Database Reference
In-Depth Information
Cumulative sum of top 300 singular values
We can see that after a certain value range for
k
(around 100 in this case), the graph flat-
tens considerably. This indicates that a number of singular values (or principal compon-
ents) equivalent to this value of
k
probably explains enough of the variation of the original
data.
Tip
Of course, if we are using dimensionality reduction to help improve the performance of
another model, we could use the same evaluation methods used for that model to help us
choose a value for
k
.
For example, we could use the AUC metric, together with cross-validation, to choose both
the model parameters for a classification model as well as the value of
k
for our dimen-
sionality reduction model. This does come at the expense of higher computation cost,
however, as we would have to recompute the full model training and testing pipeline.
