Hypothesis

Training machine learning models on event streams is an effective way to do machine learning that is more memory efficient and more flexible during deployment, without compromising accuracy measures such as F1 score, FPR/TPR, etc. We set out to test the above hypothesis with the following three evaluation goals in mind:

1. A model trained on an event stream should be competitive in performance with a model trained on a static batch of data.
2. A model trained on an event stream should take significantly less time to train than a model trained on a static batch of data.
3. A model trained on an event stream can be updated without taking training offline.