For crisp-labeled data there are many online repositories (UCI being one such example). I have recently been working on some approaches for fuzzy data classification, yet I lack the data to run the experiments on.
I could, of course, "fuzzify" the labels of crisp-labeled datasets artificially in some way, but such results would probably not be considered very relevant, as the fuzziness would be merely artificially introduced.
I am sure that there are some datasets with inherent class uncertainty out there, but I was unable to locate the proper resources.
Do you know where such benchmark data can be found?
.
I have not seen such benchmark data but this may not be so surprising as the "fuzzification" step (of the input and/or of the output) is an inherent part of the fuzzy approach to data-mining ! Data do not grow "fuzzy" or "uncertain" naturally ;)
Maybe you could consider adding well documented fuzzification steps to some publicly available ordinal datasets such as those provided to weka by Dr. Arie Ben David
http://weka.wikispaces.com/Datasets
(both input and output data are ordinal)
.
Thank You for the answer. I suspected as much.
Though I must disagree on one point - some data is uncertain and fuzzy naturally :) Though in most cases it is not really "fuzziness" but rather a probability distribution defined on the label set. It all depends on the problem context.
Additionally, "fuzziness" can be introduced either by experts or collaborative efforts of a number of people doing the labeling, it does is not necessarily an artifact of the algorithm.
However, I will probably proceed as You've suggested, by "fuzzifying" crisp data and learning from that, as it is probably sufficient for some initial research objectives.
Do You know which ways of fuzzifying the labels might be considered standard? The one that I've been meaning to employ is to observe the k-nearest neighbors and use the label proportions in the k-neighbor sets as fuzzy measures. I know this was done in the original FNN (fuzzy k-nearest neighbor) classification algorithm.
Are there some other standard/well known methods for doing this?
.
http://christianthiel.com/publications/FuzzyLVQ.pdf
describes two fuzzification methods, including the one you mentionned
see also
http://christianthiel.com/publications/Classification_on_Noised_Labels.pdf
as a side note, it seems that we do not use the same definition of "fuzzy" : if 10 persons label an item, 3 with label A and 7 with label B, the labelling is for sure "uncertain". It is still a technical matter to decide if the fuzzy membership is 3/10 for A or any other function F(3, 7) (with some natural conditions on F ...) : uncertainty may well be everywhere in nature, fuzzy or probabilistic memberships are indeed technical artifacts
.
- Badges
- Science topic