Virtual Seminar: Experimenting with Classifiers
Speaker: Martin Shepperd
1. Much machine learning research is empirical in nature (analytic solutions are intractable).
2. So we conduct experiments…
3. where the competing algorithms are *treatments*, the datasets are *experimental units* and classification performance is the *response measure*,
4. and the experimental design is typically *repeated measures*.
5. This view of machine learning research should inform the study design and analysis.
6. Failure to do so helps explain the unreliability of many published `results’.