Dimitrios Tomaras is an Assistant Professor in the Department of Computer Science at University of Nicosia in Athens. Previously, he was a Post-Doctoral Fellow at the Athens University of Economics and Business. He holds a Ph.D. (2023) and M.Sc. (2014) in Computer Science from AUEB, and a B.Sc. (2012) in Computer Science from the National and Kapodistrian University of Athens. He has held a visiting research appointment at the Competence Center Machine Learning Rhine-Ruhr (currently Lamarr Institute), TU-Dortmund, Germany. His h-index is 7, he holds 114 citations, and has an Erdős number of 3. He has participated in 4 projects funded by the European Commission.

His primary research interests include Data Analysis in Computer Systems and Sensor Networks, particularly Mobile and Sensor Networks; Big Data Management in Parallel and Distributed Architectures; Big Data Analytics; Spatio-Temporal Data Management; Serverless Computing; Machine Learning on Heterogeneous Infrastructures; Crowd Data Management; Data Privacy Management; Explainability and Fairness in AI.

Biography » CV (Web) » CV (PDF) » Research Statement » Teaching Statement »

Publications

h-index: 7   Citations: 126   Erdős #: 3   ORCID   DBLP

• Euro-Par '26 "RAGNN: A Resource-Aware System for Graph Neural Network Training at Scale", Dadamis, P., Tomaras, D., Kalogeraki V., Gunopulos D., Proceedings of the 32nd International European Conference on Parallel and Distributed Computing, Pisa, Italy, August 24–28, 2026. ▼ Abstract
  Add
  image

  Graph Neural Networks (GNNs) have attracted significant attention for supporting a wide variety of graph-based applications, including recommender systems, drug discovery and traffic forecasting. However, training GNNs at scale remains challenging, as real-world graphs impose substantial computational and memory demands while exhibiting irregular access patterns that limit hardware efficiency. The limited memory capacity of the GPUs despite their high computational throughput, often prevents large graphs from fitting on a single device. Consequently, CPU-based distributed training is commonly adopted to leverage aggregate DRAM capacity. However, this introduces additional challenges related to graph partitioning, workload imbalance and inter-node communication, which increase both runtime complexity and energy consumption. In this work, we propose RAGNN, a Resource-Aware framework for Graph Neural Network training at scale, which addresses the aforementioned challenges. RAGNN partitions the input graph into smaller subgraphs for distributed execution, enabling scheduling decisions that explicitly balance runtime performance and energy efficiency. Our extensive experiments on real-world datasets, GNN models and sampling strategies demonstrate that our approach outperforms METIS-based partitioning by up to 67.61%.
• AAAI '26 "GLANCE: Global Actions in a Nutshell for Counterfactual Explainability", Loukas Kavouras, Eleni Psaroudaki, Konstantinos Tsopelas, Dimitrios Rontogiannis, Nikolaos Theologitis, Dimitris Sacharidis, Giorgos Giannopoulos, Dimitrios Tomaras, Kleopatra Markou, Dimitrios Gunopulos, Dimitris Fotakis, Ioannis Emiris, Proceedings of the 40th Annual AAAI Conference on Artificial Intelligence, Singapore, January 20–27, 2026. ▼ Abstract
  Add
  image

  As ML models are widely deployed in high-stakes decisions, GLANCE addresses the need for globally interpretable recourse by computing a small set of counterfactual actions applicable to large population subgroups. The key challenge is balancing three competing objectives: maximising the fraction of individuals covered (effectiveness), minimising the cost of required changes, and keeping the number of actions small enough to remain interpretable. Two algorithms are proposed—C-GLANCE (clustering-based) and T-GLANCE (tree-based)—and extensive experiments show consistently superior performance over prior methods across diverse datasets.
• IEEE BigData '25 "Trustworthy Scheduling for Big Data Applications", Dimitrios Tomaras, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 13th IEEE International Conference on Big Data, IEEE, Macau SAR, China, December 8–11, 2025. ▼ Abstract
  Add
  image

