4.5 Conclusion

From the intuition behind the mathematical formulation and the results given by our chosen dataset, we reach the conclusion that different matrix factorization techniques work well towards different ends:

NMF works well for modeling non-negative data such as images. It finds sparse and parts-based representation of the data.

ICA works well for finding independent sources when the data isn’t Gaussian.

PCA has much broader application than ICA and NMF; it is ideal for pattern recognition and dimension reduction.