3
$\begingroup$

This question already has an answer here:

I have a dataset $D$ made of $m$ samples and $n$ features with $n \gg m$. For each sample I have a score $s$ which I would like to be able to predict.

As the number of features is very high (compared to the number of samples), I have performed PCA (Principal Component Analysis) and built a regression model for $s$ based on the first 3 principal components (out of 45). I cross-validated the results and am happy with the performances of my model.

Now I would like to understand what is happening in terms of my original features. In particular I would like to understand which features are relevant in predicting $s$, and select them to build a model which does not involve PCA.

How can I perform this last step (if you have R based solutions it is better, but not essential)?

$\endgroup$

marked as duplicate by gung regression Sep 9 '14 at 14:11

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • 3
    $\begingroup$ Each of your principal components will be a linear combination of your features, e.g. $PC_1 = \sum_{i=1}^n \alpha_i x_i$, where $x_i$ is the $i$th feature vector and $PC_1$ denotes the first principal component. If you look at the weights $\alpha_i$ for each feature in constructing each of the 3 principal components you've used, you should get a feel for which features might be more relevant in predicting $s$ (the ones with the largest $|\alpha_i|$). You should be able to extract the feature weights in each principal component easily enough in R depending on which library you are using. $\endgroup$ – Sam Livingstone Jan 22 '14 at 10:21
8
$\begingroup$

PCA by its own doesn't do a feature selection, but it performs a transformation that will usually include all original variates: it is rare that PCA loadings (the coefficients for the transformation/projection) are 0.

Thus, I'd say that PCA is more adapted to data where you expect the relevant information to be spread out over many/most of your variates.

There are other regularization techniques which favor coefficients becoming 0, such as LASSO - which corresponds to selecting variates.

The Elements of Statistical Learning have a nice comparison of different regularization techniques.

There is also sparse PCA which combines the idea of maximum variance projection with favoring sparse solutions.

I may add that also sparse PLS exists.

As you ask about R, several implementations of sparse PCA, sparse PLS, etc. exist. I cannot recommend any particular package as my data is usually of the type where you cannot expect a sparse solution - I therefore use normal PCA and PLS.


If you prefer to stay with the PCR you have, but would like to interpret the coefficients along the lines @SamLivingstone suggests:

  • If your data is not variance scaled, looking at the coefficients is not enough, you need to take into account the size/range of the corresponding variate ("contribution").
  • I'd resample the models (e.g. during a validation run) and compare the size of the coefficient/contribution of each variate to its spread. If you do that for the PCA, you need to take into account that PCA axes can flip (change sign). Also, you may want to consider models equal that are just rotated in the 3 dimensions you keep.
  • for a PCR, I'd look at the final PCR coefficients (those that transform the original data directly into the prediction of the dependent variable): that saves you all the hassle of aligning equivalent PCA projections. (Same for PLS)
$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.