The impact of data shape

Are different data-dependent algorithms exploiting the same properties of datasets?


The plot above compares the error rates of two data-dependent algorithms (a primary and secondary) across a variety of dataset shapes. The dataset shapes on the x-axis are arranged (left to right) in ascending order by the error rate of the primary algorithm. This allows for easy comparison with the error of the secondary algorithm, which in most cases are not correlated with that of the primary algorithm.

These results show that current data-dependent algorithms are exploiting different features of the input data; it's not the case that there are universally "easy" or "hard" datasets for the data-dependent algorithms.