Direction: Research Areas

Decision support systems (DSS) have become essential tools for decision-making in critical domains such as healthcare, finance, and defense. These systems provide valuable insights and recommendations that can help responsible personnel make more informed decisions and enhance their situational awareness. However, as Machine Learning models have become more complex, they have become increasingly opaque and difficult for users to understand. This lack of transparency can lead to a lack of trust in the models and their decisions. To address this issue, it is important for the DSS to be based on explainable AI (XAI) techniques. The ML models should become more transparent and interpretable, allowing the users to understand how a decision was yielded. This research direction includes machine learning techniques for analyzing and fusing dynamically heterogeneous information obtained from…

The Semantic Web technologies provide the means and tools to unanimously represent the meaning of Web entities and their relations to one another, in a machine-interpretable manner. While modeling and managing knowledge for web entities have long ago been established, Semantic Web technologies also show great potential in other domains, in both online and offline worlds. Two such prominent applications have to do with representing and managing context, enabling two different flavours of context-aware computing. Context awareness in pervasive computing translates to situation awareness and enables intelligent reasoning and decision making in smart spaces, in real- or close to real-time. Meanwhile, context-of-use alters the very meaning of persistent entities, such as those met in digital preservation and cultural heritage, which stimulates further research questions in topics such as ontology evolution…

Artificial intelligence (AI) technologies, and in particular the deep learning paradigm, are advancing at an astounding pace and appear to have the potential to significantly enhance the capabilities in a wide range of applications. Multimodal analytics aim to uncover the structure underlying the multimodal information of interest and organize it in an effective and efficient manner thus allowing end users to gain insights and actionable intelligence. Our research activities include the development of AI-based methods for multimodal classification, clustering, social network analysis, event and trend detection, and predictive analytics based on deep learning and machine learning techniques, ranging from unsupervised to fully supervised. [caption id="attachment_717" align="aligncenter" width="378"] Multimodal analytics dashboard that enables the extraction of useful insights from large amounts of multimodal information[/caption]

Spatial, textual and visual data are ubiquitous and are massively generated by cameras, satellites, sensors and humans. Clustering large and semi-structured data is a very popular pattern recognition problem in the Big Data era, that aims to support data management and grouping into topics. Our research focus is to identify groups of textual and/or visual content that pose a high semantic relevance within each group and low similarity between different groups of documents, i.e. news articles, short texts, social images, satellite images, etc. The estimation of the number of topics to detect is a computationally complex problem that requires effective parallel programming methods and efficient computational resources, such as cloud and High Performance Computing environments. [caption id="attachment_717" align="aligncenter" width="447"] Topic detection using keyword clustering[/caption]

A key requirement in multimodal domains is the ability to integrate the different pieces of information, to derive high-level interpretations. More precisely, in such environments, information is typically collected from multiple sources and complementary modalities, such as from multimedia streams (e.g. using video analysis and speech recognition), lifestyle and heterogeneous sensors. Though each modality is informative on specific aspects of interest, the individual pieces of information themselves are not capable of delineating complex situations. Combined pieces of information on the other hand can plausibly describe the semantics of situations, facilitating intelligent situation awareness. Similarly, social media posts usually contain short text, image captured by a mobile phone, user information, spatial and temporal details that all formulate a social media post. Satellite images oftenwise contain several bands, beyond the optical channels…

Digital images and video footages comprise one of the core data types for human-machine interaction and visual understanding in general. Human perception along with human logic is the most attractive capabilities of the human species that the research community strives to mimic. Except for the relevant perception models, the community focused also on pioneer computer vision models that could extract additional information from visual representations towards facilitating various human activities. Object recognition, face detection, pedestrian detection, style transfer and weed detection are some research fields amongst a variety of applications that can be exploited to accomplish specific objectives. Our research activities display a wide diversity of computer vision objectives in order to cover a significant number of applications and extend the state-of-the-art in these research fields. [caption id="attachment_717" align="aligncenter" width="736"]…

Big Data Collections (e.g. Earth Observation data, social media streams, large video collections, etc.) are characterized by large volumes, velocity and variety, due to their highly heterogeneous nature and timeliness. Multimedia retrieval requires efficient indexing of the available raw data and extracted metadata, such as concepts, visual features, textual features and mutually semantically linked multimodal objects. The multimodal character of Big Multimedia Data requires an effective combination of multiple modalities for similarity search and retrieval using compound queries. Our research focuses on the design and development of domain-specific and focused search engines with novel user interfaces, taking into account the AI-based extracted and indexed metadata per multimedia object. [caption id="attachment_717" align="aligncenter" width="1161"] User interface of the VERGE interactive video search engine[/caption]

The large-scale availability of user-generated content on social media platforms has opened up new possibilities for studying and understanding real-world phenomena, trends and events. Social media is a highly heterogeneous source of data that provides real-time information in specific areas of interest, which can assist public safety through the monitoring of crisis events with spatial, temporal and contextual multimodal information. Combining the discovery of social media insights with knowledge retrieved by focused crawling from traditional web sources, our research offers timely in-situ information about changes, concepts and events of interest to specific communities. [caption id="attachment_717" align="aligncenter" width="552"] Social media dashboard for visualizing, filtering, and clustering Twitter posts into events. (EOPEN EU project)[/caption] [caption id="attachment_717" align="aligncenter" width="646"] Fluctuations in the number of tweets containing flood-related keywords from 2017 to 2019, grouped…