Hi. I would like to get assistance with my Random Forest graphic which was obtained from environment sampling. I am struggling with understanding why the least abundant genus in the systems are at the top of the rank while the most of the abundant genus are shown to be less important in the system?
I would also like to get more clarity on how to go about interpreting this model in the best way.
https://towardsdatascience.com/representing-a-random-forest-with-few-decision-trees-11283e433dbb
Dear Martha Mogale
These articles and links might be a valuable asset, have a look:
1. Article Adaptive random forests for evolving data stream classification
2. https://livebook.manning.com/book/r-in-action-second-edition/chapter-17/
3. https://towardsdatascience.com/representing-a-random-forest-with-few-decision-trees-11283e433dbb
4. https://www.mdpi.com/1099-4300/23/1/18/pdf
5. https://www.ajol.info/index.php/ijest/article/view/199714/188276
Kind Regards
Qamar Ul Islam
In a nutshell, the mean decrease in Gini coefficient is a measure of the importance of a feature, and the relative abundance is a measure of the frequency of occurrence. Both of these are measuring different and unrelated attributes of each feature.
The Mean Decrease in Gini is a measure of the feature/variable's importance. Think of it as feedback from your model telling you how helpful each feature/variable was for the model to improve its prediction. So, this is largely a model based measure. If you had a different model, this metric/graph may be different. However, the relative abundance is a dataset based measure, describing your data and does not care about your model.
On an intuitive level, you may expect this to be the case where rarer/less common genus are more informative for your model. Think of it this way, if the same feature is common to all the examples, can I use it to differentiate between samples? The most discriminative feature may often be a feature that is rarer and discriminates one group from another.
In a nutshell, these 2 measures are indicators of different attribites (predictive importance versus occurrence), of different facets (model versus dataset). So there shouldn't be any expectation for both the measures to be similar.
A simple analogy can be raised from Natural Language Processing. Say you're trying to train a sentiment classifier from a set of reviews. What you'll find is that the words that occur most commonly in most reviews (like the, a , an, is, are, etc) are completely non-informative. However, the rare words (like amazing, adored, terrible, detested, etc) are very informative and discriminative. A good classifier will assign more importance to these rarer words in the reviews than the common words.
- Badges
- Science topic