Notes on: He, X., Pan, J., Jin, O., Xu, T., Liu, B., Xu, T., Shi, Y., … (2014): Practical lessons from predicting clicks on ads at facebook

Table of Contents

1 Notation

  • \(\mathbf{x}\) is the feature-transformed ad impression

2 Q & A

2.1 TODO Does it matter if we use \(\{ -1, 1 \}\) or \(\{ 0, 1 \}\) for labels?

  • In the paper they use the sign of \(y\) to provide the gradient, so \(-1\) would the negatively weight the negatives samples