Got into first roadblocks...
by BenSo I reached out to freelancers on Fiverr to talk about a potential email classifier. That made me give a closer look at the data, which has various issues:
- I don't have anything to train a model with yet! I thought I could use my personal emails, but few of them are actually labeled. So I'm going to start labeling everything I can...
- There's a huge privacy concern with reading the content of emails in order to determine how to classify them. We would probably need to remove any type of private info before sending them to our servers, especially if we're going to store some to further train the model.
- Some emails have a lot of text in them. How does the model know which ones are important? Should we send only the top n characters? Or just can it work with just the subject and some of the metadata, excluding the body entirely?
In addition, I looked into the Gmail API, and found out that they recently tacked on a new "security assessment" procedure for apps that use restricted scopes, ie anything that can read or modify emails. They require using a third party who can charge $15-75k, which is a big no-no. An alternative might be to use the IMAP protocol, but I'm not sure we'd have as much control over labels and filters.
