Data Science and machine learning laboratory
Contact: Dimitrios Rafailidis, email: draf at uth dot gr
Continuous model learning for predictive analytics
The synergy of deep and reinforcement learning strategies has already achieved remarkable performance in several real-world applications such as robotics, self-driving cars, recommendation systems, social media mining tasks and so on. However, the main challenge resides on how to build a model to continuously adapt to the dynamic evolution of an environment/application in real-time. Our group studies continuous deep graph reinforcement learning strategies to generate adaptive models with life-long learning capabilities for predictive analytics.
Traditional data-mining, predictive analytics, and statistics on business data has helped many businesses in strategic and tactical decision making. Differentiating ourselves from the traditional approach, our group aims at incorporating heterogeneous and multivariate data, treating them as streams that change over time. We develop novel online statistical models for time-series predictions, robust ensemble methods and meta-learning strategies. In addition, we develop dynamic graph-based methodologies mapping the data inter-relations in real-time. Our use cases include streaming data sources from sensors, online users’ behaviour, stock price prediction, the blockchain, estimation of energy consumption measurements, and hardware performance metrics.
User modelling and interactive recommendation engines
Current recommendation technologies model users' evolving, diverse and multi-aspect preferences to generate recommendations in various domains/applications, aiming to improve the citizens' daily life by making suggestions. The repertoire of actions now is no longer limited to the one-shot presentation of recommendation lists, which is insufficient for offering decision support for the user. To accurately provide decision support, an interactive mechanism is needed to quickly adapt to the users’ preferences through the interactions. Our group investigates whether adaptive deep learning models can be used to tailor interactive recommendation systems optimally to user preferences. We study ways to adjust and transfer recommendations based on individuals’ preferences across different domains. In addition, we perform continuous learning by self-adapting the neural network’s junctions of various deep architectures, while taking into account how users behave under different conditions. Instead of adapting humans to machines, our aim is to provide users with better recommendation services so that machines will be adapted to humans in their daily lives.
Improving viewer’s experience in video events
Over the last decade, video technologies have become the main communication solution in large enterprises. To maximize the return of investment of each video meeting, optimal video quality of experience and maximum viewer engagement are required. Our group investigates ways to (i) provide services to optimize the meeting’s return of investment, and (ii) achieve the short-term and long-term goals of the company for the employees, that is to establish efficient communication of the employees worldwide and promote the employees’ wellbeing.
Mr. Naoum Mengoudis holds a B.Sc. and a M.Sc. degree in Information Technology and Information Systems from the Aristotle University of Thessaloniki. Since 2004 he has been working in the private sector for various IT companies. His specialization is complex web applications design and development, along with useful tools/projects for automation in real time. Naoum currently works for the Cyber Crime Division of Hellenic Police (since 2010). He has more than a decade of experience in cybercrime issues in general and online children exploitation and safety in particular, as well as "hands-on" investigation experience of several cases that were brought in front of the National Courts. Naoum has attended many trainings in the field of cybercrime and he has also acted as a trainer on behalf of Europol on the topic of investigations of dissemination of child abuse material via peer-to-peer networks.
Mr. Stefanos Antaris works as an Artificial Intelligence Scientist at Hive Streaming AB, Stockholm, Sweden. In the last three years, he was also an industrial PhD candidate at KTH Royal Institute AB, Stockholm, Sweden. His main research focuses on graph neural networks and reinforcement learning on highly evolving graphs. In particular, he is interested in offline (batch) reinforcement learning and how the learned policy can be beneficial to continual reinforcement learning algorithms for graphs that significantly change over time.
Artificial Intelligence and advanced Data Analytics for Law Enforcement Agencies HORIZON2020
Roller: Graph Reinforcement Learning for Enterprise Live Video Streaming Events,
Hive Streaming AB
Live video streaming has become the main communication paradigm in large enterprises. Large enterprises rely on high-quality video content to communicate with their employees around the world. However, distributing video content in Fortune-500 companies with thousands of employees is a challenging task due to network inefficiencies. Roller aims to construct machine learning to optimally distribute the video content in the enterprise network and ensure high user experience. During this project, we combine different research areas, such as graph neural networks and reinforcement learning, to provide novel algorithms that assist large enterprises to organize live video streaming events with similar user experience as in person communication.
