Clustering analysis was done to an unlabeled dataset and then the clusters was used as label for supervised learning classification. The supervised learning produced high accuracy model. my question is, is this method appropriate and acceptable? I could not find any previous studies or research that did the same except blogpost or website. Thank you
Generally speaking - YES, it is good approach. For example, we use it, if classification data set has some missing data. But if accuracy of clustering is bad, final accuracy of classification also will be bad.
Hello Husnir,
Yes, the approach you described is acceptable. However, a few aspects shall be considered.
Such a procedure is usually performed for some part of the data for which labels are missed for some reason (you can then use for labelling the model that you have learned with the initial data).
You say that the final model is of high accuracy which means that the data undergpoing unsupervised learning (i.e., clustering) are very well-separable. You may want to have a deeper insight into that data because it may be that there is a variable which is very well correlated or even which is an indicator of the label. Let's say each sample has 3 coordinates, (xi, yi, zi), and by inspecting the data you may discover that, e.g., yi is highly correlated with the label (let's, again, say correlation of 0.95). If so, you shall consider if there is a real reason that yi may be the label itself and try using yi as the label or, even better, using only yi itself for classification and test how the results behave (in an extreme case, you may find both machine learning models unnecessary).
I think you can do that.
Deleted research item The research item mentioned here has been deleted
Dear Husnir Nasyuha , as the others already stated, this approach is totally fine. If you need more literature to back that up, here is a selection of articles discussing that topic for various applications:
https://www.scitepress.org/Papers/2018/69312/69312.pdf
Article Using clustering analysis to improve semi-supervised classification
https://www.nature.com/articles/s41598-018-24876-0
https://towardsdatascience.com/cluster-then-predict-for-classification-tasks-142fdfdc87d6
https://arxiv.org/pdf/1708.08591.pdf
Good luck with your research and happy modelling!
Annika Hamachers Ferdib Al Islam Michał Łepek Sergey Porotsky thank you all for your helpful answers!
Dear Husnir Nasyuha
These references are effective and useful for your answer
Graph-Based Semisupervised Learning
10.1109/TPAMI.2007.70765
Unsupervised and Semisupervised Learning
10.1002/9781118445112.stat08320
GOOD LUCK
Perhaps you could try the ARTMAP family of neural networks. They do exactly what you describe, and do it very effectively.
Indeed, you can use pseudo labels from clustering results to training a model for classifcation. There have been a number of studies to do this way. Nevertheless, to my knowledge, the performance by this way is often disappointed in many cases, though unsupervised learning may be a trend for the future work.
As others have stated, you can indeed use pseudo labels suggested by a clustering algorithm. But the performance of the whole model (unsupervised+supervised) is going to be largely dependent on the nature of the dataset and the performance of the clustering analysis. There are various methods for different types of clustering analysis performance measurement such as Silhouette's Coefficient, Dunn's Index, SSE, etc. You may proceed only if your clustering analysis is performing well.