I'm performing Truncated SVD on a word co-occurrence matrix for topic extraction. I'm using the scikit-learn implementation.
I have read conflicting ideas on how to interpret the sign of the loadings in Truncated SVD. On the one hand, I've seen it suggested that a negative loading indicates a lack of that particular word in the topic/component. I've also seen it suggested that they are totally arbitrary.
My question is potentially two-fold:
- Is the sign meaningless? I.e., does it not indicate the lack of a word.
- If the sign is meaningless, how do I interpret and use the loadings to identify topics?