Abstract In one of our earlier studies we noticed how straightforward cleaning of our medical data set impaired its classification results considerably with some machine learning methods, but not all of them, unexpectedly and against intuition compared to the original situation without any data cleaning. After a more precise exploration of the data, we found that the reason was the complicated variable distribution of the data although there were only two classes in it. In addition to a straightforward data cleaning method, we used an efficient way called neighbourhood cleaning that solved the problem and improved our classification accuracies 5–10%, at their best, up to 95% of all test cases. This shows how important it is first very carefully to study distributions of data sets to be classified and use different cleaning techniques in order to obtain best classification results.
- 113 Tietojenkäsittely- ja informaatiotieteet