many authors use perplexity/entropy to validate their model but I'm not fully satisfied with this. Again some author use topic coherence (Pointwise Mutual information). Can anyone suggest most accurate method to test topic model?
Here are some suggested ways to validate a topic model:
1. Use the Generalized Fowlkes-Mallows Index (requires computing the probability of intersecting events).
2. Partial class match precision (measure the probability of randomly selecting two documents from the same class taken from a randomly sampled cluster).
3. Clusterig recall (probability that a relevabt document is retrieved).
4. Single metric performance
For more about this, see the attached pdf file.
Here are some suggested ways to validate a topic model:
1. Use the Generalized Fowlkes-Mallows Index (requires computing the probability of intersecting events).
2. Partial class match precision (measure the probability of randomly selecting two documents from the same class taken from a randomly sampled cluster).
3. Clusterig recall (probability that a relevabt document is retrieved).
4. Single metric performance
For more about this, see the attached pdf file.
- Badges
- Science topic