How could t-SNE be a good dimensionality reduction approach when talking about feature engineering while it is only meant for visualization purposes? It is also not reliable in maintaining the structure/similarity in the data, so how could it be useful if it’s used to summarize a group of features that are supposed to represent important characteristics in order for the models to learn?
Is there any clarification on how it is meant to be used in this context?
Slide 28 week 2
Thanks in advance!
There are situations where t-SNE can be indirectly useful in the context of feature engineering, especially when dealing with high-dimensional data or exploring data relationships. Imagine you want to visualize and explore high-dimensional data, so t-SNE can reduce the dimensions primarily for visualization purposes. It helps reveal clusters or patterns within the data, which can indirectly inform feature engineering decisions.
I hope it makes sense for you now and feel free to ask for more clarifications.
Thank you for your response!
While I completely agree that it is precious for visualizations, it can be misleading when learning the relationships between features. Its nature makes it reliable on parameter tuning, hence different runs can show different data relations. It can expand some data and compress others at the same time, which makes it unreliable for determining the ground truth structure.
Here’s a link which explains the point better.
How to Use t-SNE Effectively (distill.pub)
Thanks for bringing this up. The staff have been notified regarding this.
You are correct. T-SNE is stochastic and so can’t be reliably used for feature engineering. It’s better suited for data visualization.