COMPX523, 92.74%
Streaming ML
CapyMOA/MOA, NumPy, concept-drift handling, online ensembles
- Built a memory-budgeted streaming ensemble (71 members: Gaussian Naive Bayes, SGD, and KNN) with recency-weighted voting and per-member drift detection (ADWIN), running at about 155KB against a 250KB budget and reaching 0.8988 accuracy on the sample stream - ahead of every published baseline tested on that stream.
- Fixed unbounded memory growth and a categorical-distance bug in a streaming KNN implementation, and cut Python/Java interop overhead on prediction from 27.8s to 20.2s.