Distributed / serverless and online operation? #2
Description
Very interesting work. When I first learned of UMAP and HDBSCAN I intuitively was curious of the whole bert embeddings -> umap -> hdbscan -> topics approach. I am quite unsurprised at the effectiveness of this approach. It is not intuitive that your approach would achieve similar results and generalizability. Congratulations this appears to be a meaningful contribution especially as it looks to be much cheaper computationally.
That cost aspect of things is really important for large scale online uses. I am working on a document classification service and there are two operational requirements for the service to be optimal in terms of operating expense.
First it needs to be online. As I add new documents the topic model should adjust. Thus there is no need to reprocess all documents to keep the model up to date as the corpus grows.
Second it needs to be distributable such that I can farm out the work of any given document or batch of documents to a serverless function, and the lighter these functions are the better.
Your approach seems like it is far more amenable to these constraints, I'd love to explore them a bit. What is your advice on this and your sense of what would be required to implement online and distributed operation of this algorithm?