Predictability of Classification Performance Measures with Meta-Learning

Aus IPD-Institutsseminar
Version vom 2. April 2019, 10:48 Uhr von Jakob Bach (Diskussion | Beiträge) (Die Seite wurde neu angelegt: „{{Vortrag |vortragender=Huijie Wang |email=upssh@student.kit.edu |vortragstyp=Proposal |betreuer=Jakob Bach |termin=Institutsseminar/2019-04-12 |kurzfassung=In…“)
(Unterschied) ← Nächstältere Version | Aktuelle Version (Unterschied) | Nächstjüngere Version → (Unterschied)
Zur Navigation springen Zur Suche springen
Vortragende(r) Huijie Wang
Vortragstyp Proposal
Betreuer(in) Jakob Bach
Termin Fr 12. April 2019
Vortragsmodus
Kurzfassung In machine learning, classification is the problem of identifying to which of a set of categories a new instance belongs. Usually, we cannot tell how the model performs until it is trained. Meta-learning, which learns about the learning algorithms themselves, can predict the performance of a model without training it based on meta-features of datasets and performance measures of previous runs. Though there is a rich variety of meta-features and performance measures on meta-learning, existing works usually focus on which meta-features are likely to correlate with model performance using one particular measure. The effect of different types of performance measures remain unclear as it is hard to draw a comparison between results of existing works, which are based on different meta-data sets as well as meta-models. The goal of this thesis is to study if certain types of performance measures can be predicted better than other ones and how much does the choice of the meta-model matter, by constructing different meta-regression models on same meta-features and different performance measures. We will use an experimental approach to evaluate our study.