Github user etrain commented on a diff in the pull request:
https://github.com/apache/spark/pull/79#discussion_r10934747
--- 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 that overflow is an issue here (particularly in the case of
sumSquares), but also agree with Manish/Hirakendu that this algorithm maintains
its ability to generate a tree in a reasonable amount of time based on this
property that we compute statistics for splits and then merge them together.
I actually do think it makes sense to maintain "(count, average,
averageSumSq)" for each partition in a way that's overflow friendly and compute
the combination as count-weighted average of both as Hirakendu suggests. This
will complicate the code but should solve the overflow problem and keep things
pretty efficient.
That said - maybe this could be taken care of in a future PR as a bugfix,
rather than in this one?
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
