Machine Learning with Spark - Second Edition

Evaluating dimensionality reduction models

Both PCA and SVD are deterministic models. That is, given a certain input dataset, they will always produce the same result. This is in contrast to many of the models we have seen so far, which depend on some random element (most often for the initialization of model weight vectors, and so on).

Both models are also guaranteed to return the top principal components or singular values, and hence, the only parameter is k. Like clustering models, increasing k always improves the model performance (for clustering, the relevant error function, while for PCA and SVD, the total amount of variability explained by the k components). Therefore, selecting a value for k is a trade-off between capturing as much ...