Bootstrap Aggregation #229
Conversation
-Added an example using bootstrap aggregation to carry out Random Forest classification. -Factored the linfa trait FromTargetArray to create an additional trait FromTargetArrayOwned.
Codecov ReportBase: 55.44% // Head: 55.11% // Decreases project coverage by
Additional details and impacted files@@ Coverage Diff @@
## master #229 +/- ##
==========================================
- Coverage 55.44% 55.11% -0.34%
==========================================
Files 95 97 +2
Lines 8774 9014 +240
==========================================
+ Hits 4865 4968 +103
- Misses 3909 4046 +137
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
| } | ||
|
|
||
| pub struct EnsembleLearnerParams<P> { | ||
| pub ensemble_size: usize, |
There was a problem hiding this comment.
Why do we need a separate field for ensemble_size? Isn't this value implied by bootstrap_proportion?
There was a problem hiding this comment.
ensemble_size gives the number of models in the ensemble while bootstrap_proportion gives the proportion of the total number of training samples that should be given to each model for training. These should be distinct parameters.
There was a problem hiding this comment.
Shouldn't bootstrap_proportion be the same as 1/ensemble_size?
There was a problem hiding this comment.
Not necessarily, each model in the ensemble just needs its own random set of samples of training data from the complete training data set. There are no constraints on the size of this set other than it being non-empty, so we let the user tune this size as a hyperparameter.
There was a problem hiding this comment.
OK so bootstrap_samples just grabs random sets of samples from the input and yields them infinitely. I thought it divided the input into random subsamples. This makes sense now.
There was a problem hiding this comment.
Can you add this behaviour to the docs, along with a general description of EnsembleLearner? We should also have top level docs in src/lib.rs like with the other crates.
| let aggregated_predictions = self.aggregate_predictions(&mut predictions); | ||
|
|
||
| for (target, output) in y_array.axis_iter_mut(Axis(0)).zip(aggregated_predictions.into_iter()) { | ||
| for (t, o) in target.into_iter().zip(output[0].0.iter()) { | ||
| *t = *o; | ||
| } | ||
| } |
There was a problem hiding this comment.
Replace with this:
// prediction map has same shape as y_array, but the elements are maps
let mut prediction_maps = y_array.map(|_| HashMap::new());
for prediction in predictions {
let p_arr = prediction.as_targets();
assert_eq!(p_arr.shape(), y_array.shape());
// Insert each prediction value into the corresponding map
Zip::from(&mut prediction_maps).and(&p_arr).for_each(|(&mut map, &val)| map.entry(val).or_insert(0) += 1);
}
// For each prediction, pick the result with the highest number of votes
y_array = prediction_maps.mapv_into(|map| map.iter().max_by_key(|(_, v)| v).0);It picks out the predictions with the highest number of votes without the complexity of aggregate_predictions
There was a problem hiding this comment.
This comment still applies I believe.
|
Hi! |
|
Merge/rebase with the latest master and address the open review comments. That's pretty much it. |
Following on from the discussion in issue #199 here is an implementation of Bootstrap Aggregation following the approach laid out there. There is also an example showing how the provided methods can be used together with the existing linfa-trees package to implement a Random Forest.
Outside of the linfa-ensemble package the only required change was to factor the trait FromTargetArray to create an additional trait FromTargetArrayOwned. This was done to allow owned data to be used without needing to specify a lifetime, which was impossible under the previous FromTargetArray trait.