Github user hirakendu commented on a diff in the pull request:
https://github.com/apache/spark/pull/79#discussion_r10926671
--- 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 Manish.
Numerical stability is the first thing that comes to mind on seeing a large
`avg = sum/count` calculation. In practice, I haven't seen any significant
difference in results or overflows with even billion sample datasets. Also,
features in machine learning are typically normalized and dynamic range is
small (bounded away from 0 and infinity).
We definitely cannot use the methods in DoubleRDDFunctions because we want
to calculate the variance of various splits, which requires the stats to be
"aggregable". But we may be able to modify the api's to use (count, avg,
avgSquares) as the stats and make the calculations more stable. E.g., to merge
(count, avg) of two parts `(c1, a1)`, `(c2, a2)`, we would have `(c1 + c2, a1 *
(c1/(c1+c2)) + a2 * (c2/(c1+c2)))`. Not too keen on that change, but let me
know if that works.
---
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.
---
