blocks' weights calculation in multi-block setting

Create issue
Issue #83 resolved
Arnaud created an issue

Hello,

Can you please explain how the weights of the blocks in a multi-block setting are calculated ?

In the manual page 134 ("predict" function) it is stated that "the weights of the blocks are calculated as the correlation between a block's components and the outcome's components". However there are typically several components whereas there's only 1 weight value per block.

Is it something like the sum of all correlations between the blocks' components and the outcome's components ? Or the sum of the squared previous correlations ?

Thanks in advance, Arnaud

Comments (6)

  1. Florian Rohart

    Hi Arnaud,

    The weight of a block is calculated as the average of the absolute values of all correlations (over 3 correlations if you have 3 components). That way we get a weight between 0 and 1.

    Hoping it's clearer!

    Florian

  2. Arnaud reporter

    Thank you Florian.

    However isn't it the case that if there are 3 components the average is over 9 correlations (the 3 components of the X block * the 3 components of the outcome Y) ?

    Thanks in advance, Arnaud

  3. Florian Rohart

    Actually, we only calculate the correlation between the same components: variate_i of the X block with variate_i of the outcome Y. So there are as many correlations as the number of components.

  4. Log in to comment