The thesis addresses the problem of missing features in regression and classification problems and its impact on the prediction loss.
To minimise the prediction loss, a recommender system for active feature acquisition is introduced. The recommender system
estimates the utility of missing features in prediction set and acquires the ones with the highest utilities.
The estimates are based on a matrix decomposition of a preference matrix (the preference matrix is composed
of calculated Shapley values for known attributes).
The developed method is evaluated for feature acquisition on learning and testing datasets.
The results indicate that the method is more effective to achieve lower prediction loss in comparison with randomly acquiring missing features. However, it is
not always the case that it performs better in comparison with unpersonalised recommendations.
|
