This will likely be especially useful when we have large feature vectors implemented (T132580) and we start working with hashing vectorization in the wild (T128087).

I talked to @Sabya in IRC. Here's the steps that I recommended.

1. Read up on methods.
2. Take our labeled data for damaging/not and split into train/test set
3. Build model on training set.
4. Run model against a random sample of revisions and take the revisions that are strongly scored (high confidence of "damaging"/not)
5. Train a new model on the training set + the strongly-labeled observations.
6. Test against the test set and see if we do better.

@Halfak which classifier algorithm should I use? Current production algorithms or HashingVector + GradientBoosting?

If it's easy to do so, I'd say "both".