Re-add mean centering and bias init #45
Conversation
Codecov Report
@@ Coverage Diff @@
## master #45 +/- ##
==========================================
+ Coverage 71.84% 73.35% +1.50%
==========================================
Files 28 28
Lines 1911 1944 +33
==========================================
+ Hits 1373 1426 +53
+ Misses 538 518 -20
Continue to review full report at Codecov.
|
| double glob_mean = 0; | ||
| for (size_t ix = 0; ix < ConfCSC.nnz; ix++) | ||
| glob_mean += (ConfCSC.values[ix] - glob_mean) / (double)(ix+1); | ||
| #pragma omp simd |
There was a problem hiding this comment.
we need to add conditional #ifdef _OPENMP as it won't compile on certain platforms
| @@ -168,7 +168,7 @@ T als_explicit(const dMappedCSC& Conf, | |||
| if(lambda > 0) { | |||
| if (with_biases) { | |||
| auto X_no_bias = X(arma::span(1, X.n_rows - 1), arma::span::all); | |||
| auto Y_no_bias = X(arma::span(1, Y.n_rows - 1), arma::span::all); | |||
| auto Y_no_bias = Y(arma::span(1, Y.n_rows - 1), arma::span::all); | |||
| arma::Col<T>& user_bias, | ||
| arma::Col<T>& item_bias, | ||
| T lambda, bool non_negative) { | ||
| double initialize_biases(const dMappedCSC& ConfCSC, |
There was a problem hiding this comment.
as we modify ConfCSC and ConfCSR in-place we need to remove const
| arma::Col<T>& user_bias, | ||
| arma::Col<T>& item_bias, | ||
| T lambda, bool non_negative) { | ||
| /* Robust mean calculation */ |
There was a problem hiding this comment.
Where can I get more information about "Robust mean"? Why is this better than just standard mean?
There was a problem hiding this comment.
Here's some discussion: https://stackoverflow.com/questions/7552443/whats-the-numerically-best-way-to-calculate-the-average
It gets a mean estimation which is correct to higher numerical precision that sum(x)/n. But it's not better than R's mean function if that's what you mean to ask (which uses a compensated summation).
In this case the sum is expected to be over a potentially very large number of elements, so a simpler mean calculation can lose precision.
This PR re-adds mean centering and better bias initialization for the explicit-feedback model.
However, I see that for the loss calculation, it takes a full row of all-ones in the factors into account, which should be excluded (that is, the matrix being optimized for should have 1 more row added into the regularized loss than the other).