National and Kapodistrian University of Athens
Anomaly detection on temporal data with applications to social networks and healthcare
Abstract
dc:descriptionNowadays, online services and applications are omnipresent, and the amount of data that is being produced and collected is enormous. Data is presumed as a powerful raw material that can impact multidisciplinary research endeavors, government and business performance. By analyzing data, we gain knowledge and understanding. Detecting unusual data patterns not only helps in preventing hazardous situations but also tracks the changes in the trends of whatever is being analyzed. The process of identifying unusual or anomalous behavior is called anomaly detection. In this dissertation, we build novel algorithms and tools for anomaly detection with applications in various domains; (1) social media, (2) news sources, and (3) healthcare. Social media platforms are a valuable source of information for those that can analyze it. We need techniques that are more expressive in terms of the returned insights, and that allow us to better understand the data at hand. We define the problem of event detection and delineation (to sub-events) and we present techniques that (i) detect events from the massive streams of user generated posts, and (ii) describe and summarize each event by creating a timeline of its important highlights. News portals are another source of information, more structured than social media, and with a more strict and limited vocabulary. Nowadays, there is a huge increase in the number of news sources that exist and the number of daily news stories that are being produced. We are in need of news recommendation and personalization systems that will make it easier for the reader to receive a curated list of news. For this reason, we propose a method to rank all notable news stories and build a framework to monitor, analyze and extract insights from multiple news sources. Healthcare data is becoming more and more available in the last few years. They contain rich information that if used in the right way, it can improve people's health. However, there is a gap in research as medical practitioners know the domain and the research questions, while computer scientists are experts in analyzing data and building predictive models. As a result, we need to bring together these two research fields in order to achieve the best possible outcomes. In our case, we study the problem of infant mortality prediction using data from birth certificate records. Our focus on socio-economic factors and how those can affect the predictive performance. We characterize a wide range of machine learning models trained on a variety of features sets extracted from the available input data. We show ways to improve the current state-of-the-art that is only utilizing two features, birth weight and APGAR score.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Saravanou, Antonia
- Σαραβάνου, Αντωνία
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/48710