The course you refer to and the paper are somewhat different. The course simply says that for a given problem, there are good datasets and bad datasets, and the

The course you refer to and the paper are somewhat different. The course simply says that for a given problem, there are good datasets and bad datasets, and the bad data may be bad no matter how big it is. Bad means biased here. The paper considers a somewhat different situation. There you have a single dataset, for which you know in advance[1] that some points in it, and you know exactly which, are more noisy than others. The question is then whether to use these points, and how. [1] But you know this because you assume a model. A particular timeseries model in this case. Its not necessarily in the data itself.