Members of the laboratory supervise undergraduate theses in the subject areas of the laboratory. Interested students may contact the supervisor directly, or the laboratory at isdlab@ionio.gr.
Proposed topics
The topics below come from the research of the lab’s members and draw on its published work, software and datasets. The scope and method of each topic are adapted to the student’s interests and background in discussion with the supervisor.
Professor Vasileios Karyotis
Link prediction in complex networks with similarity indices and machine learning
Link prediction asks which connections are missing from a network or are likely to appear next, with uses in social networks, recommender systems and biological interaction networks. The student will implement classic local similarity indices (common neighbours, Adamic-Adar, resource allocation), random-walk methods and a learning-based approach such as node embeddings with a classifier, and compare them on public networks from repositories such as SNAP using AUC and precision@k. The thesis builds on the lab’s work on the Heuristic Custom Similarity Index (HCSI) and on random-walk link prediction.
Traffic matrix estimation from link measurements with deep learning
A traffic matrix records how much traffic flows between every pair of nodes in a network and underpins capacity planning and traffic engineering, but it can rarely be measured directly. Network tomography estimates it from link loads and routing information, an ill-posed inverse problem. The student will use public traffic matrices from backbone networks such as Abilene and GÉANT to compare classical estimators (least squares, tomogravity) with neural models, and study how the estimation error changes with network size and measurement noise. The thesis builds on the lab’s work on generative deep learning and conditional invertible neural networks for traffic matrix estimation.
Community detection in large graphs with hyperbolic embedding and a graph database
Community detection groups the nodes of a graph into densely connected clusters and is a standard first step in analysing social, knowledge and sensor data graphs. The student will build a pipeline that stores a large public graph in a graph database (Neo4j), computes a hyperbolic embedding of its nodes and derives communities from the embedding coordinates, then compares the result with the Louvain and Leiden algorithms in terms of modularity, running time and interpretability of the visualisation. The thesis builds on the lab’s work on scalable community detection via hyperbolic network embedding and graph databases.
Simulating malware propagation in multilayer networks
Malware in cyber-physical systems spreads across interdependent layers, for example a communication network and the physical devices it controls, which flat single-layer models do not capture. The student will develop a Python simulator for epidemic-type models (SIS, SIR) on two-layer networks, compare Monte Carlo results with the predictions of Markov-chain models, and study how topology, inter-layer links and countermeasures such as patching or isolating nodes affect the spread. The thesis builds on the lab’s work on Markov-based malware propagation modelling in multi-layer networks.
A low-cost IoT safety application with off-the-shelf components
Off-the-shelf microcontrollers and sensors make it possible to build useful safety systems at very low cost. The student will design and build an IoT prototype on a platform such as Arduino or ESP32 that detects a hazardous situation from sensor data and alerts the user through a mobile or desktop application, and will evaluate detection reliability, alert latency and energy consumption. Possible applications include presence detection in vehicles, fall detection for older people and indoor air quality in classrooms. The thesis builds on the lab’s SafeCuddle prototype for preventing the forgotten baby syndrome.
Professor Panagiotis Kourouthanassis
Digital transformation of Greek hotels, a data-driven assessment
Hotels adopt digital technologies at very different speeds, and the technological, organisational and environmental (TOE) factors behind these differences remain an open question. The student will build a dataset of hotels in a Greek destination, such as the Ionian Islands, recording the digital services visible on their websites and booking channels (online booking engine, chatbot, mobile app, contactless check-in, social media presence) together with star rating, size and guest review scores, and will analyse the relationships with statistical models. The thesis builds on the lab’s work on the determinants of digital transformation in the hospitality industry and on digital technology implementation and hotel star rating.
AidTech, matching donations to needs in humanitarian aid campaigns
Humanitarian campaigns increasingly require transparency and zero waste, which calls for information systems that go beyond conventional supply chain software. The student will design and implement a web prototype that registers the needs of beneficiary organisations and the offers of donors, matches them with an assignment algorithm (a greedy heuristic and a genetic algorithm) and tracks each donation through a transparency dashboard, and will evaluate the matching on simulated campaigns. The thesis builds on the lab’s AidTech architecture for humanitarian supply chains and its genetic algorithm for the generalized assignment problem with non-linear costs.
