See inline comment about the bug. I also did some code clean-up. dbtsai I moved `update` to a private method of `MultivariateOnlineSummarizer`. I don't think it will cause performance regression, but it would be great if you have some time to test.
Author: Xiangrui Meng <meng@databricks.com>
Closes#3220 from mengxr/SPARK-4355 and squashes the following commits:
5ef601f [Xiangrui Meng] fix OnlineSummarizer.merge when other.mean is zero and some code clean-up
SPARK-1553 added alternating nonnegative least squares to MLLib, however it's not possible to access it via the python API. This pull request resolves that.
Author: Michelangelo D'Agostino <mdagostino@civisanalytics.com>
Closes#3095 from mdagost/python_nmf and squashes the following commits:
a6743ad [Michelangelo D'Agostino] Use setters instead of static methods in PythonMLLibAPI. Remove the new static methods I added. Set seed in tests. Change ratings to ratingsRDD in both train and trainImplicit for consistency.
7cffd39 [Michelangelo D'Agostino] Swapped nonnegative and seed in a few more places.
3fdc851 [Michelangelo D'Agostino] Moved seed to the end of the python parameter list.
bdcc154 [Michelangelo D'Agostino] Change seed type to java.lang.Long so that it can handle null.
cedf043 [Michelangelo D'Agostino] Added in ability to set the seed from python and made that play nice with the nonnegative changes. Also made the python ALS tests more exact.
a72fdc9 [Michelangelo D'Agostino] Expose nonnegative ALS in the python API.
### Summary
* Made it easier to construct default Strategy and BoostingStrategy and to set parameters using simple types.
* Added Scala and Java examples for GradientBoostedTrees
* small cleanups and fixes
### Details
GradientBoosting bug fixes (“bug” = bad default options)
* Force boostingStrategy.weakLearnerParams.algo = Regression
* Force boostingStrategy.weakLearnerParams.impurity = impurity.Variance
* Only persist data if not yet persisted (since it causes an error if persisted twice)
BoostingStrategy
* numEstimators: renamed to numIterations
* removed subsamplingRate (duplicated by Strategy)
* removed categoricalFeaturesInfo since it belongs with the weak learner params (since boosting can be oblivious to feature type)
* Changed algo to var (not val) and added BeanProperty, with overload taking String argument
* Added assertValid() method
* Updated defaultParams() method and eliminated defaultWeakLearnerParams() since that belongs in Strategy
Strategy (for DecisionTree)
* Changed algo to var (not val) and added BeanProperty, with overload taking String argument
* Added setCategoricalFeaturesInfo method taking Java Map.
* Cleaned up assertValid
* Changed val’s to def’s since parameters can now be changed.
CC: manishamde mengxr codedeft
Author: Joseph K. Bradley <joseph@databricks.com>
Closes#3094 from jkbradley/gbt-api and squashes the following commits:
7a27e22 [Joseph K. Bradley] scalastyle fix
52013d5 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into gbt-api
e9b8410 [Joseph K. Bradley] Summary of changes
```
pyspark.mllib.stat.StatisticschiSqTest(observed, expected=None)
:: Experimental ::
If `observed` is Vector, conduct Pearson's chi-squared goodness
of fit test of the observed data against the expected distribution,
or againt the uniform distribution (by default), with each category
having an expected frequency of `1 / len(observed)`.
(Note: `observed` cannot contain negative values)
If `observed` is matrix, conduct Pearson's independence test on the
input contingency matrix, which cannot contain negative entries or
columns or rows that sum up to 0.
If `observed` is an RDD of LabeledPoint, conduct Pearson's independence
test for every feature against the label across the input RDD.
For each feature, the (feature, label) pairs are converted into a
contingency matrix for which the chi-squared statistic is computed.
All label and feature values must be categorical.
:param observed: it could be a vector containing the observed categorical
counts/relative frequencies, or the contingency matrix
(containing either counts or relative frequencies),
or an RDD of LabeledPoint containing the labeled dataset
with categorical features. Real-valued features will be
treated as categorical for each distinct value.
:param expected: Vector containing the expected categorical counts/relative
frequencies. `expected` is rescaled if the `expected` sum
differs from the `observed` sum.
:return: ChiSquaredTest object containing the test statistic, degrees
of freedom, p-value, the method used, and the null hypothesis.
```
Author: Davies Liu <davies@databricks.com>
Closes#3091 from davies/his and squashes the following commits:
145d16c [Davies Liu] address comments
0ab0764 [Davies Liu] fix float
5097d54 [Davies Liu] add Hypothesis test Python API
Author: Niklas Wilcke <1wilcke@informatik.uni-hamburg.de>
Closes#2907 from numbnut/master and squashes the following commits:
7f7c767 [Niklas Wilcke] [Spark-4060] [MLlib] exposing special rdd functions to the public, #2907
This PR simplify serializer, always use batched serializer (AutoBatchedSerializer as default), even batch size is 1.