  This paper investigates trustworthy scheduling for big data workloads on heterogeneous cloud and edge infrastructures, where fairness and transparency in resource allocation are first-class requirements alongside performance. The work introduces scheduling policies that offer verifiable guarantees on job execution order and resource share, enabling operators to audit and explain scheduling decisions. Evaluations on real-world big data benchmarks demonstrate that trustworthy scheduling can be achieved without significant throughput penalties.
• ICDCS '25 "AutoFairML: An Automated Middleware for Fairness Auditing in Real-world AI Pipelines", Dimitrios Tomaras, Vana Kalogeraki, Christos Varytimidis, Evangelos Anagnostopoulos, Rahul Nair, Elizabeth M. Daly, Jakub Marecek, Dimitrios Gunopulos, Proceedings of the 45th IEEE International Conference on Distributed Computing Systems, IEEE, Glasgow, Scotland, UK, July 20–23, 2025. ▼ Abstract
  Add
  image

  AutoFairML is a middleware system that automates the continuous fairness auditing of AI pipelines deployed in production environments. It monitors model predictions in real time, detects fairness violations across protected attributes, and triggers configurable mitigation actions without requiring manual intervention. The system is validated on real-world datasets from financial and social domains, demonstrating low latency overhead while maintaining effective fairness guarantees.
• DEBS '25 "IgNITE: Scheduling pipeline-parallel DNN training jobs on heterogeneous infrastructures", Phivos Dadamis, Dimitrios Tomaras, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 19th ACM International Conference on Distributed and Event-based Systems, ACM, Gothenburg, Sweden, June 10–13, 2025. ▼ Abstract
  Add
  image

  IgNITE addresses the scheduling of pipeline-parallel deep neural network training across heterogeneous GPU/CPU infrastructures, where different pipeline stages have vastly different compute requirements. The system models the training job as a directed acyclic graph of stages and uses a cost-aware scheduler to assign stages to heterogeneous nodes while minimising end-to-end training time. Experiments on real DNN architectures show significant reductions in training time compared to homogeneous and naive heterogeneous baselines.
• Middleware '24 "STRATA: Random Forests going Serverless", Dimitrios Tomaras, Sebastian Buschjäger, Vana Kalogeraki, Katharina Morik, Dimitrios Gunopulos, Proceedings of the Middleware Conference, ACM, Hong Kong, December 2–6, 2024. ▼ Abstract
  Add
  image

  STRATA deploys Random Forest inference as a serverless Function-as-a-Service application, exploiting the elasticity of cloud functions to serve fluctuating prediction workloads at low cost. The framework decomposes a forest into independent tree-groups that can be invoked in parallel, and includes an adaptive batching policy that dynamically sizes invocations to meet latency SLOs. Evaluations against container- and VM-based baselines show that STRATA achieves comparable accuracy with substantially lower cost under bursty traffic.
• IEEE MDM '24 "TIMBER: On supporting data pipelines in Mobile Cloud Environments", Dimitrios Tomaras, Michail Tsenos, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 25th IEEE International Conference on Mobile Data Management (MDM'24), IEEE Computer Society, Brussels, Belgium, June 24–27, 2024. ▼ Abstract
  Add
  image

  TIMBER tackles the challenge of executing end-to-end data processing pipelines in Mobile Edge Cloud (MEC) infrastructures, where IoT sensors and smart-city applications generate continuous streams of raw data near the network edge. The system models pipelines as graphs of serverless functions and uses a graph-edit-distance estimator to provision the right number of containers per pipeline stage, satisfying latency SLOs under dynamic workload patterns. Experiments on a real Apache Mesos cluster demonstrate that TIMBER achieves high SLO compliance with efficient resource utilisation.
• IEEE MDM '24 "On Urban Data Analytics and Applications in the Big Data Era", Dimitrios Tomaras, Proceedings of the 25th IEEE MDM'24 Doctoral Symposium, IEEE Computer Society, Brussels, Belgium, June 24–27, 2024. ▼ Abstract
  Add
  image

  This doctoral symposium paper surveys the author's research programme on urban data analytics, covering spatio-temporal mobility modelling, privacy-preserving mobile cloud computing, serverless ML inference, and AI fairness. It situates these contributions within the broader challenges of Big Data management in distributed and mobile environments, highlighting the unifying theme of bridging scalable system design with data-driven urban intelligence. The paper also outlines future directions in energy-aware scheduling and federated learning for sensor networks.
• FAIR'24@ICDE "A Framework for Feasible Counterfactual Exploration incorporating Causality, Sparsity and Density", Kleopatra Markou, Dimitrios Tomaras, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 40th IEEE ICDE'24, IEEE Computer Society, Utrecht, Netherlands, May 13–17, 2024. ▼ Abstract
  Add
  image

