University of Patras
Αλγόριθμοι ανίχνευσης ακραίων τιμών και εφαρμογή σε δεδομένα που αφορούν τη διαχείριση ανθρώπινου δυναμικού
Abstract
dc:descriptionThe present thesis focuses on the detection of outliers in time series, that is, the identification of events or observations that do not conform to an expected pattern relative to the overall dataset. Their timely identification can contribute both to the understanding of the data and to the prediction of future developments. In the first part of the thesis, three new algorithms for outlier detection in univariate time series are proposed, which are based on transforming the data into graphs and studying the angles formed between successive values. The proposed algorithms (Angles_Per, Angles_Lim, and Angles_Graph) were evaluated on synthetic and real datasets, demonstrating high accuracy, a significant reduction in false positive results, and competitive or superior performance compared to existing well-known methods. The application of the proposed methods to human resources data highlights the importance of outlier detection in variables such as employee absences or working hours. Detecting such outliers can serve as an indicator of organizational risks, such as burnout or unequal workload distribution, providing an early warning tool for management. In this way, outlier detection is directly linked to the decision-making process. Subsequently, the analysis extends beyond the individual detection of outliers, proposing a methodology for extracting associations between outliers that occur in different time series. The methodology is based on detecting and grouping outliers according to their presence or absence within a time interval and on applying the Apriori algorithm to extract association rules. These rules are utilized to identify connections between time series as well as for the short-term prediction of outliers in a time series. Finally, a methodology for discovering future patterns in time series is proposed, based on the symbolic representation of data. Through the transformation of numerical values and their differences into symbolic sequences, it becomes possible to identify recurring structures, which are then utilized for the prediction of future patterns. The experimental results on real datasets from different domains confirm the high accuracy of the proposed method.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Katsouda, Maria
- Κατσούδα, Μαρία
Subjects
dc:subject × 26- Ακραίες τιμές
- Χρονοσειρές
- Ανίχνευση ακραίων τιμών
- Ανθρώπινο δυναμικό
- Εξόρυξη συσχετίσεων
- Αναπαράσταση με γράφους
- Εξόρυξη μοτίβων
- Outliers
- Time series
- Anomaly detection
- Human resources
- Association rule mining
- Graph-based representation
- Pattern mining
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Διεπιστημονικές εφαρμογές επιστήμης ηλεκτρονικών υπολογιστών
- Οικονομικά και Επιχειρήσεις
- Κοινωνικές Επιστήμες
- Οργανωσιακή συμπεριφορά και Διαχείριση ανθρώπινου δυναμικού
- Computer and Information Sciences
- Natural Sciences
- Computer Science Interdisciplinary Applications
- Economics and Business
- Social Sciences
- Organizational Behavior and Human Resource Management
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62605