Author: Davies Liu <davies@databricks.com>
This patch had conflicts when merged, resolved by
Committer: Josh Rosen <joshrosen@databricks.com>
Closes#2920 from davies/fix_autobatch and squashes the following commits:
e544ef9 [Davies Liu] revert unrelated change
6880b14 [Davies Liu] Merge branch 'master' of github.com:apache/spark into fix_autobatch
1d557fc [Davies Liu] fix tests
8180907 [Davies Liu] Merge branch 'master' of github.com:apache/spark into fix_autobatch
76abdce [Davies Liu] clean up
53fa60b [Davies Liu] Merge branch 'master' of github.com:apache/spark into fix_autobatch
d7ac751 [Davies Liu] Merge branch 'master' of github.com:apache/spark into fix_autobatch
2cc2497 [Davies Liu] Merge branch 'master' of github.com:apache/spark into fix_autobatch
b4292ce [Davies Liu] fix bug in master
d79744c [Davies Liu] recover hive tests
be37ece [Davies Liu] refactor
eb3938d [Davies Liu] refactor serializer in scala
8d77ef2 [Davies Liu] simplify serializer, use AutoBatchedSerializer by default.
Register MLlib's Vector as a SQL user-defined type (UDT) in both Scala and Python. With this PR, we can easily map a RDD[LabeledPoint] to a SchemaRDD, and then select columns or save to a Parquet file. Examples in Scala/Python are attached. The Scala code was copied from jkbradley.
~~This PR contains the changes from #3068 . I will rebase after #3068 is merged.~~
marmbrus jkbradley
Author: Xiangrui Meng <meng@databricks.com>
Closes#3070 from mengxr/SPARK-3573 and squashes the following commits:
3a0b6e5 [Xiangrui Meng] organize imports
236f0a0 [Xiangrui Meng] register vector as UDT and provide dataset examples
Saw Jenkins test failures due to random seeds.
jkbradley manishamde
Author: Xiangrui Meng <meng@databricks.com>
Closes#3084 from mengxr/fix-baggedpoint-suite and squashes the following commits:
f735a43 [Xiangrui Meng] fix seed in BaggedPointSuite
...sion trees. jkbradley mengxr chouqin Please review this.
Author: Sung Chung <schung@alpinenow.com>
Closes#2868 from codedeft/SPARK-3161 and squashes the following commits:
5f5a156 [Sung Chung] [SPARK-3161][MLLIB] Adding a node Id caching mechanism for training decision trees.
In #2928 , we shade commons-math3 to prevent future conflicts with hadoop. It caused problems with our Jenkins master build with maven. Some tests used local-cluster mode, where the assembly jar contains relocated math3 classes, while mllib test code still compiles with core and the untouched math3 classes.
This PR sets commons-math3 version based on hadoop profiles.
pwendell JoshRosen srowen
Author: Xiangrui Meng <meng@databricks.com>
Closes#3023 from mengxr/SPARK-4121-alt and squashes the following commits:
580f6d9 [Xiangrui Meng] replace tab by spaces
7f71f08 [Xiangrui Meng] revert changes to PoissonSampler to avoid conflicts
d3353d9 [Xiangrui Meng] do not shade commons-math3
b4180dc [Xiangrui Meng] temp work
This adds a Streaming KMeans algorithm to MLlib. It uses an update rule that generalizes the mini-batch KMeans update to incorporate a decay factor, which allows past data to be forgotten. The decay factor can be specified explicitly, or via a more intuitive "fractional decay" setting, in units of either data points or batches.
The PR includes:
- StreamingKMeans algorithm with decay factor settings
- Usage example
- Additions to documentation clustering page
- Unit tests of basic behavior and decay behaviors
tdas mengxr rezazadeh
Author: freeman <the.freeman.lab@gmail.com>
Author: Jeremy Freeman <the.freeman.lab@gmail.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#2942 from freeman-lab/streaming-kmeans and squashes the following commits:
b2e5b4a [freeman] Fixes to docs / examples
078617c [Jeremy Freeman] Merge pull request #1 from mengxr/SPARK-3254
2e682c0 [Xiangrui Meng] take discount on previous weights; use BLAS; detect dying clusters
0411bf5 [freeman] Change decay parameterization
9f7aea9 [freeman] Style fixes
374a706 [freeman] Formatting
ad9bdc2 [freeman] Use labeled points and predictOnValues in examples
77dbd3f [freeman] Make initialization check an assertion
9cfc301 [freeman] Make random seed an argument
44050a9 [freeman] Simpler constructor
c7050d5 [freeman] Fix spacing
2899623 [freeman] Use pattern matching for clarity
a4a316b [freeman] Use collect
1472ec5 [freeman] Doc formatting
ea22ec8 [freeman] Fix imports
2086bdc [freeman] Log cluster center updates
ea9877c [freeman] More documentation
9facbe3 [freeman] Bug fix
5db7074 [freeman] Example usage for StreamingKMeans
f33684b [freeman] Add explanation and example to docs
b5b5f8d [freeman] Add better documentation
a0fd790 [freeman] Merge remote-tracking branch 'upstream/master' into streaming-kmeans
9fd9c15 [freeman] Merge remote-tracking branch 'upstream/master' into streaming-kmeans
b93350f [freeman] Streaming KMeans with decay
Given the popular demand for gradient boosting and AdaBoost in MLlib, I am creating a WIP branch for early feedback on gradient boosting with AdaBoost to follow soon after this PR is accepted. This is based on work done along with hirakendu that was pending due to decision tree optimizations and random forests work.
Ideally, boosting algorithms should work with any base learners. This will soon be possible once the MLlib API is finalized -- we want to ensure we use a consistent interface for the underlying base learners. In the meantime, this PR uses decision trees as base learners for the gradient boosting algorithm. The current PR allows "pluggable" loss functions and provides least squares error and least absolute error by default.
