roughly equal counts were labelled neutral.
Methods
  • Preprocessing: punctuation removal, case folding, contraction expansion, tokenisation, stop-word removal and stemming
  • TF–IDF term weighting, implemented directly
  • Three-class Naive Bayes with Laplace smoothing, implemented from scratch in Python 3.7 with NumPy and Pandas (NLTK used for the stop-word list and stemming)
  • 5-fold cross-validation
  • Accuracy, precision, recall and F-measure from multi-class confusion matrices, averaged across folds
  • Experiment comparing the standard NLTK stop-word list with a reduced list that keeps negation-related words (“not”, “no”, “but”, “if”)
Main findings
  • Baseline system averaged 69% accuracy, 70% precision, 70% recall and a 70% F-measure across the five folds.
  • Modified stop-word handling raised the averages to 72% accuracy, 75% precision, 72% recall and a 73% F-measure.
  • Error analysis neutral reviews, which mix positive and negative language, were the hardest to classify correctly.
My contribution
Sole author: collected and labelled the dataset, designed the pipeline, implemented TF–IDF weighting and the Naive Bayes classifier without pre-built machine-learning estimators, and ran the evaluation.
Limitations
  • A small dataset (750 reviews) drawn from Indonesian listings in a single category.
  • Labels come from a manual word-counting rule, so label quality depends on that rule.
  • Only one classifier was evaluated; the thesis recommends comparing alternative methods.
  • Additional signals such as star ratings were not available on the platform at the time.
Status
Completed — BSc thesis (degree awarded 2019)
Keywords
  • opinion mining
  • Naive Bayes
  • TF–IDF
  • text classification
  • customer reviews
Course Research Project2023Co-authored course research project — not a peer-reviewed publication

An Empirical Analysis of Factors Impacting on Political Stability

Corvinus University of Budapest · Course: Data Analysis for World Economics

Research question

Which economic, educational, political and institutional factors are associated with differences in political stability across countries?

Motivation
To examine political stability with the same panel-data tools used for other cross-country outcomes, combining economic, educational and institutional indicators in a single framework.
Data
Country-level panel data from the V-Dem dataset, which also carries the World Bank Worldwide Governance Indicators estimate of political stability used as the outcome.
Sample
Country-level panel covering 202 countries.
Methods
  • Fixed-effects panel regression with year effects
  • Robust standard errors
  • Population weights
  • Step-wise model comparison
  • Separate models by world region
  • Robustness checks
Main findings
  • Positive associations with GDP per capita, domestic autonomy and educational equality.
  • Negative associations with suffrage and with coup attempts.
  • Regional variation the relationships differed across world regions.
My contribution
Contributed to the fixed-effects panel modelling, including the population weighting, robust standard errors, model comparisons, regional analysis and robustness checks. The allocation of other tasks within the group is not documented here.
Status
Completed course project (2023) — not peer reviewed
Keywords
  • political stability
  • panel data
  • fixed effects
  • institutions
  • V-Dem
View Code