Announcements

We are honored to announce that Dr. Aristeidis Karras has been selected as the winner of the Future Internet Best Paper Award 2024.

The following is an interview with Dr. Aristeidis Karras:

1. Congratulations on winning the Future Internet 2024 Best Paper Award! Could you please briefly introduce yourself and your research background?
Thank you very much. It is a great honor for me and for all the authors to receive this award.
My research focuses on efficient Big Data management in large-scale Internet of Things systems. More specifically, I work in areas including distributed and edge computing, TinyML, machine learning, data engineering, and intelligent processing on resource-constrained devices.
A central question in my research is how we can process data closer to where it is generated. In a large IoT environment, transmitting every raw measurement to a central server is often inefficient. It increases communication overhead, latency, and storage requirements. Edge computing and TinyML provide opportunities to clean, analyze, compress, organize, and prioritize data locally before it is transferred to a central infrastructure.
I would also like to emphasize that this award reflects the collective work of the entire research team. All co-authors contributed to the study's conception, experimental work, result analysis, and manuscript preparation.

2. Could you give a brief overview of the key findings of this award-winning paper?
The main purpose of the paper was to examine whether TinyML could support not only machine learning inference but also broader data-management operations in large-scale IoT systems.
We proposed five algorithms, each addressing a different part of the IoT data-management process. TinyCleanEDF uses federated learning for distributed data cleaning and anomaly detection and employs autoencoders for feature extraction. EdgeClusterML combines reinforcement learning with self-organizing maps for adaptive clustering at the edge. CompressEdgeML performs neural-network-based adaptive data compression. CacheEdgeML uses predictive analytics and tiered caching to improve access to frequently requested data. Finally, TinyHybridSenseQ assesses data quality and decides whether data should be stored locally, in the cloud, or in a central database according to their quality and priority.
The algorithms were evaluated using different configurations ranging from one to ten Raspberry Pi devices and a heterogeneous collection of sensor data. The dataset included environmental, motion, orientation, light, distance, and soil-moisture measurements and contained more than one terabyte of raw data.
The results showed that TinyML can support several important data-management functions directly at the edge. EdgeClusterML achieved approximately 90% clustering accuracy. CompressEdgeML maintained data integrity above 90% while reducing compression time from approximately 1200 milliseconds with one device to around 300 milliseconds with ten devices. CacheEdgeML increased its cache hit rate from 85% to 92%. TinyHybridSenseQ maintained a data-quality score between 90% and 95%, increased storage efficiency from 85% to 91%, and reduced data-transfer latency from 250 milliseconds to 160 milliseconds.
An important observation is that adding more devices did not improve every metric equally. Some measurements, such as compression efficiency and data-quality score, decreased slightly as the system scaled, although they remained at high levels. This illustrates the practical trade-offs involved in distributed edge processing.
Overall, the study demonstrated that TinyML can become an active part of the data-management architecture. It can help decide which data should be cleaned, compressed, cached, prioritized, stored, or transmitted before that data reaches the central Big Data infrastructure.

3. What were the biggest challenges you faced during this research, and how did you overcome them?
One of the main challenges was the diversity and volume of the data. Our experimental environment included several types of sensors producing data with different characteristics and collection frequencies. Managing more than one terabyte of raw sensor data required careful preprocessing, including data cleaning and normalization, before the algorithms could be evaluated consistently.
A second challenge was the distributed experimental environment. We needed to evaluate the algorithms on resource-constrained edge devices while also considering communication with a central computing infrastructure. The system included Raspberry Pi devices, an HPC cluster, local model training, model aggregation, and communication under both Ethernet and Wi-Fi conditions.
Scalability was another important challenge. Demonstrating that an algorithm works on one device is not sufficient for a study concerning large-scale IoT systems. We therefore evaluated the algorithms using one, two, five, and ten devices and examined how accuracy, processing time, communication efficiency, resource utilization, caching, storage, and latency changed as the system expanded.
We addressed these challenges by designing the system in a modular way. Each algorithm was responsible for a clearly defined data-management function. We also used a common architecture connecting the edge devices to the central Big Data system and applied consistent evaluation metrics across the different configurations.
Perhaps the most important lesson was that scalability always involves trade-offs. Some metrics improved significantly as the workload was distributed, while others remained stable or declined slightly. Reporting these trade-offs honestly was essential for presenting a realistic evaluation of the proposed system.

4. What advice would you give to young researchers who aspire to produce high-impact research results?
My first piece would be to begin with a real and clearly defined problem rather than with a fashionable technology. A technology is useful only when it addresses a genuine scientific or practical need.
Researchers should also try to implement and test their ideas whenever possible. A theoretical model may appear effective, but practical experimentation often reveals issues related to hardware limitations, communication overhead, data quality, scalability, and reproducibility. Even a modest prototype can provide valuable scientific evidence.
It is equally important to report limitations and negative results. A strong research paper does not need to claim that every method performs perfectly. In many cases, the most useful result is an honest explanation of the conditions under which a method works well and the situations in which it requires further development.
Young researchers should also be patient. Research involves failed experiments, repeated implementation, manuscript revisions, and critical feedback. These are not exceptions to the research process; they are part of it.
Finally, collaboration is essential. This study combined knowledge from IoT systems, Big Data management, machine learning, edge computing, and embedded platforms. Such work becomes stronger when researchers with different areas of expertise contribute to the same problem.

5. How was your experience with the editorial and peer-review process for Future Internet?
My overall experience with Future Internet was positive. The editorial communication was clear, and the manuscript progressed through the submission, review, revision, and publication stages in an organized manner.
The peer-review process gave us the opportunity to revisit the manuscript carefully and improve the clarity of the methodology, the description of the proposed algorithms, and the presentation of the experimental results. We treated the revision process as an opportunity to strengthen the paper rather than simply as a procedural requirement.
I also appreciated the journal’s willingness to consider interdisciplinary research. The paper brings together IoT, Big Data management, TinyML, machine learning, embedded computing, and cloud-edge coordination. Research of this kind needs reviewers and editors who can appreciate contributions that extend across several technical areas.
Receiving the Best Paper Award has made the publication experience particularly meaningful for the entire author team.

6. Is there anything else you would like to share—perhaps thoughts on how journals like Future Internet can better support researchers, or reflections that have not come up yet?
I believe that journals can further support researchers by encouraging reproducibility and practical validation. In fields such as IoT, TinyML, and Edge AI, it is important to report not only algorithmic results but also hardware configurations, communication conditions, data characteristics, resource limitations, and implementation details.
Journals can also encourage authors to discuss trade-offs more openly. Results that show a limitation or a decline in one metric can still be scientifically valuable when they help researchers understand the behavior of a system under realistic conditions.
Where possible, sharing code, configurations, datasets, and experimental protocols can also help other researchers reproduce and extend published work. Standardized benchmarks for TinyML and distributed edge systems would be particularly valuable because comparisons between studies are often difficult when different hardware platforms, datasets, and metrics are used.
From a research perspective, several important challenges remain. These include more accurate anomaly detection, lower energy consumption, improved cloud-edge integration, secure and privacy-preserving processing, faster analysis of streaming data, greater adaptability to different IoT applications, and improved interoperability between heterogeneous devices and platforms.
Finally, I would like to thank the Future Internet Editorial Office, the Best Paper Awards Committee, the reviewers, and all my co-authors. This award is a significant encouragement for us to continue investigating practical and efficient approaches to intelligent data management in future IoT systems.