Here is the task list:
- [x] Gradient boosting support
- [x] Pluggable loss functions
- [x] Stochastic gradient boosting support – Re-use the BaggedPoint approach used for RandomForest.
- [x] Binary classification support
- [x] Support configurable checkpointing – This approach will avoid long lineage chains.
- [x] Create classification and regression APIs
- [x] Weighted Ensemble Model -- created a WeightedEnsembleModel class that can be used by ensemble algorithms such as random forests and boosting.
- [x] Unit Tests
Future work:
+ Multi-class classification is currently not supported by this PR since it requires discussion on the best way to support "deviance" as a loss function.
+ BaggedRDD caching -- Avoid repeating feature to bin mapping for each tree estimator after standard API work is completed.
cc: jkbradley hirakendu mengxr etrain atalwalkar chouqin
Author: Manish Amde <manish9ue@gmail.com>
Author: manishamde <manish9ue@gmail.com>
Closes#2607 from manishamde/gbt and squashes the following commits:
991c7b5 [Manish Amde] public api
ff2a796 [Manish Amde] addressing comments
b4c1318 [Manish Amde] removing spaces
8476b6b [Manish Amde] fixing line length
0183cb9 [Manish Amde] fixed naming and formatting issues
1c40c33 [Manish Amde] add newline, removed spaces
e33ab61 [Manish Amde] minor comment
eadbf09 [Manish Amde] parameter renaming
035a2ed [Manish Amde] jkbradley formatting suggestions
9f7359d [Manish Amde] simplified gbt logic and added more tests
49ba107 [Manish Amde] merged from master
eff21fe [Manish Amde] Added gradient boosting tests
3fd0528 [Manish Amde] moved helper methods to new class
a32a5ab [Manish Amde] added test for subsampling without replacement
781542a [Manish Amde] added support for fractional subsampling with replacement
3a18cc1 [Manish Amde] cleaned up api for conversion to bagged point and moved tests to it's own test suite
0e81906 [Manish Amde] improving caching unpersisting logic
d971f73 [Manish Amde] moved RF code to use WeightedEnsembleModel class
fee06d3 [Manish Amde] added weighted ensemble model
1b01943 [Manish Amde] add weights for base learners
9bc6e74 [Manish Amde] adding random seed as parameter
d2c8323 [Manish Amde] Merge branch 'master' into gbt
2ae97b7 [Manish Amde] added documentation for the loss classes
9366b8f [Manish Amde] minor: using numTrees instead of trees.size
3b43896 [Manish Amde] added learning rate for prediction
9b2e35e [Manish Amde] Merge branch 'master' into gbt
6a11c02 [manishamde] fixing formatting
823691b [Manish Amde] fixing RF test
1f47941 [Manish Amde] changing access modifier
5b67102 [Manish Amde] shortened parameter list
5ab3796 [Manish Amde] minor reformatting
9155a9d [Manish Amde] consolidated boosting configuration and added public API
631baea [Manish Amde] Merge branch 'master' into gbt
2cb1258 [Manish Amde] public API support
3b8ffc0 [Manish Amde] added documentation
8e10c63 [Manish Amde] modified unpersist strategy
f62bc48 [Manish Amde] added unpersist
bdca43a [Manish Amde] added timing parameters
2fbc9c7 [Manish Amde] fixing binomial classification prediction
6dd4dd8 [Manish Amde] added support for log loss
9af0231 [Manish Amde] classification attempt
62cc000 [Manish Amde] basic checkpointing
4784091 [Manish Amde] formatting
78ed452 [Manish Amde] added newline and fixed if statement
3973dd1 [Manish Amde] minor indicating subsample is double during comparison
aa8fae7 [Manish Amde] minor refactoring
1a8031c [Manish Amde] sampling with replacement
f1c9ef7 [Manish Amde] Merge branch 'master' into gbt
cdceeef [Manish Amde] added documentation
6251fd5 [Manish Amde] modified method name
5538521 [Manish Amde] disable checkpointing for now
0ae1c0a [Manish Amde] basic gradient boosting code from earlier branches
Implementation of various multi-label classification measures, including: Hamming-loss, strict and default Accuracy, macro-averaged Precision, Recall and F1-measure based on documents and labels, micro-averaged measures: https://issues.apache.org/jira/browse/SPARK-2329
Multi-class measures are currently in the following pull request: https://github.com/apache/spark/pull/1155
Author: Alexander Ulanov <nashb@yandex.ru>
Author: avulanov <nashb@yandex.ru>
Closes#1270 from avulanov/multilabelmetrics and squashes the following commits:
fc8175e [Alexander Ulanov] Merge with previous updates
43a613e [Alexander Ulanov] Addressing reviewers comments: change Set to Array
517a594 [avulanov] Addressing reviewers comments: Scala style
cf4222bc [avulanov] Addressing reviewers comments: renaming. Added label method that returns the list of labels
1843f73 [Alexander Ulanov] Scala style fix
79e8476 [Alexander Ulanov] Replacing fold(_ + _) with sum as suggested by srowen
ca46765 [Alexander Ulanov] Cosmetic changes: Apache header and parameter explanation
40593f5 [Alexander Ulanov] Multi-label metrics: Hamming-loss, strict and normal accuracy, fix to macro measures, bunch of tests
ad62df0 [Alexander Ulanov] Comments and scala style check
154164b [Alexander Ulanov] Multilabel evaluation metics and tests: macro precision and recall averaged by docs, micro and per-class precision and recall averaged by class
More efficient sampling, based on Gap Sampling optimization:
http://erikerlandson.github.io/blog/2014/09/11/faster-random-samples-with-gap-sampling/
Author: Erik Erlandson <eerlands@redhat.com>
Closes#2455 from erikerlandson/spark-3250-pr and squashes the following commits:
72496bc [Erik Erlandson] [SPARK-3250] Implement Gap Sampling optimization for random sampling
Create several helper functions to call MLlib Java API, convert the arguments to Java type and convert return value to Python object automatically, this simplify serialization in MLlib Python API very much.
