Συντάχθηκε 11-09-2026 14:30
Τόπος:
Σύνδεσμος τηλεδιάσκεψης
Έναρξη: 14/09/2026 12:30
Λήξη: 14/09/2026 13:30
ΠΟΛΥΤΕΧΝΕΙΟ ΚΡΗΤΗΣ
Σχολή Μηχανικών Παραγωγής και Διοίκησης
Πρόγραμμα Μεταπτυχιακών Σπουδών
Διοίκηση Επιχειρήσεων
ΠΑΡΟΥΣΙΑΣΗ ΜΕΤΑΠΤΥΧΙΑΚΗΣ ΕΡΓΑΣΙΑΣ
Δευτέρα, 14 Σεπτεμβρίου 2026, 12:30
https://tuc-gr.zoom.us/j/5741492636?pwd=MTE1QnR5c1JjM3UxZUo3SU84SWJQUT09
Ονοματεπώνυμο: ΒΕΡΕΡΟΥΔΑΚΗΣ ΕΜΜΑΝΟΥΗΛ
Θέμα: Ανάπτυξη μιας Συλλογικής Μεθόδου Μηχανικής Μάθησης βασισμένης σε Πολυκριτήριες Μεθόδους Αναλυτικής – Συνθετικής Προσέγγισης
Title: Development of a Machine Learning Ensemble Method based on Multicriteria Aggregation – Disaggregation Approach
Εξεταστική Επιτροπή
- ΜΑΤΣΑΤΣΙΝΗΣ ΝΙΚΟΛΑΟΣ, Ομότιμος Καθηγητής (επιβλέπων)
- ΣΙΣΚΟΣ ΕΛΕΥΘΕΡΙΟΣ, Επίκουρος Καθηγητής
- ΦΑΦΑΛΙΟΣ ΠΑΥΛΟΣ, Επίκουρος Καθηγητής
Περίληψη
Preference disaggregation estimates a decision maker’s value function from a ranking of alternatives evaluated on several criteria. UTASTAR represents this function as a sum of monotone contributions, making the resulting preferences directly interpretable. This thesis investigates whether collections of such functions can also support accurate prediction, and when combining models is more useful than fitting a single model. The study considers two prediction settings. In a French olive-oil panel of 176 respondents, models predict the rankings of people excluded from training. A second experiment reveals part of each person’s ranking and predicts the remaining choices. On public red- and white-wine data, the study evaluates the classical setting in which new alternatives are assessed under a shared quality judgment. The study compares UTASTAR models and ensembles with simple population baselines, collaborative filtering, and tree-based machine-learning methods. For an unknown respondent described only by criterion ratings, averaging individual value functions offers little improvement over an equal-weights score. The strongest ratings-based methods achieve rank correlations of approximately 0.56 on the olive-oil panel. Although the collection contains functions that closely match individual respondents, selecting an appropriate function requires additional information about their preferences. Once several choices are revealed, weighting models by their agreement with those choices improves prediction of the remaining ranking. The wine experiments show that predictive performance depends on both the available reference wines and the evaluation measure. Additive models provide useful rankings and transparent explanations, while tree models can capture relationships that additive models cannot. Averaging can improve some measures when ranked wines are scarce, but does not consistently improve recognition of each quality class. The findings support a selective use of ensembles: their main value lies in combining existing preference information with a person’s own choices, while retaining explanations through criterion-level value functions. Keywords: preference disaggregation; UTASTAR; multiple criteria decision aiding; ensemble learning; preference prediction; interpretable machine learning.
