Tutorials

Scope and Aims

The conference includes tutorials, which aim to inform undergraduate and postgraduate students about the current state of affairs of AI research as conducted by scientists in Greece and worldwide. The organizers have selected the following tutorials, listed here by title and abstract.

Can computers understand what is happening?

Tutors: Alexander Artikis, George Paliouras & Anastasios Skarlatidis, NCSR Demokritos

Abstract:
Today's organizations collect data in various structured and unstructured digital formats, but they cannot fully utilize these data to support their resource management process. It is evident that the analysis and interpretation of the collected data needs to be automated, in order for large data volumes to be transformed into operational knowledge. Events are particularly important pieces of knowledge, as they represent activities of special significance within an organisation. Therefore, the /recognition of events/ is of outmost importance. Consider, for example, the recognition of attacks on nodes of a computer network given the exchanged TCP/IP messages, the recognition of suspicious trader behaviour given the transactions in a financial market, and the recognition of various types of cardiac arrhythmia given electrocardiographs. In this tutorial we will present the state-of-the-art on event recognition. We will present languages for expressing events and their structures, algorithms for efficient event recognition in the presence of large amounts of data, and machine learning techniques for automatically developing knowledge bases of event structures. We will also demonstrate the use of event recognition techniques in real-world applications, such as public transport management and the management of emergency rescue operations.

N-gram Graphs: A generic machine learning tool in the arsenal of NLP, Video Analysis and Adaptive Systems

Tutor: George Giannakopoulos, Trento University

Abstract:
The tutorial aims to acquaint the audience to the n-gram graph framework, which is a generic machine learning representation and a corresponding kit that goes beyond the feature vector representation. It manages to capture the semantics of proximity (or relation), regardless of domain, offering the set of algorithms and methodologies to use this semantics for such tasks as
classification, clustering, indexing, etc. In NLP, the n-gram graph framework is language independent, since it is statistical in it principles. In machine learning, it offers the ability to represent and use proximity information in an updatable, non-monotonic model. The wide applicability of the algorithms and structures makes the tutorial interesting for a variety of applications, including text analysis, optical flow, and user adaptation.

Software management & engineering in AI-based projects

Tutor: Wieslaw Pietruszkiewicz, Szczecin University of Technology (Poland) and Technical Director at SDART Ltd (Manchester, UK)

Abstract:
The tutorial's aim is to familiarise participants with the problems concerning
AI-based software project management. The unpredictable character of research projects distinguishes them from the majority of software projects and increases the demand on project management techniques, supporting the management of high risk projects. A proper approach to AI-based projects and the awareness of typical problems relating to them are major step to achieve project's success.

Style-based Text Categorization

Tutor: Efstathios Stamatatos, University of the Aegean

Abstract:
During the last decade, text categorization has been substantially developed
providing effective methods able to deal with thousands of documents and
multiple categories. The vast majority of text categorization studies deal with thematic categorization. However, style can also be used as a discriminating factor. Style-based text categorization includes the tasks of authorship attribution, genre detection, and plagiarism analysis. The plethora of available electronic text reveals that this technology can be used in several important applications including forensic analysis (identifying the author of harassing messages, linking different terrorist proclamations by authorship, etc.), intelligent search engines (incorporating the genre of web pages in search criteria and the presentation of search results), and intelligent plagiarism detection tools.
The proposed tutorial will present the state-of-the-art of this technology and highlight the main differences with topic-based text categorization. Moreover, it will present different methods for representing writing style (the so-called stylometry) as well as the most important approaches for authorship attribution, genre detection, and plagiarism analysis. Finally, it will give an overview of existing resources that can be used to build and evaluate tools for these tasks.

An Introduction to Causal Discovery

Tutors: Ioannis Tsamardinos, University of Crete, Sofia Triantafillou, ICS, Forth

Abstract:
Discovery of causal knowledge is the ultimate goal of scientific research. Yet, most of classical statistics, machine learning, and data mining algorithms deal with predictive (non-causal) models and algorithms. Such models cannot predict the effect of manipulations to the system. The induction of causal relations from observational data has traditionally been an anathema in statistics summarized in the well-known quote "correlation is not causation". Nevertheless, theories and algorithms for causal induction from observational data, or mixed observational and experimental data under different sampling conditions exist thank to the pioneering work of Spirtes, Glymour, Scheines, Pearl, Cooper, Richarson, Wermuth and others. Such theories have direct ramifications for other important machine learning tasks, such as variable selection.

The tutorial will first present the basic theory of causal discovery from a Causal Bayesian Network pespective. This includes graphical models for representing and inducing causality such as Causal Bayesian Networks, Maximal Ancestral Graphs and Partial Ancestral Graphs as well as concepts such as the Causal Markov Condition, the Faithfulness Condition, and the d-separation criterion. It will also present prototypical and state-of-the-art algorithms such as the PC, FCI and ΗΙΤΟN for learning such models (global learning) or parts of such models (local learning) from data. The tutorial will also discuss the connections of causality to variable selection (a.k.a. feature selection) and present causal-based variable selection techniques. Finally, case-studies of applications of causal discovery algorithms will be presented.