Otimização multi-objetivo em aprendizado de máquina

Author(s)

Marcos Medeiros Raimundo

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação
• Defense date: 2018-12-12

Examining board members

Fernando José Von Zuben; Ricardo Hiroshi Caldeira Takahashi; Márcio Porto Basgalupp; Guilherme Palermo Coelho; Levy Boccato

• Advisor: Fernando José Von Zuben

Abstract

Regularized multinomial logistic regression, multi-label classification, and multi-task learning are examples of machine learning problems in which conflicting objectives, such as losses and regularization penalties, should be simultaneously minimized. Therefore, the narrow perspective of looking for the learning model with the best performance should be replaced by the proposition and further exploration of multiple efficient learning models, each one characterized by a distinct trade-off among the conflicting objectives. Committee machines and a posteriori preferences of the decision-maker may be implemented to properly explore this diverse set of efficient learning models toward performance improvement. The whole multi-objective framework for machine learning is supported by three stages: (1) The multi-objective modelling of each learning problem, explicitly highlighting the conflicting objectives involved; (2) Given the multi-objective formulation of the learning problem, for instance, considering loss functions and penalty terms as conflicting objective functions, efficient solutions well-distributed along the Pareto front are obtained by a deterministic and exact solver named NISE (Non-Inferior Set Estimation); (3) Those efficient learning models are then subject to a posteriori model selection, or to ensemble filtering and aggregation. Given that NISE is restricted to two objective functions, an extension for many objectives, named MONISE (Many Objective NISE), is also proposed here, being an additional contribution and expanding the applicability of the proposed framework. To properly access the merit of our multi-objective approach, more specific investigations were conducted, restricted to regularized linear learning models: (1) What is the relative merit of the a posteriori selection of a single learning model, among the ones produced by our proposal, when compared with other single-model approaches in the literature? (2) Is the diversity level of the learning models produced by our proposal higher than the diversity level achieved by alternative approaches devoted to generating multiple learning models? (3) What about the prediction quality of ensemble filtering and aggregation of the learning models produced by our proposal on: (i) multi-class classification, (ii) unbalanced classification, (iii) multi-label classification, (iv) multi-task learning, (v) multi-view learning? The deterministic nature of NISE and MONISE, their ability to properly deal with the shape of the Pareto front in each learning problem, and the guarantee of always obtaining efficient learning models are advocated here as being responsible for the promising results achieved in all those three specific investigations (AU)

• FAPESP's process: 14/13533-0 - Multi-objective optimization in multi-task learning
• Grantee: Marcos Medeiros Raimundo
• Support Opportunities: Scholarships in Brazil - Doctorate