Department of Statistics and Actuarial Science

Permanent URI for this community

https://scholar.sun.ac.za/handle/10019.1/73

Browse

Browsing Department of Statistics and Actuarial Science by Subject "Algorithms"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Resampling algorithms for multi-label classification
    (Stellenbosch : Stellenbosch University, 2022-04) Kotze, Ulrich; Sandrock, Trudie; Stellenbosch University. Faculty of Economic and Management Sciences. Dept. of Statistics and Actuarial Science.
    ENGLISH SUMMARY: Multi-label classification is a member of the supervised learning family and represents a scenario where we wish to classify an observation into many of many classes. Therefore, in the classification paradigm an observation can belong to more than one class simultaneously. Imbalanced data is a common problem in the multi-label paradigm of learning. This project investigated resampling algorithms as a pre-processing mechanism to address the manifestation of imbalance in multi-label data to improve multi-label classification performance. Imbalance can manifest itself through a sparse data matrix at small global densities. Imbalance can also manifest itself through a disparity in local label density at larger global densities. The effect of resampling algorithms on multi-label performance is studied for both of these forms of imbalance. We specifically study the effect of these resampling algorithms on multi-label performance at changing levels of global density. The thesis made use of simulated data, five common multi-label classification techniques and seven of the most popular resampling algorithms. Three example-based, label-based and ranking-based evaluation metrics were used to assess the effect of the resampling algorithms on multi-label classification performance.