After this, the MLlib Python API does not need to deal with serialization details anymore, it's easier to add new API.
cc mengxr
Author: Davies Liu <davies@databricks.com>
Closes#2995 from davies/cleanup and squashes the following commits:
8fa6ec6 [Davies Liu] address comments
16b85a0 [Davies Liu] Merge branch 'master' of github.com:apache/spark into cleanup
43743e5 [Davies Liu] bugfix
731331f [Davies Liu] simplify serialization in MLlib Python API
Add RegressionMetrics.scala as regression metrics used for evaluation and corresponding test case RegressionMetricsSuite.scala.
Author: Yanbo Liang <yanbohappy@gmail.com>
Author: liangyanbo <liangyanbo@meituan.com>
Closes#2978 from yanbohappy/regression_metrics and squashes the following commits:
730d0a9 [Yanbo Liang] more clearly annotation
3d0bec1 [Yanbo Liang] rename and keep code style
a8ad3e3 [Yanbo Liang] simplify code for keeping style
d454909 [Yanbo Liang] rename parameter and function names, delete unused columns, add reference
2e56282 [liangyanbo] rename r2_score() and remove unused column
43bb12b [liangyanbo] add regression metrics
This simple patch filters out extra whitespace entries.
Author: Joseph E. Gonzalez <joseph.e.gonzalez@gmail.com>
Author: Joey <joseph.e.gonzalez@gmail.com>
Closes#2996 from jegonzal/loadLibSVM and squashes the following commits:
e0227ab [Joey] improving readability
e028e84 [Joseph E. Gonzalez] fixing whitespace bug in loadLibSVMFile when parsing libSVM files
In MultivariateOnlineSummarizer, breeze's activeIterator is used
to loop through the nonZero elements in the vector. However,
activeIterator doesn't perform well due to lots of overhead.
In this PR, native while loop is used for both DenseVector and SparseVector.
The benchmark result with 20 executors using mnist8m dataset:
Before:
DenseVector: 48.2 seconds
SparseVector: 16.3 seconds
After:
DenseVector: 17.8 seconds
SparseVector: 11.2 seconds
Since MultivariateOnlineSummarizer is used in several places,
the overall performance gain in mllib library will be significant with this PR.
Author: DB Tsai <dbtsai@alpinenow.com>
Closes#2992 from dbtsai/SPARK-4129 and squashes the following commits:
b99db6c [DB Tsai] fixed java.lang.ArrayIndexOutOfBoundsException
2b5e882 [DB Tsai] small refactoring
ebe3e74 [DB Tsai] First commit
Also update step parameter to pass the proposed test
Author: coderxiang <shuoxiangpub@gmail.com>
Closes#2965 from coderxiang/nnls-test and squashes the following commits:
24b06f9 [coderxiang] add test case on objective value for NNLS; update step parameter to pass the test
This change replaces usages of colt with commons-math3 equivalents, and makes some minor necessary adjustments to related code and tests to match.
Author: Sean Owen <sowen@cloudera.com>
Closes#2928 from srowen/SPARK-4022 and squashes the following commits:
61a232f [Sean Owen] Fix failure due to different sampling in JavaAPISuite.sample()
16d66b8 [Sean Owen] Simplify seeding with call to reseedRandomGenerator
a1a78e0 [Sean Owen] Use Well19937c
31c7641 [Sean Owen] Fix Python Poisson test by choosing a different seed; about 88% of seeds should work but 1 didn't, it seems
5c9c67f [Sean Owen] Additional test fixes from review
d8f88e0 [Sean Owen] Replace colt with commons-math3. Some tests do not pass yet.
This follows https://github.com/apache/spark/pull/2893 , but does not completely fix SPARK-3359 either. This fixes minor scaladoc/javadoc issues that Javadoc 8 will treat as errors.
Author: Sean Owen <sowen@cloudera.com>
Closes#2909 from srowen/SPARK-3359 and squashes the following commits:
f62c347 [Sean Owen] Fix some javadoc issues that javadoc 8 considers errors. This is not all of the errors turned up when javadoc 8 runs on output of genjavadoc.
Thare are some inconsistent spellings 'MLlib' and 'MLLib' in some documents and source codes.