  This work presents a framework for generating counterfactual explanations that are not only algorithmically valid but also causally consistent, sparse, and located in high-density data regions—making them realistic and actionable for end users. The framework integrates a causal graph to enforce cause-effect constraints, a sparsity penalty to minimise the number of attribute changes, and a density criterion to ensure the suggested counterfactual lies in a plausible region of the feature space. Empirical results show that the combined constraints yield explanations that are significantly more feasible in practice than those produced by unconstrained methods.
• FAIR'24@ICDE "Pricefair: On fair scheduling of heterogeneous resources", Aristotelis Peri, Dimitrios Tomaras, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 40th IEEE ICDE'24, IEEE Computer Society, Utrecht, Netherlands, May 13–17, 2024. ▼ Abstract
  Add
  image

  Pricefair addresses the problem of fair resource allocation in shared heterogeneous computing clusters, where jobs compete for CPUs, GPUs, and memory with different price-performance characteristics. The paper proposes a pricing-based fairness mechanism that assigns shadow prices to resources to disincentivise greedy consumption, while guaranteeing that each tenant receives a share proportional to its entitlement over time. Experiments on mixed CPU/GPU workloads show that Pricefair reduces unfairness metrics significantly compared to dominant-resource fairness baselines.
• ACSOS '23 "Prediction-driven resource provisioning for serverless container runtimes", Dimitrios Tomaras, Michail Tsenos, Vana Kalogeraki, IEEE International Conference on Autonomic Computing and Self-Organizing Systems (ACSOS), pp. 1–6, doi:10.1109/ACSOS58161.2023.00033, 2023. ▼ Abstract
  Add
  image

  This paper presents SLOPE, a prediction framework for serverless Function-as-a-Service (FaaS) platforms that estimates the number of container instances required to meet application SLOs under dynamic request patterns. SLOPE trains a neural network on historical invocation traces to forecast future load, and falls back to a graph-edit-distance similarity approach when prior knowledge is absent. Evaluations on real FaaS workloads demonstrate that SLOPE reduces SLO violations and container over-provisioning compared to reactive autoscaling policies.
• Elsevier Information Systems '23 "A holistic approach for modeling and predicting bike demand", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, Elsevier, Information Systems, Volume 111, pp. 102–129, doi:10.1016/j.is.2022.102129, 2023. ▼ Abstract
  Add
  image

  This journal paper presents a comprehensive demand prediction model for urban bike-sharing systems that jointly captures temporal patterns, spatial context, and exogenous factors such as weather and points of interest. The model combines time-series forecasting with spatial graph embeddings derived from the station network, enabling accurate short- and medium-term predictions across an entire city. Extensive evaluation on real data from two bike-sharing systems demonstrates improvements over prior single-modality baselines in both accuracy and generalisation.
• IEEE MDM '22 "Practical Privacy Preservation in a Mobile Cloud Environment", Dimitrios Tomaras, Michail Tsenos, Vana Kalogeraki, Proceedings of the 23rd IEEE MDM'22, IEEE Computer Society, Paphos, Cyprus, June 6–9, 2022. ▼ Abstract
  Add
  image

  This paper addresses privacy preservation for mobile cloud applications where user data is processed across edge nodes with limited computational resources and intermittent connectivity. The proposed framework applies lightweight anonymisation and perturbation techniques at the edge before data is uploaded to the cloud, preserving query expressiveness while bounding privacy leakage. Experiments on real mobile traces show that the framework provides strong privacy guarantees with minimal impact on query accuracy and latency.
• IEEE MDM'22 (Demo) "A Framework for Supporting Privacy Preservation Functions in a Mobile Cloud Environment", Dimitrios Tomaras, Michail Tsenos, Vana Kalogeraki, Proceedings of the 23rd IEEE MDM'22, IEEE Computer Society, Paphos, Cyprus, June 6–9, 2022. ▼ Abstract
  Add
  image

  This demonstration paper presents the system prototype implementing the privacy preservation framework for mobile cloud environments, showcasing an end-to-end pipeline from data collection on mobile devices to privacy-protected storage and querying in the cloud. The demo illustrates how different anonymisation functions (k-anonymity, differential privacy, data perturbation) can be configured and composed within the framework via a live interactive interface. Attendees can observe the trade-offs between privacy budget, query accuracy, and processing latency in real time.
• IEEE MDM'18 "Crowd-Based Ecofriendly Trip Planning", Dimitrios Tomaras, Vana Kalogeraki, Thomas Liebig, Dimitrios Gunopulos, 19th IEEE International Conference on Mobile Data Management (MDM'18), Aalborg, Denmark, June 25–28, 2018. ▼ Abstract
  Add
  image