An information system for small logistics companies and last-mile delivery
Small and medium courier and logistics companies rarely have access to enterprise systems, while customers expect flexible delivery options such as time windows, parcel lockers and live tracking. The student will develop a web and mobile application covering order registration, route planning with open-source routing software (OSRM, Google OR-Tools), customer delivery preferences and shipment tracking, and will evaluate it on realistic delivery scenarios. The thesis builds on the lab’s framework for a holistic information system for small-medium logistics enterprises and its study of last-mile delivery options.
A mobile augmented reality guide for the Old Town of Corfu
Augmented reality can enrich a visit to a historic town by overlaying information on what the visitor sees, but its adoption depends on emotions and perceived value as much as on technical quality. The student will design and develop a mobile AR application (for example with Unity AR Foundation or ARCore) presenting points of interest in the Old Town of Corfu, a UNESCO World Heritage Site, and will evaluate usability, emotional responses and intention to use through a field study with visitors. The thesis builds on the lab’s work on the design of mobile augmented reality applications and on tourists’ responses to mobile AR travel guides.
Explaining the user experience of mobile applications with fsQCA
Satisfaction with a mobile app rarely has a single cause, and different combinations of usability, enjoyment, trust and emotions can lead to the same outcome. The student will run a questionnaire survey among users of one class of apps (for example food delivery, mobile games or mobile banking), analyse which configurations of factors lead to high satisfaction and continued use with fuzzy-set qualitative comparative analysis (fsQCA), and compare the findings with a regression model. The thesis builds on the lab’s work on user experience in mobile gaming applications and on online shopping behaviour with fsQCA.
Professor Nikolaos Korfiatis
Collection and analysis of unstructured data using RAG methods
Retrieval-augmented generation (RAG) lets a large language model answer questions from a specific document collection rather than only from its general training data. The student will build a RAG system over a collection of unstructured Greek documents, such as study regulations, course outlines or public administration announcements, covering document ingestion and chunking, embedding models and a vector store, retrieval and answer generation, and will evaluate retrieval quality and answer faithfulness on a set of annotated questions. UniBot, a completed thesis of the lab, applied this approach to curriculum guidance.
Measuring service quality from online reviews with topic models
Online reviews on platforms such as TripAdvisor describe in customers’ own words what went well or badly in a service. The student will work with a dataset of reviews for one service sector (museums, hotels or airlines), extract their themes with a classical topic model (LDA) and an embedding-based one (BERTopic), and test whether topic membership improves the prediction of the star rating with gradient boosting, comparing topic coherence and predictive accuracy. The thesis builds on the lab’s work on measuring service quality from airline passengers’ reviews and on topic membership in review rating prediction.
Long-tailed multi-label recognition on the G3 Corfu Butterflies benchmark
Many real image collections are small and heavily imbalanced, with a few common attributes and many rare ones. G3 Corfu Butterflies is the lab’s public benchmark of butterfly photographs from Corfu annotated with colour, pattern and other visual attributes. The student will train and compare baselines on it, such as fine-tuned convolutional networks and vision transformers, loss functions designed for imbalance (focal and asymmetric loss) and zero-shot or few-shot recognition with vision-language models such as CLIP, reporting per-attribute results with emphasis on the rare attributes. The thesis uses the G3 Corfu Butterflies dataset published by the lab on Harvard Dataverse.
Sentiment, on-chain activity and the price of Bitcoin, a Granger causality study
Cryptocurrency prices react both to market sentiment and to activity recorded on the blockchain itself, but the direction of influence between them is not obvious. The student will assemble daily time series of on-chain metrics (transaction counts, fees, transferred value), public sentiment indicators and Bitcoin returns, test lead-lag relationships with Granger causality using the lab’s R package grangersearch, and check whether the relationships found improve out-of-sample forecasts. The thesis builds on the lab’s work on sentiment embedded in blockchain transactional data.
AutoML and tabular foundation models for supply chain prediction
Automated machine learning frameworks and pretrained foundation models for tabular data promise accurate predictions with little manual tuning. The student will compare AutoGluon and TabPFN with carefully tuned gradient boosting models (XGBoost, LightGBM) on public supply chain datasets, for example demand forecasting or late-delivery prediction, in terms of accuracy, training and inference time and explainability, and will identify the data sizes and problem types for which each approach is preferable. The thesis builds on the lab’s work on AutoGluon and TabPFN for supply chain optimization.