Author: Kousuke Saruta <sarutak@oss.nttdata.co.jp>
Closes#2903 from sarutak/SPARK-4055 and squashes the following commits:
b031640 [Kousuke Saruta] Fixed inconsistent spelling "MLlib and MLLib"
Add common metrics for ranking algorithms (http://www-nlp.stanford.edu/IR-book/), including:
- Mean Average Precision
- Precisionn: top-n precision
- Discounted cumulative gain (DCG) and NDCG
The following methods and the corresponding tests are implemented:
```
class RankingMetrics[T](predictionAndLabels: RDD[(Array[T], Array[T])]) {
/* Returns the precsionk for each query */
lazy val precAtK: RDD[Array[Double]]
/**
* param k the position to compute the truncated precision
* return the average precision at the first k ranking positions
*/
def precision(k: Int): Double
/* Returns the average precision for each query */
lazy val avePrec: RDD[Double]
/*Returns the mean average precision (MAP) of all the queries*/
lazy val meanAvePrec: Double
/*Returns the normalized discounted cumulative gain for each query */
lazy val ndcgAtK: RDD[Array[Double]]
/**
* param k the position to compute the truncated ndcg
* return the average ndcg at the first k ranking positions
*/
def ndcg(k: Int): Double
}
```
Author: coderxiang <shuoxiangpub@gmail.com>
Closes#2667 from coderxiang/rankingmetrics and squashes the following commits:
d881097 [coderxiang] update doc
14d9cd9 [coderxiang] remove unexpected files
d7fb93f [coderxiang] style change and remove ignored files
f113ee1 [coderxiang] modify doc for displaying superscript and subscript
f626896 [coderxiang] improve doc and remove unnecessary computation while labSet is empty
be6645e [coderxiang] set the precision of empty labset to 0.0
d64c120 [coderxiang] add logWarning for empty ground truth set
dfae292 [coderxiang] handle empty labSet for map. add test
62047c4 [coderxiang] style change and add documentation
f66612d [coderxiang] add additional test of precisionAt
b794cb2 [coderxiang] move private members precAtK, ndcgAtK into public methods. style change
77c9e5d [coderxiang] set precAtK and ndcgAtK as private member. Improve documentation
5f87bce [coderxiang] add API to calculate precision and ndcg at each ranking position
b7851cc [coderxiang] Use generic type to represent IDs
e443fee [coderxiang] change style and use alternative builtin methods
3a5a6ff [coderxiang] add ranking metrics
https://issues.apache.org/jira/browse/SPARK-3770
We need access to the underlying latent user features from python. However, the userFeatures RDD from the MatrixFactorizationModel isn't accessible from the python bindings. I've added a method to the underlying scala class to turn the RDD[(Int, Array[Double])] to an RDD[String]. This is then accessed from the python recommendation.py
Author: Michelangelo D'Agostino <mdagostino@civisanalytics.com>
Closes#2636 from mdagost/mf_user_features and squashes the following commits:
c98f9e2 [Michelangelo D'Agostino] Added unit tests for userFeatures and productFeatures and merged master.
d5eadf8 [Michelangelo D'Agostino] Merge branch 'master' into mf_user_features
2481a2a [Michelangelo D'Agostino] Merged master and resolved conflict.
a6ffb96 [Michelangelo D'Agostino] Eliminated a function from our first approach to this problem that is no longer needed now that we added the fromTuple2RDD function.
2aa1bf8 [Michelangelo D'Agostino] Implemented a function called fromTuple2RDD in PythonMLLibAPI and used it to expose the MF userFeatures and productFeatures in python.
34cb2a2 [Michelangelo D'Agostino] A couple of lint cleanups and a comment.
cdd98e3 [Michelangelo D'Agostino] It's working now.
e1fbe5e [Michelangelo D'Agostino] Added scala function to stringify userFeatures for access in python.
DecisionTree splits on continuous features by choosing an array of values from a subsample of the data.
Currently, it does not check for identical values in the subsample, so it could end up having multiple copies of the same split. In this PR, we choose splits for a continuous feature in 3 steps:
1. Sort sample values for this feature
2. Get number of occurrence of each distinct value
3. Iterate the value count array computed in step 2 to choose splits.
After find splits, `numSplits` and `numBins` in metadata will be updated.
CC: mengxr manishamde jkbradley, please help me review this, thanks.
Author: Qiping Li <liqiping1991@gmail.com>
Author: chouqin <liqiping1991@gmail.com>
Author: liqi <liqiping1991@gmail.com>
Author: qiping.lqp <qiping.lqp@alibaba-inc.com>
Closes#2780 from chouqin/dt-findsplits and squashes the following commits:
18d0301 [Qiping Li] check explicitly findsplits return distinct splits
8dc28ab [chouqin] remove blank lines
ffc920f [chouqin] adjust code based on comments and add more test cases
9857039 [chouqin] Merge branch 'master' of https://github.com/apache/spark into dt-findsplits
d353596 [qiping.lqp] fix pyspark doc test
9e64699 [Qiping Li] fix random forest unit test
3c72913 [Qiping Li] fix random forest unit test
092efcb [Qiping Li] fix bug
f69f47f [Qiping Li] fix bug
ab303a4 [Qiping Li] fix bug
af6dc97 [Qiping Li] fix bug
2a8267a [Qiping Li] fix bug
c339a61 [Qiping Li] fix bug
369f812 [Qiping Li] fix style
8f46af6 [Qiping Li] add comments and unit test
9e7138e [Qiping Li] Merge branch 'dt-findsplits' of https://github.com/chouqin/spark into dt-findsplits
1b25a35 [Qiping Li] Merge branch 'master' of https://github.com/apache/spark into dt-findsplits
0cd744a [liqi] fix bug
3652823 [Qiping Li] fix bug
af7cb79 [Qiping Li] Choose splits for continuous features in DecisionTree more adaptively
SPARK-3934: When run with a mix of unordered categorical and continuous features, on multiclass classification, RandomForest fails. The bug is in the sanity checks in getFeatureOffset and getLeftRightFeatureOffsets, which use the wrong indices for checking whether features are unordered.
