Publications
The finished work: dated, reviewed, with an author accountable for it and a statement of what it does not establish.
A publication here is not a longer note. It is the one layer where the build refuses to publish without a date, a review status and an accessibility status, because a conclusion nobody is accountable for should not be presented as reviewed work.
Everything else on this site is working material and says so. That is the distinction this page exists to make: not everything written is finished, and the finished set is deliberately small.
| Title | Summary | Kind | Date |
|---|---|---|---|
| Choosing a recommender by evaluating the evaluation | Three approaches were run against the same play-count dataset: a popularity baseline, user-user collaborative filtering, and item-item collaborative filtering. Both similarity models beat the baseline, by a margin small enough that the evaluation could not clearly distinguish them from it. The conclusion is that the metric was the problem: accuracy against held-out play counts rewards reproducing a concentrated distribution, which all three models can do. The recommendation is to fix the evaluation before choosing a model. | Case study |