[WIP] first cut at LambdaMART

examples/applications/letor_with_lambdamart.py

@@ -0,0 +1,48 @@
+#!/usr/bin/env python
+"""
+================================
+Learning to Rank with LambdaMART
+================================
+
+LambdaMART is a state-of-the-art algorithm for learn to rank problems.
+This example uses the MQ2008 dataset and compares LambdaMART with
+linear regression and least squares gradient boosting.
+
+The MQ2008 dataset is available at the following links,
+download MQ2008.rar and extract the files to 'data' folder.
+
+https://research.microsoft.com/en-us/um/beijing/projects/letor/letor4download.aspx
+https://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ2008.rar
+"""
+
+# Author: Jacques Kvam <jwkvam@gmail.com>
+# License: BSD 3 clause
+
+from sklearn.datasets import load_svmlight_file
+from sklearn.ensemble import LambdaMART
+from sklearn.ensemble import GradientBoostingRegressor
+from sklearn.linear_model import LinearRegression
+from time import time
+
+X_train, y_train, group_train = load_svmlight_file('data/train.txt',
+                                                   query_id=True)
+X_valid, y_valid, group_valid = load_svmlight_file('data/vali.txt',
+                                                   query_id=True)
+
+X_train = X_train.todense()
+X_valid = X_valid.todense()
+
+# transform labels to NDCG relevance values
+y_train = 2**y_train - 1
+y_valid = 2**y_valid - 1
+
+t0 = time()
+lm = LambdaMART().fit(X_train, y_train, group=group_train)
+print("LambdaMART fit in %fs" % (time() - t0))
+gb = GradientBoostingRegressor().fit(X_train, y_train)
+lr = LinearRegression().fit(X_train, y_train)
+for reg in [lm, gb, lr]:
+    print("%s training score is %f" % (reg.__class__.__name__,
+          lm.loss_(y_train, reg.predict(X_train)[:, None], group_train)))
+    print("%s validation score is %f" % (reg.__class__.__name__,
+          lm.loss_(y_valid, reg.predict(X_valid)[:, None], group_valid)))

sklearn/ensemble/__init__.py

@@ -15,6 +15,7 @@
 from .weight_boosting import AdaBoostRegressor
 from .gradient_boosting import GradientBoostingClassifier
 from .gradient_boosting import GradientBoostingRegressor
+from .gradient_boosting import LambdaMART
 
 from . import bagging
 from . import forest
@@ -27,6 +28,7 @@
            "RandomTreesEmbedding", "ExtraTreesClassifier",
            "ExtraTreesRegressor", "BaggingClassifier",
            "BaggingRegressor", "GradientBoostingClassifier",
-           "GradientBoostingRegressor", "AdaBoostClassifier",
-           "AdaBoostRegressor", "bagging", "forest", "gradient_boosting",
+           "GradientBoostingRegressor", "LambdaMART",
+           "AdaBoostClassifier", "AdaBoostRegressor", "bagging",
+           "forest", "gradient_boosting",
            "partial_dependence", "weight_boosting"]