Fix: Remove the sanity checks since they are not really needed, and since they would require DTStatsAggregator to keep track of an extra set of indices (for the feature subset).
Added test to RandomForestSuite which failed with old version but now works.
SPARK-3918: Added baggedInput.unpersist at end of training.
Also:
* I removed DTStatsAggregator.isUnordered since it is no longer used.
* DecisionTreeMetadata: Added logWarning when maxBins is automatically reduced.
* Updated DecisionTreeRunner to explicitly fix the test data to have the same number of features as the training data. This is a temporary fix which should eventually be replaced by pre-indexing both datasets.
* RandomForestModel: Updated toString to print total number of nodes in forest.
* Changed Predict class to be public DeveloperApi. This was necessary to allow users to create their own trees by hand (for testing).
CC: mengxr manishamde chouqin codedeft Just notifying you of these small bug fixes.
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Closes#2785 from jkbradley/dtrunner-update and squashes the following commits:
9132321 [Joseph K. Bradley] merged with master, fixed imports
9dbd000 [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dtrunner-update
e116473 [Joseph K. Bradley] Changed Predict class to be public DeveloperApi.
f502e65 [Joseph K. Bradley] bug fix for SPARK-3934
7f3d60f [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dtrunner-update
ba567ab [Joseph K. Bradley] Changed DTRunner to load test data using same number of features as in training data.
4e88c1f [Joseph K. Bradley] changed RF toString to print total number of nodes
Customized pickler should be registered before unpickling, but in executor, there is no way to register the picklers before run the tasks.
So, we need to register the picklers in the tasks itself, duplicate the javaToPython() and pythonToJava() in MLlib, call SerDe.initialize() before pickling or unpickling.
Author: Davies Liu <davies.liu@gmail.com>
Closes#2830 from davies/fix_pickle and squashes the following commits:
0c85fb9 [Davies Liu] revert the privacy change
6b94e15 [Davies Liu] use JavaConverters instead of JavaConversions
0f02050 [Davies Liu] hotfix: Customized pickler does not work in cluster
Avoid overflow in computing n*(n+1)/2 as much as possible; throw explicit error when Gramian computation will fail due to negative array size; warn about large result when computing Gramian too
Author: Sean Owen <sowen@cloudera.com>
Closes#2801 from srowen/SPARK-3803 and squashes the following commits:
b4e6d92 [Sean Owen] Avoid overflow in computing n*(n+1)/2 as much as possible; throw explicit error when Gramian computation will fail due to negative array size; warn about large result when computing Gramian too
During trainning Gradient Boosting Decision Tree on large-scale sparse data, spark spill hundreds of data onto disk. And find the bug below:
In version 1.1.0 DecisionTree.scala, train Method, treeInput has been persisted in Memory, but without unpersist. It caused heavy DISK usage.
In github version(1.2.0 maybe), RandomForest.scala, train Method, baggedInput has been persisted but without unpersisted too.
After added unpersist, it works right.
https://issues.apache.org/jira/browse/SPARK-3918
Author: omgteam <Kimlong.Liu@gmail.com>
Closes#2775 from omgteam/master and squashes the following commits:
815d543 [omgteam] adjust tab to spaces
1a36f83 [omgteam] Bug: fix without unpersist baggedInput in RandomForest.scala
I noticed a few issues with how temp directories are created and deleted:
*Minor*
* Guava's `Files.createTempDir()` plus `File.deleteOnExit()` is used in many tests to make a temp dir, but `Utils.createTempDir()` seems to be the standard Spark mechanism
* Call to `File.deleteOnExit()` could be pushed into `Utils.createTempDir()` as well, along with this replacement
* _I messed up the message in an exception in `Utils` in SPARK-3794; fixed here_
*Bit Less Minor*
* `Utils.deleteRecursively()` fails immediately if any `IOException` occurs, instead of trying to delete any remaining files and subdirectories. I've observed this leave temp dirs around. I suggest changing it to continue in the face of an exception and throw one of the possibly several exceptions that occur at the end.
* `Utils.createTempDir()` will add a JVM shutdown hook every time the method is called. Even if the subdir is the parent of another parent dir, since this check is inside the hook. However `Utils` manages a set of all dirs to delete on shutdown already, called `shutdownDeletePaths`. A single hook can be registered to delete all of these on exit. This is how Tachyon temp paths are cleaned up in `TachyonBlockManager`.
I noticed a few other things that might be changed but wanted to ask first:
* Shouldn't the set of dirs to delete be `File`, not just `String` paths?
* `Utils` manages the set of `TachyonFile` that have been registered for deletion, but the shutdown hook is managed in `TachyonBlockManager`. Should this logic not live together, and not in `Utils`? it's more specific to Tachyon, and looks a slight bit odd to import in such a generic place.
