Frequently Asked Questions (FAQ)
This page answers common questions about SmartKNN’s design, behavior, and intended usage.
If you are new to SmartKNN, this section helps clarify why certain choices were made and what to expect in practice.
What makes SmartKNN different from standard KNN?
SmartKNN treats nearest-neighbor learning as a system, not a single algorithm.
Key differences include: - Learned feature importance instead of equal weighting - Feature pruning via weight thresholding - Distance-weighted regression and classification - Automatic backend selection (brute vs ANN) - Deterministic, production-safe inference
Standard KNN typically applies a fixed distance metric and naive aggregation.
Does SmartKNN require manual feature scaling or preprocessing?
No.
SmartKNN handles internally: - Feature normalization - NaN and Inf sanitization - Numerical stability safeguards
Users can provide raw numeric features directly.
Is SmartKNN a lazy learner?
Yes — but with structure.
Like classical KNN, SmartKNN stores the training data and performs computation at inference time.
However, SmartKNN adds an explicit configuration phase where feature importance, pruning, and backend strategy are learned and frozen.
Inference itself is deterministic and optimized.
Does SmartKNN perform online or incremental learning?
No.
SmartKNN intentionally does not support online or continual learning.
All learning occurs during fit().
Inference uses a frozen configuration to guarantee:
- Predictable latency
- Deterministic outputs
- Stable interpretability
For streaming or continuously adapting systems, SmartKNN is not the right tool.
How does SmartKNN handle large datasets?
SmartKNN supports large datasets through: - Automatic backend selection - Approximate Nearest Neighbor (ANN) backends - Safe fallback to brute-force when ANN quality is insufficient
Backend choice affects how neighbors are retrieved, not how predictions are computed.
Is ANN always used for large datasets?
Not always.
SmartKNN may: - Use brute-force for small or medium datasets - Use ANN for large datasets - Fall back to brute-force if ANN quality checks fail
This behavior is automatic and designed to preserve prediction correctness.
Does ANN change prediction behavior?
No.
ANN affects candidate retrieval only.
Once neighbors are retrieved: - Distance computation - Prediction aggregation - Interpretability behavior
remain identical to brute-force execution.
How are predictions computed for regression?
Regression uses distance-weighted local regression.
- Feature weights influence distance computation
- Neighbor influence is proportional to inverse distance
- Predictions are normalized and numerically stable
This produces smooth, locally adaptive predictions.
How are predictions computed for classification?
Classification uses distance-weighted class voting.
- Each neighbor contributes a vote weighted by distance
- Votes are accumulated per class
- The class with the highest total weight is selected
This avoids brittle majority voting and improves minority-class recall.
Is SmartKNN deterministic?
Yes.
Given: - Identical data - Identical configuration - Identical backend
SmartKNN produces identical predictions across runs.
No randomness or hidden state is introduced during inference.
Is SmartKNN compatible with scikit-learn?
Yes.
SmartKNN implements standard sklearn-style methods:
- fit
- predict
- get_params
- set_params
It can be used in sklearn pipelines and evaluation workflows.
Can SmartKNN run without a GPU?
Yes.
SmartKNN is CPU-first by design.
GPU support, when available, is limited to neighbor search in ANN backends and is optional.
What types of data does SmartKNN support?
SmartKNN is best suited for: - Numeric tabular data - Structured datasets - Problems with meaningful feature similarity
It is not designed for raw text, images, or unstructured data without prior feature extraction.
When should I NOT use SmartKNN?
Avoid SmartKNN if: - You require online learning - Your data is extremely high-dimensional and dense - You need a very small model footprint - Strong global generalization is required
See When to Use / When Not Use for details.
Is SmartKNN production-ready?
Yes.
SmartKNN is designed with: - Deterministic inference - Memory safety checks - Explicit backend control - Serialization-safe model state
It is suitable for real-world production systems.
Where should I start?
- Read Design Goals and Core Concepts for intuition
- Explore runnable examples in the repository
- Use Quickstart for minimal setup
- Review Benchmarks before deployment