SmartKNN started as a crazy idea:
How far can we push K-Nearest Neighbors if we remove every bottleneck… instead of accepting them?
V1 proved something important — KNN can still be used in production. With weighted features, prototype compression, optimized distance rules and SmartKNN became a real production-ready tabular model.
Now it’s time for v2.
And v2 is not just an upgrade… it’s the beginning of the next era.
What’s coming in SmartKNN v2?
1. Native Classification Support
No more “only regression” limitation. SmartKNN v2 automatically detects the target type and switches intelligently between:
Classification Regression
One fit() -> SmartKNN knows what to do. No configs. No confusion.
2. Faster Prediction (Like… way faster)
…
SmartKNN started as a crazy idea:
How far can we push K-Nearest Neighbors if we remove every bottleneck… instead of accepting them?
V1 proved something important — KNN can still be used in production. With weighted features, prototype compression, optimized distance rules and SmartKNN became a real production-ready tabular model.
Now it’s time for v2.
And v2 is not just an upgrade… it’s the beginning of the next era.
What’s coming in SmartKNN v2?
1. Native Classification Support
No more “only regression” limitation. SmartKNN v2 automatically detects the target type and switches intelligently between:
Classification Regression
One fit() -> SmartKNN knows what to do. No configs. No confusion.
2. Faster Prediction (Like… way faster)
SmartKNN v2 integrates an ANN backend - but here’s the twist:
ANN finds the nearest candidates ->> SmartKNN’s brain finishes the job.
So we get: ANN-level speed SmartKNN-level accuracy
Best of both worlds. KNN doesn’t have to be slow anymore.
3. Full Vectorization
We’re removing raw Python loops (the main slowdown of KNN everywhere).
v2 uses full vectorization for distance math, weight computation and voting.
Result:
- Faster training
- Faster prediction
- Scales better on bigger datasets
4. Smarter Feature Weights (Rethinking MI + LR + RF)
Right now SmartKNN uses a mix of:
- Mutual Information
- Linear Regression
- Random Forest importance
It works insanely well… but it doesn’t scale to millions of rows.
For v2 we are experimenting with a new mathematical formula that:
- doesn’t need huge computation
- scales to massive datasets
- keeps (or improves) accuracy
Not guaranteed to land in v2. But research is happening.
The Philosophy of SmartKNN (still unchanged)
Even with ANN support and huge optimizations — SmartKNN v2 is not chasing maximum speed at the cost of accuracy.
The priority stays: accuracy + interpretability + simplicity with better speed as a bonus.
This is just the classification entry + first speed boost. The future will go even further.
Beyond v2 — What’s next?
After v2, SmartKNN will eventually ship with its own special ANN backend built only for SmartKNN conditions (not a generic ANN library).
That’s where we expect:
- 10× prediction speed
- Near Boosting models latency on CPU
- Full explainability (unlike tree models)
The long-term mission:
Make a neighbour-based model that is fully production-grade for tabular ML.
Final Words
SmartKNN v2 is not about reinventing KNN.
It’s about proving that:
with the right engineering + math + shortcuts + structure, KNN can be a modern production model — not a toy baseline.
Thank you to everyone supporting SmartKNN so far. v2 is on the way. Stay tuned.
Stay Updated
Install SmartKNN
pip install smart-knn
GitHub Repository https://github.com/thatipamula-jashwanth/smart-knn
Benchmarks Page https://www.kaggle.com/jashwanththatipamula
Release Announcements (You can follow the repo to get notified when v2 drops)
Follow SmartKNN on GitHub
Jashwanth Thatipamula - Creator SmartKNN