[ML] Return total SHAP per feature as a new result type #1387
Conversation
|
For v7.9.0 only if we manage to have a Java parser implemented timely. |
There was a problem hiding this comment.
I made a couple of style suggestions. My main comments are I think:
- We should normalise by the document count. (You should be able to switch the map's value type to a
CBasicStatistics::SSampleMean<TVector>::TAccumulatorto do this. Although note you have to initialise this with a zero vector of the correct size.) - It would be nice to compute these quantities accurately, i.e. not just summing over the top importances for each document.
|
We agreed to defer this from 7.9 to 7.10, so I altered the labels. |
| writer.Double(item.second(0)); | ||
| } else { | ||
| for (int j = 0; j < item.second.size() && j < numberClasses; ++j) { | ||
| writer.Key(classValues[j]); |
There was a problem hiding this comment.
This will not work for storage in ES. This index stores the information for all trained models and indexing the class names for the feature importances will not scale.
I propose this format:
{
"feature_name": "c4",
"importance": 0.4810469375580312,
"class_importance": [
{
"class_name": "foo",
"importance": 0.24052346877901588
},
{
"class_name": "baz",
"importance": 0.19615020783390645
},
{
"class_name": "bar",
"importance": 0.04437326094510882
}
]
}
class_importance will be a nested data type that allows aggregations and searches for specific models and classnames.
|
Java side parsing: elastic/elasticsearch#59725 We will probably have to mute integration tests, merge C++ side, then unmute and merge the parsing java side. |
|
@valeriy42 the new format has |
@benwtrent I cannot think of anything more. |
|
@valeriy42 it would be nice if the per class importance looked like the regression importance. That way we have consistent JSON objects. |
|
@benwtrent you are right! I overlooked at in my past commit. I'll fix it accordingly. |
This PR add computation of the total feature importance values.
This PR add computation of the total feature importance values. Backport of #1387.
This updates the feature_importance mapping change from elastic/ml-cpp#1387
This updates the feature_importance mapping change from elastic/ml-cpp#1387
This updates the feature_importance mapping change from elastic/ml-cpp#1387
Activate the output of the model metadata and the corresponding unit tests for total feature importance. The implementation itself was introduced in #1387 however, I need to fix the documentation, it was originally attributed to v7.10. Hence, I mark this PR as enhancement to rectify the docs.
Activate the output of the model metadata and the corresponding unit tests for total feature importance. The implementation itself was introduced in #1387 however, I need to fix the documentation, it was originally attributed to v7.10. Hence, I mark this PR as enhancement to rectify the docs. Co-authored-by: Valeriy Khakhutskyy <1292899+valeriy42@users.noreply.github.com>
Activate the output of the model metadata and the corresponding unit tests for total feature importance. The implementation itself was introduced in elastic#1387 however, I need to fix the documentation, it was originally attributed to v7.10. Hence, I mark this PR as enhancement to rectify the docs.
This PR add computation of the total feature importance values and outputs it as a new result type.
Example outputs:
{ "row_results": { "checksum": 0, "results": { "ml": { "target_prediction": 1358.2039794921876, "is_training": true, "feature_importance": [ { "feature_name": "c1", "importance": 448.3700609767421 }, { "feature_name": "c2", "importance": 1146.2576129663276 }, { "feature_name": "c3", "importance": -388.896988459571 }, { "feature_name": "c4", "importance": 158.7185644712811 } ] } } } }, { "model_metadata": { "total_feature_importance": [ { "feature_name": "c4", "importance": { "mean_magnitude": 233.4600671221131, "min": -565.7664157184156, "max": 468.8953979253651 } }, { "feature_name": "c3", "importance": { "mean_magnitude": 227.52681995349807, "min": -474.187447119175, "max": 500.79764582176218 } }, { "feature_name": "c1", "importance": { "mean_magnitude": 479.8491325534919, "min": -584.6059620924166, "max": 601.3424189083114 } }, { "feature_name": "c2", "importance": { "mean_magnitude": 729.7375145579323, "min": -1438.469059491588, "max": 1428.738023747545 } } ] } }{ "row_results": { "checksum": 0, "results": { "ml": { "target_prediction": "foo", "prediction_probability": 0.9632688006864724, "prediction_score": 0.9632688006864724, "is_training": true, "feature_importance": [ { "feature_name": "c1", "classes": [ { "class_name": "foo", "importance": -0.050036180361561228 }, { "class_name": "bar", "importance": 0.050036180361561228 } ] }, { "feature_name": "c2", "classes": [ { "class_name": "foo", "importance": -2.787898169333443 }, { "class_name": "bar", "importance": 2.787898169333443 } ] }, { "feature_name": "c3", "classes": [ { "class_name": "foo", "importance": -0.9016447487592819 }, { "class_name": "bar", "importance": 0.9016447487592819 } ] }, { "feature_name": "c4", "classes": [ { "class_name": "foo", "importance": 0.4345632399908005 }, { "class_name": "bar", "importance": -0.4345632399908005 } ] } ] } } } }, { "model_metadata": { "total_feature_importance": [ { "feature_name": "c4", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 0.5077140893490363, "min": -1.2245953772608847, "max": 1.2245953772608847 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.5077140893490363, "min": -1.2245953772608847, "max": 1.2245953772608847 } } ] }, { "feature_name": "c3", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 0.37436172343432769, "min": -1.3221622827321056, "max": 1.3221622827321056 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.37436172343432769, "min": -1.3221622827321056, "max": 1.3221622827321056 } } ] }, { "feature_name": "c1", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 1.0116256529234005, "min": -2.4239089033397378, "max": 2.4239089033397378 } }, { "class_name": "bar", "importance": { "mean_magnitude": 1.0116256529234005, "min": -2.4239089033397378, "max": 2.4239089033397378 } } ] }, { "feature_name": "c2", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 1.878800695094461, "min": -3.288343526748284, "max": 3.288343526748284 } }, { "class_name": "bar", "importance": { "mean_magnitude": 1.878800695094461, "min": -3.288343526748284, "max": 3.288343526748284 } } ] } ] } }{ "row_results": { "checksum": 0, "results": { "ml": { "target_prediction": "foo", "prediction_probability": 0.9462761477876273, "prediction_score": 0.17141987121246278, "is_training": true, "top_classes": [ { "class_name": "foo", "class_probability": 0.9462761477876273, "class_score": 0.17141987121246278 }, { "class_name": "bar", "class_probability": 0.034511190424692039, "class_score": 0.034511190424692039 }, { "class_name": "baz", "class_probability": 0.019212661787680529, "class_score": 0.003818541178138616 } ], "feature_importance": [ { "feature_name": "c1", "classes": [ { "class_name": "foo", "importance": 0.27949101986590138 }, { "class_name": "baz", "importance": -0.12386717688503159 }, { "class_name": "bar", "importance": -0.1556238429808595 } ] }, { "feature_name": "c2", "classes": [ { "class_name": "foo", "importance": 1.663225619115169 }, { "class_name": "baz", "importance": -1.72288680107119 }, { "class_name": "bar", "importance": 0.05966118195592636 } ] }, { "feature_name": "c3", "classes": [ { "class_name": "foo", "importance": 0.22379061504358678 }, { "class_name": "baz", "importance": -0.23352199623126147 }, { "class_name": "bar", "importance": 0.00973138118767497 } ] }, { "feature_name": "c4", "classes": [ { "class_name": "foo", "importance": -0.24052346877901588 }, { "class_name": "baz", "importance": 0.19615020783390645 }, { "class_name": "bar", "importance": 0.04437326094510882 } ] } ] } } } }, { "model_metadata": { "total_feature_importance": [ { "feature_name": "c4", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 0.24209656030374336, "min": -0.5757885311922144, "max": 0.6352558320805585 } }, { "class_name": "baz", "importance": { "mean_magnitude": 0.21362926518754464, "min": -0.6975561926823535, "max": 0.5758437812831863 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.0346461346585683, "min": -0.15232748182282736, "max": 0.10645868140524567 } } ] }, { "feature_name": "c3", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 0.30818045910633587, "min": -0.7931796597779941, "max": 0.3339785961510332 } }, { "class_name": "baz", "importance": { "mean_magnitude": 0.3302457015751672, "min": -0.45783991999546966, "max": 0.8242004300074223 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.05158362329758712, "min": -0.31757994080335757, "max": 0.2435538443329867 } } ] }, { "feature_name": "c1", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 0.6477535654120877, "min": -1.9137505247875509, "max": 1.287337819860563 } }, { "class_name": "baz", "importance": { "mean_magnitude": 0.7520521962038734, "min": -1.531931414792879, "max": 1.6810760229277138 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.16574602557111424, "min": -0.3434321409380257, "max": 0.4223094459672269 } } ] }, { "feature_name": "c2", "classes": [ { "class_name": "foo", "importance": { "mean_magnitude": 1.101426013839722, "min": -2.2925533638349937, "max": 1.7987193407562752 } }, { "class_name": "baz", "importance": { "mean_magnitude": 1.1717215530182037, "min": -1.8971843995291227, "max": 2.6284289404335188 } }, { "class_name": "bar", "importance": { "mean_magnitude": 0.18929675306403358, "min": -0.43649354351400246, "max": 0.612025928728523 } } ] } ] } }Closes #974
EDIT: I updated the format example above.