Author: Sean Owen <sowen@cloudera.com>
Closes#2670 from srowen/SPARK-3811 and squashes the following commits:
071ae60 [Sean Owen] Update per @vanzin's review
da0146d [Sean Owen] Make Utils.deleteRecursively try to delete all paths even when an exception occurs; use one shutdown hook instead of one per method call to delete temp dirs
3a0faa4 [Sean Owen] Standardize on Utils.createTempDir instead of Files.createTempDir
cc mengxr
Author: GuoQiang Li <witgo@qq.com>
Closes#2730 from witgo/SPARK-3856 and squashes the following commits:
2cffce1 [GuoQiang Li] use norm operator after breeze 0.10 upgrade
Currently, the implementation does one unnecessary aggregation step. The aggregation step for level L (to choose splits) gives enough information to set the predictions of any leaf nodes at level L+1. We can use that info and skip the aggregation step for the last level of the tree (which only has leaf nodes).
### Implementation Details
Each node now has a `impurity` field and the `predict` is changed from type `Double` to type `Predict`(this can be used to compute predict probability in the future) When compute best splits for each node, we also compute impurity and predict for the child nodes, which is used to constructed newly allocated child nodes. So at level L, we have set impurity and predict for nodes at level L +1.
If level L+1 is the last level, then we can avoid aggregation. What's more, calculation of parent impurity in
Top nodes for each tree needs to be treated differently because we have to compute impurity and predict for them first. In `binsToBestSplit`, if current node is top node(level == 0), we calculate impurity and predict first.
after finding best split, top node's predict and impurity is set to the calculated value. Non-top nodes's impurity and predict are already calculated and don't need to be recalculated again. I have considered to add a initialization step to set top nodes' impurity and predict and then we can treat all nodes in the same way, but this will need a lot of duplication of code(all the code to do seq operation(BinSeqOp) needs to be duplicated), so I choose the current way.
CC mengxr manishamde jkbradley, please help me review this, thanks.
Author: Qiping Li <liqiping1991@gmail.com>
Closes#2708 from chouqin/avoid-agg and squashes the following commits:
8e269ea [Qiping Li] adjust code and comments
eefeef1 [Qiping Li] adjust comments and check child nodes' impurity
c41b1b6 [Qiping Li] fix pyspark unit test
7ad7a71 [Qiping Li] fix unit test
822c912 [Qiping Li] add comments and unit test
e41d715 [Qiping Li] fix bug in test suite
6cc0333 [Qiping Li] SPARK-3158: Avoid 1 extra aggregation for DecisionTree training
Got warning msg:
~~~
[warn] /Users/meng/src/spark/mllib/src/main/scala/org/apache/spark/mllib/feature/Normalizer.scala:50: method norm in trait NumericOps is deprecated: Use norm(XXX) instead of XXX.norm
[warn] var norm = vector.toBreeze.norm(p)
~~~
dbtsai
Author: Xiangrui Meng <meng@databricks.com>
Closes#2718 from mengxr/SPARK-3856 and squashes the following commits:
4f38169 [Xiangrui Meng] use norm operator
In Breeze 0.10, the L1regParam can be configured through anonymous function in OWLQN, and each component can be penalized differently. This is required for GLMNET in MLlib with L1/L2 regularization.
2570911026
Author: DB Tsai <dbtsai@dbtsai.com>
Closes#2693 from dbtsai/breeze0.10 and squashes the following commits:
7a0c45c [DB Tsai] In Breeze 0.10, the L1regParam can be configured through anonymous function in OWLQN, and each component can be penalized differently. This is required for GLMNET in MLlib with L1/L2 regularization. 2570911026
mengxr
Added PySpark support for Word2Vec
Change list
(1) PySpark support for Word2Vec
(2) SerDe support of string sequence both on python side and JVM side
(3) Test for SerDe of string sequence on JVM side
Author: Liquan Pei <liquanpei@gmail.com>
Closes#2356 from Ishiihara/Word2Vec-python and squashes the following commits:
476ea34 [Liquan Pei] style fixes
b13a0b9 [Liquan Pei] resolve merge conflicts and minor fixes
8671eba [Liquan Pei] Merge remote-tracking branch 'upstream/master' into Word2Vec-python
daf88a6 [Liquan Pei] modification according to feedback
a73fa19 [Liquan Pei] clean up
3d8007b [Liquan Pei] fix findSynonyms for vector
1bdcd2e [Liquan Pei] minor fixes
cdef9f4 [Liquan Pei] add missing comments
b7447eb [Liquan Pei] modify according to feedback
b9a7383 [Liquan Pei] cache words RDD in fit
89490bf [Liquan Pei] add tests and Word2VecModelWrapper
78bbb53 [Liquan Pei] use pickle for seq string SerDe
a264b08 [Liquan Pei] Merge remote-tracking branch 'upstream/master' into Word2Vec-python
ca1e5ff [Liquan Pei] fix test
68e7276 [Liquan Pei] minor style fixes
48d5e72 [Liquan Pei] Functionality improvement
0ad3ac1 [Liquan Pei] minor fix
c867fdf [Liquan Pei] add Word2Vec to pyspark
Add a toString method to GeneralizedLinearModel, also change `__str__` to `__repr__` for some classes, to provide better message in repr.
