Automatic nuisance recognition #28
Description
This issue is a brain dump for ideas to get ekill to the next level of usability.
Proposition
Instead of relying on a static list of DOMStrings, what if ekill could automatically detect and kill off offending elements, potentially without any previous interaction?
This could be approached as a machine learning classification problem, in the sense of answering the question; What elements on a page is useful and what is unwanted content?
Training
Potentially, every time a user kills off an element, that information could be used as a datapoint for training a classification model. Given enough users, this might sufficient to train a classifier which could be used in practice.
Work
Some tasks to get there would probably include:
- A central service to gather and work the data.
- Intake api.
- ML harness for training / validation.
- Model distribution.
- Privacy management, (GDPR?)
- Extension augmentation.
- Fetching/Updating and using an ML model as basis for the Hit List.
- Sending kill data back to the "mothership" to part-take in model training.
- Add a separate UI flow for indicating false positives. This might entail adding a toggle to see a page in a before and after state, with indications of what elements was removed.