  This paper proposes a crowd-sourced trip planning system that recommends eco-friendly urban routes by leveraging real-time mobility data contributed by city users. The system aggregates crowd signals—vehicle speeds, congestion reports, and emission proxies—to build a dynamic environmental cost graph over the city road network, and uses it to compute routes that minimise carbon impact alongside travel time. Evaluations on a real metropolitan dataset demonstrate measurable reductions in estimated emissions compared to shortest-path and fastest-route baselines.
• IEEE PerCom'18 "Modeling and Predicting Bike Demand in Large City Situations", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, IEEE International Conference on Pervasive Computing and Communications (PerCom'18), Athens, Greece, March 19–23, 2018. ▼ Abstract
  Add
  image

  This paper presents an early demand prediction model for large-scale bike-sharing systems that captures both temporal periodicity and spatial dependencies among docking stations. The approach incorporates contextual features—time of day, day of week, weather, and neighbourhood characteristics—into a regression model trained on historical trip data from a major metropolitan system. Results show that spatial context significantly improves prediction accuracy over purely temporal baselines, particularly for stations near transit hubs.
• IEEE WoWMoM'18 "Evaluating the Health State of Urban Areas Using Multi-source Heterogeneous Data", Dimitrios Tomaras, Vana Kalogeraki, Nikolas Zygouras, Nikolaos Panagiotou, Dimitrios Gunopulos, 19th IEEE International Symposium WoWMoM'18, Chania, Greece, June 12–15, 2018. ▼ Abstract
  Add
  image

  This paper develops a methodology for assessing the environmental health state of urban neighbourhoods by fusing heterogeneous data streams including air-quality sensors, social media activity, mobility traces, and points of interest. A multi-source data fusion model is proposed that assigns health scores to city zones at fine spatial and temporal granularity, supporting urban planners and citizens in identifying pollution hotspots and high-risk areas. Experiments on data from a real European city demonstrate the system's ability to detect known pollution events with high precision.
• LocalRec@GIS'17 "An efficient technique for event location identification using multiple sources of urban data", Dimitrios Tomaras, Vana Kalogeraki, Nikolas Zygouras, Dimitrios Gunopulos, Proceedings of the 1st ACM SIGSPATIAL Workshop on Recommendations (LocalRec@GIS'17), Redondo Beach, CA, USA, November 7–10, 2017. ▼ Abstract
  Add
  image

  This paper proposes an efficient technique for identifying the precise geographic location of urban events—such as concerts, accidents, or demonstrations—by cross-referencing signals from social media, mobility data, and point-of-interest databases. The method formulates location identification as an optimisation problem that ranks candidate locations by a combined evidence score derived from the heterogeneous sources. Evaluations on real city events demonstrate high accuracy and low latency, making the approach suitable for near-real-time urban awareness applications.
• PETRA'17 "Lessons Learnt from the analysis of a bike sharing system", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, Proceedings of the 10th International Conference on PErvasive Technologies Related to Assistive Environments (PETRA'17), ACM, 2017. ▼ Abstract
  Click
  to zoom

  This paper presents an in-depth analysis of a large bike-sharing system's operational data, surfacing patterns in station imbalance, usage seasonality, and user trip behaviour across different city zones. The analysis yields actionable insights for system operators—such as optimal rebalancing schedules and high-demand station clusters—demonstrating the value of data-driven management over heuristic approaches. The lessons learned directly motivated the subsequent demand prediction work published at PerCom 2018 and the Information Systems journal.
• GIS'16 "LOCAl: a personalized cache mechanism for location-based social networks", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, Dimitrios Gunopulos, Proceedings of the 24th ACM SIGSPATIAL GIS'16, Burlingame, CA, USA, October 31–November 3, 2016. ▼ Abstract
  Add
  image

