language: - en tags: - tabular - nearest-neighbors - knn - classification - regression - cpu - low-latency - interpretable library_name: smart-knn license: mit pipeline_tag: tabular-classification model_name: SmartKNN v2
SmartKNN v2
SmartKNN v2 is a high-performance, CPU-first nearest-neighbors model designed for low-latency production inference on real-world tabular data.
It delivers competitive accuracy with gradient-boosted models while maintaining sub-millisecond single-prediction latency (p95) on CPU-only systems.
SmartKNN v2 is part of the SmartEco ecosystem.
Model Details
- Model type: Distance-weighted K-Nearest Neighbors
- Tasks: Classification, Regression
- Backend: Adaptive (Brute-force + ANN)
- Hardware: CPU-only (GPU not required)
- Focus: Low latency, interpretability, production readiness
Unlike classical KNN, SmartKNN v2 learns feature importance, adapts execution strategy based on data size, and uses optimized distance kernels for fast inference.
What’s New in v2
- Full classification support restored
- ANN backend introduced for scalable neighbor search
- Automatic backend selection (small → brute, large → ANN)
- Distance-weighted voting for improved accuracy
- Interpretable neighbor influence statistics
- Foundation for adaptive-K strategies
Architecture Overview
- Feature Weighting
- Backend Selector
- Brute Backend (small datasets)
- ANN Backend (large datasets)
- Distance Kernel
- Weighted Voting
- Prediction
This hybrid architecture ensures consistent low latency across dataset sizes.
Performance (Internal Evaluation)
Public benchmarks will be released soon.
From internal testing on real-world tabular datasets:
- Accuracy comparable to XGBoost / LightGBM / CatBoost
- Single-prediction latency:
- Median: sub-millisecond
- p95: consistently low on CPU
- Predictable batch inference scaling
SmartKNN v2 has not yet reached its performance ceiling. Future releases will further optimize speed and accuracy.
Limitations
- Not designed for unstructured data (text, images)
- ANN backend focuses on CPU efficiency, not GPU acceleration
- Best suited for tabular datasets
Future Work
- Adaptive-K accuracy optimization
- Kernel-level speed improvements
- Custom ANN backend