A major research direction of our group is graph machine learning, with a particular focus on learning from graphs when node features are absent, incomplete, or unreliable. We develop efficient representation-learning methods that exploit the structural information inherent in graphs, reducing dependence on conventional node features and computationally expensive graph encoders. Our recent work includes semi-supervised and feature-agnostic graph representation learning, combining structural information, available labels, and partial supervision to learn useful node representations efficiently. Building on this foundation, we are expanding our work towards several emerging problems in graph and hypergraph machine learning, including graph condensation, generalized graph anomaly detection, and learning on heterophilic graphs, where connected nodes may have very different properties or labels. We are particularly interested in developing local, scalable, and feature-agnostic learning approaches that can capture diverse structural patterns and remain effective when standard assumptions about graph data do not hold.

We developed Mean Shift Density Enhancement (MSDE), an unsupervised anomaly detection framework that identifies anomalous observations based on their geometric response to iterative density enhancement. Rather than assuming a particular type of anomaly, MSDE examines how observations move towards nearby density modes: well-supported observations tend to remain relatively stable, while anomalous observations exhibit larger displacements. The method was evaluated across 46 real-world tabular datasets, different anomaly-generation mechanisms, and varying levels of noise. We are now extending this framework towards multimodal anomaly detection, with a focus on integrating information across image and graph modalities. This direction aims to develop anomaly detection approaches that can exploit complementary structural and geometric information across modalities, while remaining largely unsupervised and adaptable to different types of anomalous patterns.

Spike-Timing-Dependent Plasticity (STDP) provides a biologically grounded learning rule for spiking neural networks (SNNs), but its reliance on precise spike timing and pairwise updates limits fast learning of weights. We introduce Spike Agreement-Dependent Plasticity (SADP), which replaces pairwise spike-timing comparisons with population-level agreement metrics such as Cohen’s Kappa. The proposed learning rule preserves strict synaptic locality, admits linear-time complexity, and enables efficient learning without backpropagation, surrogate gradients. We are now exploring development of diverse neural network architectures assisted by the SADP learning paradigm. The main goal of this project is to develop fast learning algorithms for low power computation. 

One of our key research directions is context-aware synthetic tabular data generation. We view synthetic data not as a single generic artifact, but as a resource whose notion of quality depends strongly on its intended use. Applications such as classification, privacy-preserving data sharing, patient stratification, or exploratory analysis impose different requirements on what aspects of the data distribution should be preserved. Consequently, effective synthetic data generation must adapt to the downstream context rather than aiming for uniform realism across all settings. Our work focuses on developing generative strategies that explicitly incorporate contextual information such as task objectives, dependency structure among features, and domain-specific constraints into the data generation process.

In this project, we maintain a close association with the group of Prof. Olaf Wolkenhauer, at the University of Rostock, Germany.