MSc Thesis Defense: İclal Zeynep Özgüç, A MIP FORMULATION FOR PREFERENCE LEARNING WITH MULTIPLE SEGMENTS, Date & Time: 21 July, 2026 – 1:30 PM, Place: FENS L063

A MIP FORMULATION FOR PREFERENCE LEARNING WITH MULTIPLE SEGMENTS

İclal Zeynep Özgüç
Data Science, MSc Thesis, 2026

Thesis Jury

     Asst. Prof. Ezgi Karabulut Türkseven (Thesis Advisor)

  Assoc. Prof. Kemal Kılıç

  Asst. Prof. Hande Küçükaydın

Date & Time: July 21st, 2026 – 1.30 PM

Place: FENS L063

Keywords : Preference Learning, Mixed-integer Programming, Multiple Utility Functions, SVM

Abstract

Preference learning aims to estimate the utility function that explains observed choices among alternatives. For linear utility functions, learning from preference observations can be formulated as finding a separating hyperplane. This formulation assumes that all comparisons are generated by one utility function. However, a single hyperplane does not suffice when the observed preferences are generated by multiple utility functions. This thesis proposes a mixed-integer programming (MIP) formulation for estimating multiple linear utility functions from preference observations. The formulation is adapted to deterministic and probabilistic preference observations, since an observed choice does not always agree with the underlying utility. The formulation is solved on samples of the training data, since solving it on the full training set is computationally expensive. Each sample produces its own estimate, therefore five post-sampling methods are proposed to select or aggregate the resulting sample-level estimates. Computational experiments are conducted on problems of dimension three to five. The estimated weight vectors closely match the underlying vectors, with accuracy between 95% and 99.8%. Accuracy decreases as the dimension grows, from 99.8% in dimension three to 95% in dimension five, which indicates that higher dimensional settings are harder to estimate. The results also show that the sampling approach is effective when the full training set is too large to solve directly.