Assistant Professor Georgios Dimitrakopoulos
Explainable detection of mental fatigue from EEG signals
Mental fatigue while driving or monitoring systems is a serious safety risk, and EEG makes it possible to detect it from brain activity, provided the models can explain which signals they rely on. The student will use a public EEG dataset recorded during simulated driving (for example SEED-VIG), extract spectral features and functional connectivity measures between brain regions, train sparse (LASSO) and tree-based classifiers, and identify the features and channels that drive the predictions. The thesis builds on the lab’s work on feature-level explainability in EEG-based fatigue detection and on brain connectivity patterns in driving fatigue.
Feature selection for single-cell RNA sequencing data
Single-cell RNA sequencing measures the expression of thousands of genes in each of thousands of cells, and the choice of genes to keep strongly affects every later step of the analysis. The student will compare feature selection methods, such as highly variable genes, minimum redundancy maximum relevance (mRMR), LASSO and random-forest importance, on public single-cell datasets using Python (Scanpy), and measure their effect on cell clustering (ARI, NMI) and cell-type classification. The thesis builds on the lab’s comprehensive evaluation of feature selection in single-cell RNA sequencing data and its application of mRMR to a single-cell dataset.
Gene regulatory network inference with machine learning
Gene regulatory networks describe which transcription factors control which target genes, and inferring them from expression data is a long-standing problem in computational biology. The student will implement tree-based inference methods (in the spirit of GENIE3 and boosted-tree regression), compare them with correlation and mutual-information baselines on the BEELINE benchmark datasets, which come with reference networks, using precision and recall measures, and may add prior knowledge about transcription factors. The thesis builds on the lab’s XGRN method and its work on integrating machine learning and biological context for single-cell network inference.
A graph database of drugs and proteins for drug repurposing
Drug repurposing looks for existing drugs that may act on a different disease, which is faster and cheaper than developing a new compound. The student will integrate open data on drugs, targets and protein interactions (for example ChEMBL, STRING and Open Targets) into a graph database (Neo4j), rank candidate drugs for a disease such as Alzheimer’s disease with network proximity measures and link prediction on graph embeddings, and check the top candidates against the literature. The thesis builds on the lab’s work on drug and protein interaction networks for drug repurposing in Alzheimer’s disease and on graph-based integration of large-scale drug and protein data.
Lip reading with deep learning
Visual speech recognition, or lip reading, recognises spoken words from the movement of the lips alone, with applications in assistive technology and noisy environments. The student will build the full pipeline on a freely available audiovisual corpus such as GRID, from face and mouth-region detection (for example with MediaPipe) to a model that combines 3D convolutional layers with recurrent or transformer layers, evaluate word and sentence accuracy, and develop a small demonstration application. The thesis builds on the lab’s work on 3D CNN and LSTM fusion for lip reading.
In progress
| Student | Title | Supervisor |
|---|---|---|
| Ioannis Katsioulis | Artificial intelligence in primary education, a PRISMA systematic review | Prof. Nikolaos Korfiatis |
| Margarita Soupioni | Maritime Management System, an integrated application for crews, vessels and marinas | Prof. Nikolaos Korfiatis |
| Mohammad-Matin Marzie | Personalized Language Learning Application | Prof. Nikolaos Korfiatis |
Completed
| Year | Student | Title | Supervisor |
|---|---|---|---|
| 2026 | Agni Tatsi | UniBot: An Agent and RAG-Based Framework for Greek University Curriculum Guidance with Multi-Judge Evaluation | Prof. Nikolaos Korfiatis |
| 2026 | Andreas Klimentidis | Cross-Sectional Return Prediction of Greek Equities with Machine Learning | Prof. Nikolaos Korfiatis |
| 2026 | Konstantinos Nikolaos Perros | Ανάλυση Καινοτομίας και της χρήσης Πληροφορικής στον Κλάδο των Υπηρεσιών | Prof. Panagiotis Kourouthanassis |