Even if many algorithms are in principle general purpose, we know that testing the algorithm on different data set (that is on different problem) the accuracy can change a lot. The training hasn’t been done properly: consider for example datasets having high dimensionality, how can you be sure that your classifier hasn’t affected by overfitting? The classifier it self cannot solve such problem: e.
![ADA boost: a way to improve a weak classifier](https://cdn-ak-scissors.b.st-hatena.com/image/square/a688041aa05f26849c7c3f1a63a5d13ab31e2cdb/height=288;version=1;width=512/https%3A%2F%2Fblogger.googleusercontent.com%2Fimg%2Fb%2FR29vZ2xl%2FAVvXsEiUTGnXC0sZDjmTpK8A1Jeg-AyaUVxCK_Snv7ZIZ2TxTcPfWna-jYpZP08gDJORo3Co5yHC9S2DBLDh02nbK7o57HAZjOg-aDnNrb3pMRX2E8BxYturALKP108WlGVq-LD087j3S_3jEjM%2Fw1200-h630-p-k-no-nu%2FHyperplanesDivided.bmp)