METRIC LEARNING FOR INCORPORATING PRIVILEGED INFORMATION IN PROTOTYPE-BASED MODELS

Prototype-based classification models, and particularly Learning Vector Quantization (LVQ)
frameworks with adaptive metrics, are powerful supervised classification techniques with good
generalization behaviour. This thesis proposes three advanced learning methodologies, in the
context of LVQ, aiming at better classification performance under various classification settings.
The first contribution presents a direct and novel methodology for incorporating valuable
privileged knowledge in the LVQ training phase, but not in testing. This is done by manipulat-
ing the global metric in the input space, based on distance relations revealed by the privileged
information. Several experiments have been conducted that serve as illustration, and demon-
strate the benefit of incorporating privileged information on the classification accuracy.
Subsequently, the thesis presents a relevant extension of LVQ models, with metric learning,
to the case of ordinal classification problems. Unlike in existing nominal LVQ, in ordinal LVQ
the class order information is explicitly utilized during training. Competitive results have been
obtained on several benchmarks, which improve upon standard LVQ as well as benchmark
ordinal classifiers.
Finally, a novel ordinal-based metric learning methodology is presented that is principally
intended to incorporate privileged information in ordinal classification tasks. The model has
been verified experimentally through a number of benchmark and real-world data sets.