18-10-2025, 11:37 AM
(This post was last modified: 18-10-2025, 11:55 AM by quimqu.
Edit Reason: added PCA plots
)
OK, so after all the morning investigating, finally I got it clear.
First: headlines topic distribution finally scored better NMF model, which gives discrete values for each topic. LDA topic distribution (the one that scored better in the Voynich) gives a percent on each topic, that adds up to 1. That's why both plots were so different (clouds for headlines NMF, sneaky fiures for Voynich LDA)
Remember that the plots indicate the topic distribution per paragraph, and that the topic distribution is given by the model. So for K=2, UMAP is exploding the differences of a single line (from distributions (0,1), (0.01,0.99)... (1,0), accentuating the differences of topic distribution.
To understand it better, I recalculated everything to K=3 topics with LDA (the second best score). For K=3 the vectors would look like (0,0.99,0.01) and UMAP reduces them to a bidimensional vector. I keep UMAP because it helps visualize the distribution of the topics assigned in my opinion more clearly as PCA. For this post I changed a bit the parameters so they are better separated. I post also PCA plots, so you can see the differences ploting UMAP vs. PCA.
UMAP for K=2 (model finds 2 topics) and by paragraph:
PCA for K=2 by paragraph:
UMAP for K=3 (model finds 3 topics) and by paragraph:
PCA for K=3 by paragraph:
First: headlines topic distribution finally scored better NMF model, which gives discrete values for each topic. LDA topic distribution (the one that scored better in the Voynich) gives a percent on each topic, that adds up to 1. That's why both plots were so different (clouds for headlines NMF, sneaky fiures for Voynich LDA)
Remember that the plots indicate the topic distribution per paragraph, and that the topic distribution is given by the model. So for K=2, UMAP is exploding the differences of a single line (from distributions (0,1), (0.01,0.99)... (1,0), accentuating the differences of topic distribution.
To understand it better, I recalculated everything to K=3 topics with LDA (the second best score). For K=3 the vectors would look like (0,0.99,0.01) and UMAP reduces them to a bidimensional vector. I keep UMAP because it helps visualize the distribution of the topics assigned in my opinion more clearly as PCA. For this post I changed a bit the parameters so they are better separated. I post also PCA plots, so you can see the differences ploting UMAP vs. PCA.
UMAP for K=2 (model finds 2 topics) and by paragraph:
PCA for K=2 by paragraph:
UMAP for K=3 (model finds 3 topics) and by paragraph:
PCA for K=3 by paragraph: