k-Nearest Neighbours for Data Streams

Online, single-pass kNN classifiers for evolving data streams, built on CapyMOA / MOA. University assignment for COMPX523 (Machine Learning for Data Streams). The interesting part isn't kNN itself — it's how the windowing strategy changes behaviour under class imbalance and concept drift.

Three windowing strategies

All three are incremental classifiers that test-then-train on each instance with a bounded memory window:

Variant	File class	Idea
Base kNN	KNNBase	One global FIFO sliding window of the last w instances.
Class-weighted (cw)	KNN_cw_1689664	A separate window per class, so minority classes keep their own recent history instead of being flushed out by a dominant class.
Age-aware (asw)	KNN_asw_1689664	A single window where, on overflow, an instance is probabilistically ejected with weight ∝ 1.2^age (log-sum-exp for numerical stability) — old instances leave first, adapting to drift while keeping a soft recency bias.

Results

Tested on the electricity stream and a synthetic RBF stream with concept drift, across window sizes {100, 500} and k {3, 21}. Accuracy (%):

Variant	electricity (best)	RBF (best)
Base kNN	86.2 (w=100, k=3)	87.4 (w=500, k=3)
Class-weighted	87.0 (w=100, k=3)	87.5 (w=500, k=3)
Age-aware	86.1 (w=100, k=3)	87.4 (w=500, k=3)

Key finding: the class-weighted variant is the most robust. On the imbalanced/drifting RBF stream, plain kNN with a small window collapsed to ~66%, while per-class windows held at ~85%, because each class retained enough neighbours to vote with. Larger k (21) consistently hurt on these streams — the local decision boundary is better served by a tight neighbourhood. Full numbers in results.csv; methodology and discussion in REPORT.pdf.

Repository layout

knnAssignment.py      The three kNN stream classifiers (Base, cw, asw)
assignment1.py        Final submitted experiment driver
run_experiments.py    Sweeps datasets × window sizes × k and writes results.csv
experiments.ipynb     Exploratory notebook (outputs cleared)
generate_report.py    Programmatic PDF report generator (ReportLab)
results.csv           Full metrics: accuracy, recall, precision, wall-clock
REPORT.pdf            Written report

Data not included. The electricity and RBF ARFF streams are MOA-provided datasets; point the scripts at your own MOA data directory to reproduce.

Tech

Python · CapyMOA / MOA (JPype) · single-pass streaming evaluation.

---

University coursework (COMPX523). Shared as a portfolio sample.