[DOCS] Add feature importance to classification example #1359
Conversation
lcawl
changed the title
lcawl
force-pushed
the
feature-importance
branch
from
8960791
to
4ba1595
Compare
| in your destination index. See the | ||
| {ml-docs}/flightdata-classification.html#flightdata-classification-results[Viewing {classification} results] | ||
| section in the {classification} example. | ||
| in your destination index. | ||
|
|
||
| [[dfa-classification-class-score]] | ||
| === `class_score` | ||
|
|
||
| The value of `class_score` controls the probability at which a class label is |
There was a problem hiding this comment.
class_score is definitely not a probability, since if I choose k very very small, class_score may be arbitrary large, while probability is always between 0 and 1. It's better to call it a "likelihood". And also it doesn't "control" it but simple "shows". It is controlled by the threshold k, which we estimate automagically based on class_assignment_objective configuration.
| values. A higher number means that the model is more confident. | ||
| If you want to understand how certain the model is about each prediction, you | ||
| can examine its probability and score (`ml.prediction_probability` and | ||
| `ml.prediction_score`). These values range between 0 and 1; the higher the |
There was a problem hiding this comment.
Strictly speaking, class_score can be larger than 1 in some degenerated cases. So it's defined as larger or equal to 0.
| //Does this mean the sum of the feature importance values for false in this | ||
| example should equal the logit(p), where p is the class_probability for false? |
There was a problem hiding this comment.
This is correct up to a constant. There is also a datapoint-independent constant -- an average log-odd overall all training points, which we add to the sum of feature importance before taking the inverse-logit to compute the probabilities.
| any class. | ||
| //Does this mean the sum of the feature importance values for false in this | ||
| example should equal the logit(p), where p is the class_probability for false? | ||
| //Does this imply that the feature importance value itself is the result of a logit function? Or that we use the function to merely represent the distribution of feature importance values? |
There was a problem hiding this comment.
What happens is that the decision forest predicted the log-odds directly and then we compute feature importance on the log-odds values. When we evaluate a data point, we take the log-odds predicted by the decision forest and then apply the inverse of the logit function to get the class probability.
| ==== | ||
| While the probability of a class ranges between 0 and 1, its log-odds range | ||
| between negative and positive infinity. In {kib}, the decision path for each | ||
| class starts near zero, which represents a class probability of 0.5. |
There was a problem hiding this comment.
This is unfortunately a bit more complicated: 0 would represent the class probability of a constant baseline. It relates to the average class probability for the selected class (in Kibana UI) over entire training set.
If you select Canceled as the target variable in flight data, this nuance becomes obvious. Since there are many more data points with Canceled = False, let's assume that the average class probability over the entire training set would be something like 0.92. This means that if the class probability of a data point is larger than 0.92 (for example 0.98) than the decision path will go to the right (sum of feature importances is positive). On the other hand, if the class probability is smaller than 0.92 (for example 0.84), than the decision path will go to the left (sum of feature importances is negative).
| While the probability of a class ranges between 0 and 1, its log-odds range | ||
| between negative and positive infinity. In {kib}, the decision path for each | ||
| class starts near zero, which represents a class probability of 0.5. | ||
| // Is this true for multi-class classification or just binary classification? |
There was a problem hiding this comment.
Yes, it's true for both multi-class and binary since in the UI you are selecting a class of interest from a drop-down menu.
lcawl
force-pushed
the
feature-importance
branch
2 times, most recently
from
5d8027e
to
691f8df
Compare
lcawl
force-pushed
the
feature-importance
branch
from
691f8df
to
cac81f1
Compare
lcawl
force-pushed
the
feature-importance
branch
from
ea7183a
to
6947f68
Compare
Related to elastic/kibana#73561
This PR drafts changes to the classification example such that it includes feature importance explanations.
It will be backported to 7.10 and does not take into consideration changes in 7.11 and later for elastic/kibana#77874
Preview
https://stack-docs_1359.docs-preview.app.elstc.co/guide/en/machine-learning/master/flightdata-classification.html