elasticsearch-learning-to-rank
elasticsearch-learning-to-rank copied to clipboard
o19s
XGBoostJsonParser not working well with 'binary features'
The current setup of the plugin requires a feature map to be used for creating serialized xgboost json model file (for an example of feature map see this).
In the feature map, each feature can be assigned 3 possible data types: q (quantitate), i (binary) and int (integer).
When the data type is int or q, each split node will be serialized to look like below:
{ "nodeid": 6, "depth": 2, "split": "f1", "split_condition": 5, "yes": 13, "no": 14, "missing": 14, "children": [
{ "nodeid": 13, "leaf": 0.000920585 },
{ "nodeid": 14, "leaf": -0.044742 }
]}
However, when data type is i, each split node would look like this after serialization:
{ "nodeid": 4, "depth": 2, "split": "f2", "yes": 9, "no": 10, "children": [
{ "nodeid": 9, "leaf": 0.138548 },
{ "nodeid": 10, "leaf": -0.0143873 }
]}
Basically, there will be no field for 'missing' and 'split_condition'.
The current XGBoostJsonParser though, explicitely checks for existence of split conditions and therefore throws exceptions when parsing binary nodes. (The code below is copied from here:)
boolean splitHasAllFields() {
return nodeId != null && threshold != null && split != null && leftNodeId != null && rightNodeId != null && depth != null
&& children != null && children.size() == 2;
}
What I suggest for the fix:
- In the short term, define all binary features into integer features and notify users of this limitation somewhere in the documentation.
- In the long run, revise splitHasAllFields() to account for the data type of the split nodes, or just eliminate the check on split conditions and threshold, or provide default values for binary split nodes.
Hey @o19s-admin, This seems to be an old issue, do we have an update/fix on it? If not, then it must be at least mentioned somewhere in the docs that boolean may not be supported and the training script must take this into consideration.
Is this the demo you are looking for? http://es-learn-to-rank.labs.o19s.com/
Branches for older versions of elastic are kept, so if you ever need "older" material that is a good way to find it. Aside from that we want to improve the documetation around xgboost models and would appreciate any help from the community
