[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15965244#comment-15965244 ] Yan Facai (颜发才) commented on SPARK-10788: - [~josephkb] As categories A, B and C are independent, why not collect statistics only for cateogry? Splits are calculated in the last step in `binsToBestSplit`. So communication cost is N bins. > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley >Assignee: Seth Hendrickson >Priority: Minor > Fix For: 2.0.0 > > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.15#6346) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14990576#comment-14990576 ] Apache Spark commented on SPARK-10788: -- User 'sethah' has created a pull request for this issue: https://github.com/apache/spark/pull/9474 > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley >Priority: Minor > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14940345#comment-14940345 ] Joseph K. Bradley commented on SPARK-10788: --- Though I should say: I should probably put this as Minor priority. It's not a huge savings, and it's likely a somewhat complex change. If you have other things you're working on, I'd prioritize those instead. > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14940343#comment-14940343 ] Joseph K. Bradley commented on SPARK-10788: --- OK, thanks! > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14940316#comment-14940316 ] Seth Hendrickson commented on SPARK-10788: -- Yes, much clearer. I can work on this task. > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14940196#comment-14940196 ] Joseph K. Bradley commented on SPARK-10788: --- Updated. Does it make more sense now? > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) communicate twice as much > data as needed for unordered categorical features. Here's an example. > Say there are 3 categories A, B, C. We consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we collect statistics for each of the 6 subsets of categories (3 * > 2 = 6). However, we could instead collect statistics for the 3 subsets on > the left-hand side of the 3 possible splits: A and A,B and A,C. If we also > have stats for the entire node, then we can compute the stats for the 3 > subsets on the right-hand side of the splits. In pseudomath: {{stats(B,C) = > stats(A,B,C) - stats(A)}}. > We should eliminate these extra bins within the spark.ml implementation since > the spark.mllib implementation will be removed before long (and will instead > call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14940186#comment-14940186 ] Joseph K. Bradley commented on SPARK-10788: --- Reading what I wrote now, I realize I didn't actually phrase it correctly. I'll update the description. > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) effectively creates a second > copy of each split. E.g., if there are 3 categories A, B, C, then we should > consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we also consider the 3 flipped splits: > * B,C vs. A > * C vs. A, B > * B vs. A, C > This means we communicate twice as much data as needed for these features. > We should eliminate these duplicate splits within the spark.ml implementation > since the spark.mllib implementation will be removed before long (and will > instead call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-10788) Decision Tree duplicates bins for unordered categorical features
[ https://issues.apache.org/jira/browse/SPARK-10788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14936219#comment-14936219 ] Seth Hendrickson commented on SPARK-10788: -- [~josephkb] I'm interested in working on this issue, but I'm not sure I see the problem. Looking through ML RandomForest implementation I found that {{numBins}} for unordered features is {{def numUnorderedBins(arity: Int): Int = 2 * ((1 << arity - 1) - 1)}} and that {{numSplits}} is just {{numBins / 2}}. In the 3 category example: {{numBins = 2 * (( 1 << (3 - 1)) - 1) = 6}} and so the number of splits considered is {{numSplits = 6 / 2 = 3}}. This seems to be the same as in the MLlib implementation. Perhaps I am overlooking something. I'd appreciate any feedback... > Decision Tree duplicates bins for unordered categorical features > > > Key: SPARK-10788 > URL: https://issues.apache.org/jira/browse/SPARK-10788 > Project: Spark > Issue Type: Improvement > Components: ML >Reporter: Joseph K. Bradley > > Decision trees in spark.ml (RandomForest.scala) effectively creates a second > copy of each split. E.g., if there are 3 categories A, B, C, then we should > consider 3 splits: > * A vs. B, C > * A, B vs. C > * A, C vs. B > Currently, we also consider the 3 flipped splits: > * B,C vs. A > * C vs. A, B > * B vs. A, C > This means we communicate twice as much data as needed for these features. > We should eliminate these duplicate splits within the spark.ml implementation > since the spark.mllib implementation will be removed before long (and will > instead call into spark.ml). -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org