  LOCAl is a personalised caching mechanism for location-based social networks (LBSNs) that reduces repeated server queries by maintaining a user-specific cache of frequently accessed geo-tagged content. The cache replacement policy combines user movement predictions with content popularity to pre-fetch items likely to be requested in the near future, adapting to changing mobility patterns over time. Experiments on real LBSN check-in data show significant reductions in cache miss rates and server load compared to standard LRU and popularity-based caching schemes.
• PerComWorkshops'15 "Travel time estimation in real-time using buses as speed probes", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, International Conference on Pervasive Computing and Communication Workshops (IQ2S'15), IEEE Computer Society, St. Louis, MO, USA, March 23–27, 2015. ▼ Abstract
  Add
  image

  This paper proposes using public bus GPS traces as low-cost speed probes to estimate real-time travel times on urban road segments, exploiting the dense and regular coverage buses provide across city networks. A spatio-temporal interpolation method maps bus speed readings to road segments and fuses them with historical traffic patterns to produce fine-grained estimates even on non-bus-covered roads. Evaluations in a real city show that bus-probe estimates closely match ground-truth travel times recorded by dedicated sensors, at a fraction of the infrastructure cost.
• PETRA '14 "Towards real-time emergency response using crowdsourcing", Dimitrios Tomaras, Ioannis Boutsis, Vana Kalogeraki, Proceedings of the 7th International Conference on PErvasive Technologies Related to Assistive Environments (PETRA'14), ACM, Island of Rhodes, Greece, May 27–30, 2014. ▼ Abstract
  Add
  image

  This early paper explores how crowd-sourced reports from mobile devices can be harnessed to accelerate emergency response in urban areas, enabling faster situational awareness than traditional infrastructure-based channels. A mobile platform is presented that collects, filters, and aggregates citizen reports in real time, distinguishing credible emergency signals from noise using a lightweight trust and relevance model. Experiments simulating disaster scenarios demonstrate that the crowdsourced system reduces the time to first responder notification compared to traditional 911-style reporting pipelines.

Projects & Software

Conference Service More »

• Euro-Par'26 Technical Program Committee, Euro-Par 2026 — 32nd International European Conference on Parallel and Distributed Computing, Pisa, Italy, August 24–28, 2026.
• MDM'26 Local Organization Chair, IEEE MDM 2026 — 27th IEEE International Conference on Mobile Data Management, Athens, Greece, June 29–July 2, 2026.
• MDM'26 Technical Program Committee, IEEE MDM 2026 — 27th IEEE International Conference on Mobile Data Management, Athens, Greece, June 29–July 2, 2026.
• Mobiquitous'24 Technical Program Committee, EAI MobiQuitous 2024, Oslo, Norway, November 12–15, 2024.
• MDM'24 Session Chair (ML & Analytics II — Spatio-Temporal Analysis), 25th Conference on Mobile Data Management, Brussels, Belgium, June 24–27, 2024.
• FAIR'24@ICDE Co-Chair, 1st International Workshop on Fairness in AI @ ICDE, Utrecht, Netherlands, May 13, 2024.
• EuroSys'24 Publicity Chair, 19th European Conference on Computer Systems, Athens, Greece, April 22–26, 2024.
• EuroSys'24 Doctoral Workshop Session Chair, 19th European Conference on Computer Systems, Athens, Greece, April 22–26, 2024.
• Mobiquitous'23 Technical Program Committee, EAI MobiQuitous 2023, Melbourne, Australia, November 14–17, 2023.

Grants More »

• AUEB'24 AUEB Travel Grant, for Middleware Conference, Hong Kong, China, December 2–6, 2024.
• GEC'22 GEC'22 Travel Grant, 4th Summit on Gender Equality in Computing, Aristotle University Research Dissemination Center, Thessaloniki, Greece, June 17–18, 2022.
• ACSOS'21 ACSOS'21 Student Registration Grant, 2nd IEEE International Conference on Autonomic Computing and Self-Organizing Systems, virtual, 2021.

Teaching More » Teaching Statement (PDF)

• Graduate Distributed Systems, Big Data and Cloud Computing, MSc in Computer Science — Guest Lectures
• Graduate Digital Infrastructure Technologies, MSc in Information Systems Development and Security — Guest Lectures
• Undergraduate Distributed Systems (Spring) — Teaching Assistant / Instructor
• Undergraduate Compilers (Fall) — Teaching Assistant

Random Links

• SmartStudent SmartStudent for Distributed Systems Course.
• Github I support open-source code development.
• Music My favorite hobby besides learning :-)

Contact