New methods to compare with:
- Sneyd and Stevenson, SIGIR 2021. Only implement IP-P, code is available here. Should be pretty easy to do.
- Lewis et al., CIKM 2021. They propose Quantile Point Estimate Threshold (QPET), and Quantile Binomial Confidence Bound (QBCB). Code is not available :(. I would only implement QBCB, since it's also the one the authors mostly experiment with. Asked E. Yang for the code, if we cannot have it/republish it I would honestly only implement Sneyd and Stevenson's.
Active learning:
- One of the reviewers asked to test ALvUS as well. This is not relevant for TAR, but could still be useful, since
$\mathrm{SAL}_{\tau}$ can work in any active learning context.