Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.781806
Title: Privileged Learning using Unselected Features
Author: Taylor, Joseph Gerard
Awarding Body: University of Sussex
Current Institution: University of Sussex
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
This thesis proposes a novel machine learning paradigm called Learning using Unselected Features (LUFe), which front-loads computation to training time in order to improve classifier performance, without additional cost at deployment. This is achieved by repurposing and combining techniques from feature selection and Learning Using Privileged Information (LUPI). Feature selection is a means of reducing model complexity, which enables deployment in devices with limited computational power, but this can waste additional resources which may be available at training time. LUPI is a paradigm that allows extra information about the training data to be harnessed by the learner, but this requires an additional set of highly informative attributes. In the LUFe setting, feature selection is used to partition datasets into primary and secondary subsets, instead of discarding the features which are unselected. Both datasets are then passed to a LUPI algorithm, enabling the secondary feature-set to provide additional guidance at training time only, in place of `privileged' information. Only the selected features are used at train time, maintaining low-cost deployment while exploiting train-time resources. Experimental results on a large number of datasets demonstrate that LUFe facilitates an improvement in classification accuracy over standard feature selection approaches in a majority of cases. This performance boost is consistent across a range of feature selection approaches, and is largest when the SVM+ algorithm is used for implementation. This effect is shown to be partially dependent on the usage of information in the unselected features, as well as resulting from the presence of additional constraints on the function space searched for the model. The enhancement by LUFe is shown to be inversely correlated with the performance of standard feature selection and mediated by a further reduction in model variance, beyond that provided by standard feature selection. Aside from demonstrating the direct practical benefit of LUFe, this work makes the contribution of broadening the scope of applications for the LUPI framework.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.781806  DOI: Not available
Keywords: Q0325.5 Machine learning
Share: