Department Seminar [MGMT] 04-06-2019 SPEAKER: Dr. Manish Aggarwal, IIM, Ahmedabad. TITLE: “Preference-based Learning of Ideal Solutions”

SPEAKER: Dr. Manish Aggarwal, IIM, Ahmedabad.

TITLE: “Preference-based Learning of Ideal Solutions”


DAY & DATE: Tuesday 4th June 2019

VENUE: Seminar Hall, Department of Management Studies, IISc.



This work combines the established modelling techniques from multiple-criteria decision aiding (MCDA) with recent algorithmic advances in the emerg-ing field of preference learning (PL). By this, we mean the idea of applying machine learning methods for a data-driven construction of MCDA models on the basis of observed preference information. On the basis of exemplary pairwise comparisons between alternatives, our method seeks to induce an ‘ideal solution’ (IS) that, in conjunction with aweight factor for each criterion, represents the preferences of the decision maker. Different candidate solutions are evaluated according to their distance from this reference point. The closer a solution is to the IS, the more preferred it is. To this end, we resort to probabilistic models of discrete choice and make use of maximum likelihood inference. First experimental results on suitable preference data suggest that our approach is not only intuitively appealing and interesting from an interpretation point of view but also competitive to state- of-the-art preference learning methods in terms of prediction accuracy. The proposed method (PLIS) can also be seen as an adaptive version of the popular MCDA approach TOPSIS, the technique for order preference by similarity to ideal solution decision model. Our concrete idea is to learn the ideal (positive) solution from observed preference information instead of defining this solution in a more or less arbitrary way. Here, preference information is supposed to be given in the form of pairwise comparisons between candidate solutions. Roughly speaking, the learning algorithm is supposed to answer the following question: Given the pairwise preferences revealed by a DM, what could be the ideal solution this DM has in mind? Because PLIS positions itself in-between machine learning and MCDA, its evaluation requires data sets that meet the MCDA requirements. In particular, it should be possible to define, in a semantically meaningful way, a ranking of the data and any subset thereof. We tested the validity of our approach on 12 such datasets, originating from the UCI repository, and EKA machine learning framework. The datasets also include real decision making data on evaluation of journals, and the houses in Den Bosch city. The experimental study suggests that PLIS is competitive to state-of-the-art methods for learning to rank. From an MCDA point of view, PLIS suggests interesting extensions of the original TOPSIS method, making it more flexible and allowing one to adapt the decision model to the preferences of a specific DM. From a machine learning perspective, our approach offers a simple yet intuitively plausible and interpretable model in the context of preference learning: The ideal solution shows what the DM is aiming for, and the attribute weights inform about the importance of the different criteria

About the Speaker:

Dr. Manish Aggarwal received his Ph.D. in information technology in 2013 from Indian Institute of Technology Delhi, New Delhi, India. He is currently an assistant professor at Indian Institute of Management Ahmedabad. During his academic career, he has been associated with well-known universities in India and abroad as visiting faculty, and as research fellow. He also has a decade of industrial experience in the roles of senior subject matter expert, and lead business analyst in renowned global corporations. His research interests include multicriteria decision aiding (MCDA), preference learning, evolutionary optimization, and fuzzy decision analysis