Interpretability of Non-linear DR Methods

[malikeh1375]

Hi,
I wonder if your interpretability technique is applicable to non-linear DR approaches such as t-SNE or UMAP.

[avrambardas]

Hello, It is applicable to non-linear DR techniques, as long as they can dimensionally reduce new arbitrary instances without the need to retrain them (retrain them on the whole dataset + the new instance). t-SNE for example, does not offer this functionality, since it can only dimensionally reduce data that has been trained on. Στις Δευ 9 Μαΐ 2022 στις 8:24 μ.μ., ο/η Malikeh Ehghaghi < ***@***.***> έγραψε:

…

Hi, I wonder if your interpretability technique is applicable to non-linear DR approaches such as t-SNE or UMAP. — Reply to this email directly, view it on GitHub <#1>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AI3HTS5TTD73ZR5KQAXS6FLVJFC6DANCNFSM5VO6QXFQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[malikeh1375]

Awesome! Thanks a lot for your quick response.

In my case, what I intend to do is to apply this approach to UMAP embeddings. I want to find the main differentiating features between the clusters of points appear after DR. I have a combined dataset of multiple labels and I want to show what features or feature categories are accountable for clustering the points within distinct clusters after dimension reduction. Is it applicable in this case?

Also, can you please name a couple of non-linear DR approaches which are generalizable to new instances?

[avrambardas]

The most common non-linear DR approach that comes to my mind is Kernel PCA with non-linear kernels ("poly", "rbf"...). Generally, the DR technique should provide a transform method in addition to the fit or fit_transform methods so you can just dimensionally reduce new unseen instances without the need to retrain the model all over again. I think UMAP provides a transform method so it should be applicable. Στις Δευ 9 Μαΐ 2022 στις 9:44 μ.μ., ο/η Malikeh Ehghaghi < ***@***.***> έγραψε:

…

Awesome! Thanks a lot for your quick response. In my case, what I intend to do is to apply this approach to UMAP embeddings. I want to find the main differentiating features between the clusters of points appear after DR. I have a combined dataset of multiple labels and I want to show what features or feature categories are accountable for clustering the points within distinct clusters after dimension reduction. Is it applicable in this case? Also, can you please name a couple of non-linear DR approaches which are generalizable to new instances? — Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AI3HTS4PGMGECLGIB2B6AILVJFMH7ANCNFSM5VO6QXFQ> . You are receiving this because you commented.Message ID: <avrambardas/Interpretable-Unsupervised-Learning/issues/1/1121449741@ github.com>

[malikeh1375]

Great! Thank you so much for your response.