Github user manishamde commented on a diff in the pull request:
https://github.com/apache/spark/pull/79#discussion_r10959264
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/tree/impurity/Impurity.scala ---
@@ -0,0 +1,42 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.mllib.tree.impurity
+
+/**
+ * Trail for calculating information gain
+ */
+trait Impurity extends Serializable {
+
+ /**
+ * information calculation for binary classification
+ * @param c0 count of instances with label 0
+ * @param c1 count of instances with label 1
+ * @return information value
+ */
+ def calculate(c0 : Double, c1 : Double): Double
+
+ /**
+ * information calculation for regression
+ * @param count number of instances
+ * @param sum sum of labels
+ * @param sumSquares summation of squares of the labels
+ * @return information value
+ */
+ def calculate(count: Double, sum: Double, sumSquares: Double): Double
--- End diff --
I agree with @mengxr that the design needs to be made more generic for
extensibility -- the signature collision is an issue we will encounter soon. As
@mengxr mentioned, the impurity class should specify three methods: a) a method
to calculate impurity (or an intermediate value) for a single label, b) the
```binSeqOp``` method and c) the ```binCombOp``` method.
We will encounter a loss in performance if we calculate the stats per
instance and then "merge" them and a loss in precision (as @mengxr pointed out)
if we calculate stats after accumulating large numbers. I wonder whether the
sweet spot is to calculating the stats (impurity) per partition without
encountering a significant loss in performance and precision.
The important question is that whether this issue "blocks" this PR or
belongs to a separate PR. I vote for the postponing it to a future PR. We will
soon encounter more loss functions during ensemble implementations (GBT for
example) so it might be good to handle it then.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