Prediction and Visual Intelligence for Security Information
Design and Specifications of an EU-wide Mobile Application for the protection of minors online under EMPACT policy cycle, 2018-2021
Automated Auditing and Reporting (A2R) Project in Military Academy
Cryptocurrencies predictive model
A model that predicts the directional movement of time-series has been developed. It is based on variance minimization techniques. It is data, time-period, and sampling agnostic. It uses statistical tests in the form of bootstrap, jackknife, and it is implemented as an algorithmic trading strategy (short/long signals). The project involved full-stack research and development of market making and spot trading bots, for crypto-currencies markets.
European Organization for Nuclear Research
MediaStream: Highly-scaleble cloud computing framework for multimedia stream processing 2012-2015
CUBRIK: Human-enhanced time-aware multimedia search
I-SEARCH: A unified framework for multimodal content SEARCH ITI-CERTH, FP7
I-SEARCH develops a novel generation of multimodal search engines providing users with natural and expressive interfaces. Additionally, I-SEARCH presents novel solutions for relevance feedback, based on users’ social behaviour and recommendations. The above result in a highly user-centric search engine, able to deliver to the end-users only the content of interest, satisfying their information needs and preferences. I-SEARCH also introduces efficient tools for visualising the search results in order to enhance the presentation layer of search engines. Several aspects, such as user profile, end-user terminal, available network bandwidth, interaction modality preference, are taken into account to achieve the optimal presentation result. Finally, the search engine is dynamically adapted to end-user’s device, which vary from a simple mobile phone to a high-performance PC.
Social Networking Trends and Dynamics Detection via a Cloud-based Framework Design, AUTH, FP7, 2010-2012
As a pilot application of the Venus-C project, Cloud4Trends provides a cloud-based framework to analyze in real-time the geolocated content of the microblogging and blogging platforms, such as Twitter, Blogger, and so on. Cloud4Trends aims to construct a service for researchers and practitioners to detect trending topics and events that are discussed in specified geographic locations. Such a service is beneficial in various domains, such as defining the marketing strategy in certain geographical regions, detecting outburst events, and predicting the users’ opinion. Moreover, the goal of this project is to verify the suitability of the cloud infrastructure to support real-time data analysis services.
SMS marketing services for large enterprises and SMEs
Isosoft (IBC OE)
Virtual Museum over a Sensor Web (iMuse) EEA Financial mechanism
Broadband Weather Representation Services, project ΕΥΡΗΜΑ ΓΓΕΤ
The aim of this project is to create a web based and mobile application that displays weather radar information in almost real time. In particular, two weather radars located close to Thessaloniki and Volos are feeding cloud reflectivity information to a centralized server running the TITAN weather software. TITAN is performing storm and hail predictions in terms of probability and trajectory. The cloud reflectivity data is transformed, merged, and unified in a database. Afterwards, a novel image synthesis application is feeding animated images to a web and mobile application.
An interactive Environment for Automated Design of Elevators, ΠΑΒΕΤ-2005, ΓΓΕΤ
The main activity of the project is to conduct Industrial research for the design and development of an interactive environment for a fully automated elevator design. Research is carried out on the design and development of an innovative environment for interactive data entry (mainly dimensional) and fully automated design (using an existing CAD system, but without human intervention). The objectives of the project are: The elimination of time-consuming stages in the execution of an order, the drastic reduction of the cost of ordering, the upgrading of the "image" of the company to potential customers, the minimization of design errors, the reduction of particularly time-consuming and unnecessary of the production process, and ultimately the improvement of the productivity and competitiveness of the company.
T.ARCH.H.N.A. Towards ARCheological Heritage New Accessibility, CULTURE 2000, EU 2004-2007
ML-based chess engine and end-game analysis
A Fast Bitboard engine has been developed in C which was used to produced chess game positions and then transition graphs. The produced graphs are represented as a kernel space and the Laplacian eigenmap technique has been applied. The low dimensional representation of the graphs is then visualized by coloring each position by the distance to mate/draw/stalemate. The results indicate that there might be a continuous smooth gradient to final states by following an appropriate vector in the represented space. In other words, instead of searching the game tree, one may sample the tree, produce a low dimensional representation and then select moves in the gradient’s direction that may lead to the desired outcome.