[sanxiyn]

I actually think chat conversation from a decade ago would be quite relevant. One baseline recommendation system is "people who bought X also bought Y". Consider "people whose conversation is in cluster X generally liked people in cluster Y". If chat conversation can be usefully used to cluster users for better matching (and I think it can), it would be valuable to keep even if content is of no interest.

[adrian1973]

> even if content is of no interest.

Can't they just keep (at most) the metadata?

[sanxiyn]

As a data scientist, I think losing actual words would be a loss. Words would be only used by word embeddings like word2vec, but actual words let you switch to better word embedding later.