Issues specific to unsupervised learning
The following are some issues that pertain to unsupervised learning techniques:
- Parameter setting: Deciding on number of features, usefulness of features, number of clusters, shapes of clusters, and so on, pose enormous challenges to certain unsupervised methods
- Evaluation methods: Since unsupervised learning methods are ill-posed due to lack of ground-truth, evaluation of algorithms becomes very subjective.
- Hard or soft labeling: Many unsupervised learning problems require giving labels to the data in an exclusive or probabilistic manner. This poses a problem for many algorithms
- Interpretability of results and models: Unlike supervised learning, the lack of ground truth and the nature of some algorithms make interpreting the results from both model and labeling even more difficult
