This PR adds RandomForest to MLlib. The implementation is basic, and future performance optimizations will be important. (Note: RFs = Random Forests.)
# Overview
## RandomForest
* trains multiple trees at once to reduce the number of passes over the data
* allows feature subsets at each node
* uses a queue of nodes instead of fixed groups for each level
This implementation is based an implementation by manishamde and the [Alpine Labs Sequoia Forest](https://github.com/AlpineNow/SparkML2) by codedeft (in particular, the TreePoint, BaggedPoint, and node queue implementations). Thank you for your inputs!
## Testing
Correctness: This has been tested for correctness with the test suites and with DecisionTreeRunner on example datasets.
Performance: This has been performance tested using [this branch of spark-perf](https://github.com/jkbradley/spark-perf/tree/rfs). Results below.
### Regression tests for DecisionTree
Summary: For training 1 tree, there are small regressions, especially from feature subsampling.
In the table below, each row is a single (random) dataset. The 2 different sets of result columns are for 2 different RF implementations:
* (numTrees): This is from an earlier commit, after implementing RandomForest to train multiple trees at once. It does not include any code for feature subsampling.
* (feature subsets): This is from this current PR's code, after implementing feature subsampling.
These tests were to identify regressions in DecisionTree, so they are training 1 tree with all of the features (i.e., no feature subsampling).
These were run on an EC2 cluster with 15 workers, training 1 tree with maxDepth = 5 (= 6 levels). Speedup values < 1 indicate slowdowns from the old DecisionTree implementation.
numInstances | numFeatures | runtime (sec) | speedup | runtime (sec) | speedup
---- | ---- | ---- | ---- | ---- | ----
| | (numTrees) | (numTrees) | (feature subsets) | (feature subsets)
20000 | 100 | 4.051 | 1.044433473 | 4.478 | 0.9448414471
20000 | 500 | 8.472 | 1.104461756 | 9.315 | 1.004508857
20000 | 1500 | 19.354 | 1.05854087 | 20.863 | 0.9819776638
20000 | 3500 | 43.674 | 1.072033704 | 45.887 | 1.020332556
200000 | 100 | 4.196 | 1.171830315 | 4.848 | 1.014232673
200000 | 500 | 8.926 | 1.082791844 | 9.771 | 0.989151571
200000 | 1500 | 20.58 | 1.068415938 | 22.134 | 0.9934038131
200000 | 3500 | 48.043 | 1.075203464 | 52.249 | 0.9886505005
2000000 | 100 | 4.944 | 1.01355178 | 5.796 | 0.8645617667
2000000 | 500 | 11.11 | 1.016831683 | 12.482 | 0.9050632911
2000000 | 1500 | 31.144 | 1.017852556 | 35.274 | 0.8986789136
2000000 | 3500 | 79.981 | 1.085382778 | 101.105 | 0.8586123337
20000000 | 100 | 8.304 | 0.9270231214 | 9.073 | 0.8484514494
20000000 | 500 | 28.174 | 1.083268262 | 34.236 | 0.8914592826
20000000 | 1500 | 143.97 | 0.9579634646 | 159.275 | 0.8659111599
### Tests for forests
I have run other tests with numTrees=10 and with sqrt(numFeatures), and those indicate that multi-model training and feature subsets can speed up training for forests, especially when training deeper trees.
# Details on specific classes
## Changes to DecisionTree
* Main train() method is now in RandomForest.
* findBestSplits() is no longer needed. (It split levels into groups, but we now use a queue of nodes.)
* Many small changes to support RFs. (Note: These methods should be moved to RandomForest.scala in a later PR, but are in DecisionTree.scala to make code comparison easier.)
## RandomForest
* Main train() method is from old DecisionTree.
* selectNodesToSplit: Note that it selects nodes and feature subsets jointly to track memory usage.
## RandomForestModel
* Stores an Array[DecisionTreeModel]
* Prediction:
* For classification, most common label. For regression, mean.
* We could support other methods later.
## examples/.../DecisionTreeRunner
* This now takes numTrees and featureSubsetStrategy, to support RFs.
## DTStatsAggregator
* 2 types of functionality (w/ and w/o subsampling features): These require different indexing methods. (We could treat both as subsampling, but this is less efficient
DTStatsAggregator is now abstract, and 2 child classes implement these 2 types of functionality.
## impurities
* These now take instance weights.
## Node
* Some vals changed to vars.
* This is unfortunately a public API change (DeveloperApi). This could be avoided by creating a LearningNode struct, but would be awkward.
## RandomForestSuite
Please let me know if there are missing tests!
## BaggedPoint
This wraps TreePoint and holds bootstrap weights/counts.
# Design decisions
* BaggedPoint: BaggedPoint is separate from TreePoint since it may be useful for other bagging algorithms later on.
* RandomForest public API: What options should be easily supported by the train* methods? Should ALL options be in the Java-friendly constructors? Should there be a constructor taking Strategy?
* Feature subsampling options: What options should be supported? scikit-learn supports the same options, except for "onethird." One option would be to allow users to specific fractions ("0.1"): the current options could be supported, and any unrecognized values would be parsed as Doubles in [0,1].
