The dirty secret of ML research is almost all of it is shit (much worse than Sturgeon's law), and something like KNN with minor tweaks will almost always win.

The dirty secret of ML research is almost all of it is shit (much worse than Sturgeon's law), and something like KNN with minor tweaks will almost always win. I'm out in the trenches trying to turn ding dong papers into tools that does something useful, and the "muh mape 50% better" papers are effectively always something that should be ignored. Actually better algorithms and implementations (something like gradient boost machines/xgboost) are rare enough attempting to find them in the literature is futile.