Off-line training with dirty label data with varying noise level from 0% to 100%. For the IoT attacks (thermostat) dataset and Cluster task failures detection (task) datasets, all experiments are repeated 10 times. For the face dataset with MLP algorithm, we repeat experiments 10 times. For the face dataset with VGG algorithm, we repeat experiments 3 times due to the higher training time. The experiments using nearest centroid on the task dataset have greatly varying results between runs. We repeat these experiments 100 times.
To run the experiment, user can simply run:
python assessment_*.py
The datasets are available on google drive.