RESEARCH & PUBLICATIONS

Understanding difficult data.
Building more reliable AI.

My research takes a data-first view of machine learning: understand the data, characterize its difficulties, transform it where necessary, and evaluate whether the resulting learning system is reliable. Below are my research themes, key projects and the full publication record.

RESEARCH MAP

Difficult datascarcity · skew · noise
Characterizationstructure · overlap · anomalies
Data qualitycleaning · representation
Transformationresampling · synthesis
Reliable MLevaluation · robustness
AI automationgenerative · agentic

RESEARCH EVOLUTION

From early experiments to data-centric & reliable ML.

A research trajectory shaped by a recurring question: how can better understanding and transformation of data improve machine learning?

2019–2021
FOUNDATIONS

Behavioral biometrics · synthetic data · computer vision

Early work explored keystroke dynamics, data generation and supervised computer-vision problems, building a foundation in applied machine learning and data preparation.

2021–2024
DATA-LEVEL LEARNING

Class imbalance · resampling · GANs · dataset characterization

The research focus moved toward difficult class distributions, synthetic minority data, directed undersampling and advanced resampling strategies.

2024–2026
RELIABILITY

Advanced resampling · anomaly detection · intrinsic evaluation

Recent work extends the data-centric perspective beyond class imbalance toward anomaly separation, overlap and evaluation when reliable labels are unavailable.

NOW
CURRENT RESEARCH IDENTITY

Data-Centric & Reliable Machine Learning

The current programme connects data characterization, data quality, imbalance, synthetic data, anomaly detection and reliable evaluation into a coherent data-first research direction.

EMERGING
FUTURE DIRECTION

Generative AI · Agentic AI · automated data-centric ML

An emerging direction is to use generative and agentic systems to diagnose datasets, identify difficult regions, select data transformations, generate and validate synthetic data, and support end-to-end ML workflows.

01 · DATA-CENTRIC AI

Make the dataset part of the model design.

Data-centric machine learning asks what can be learned by understanding and improving the dataset itself. My work connects dataset structure, class distributions, overlap, preprocessing and representation with downstream learning behavior.

02 · SYNTHETIC DATA

Generate data that is useful, not merely plausible.

I have investigated synthetic data generation for biomedical and imbalanced classification problems, including GAN-based tabular generation and adaptive sample generation around minority regions.

03 · RELIABLE ML

Evaluate learning systems under imperfect assumptions.

Reliability in my research includes appropriate evaluation, robustness to difficult distributions, and understanding when conventional metrics or labels are insufficient.

04 · IMBALANCED LEARNING

Learn from rare and underrepresented patterns.

My doctoral research focused on advanced data-level solutions for class imbalance, including adaptive oversampling, directed undersampling, minority-region discovery and synthetic sample generation.

05 · ANOMALY DETECTION

Measure anomalies when labels are missing.

The ASOI work extends this data-centric perspective to unsupervised anomaly detection by introducing intrinsic measures of separation and distributional overlap that do not require ground-truth anomaly labels.

06 · EMERGING

Generative & agentic AI for data-centric workflows.

This is an emerging direction: using generative and agentic AI to automate dataset diagnosis, data-quality analysis, synthetic-data generation, model development and evaluation. It is a future research direction rather than a claim of an established project.

PROJECTS & RESEARCH PROGRAMMES

Key research activities and outcomes.

COMPLETED · DOCTORAL RESEARCH

Class Imbalance Learning: Enhancing Classification Through Advanced Resampling Solutions

Doctoral research focused on improving classification under class imbalance through advanced data-level methods, including resampling, synthetic data generation, minority-region identification and evaluation.

ELTE · Data Science & EngineeringRole: PhD researcherDissertation record ↗
RESEARCH ACTIVITY

Energy-aware network orchestration & simulation

Research activity connected to Telekom Innovation Laboratories at ELTE, involving predictive and energy-aware orchestration of network Point-of-Presence resources and simulation of resource utilization and power behavior.

Telekom Innovation Laboratories / ELTERole: researcher
COMPLETED · METHOD

AROSS — Adaptive cluster-based oversampling

AROSS combines clustering, representative areas, safe/half-safe region identification and synthetic generation to address difficult minority distributions in imbalanced learning.

Publication: Intelligent Systems with Applications, 2024DOI ↗
COMPLETED · METHOD

ASOI — intrinsic evaluation for anomaly detection

ASOI introduces intrinsic evaluation metrics for unsupervised anomaly detection, combining separation and distributional overlap to assess detected anomalies without ground-truth labels.

Complex & Intelligent Systems, 2026DOI ↗
ACTIVE RESEARCH CONTRIBUTION

IslamTrust benchmark

A multilingual benchmark for evaluating LLM alignment with Islamic values, developed with colleagues at ELTE and an external collaborator. The project includes open code and a dataset.

MusIML Workshop · NeurIPS 2025Paper ↗Code ↗
EMERGING

Agentic Data-Centric AI

An emerging research direction exploring how AI agents could diagnose datasets, identify difficult regions, select data transformations, generate and validate synthetic data, and evaluate downstream models.

Research direction · not presented as a funded project

RESEARCH RECORD

Publications

Use the filters to explore the research trajectory by year or theme. Each entry links to an available publisher, DOI, proceedings or project record.

2019–2026research period
10+publications
6research themes

BIBLIOGRAPHY

Publications

This is a curated website record rather than a claim of completeness. Publication metadata is progressively reconciled against publisher, DOI, institutional and bibliographic records. Use the linked publisher/DOI record as the authoritative source for citation details.