* Splits and bins are computed before bootstrapping, so all trees use the same discretization.
* One queue, instead of one queue per tree.
CC: mengxr manishamde codedeft chouqin Please let me know if you have suggestions---thanks!
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Author: qiping.lqp <qiping.lqp@alibaba-inc.com>
Author: chouqin <liqiping1991@gmail.com>
Closes#2435 from jkbradley/rfs-new and squashes the following commits:
c694174 [Joseph K. Bradley] Fixed typo
cc59d78 [Joseph K. Bradley] fixed imports
e25909f [Joseph K. Bradley] Simplified node group maps. Specifically, created NodeIndexInfo to store node index in agg and feature subsets, and no longer create extra maps in findBestSplits
fbe9a1e [Joseph K. Bradley] Changed default featureSubsetStrategy to be sqrt for classification, onethird for regression. Updated docs with references.
ef7c293 [Joseph K. Bradley] Updates based on code review. Most substantial changes: * Simplified DTStatsAggregator * Made RandomForestModel.trees public * Added test for regression to RandomForestSuite
593b13c [Joseph K. Bradley] Fixed bug in metadata for computing log2(num features). Now it checks >= 1.
a1a08df [Joseph K. Bradley] Removed old comments
866e766 [Joseph K. Bradley] Changed RandomForestSuite randomized tests to use multiple fixed random seeds.
ff8bb96 [Joseph K. Bradley] removed usage of null from RandomForest and replaced with Option
bf1a4c5 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into rfs-new
6b79c07 [Joseph K. Bradley] Added RandomForestSuite, and fixed small bugs, style issues.
d7753d4 [Joseph K. Bradley] Added numTrees and featureSubsetStrategy to DecisionTreeRunner (to support RandomForest). Fixed bugs so that RandomForest now runs.
746d43c [Joseph K. Bradley] Implemented feature subsampling. Tested DecisionTree but not RandomForest.
6309d1d [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into rfs-new. Added RandomForestModel.toString
b7ae594 [Joseph K. Bradley] Updated docs. Small fix for bug which does not cause errors: No longer allocate unused child nodes for leaf nodes.
121c74e [Joseph K. Bradley] Basic random forests are implemented. Random features per node not yet implemented. Test suite not implemented.
325d18a [Joseph K. Bradley] Merge branch 'chouqin-dt-preprune' into rfs-new
4ef9bf1 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into rfs-new
61b2e72 [Joseph K. Bradley] Added max of 10GB for maxMemoryInMB in Strategy.
a95e7c8 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into chouqin-dt-preprune
6da8571 [Joseph K. Bradley] RFs partly implemented, not done yet
eddd1eb [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into rfs-new
5c4ac33 [Joseph K. Bradley] Added check in Strategy to make sure minInstancesPerNode >= 1
0dd4d87 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-spark-3160
95c479d [Joseph K. Bradley] * Fixed typo in tree suite test "do not choose split that does not satisfy min instance per node requirements" * small style fixes
e2628b6 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into chouqin-dt-preprune
19b01af [Joseph K. Bradley] Merge remote-tracking branch 'chouqin/dt-preprune' into chouqin-dt-preprune
f1d11d1 [chouqin] fix typo
c7ebaf1 [chouqin] fix typo
39f9b60 [chouqin] change edge `minInstancesPerNode` to 2 and add one more test
c6e2dfc [Joseph K. Bradley] Added minInstancesPerNode and minInfoGain parameters to DecisionTreeRunner.scala and to Python API in tree.py
306120f [Joseph K. Bradley] Fixed typo in DecisionTreeModel.scala doc
eaa1dcf [Joseph K. Bradley] Added topNode doc in DecisionTree and scalastyle fix
d4d7864 [Joseph K. Bradley] Marked Node.build as deprecated
d4dbb99 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dt-spark-3160
1a8f0ad [Joseph K. Bradley] Eliminated pre-allocated nodes array in main train() method. * Nodes are constructed and added to the tree structure as needed during training.
0278a11 [chouqin] remove `noSplit` and set `Predict` private to tree
d593ec7 [chouqin] fix docs and change minInstancesPerNode to 1
2ab763b [Joseph K. Bradley] Simplifications to DecisionTree code:
efcc736 [qiping.lqp] fix bug
10b8012 [qiping.lqp] fix style
6728fad [qiping.lqp] minor fix: remove empty lines
bb465ca [qiping.lqp] Merge branch 'master' of https://github.com/apache/spark into dt-preprune
cadd569 [qiping.lqp] add api docs
46b891f [qiping.lqp] fix bug
e72c7e4 [qiping.lqp] add comments
845c6fa [qiping.lqp] fix style
f195e83 [qiping.lqp] fix style
987cbf4 [qiping.lqp] fix bug
ff34845 [qiping.lqp] separate calculation of predict of node from calculation of info gain
ac42378 [qiping.lqp] add min info gain and min instances per node parameters in decision tree
0dc2b6361d