NMF

Non-Negative Matrix Factorization (NMF) decomposes data into non-negative components, producing additive, parts-based representations. It is popular for topic modeling, image decomposition, and any domain where negative components are not meaningful.

When to use:

• Topic modeling on document-term matrices
• Image part extraction (e.g., face parts from face images)
• Any application where features and components must be non-negative

Input: Non-negative tabular data with the feature columns defined during training Output: Component activations for each row (all non-negative)

Model Settings (set during training, used at inference)

N Components (default: 2) Number of non-negative components.

Initialization (default: null — auto) Method for factor matrix initialization. nndsvd is good for sparseness; random is general-purpose.

Max Iterations (default: 200) Maximum alternating least squares iterations.

Tolerance (default: 0.0001) Convergence threshold.

Inference Settings

No dedicated inference-time settings. The trained basis matrix projects new non-negative data into the component space.