This PR is based on #1388, thanks to sryza!
closes#1388
Author: Sandy Ryza <sandy@cloudera.com>
Author: Davies Liu <davies.liu@gmail.com>
Closes#2625 from davies/string and squashes the following commits:
3544aad [Davies Liu] fix LinearModel
0bcd642 [Davies Liu] Merge branch 'sandy-spark-2461' of github.com:sryza/spark
1ce5c2d [Sandy Ryza] __repr__ back to __str__ in a couple places
aa9e962 [Sandy Ryza] Switch __str__ to __repr__
a0c5041 [Sandy Ryza] Add labels back in
1aa17f5 [Sandy Ryza] Match existing conventions
fac1bc4 [Sandy Ryza] Fix PEP8 error
f7b58ed [Sandy Ryza] SPARK-2461. Add a toString method to GeneralizedLinearModel
Currently, all best splits are computed on the driver, which makes the driver a bottleneck for both communication and computation. This PR fix this problem by computed best splits on executors.
Instead of send all aggregate stats to the driver node, we can send aggregate stats for a node to a particular executor, using `reduceByKey` operation, then we can compute best split for this node there.
Implementation details:
Each node now has a nodeStatsAggregator, which save aggregate stats for all features and bins.
First use mapPartition to compute node aggregate stats for all nodes in each partition.
Then transform node aggregate stats to (nodeIndex, nodeStatsAggregator) pairs and use to `reduceByKey` operation to combine nodeStatsAggregator for the same node.
After all stats have been combined, best splits can be computed for each node based on the node aggregate stats. Best split result is collected to driver to construct the decision tree.
CC: mengxr manishamde jkbradley, please help me review this, thanks.
Author: qiping.lqp <qiping.lqp@alibaba-inc.com>
Author: chouqin <liqiping1991@gmail.com>
Closes#2595 from chouqin/dt-dist-agg and squashes the following commits:
db0d24a [chouqin] fix a minor bug and adjust code
a0d9de3 [chouqin] adjust code based on comments
9f201a6 [chouqin] fix bug: statsSize -> allStatsSize
a8a7ed0 [chouqin] Merge branch 'master' of https://github.com/apache/spark into dt-dist-agg
f13b346 [chouqin] adjust randomforest comments
c32636e [chouqin] adjust code based on comments
ac6a505 [chouqin] adjust code based on comments
7bbb787 [chouqin] add comments
bdd2a63 [qiping.lqp] fix test suite
a75df27 [qiping.lqp] fix test suite
b5b0bc2 [qiping.lqp] fix style
e76414f [qiping.lqp] fix testsuite
748bd45 [qiping.lqp] fix type-mismatch bug
24eacd8 [qiping.lqp] fix type-mismatch bug
5f63d6c [qiping.lqp] add multiclassification using One-Vs-All strategy
4f56496 [qiping.lqp] fix bug
f00fc22 [qiping.lqp] fix bug
532993a [qiping.lqp] Compute best splits distributively in decision tree
Thread names are useful for correlating failures.
Author: Reynold Xin <rxin@apache.org>
Closes#2600 from rxin/log4j and squashes the following commits:
83ffe88 [Reynold Xin] [SPARK-3748] Log thread name in unit test logs
DecisionTreeRunner functionality additions:
* Allow user to pass in a test dataset
* Do not print full model if the model is too large.
As part of this, modify DecisionTreeModel and RandomForestModel to allow printing less info. Proposed updates:
* toString: prints model summary
* toDebugString: prints full model (named after RDD.toDebugString)
Similar update to Python API:
* __repr__() now prints a model summary
* toDebugString() now prints the full model
CC: mengxr chouqin manishamde codedeft Small update (whomever can take a look). Thanks!
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Closes#2604 from jkbradley/dtrunner-update and squashes the following commits:
b2b3c60 [Joseph K. Bradley] re-added python sql doc test, temporarily removed before
07b1fae [Joseph K. Bradley] repr() now prints a model summary toDebugString() now prints the full model
1d0d93d [Joseph K. Bradley] Updated DT and RF to print less when toString is called. Added toDebugString for verbose printing.
22eac8c [Joseph K. Bradley] Merge remote-tracking branch 'upstream/master' into dtrunner-update
e007a95 [Joseph K. Bradley] Updated DecisionTreeRunner to accept a test dataset.
1. doc updates
2. simple checks on vector dimensions
3. use column major for matrices
davies jkbradley
Author: Xiangrui Meng <meng@databricks.com>
Closes#2548 from mengxr/mllib-py-clean and squashes the following commits:
6dce2df [Xiangrui Meng] address comments
116b5db [Xiangrui Meng] use np.dot instead of array.dot
75f2fcc [Xiangrui Meng] fix python style
fefce00 [Xiangrui Meng] better check of vector size with more tests
067ef71 [Xiangrui Meng] majored -> major
ef853f9 [Xiangrui Meng] update python linalg api and small fixes
# All-pairs similarity via DIMSUM
Compute all pairs of similar vectors using brute force approach, and also DIMSUM sampling approach.
Laying down some notation: we are looking for all pairs of similar columns in an m x n RowMatrix whose entries are denoted a_ij, with the i’th row denoted r_i and the j’th column denoted c_j. There is an oversampling parameter labeled ɣ that should be set to 4 log(n)/s to get provably correct results (with high probability), where s is the similarity threshold.
The algorithm is stated with a Map and Reduce, with proofs of correctness and efficiency in published papers [1] [2]. The reducer is simply the summation reducer. The mapper is more interesting, and is also the heart of the scheme. As an exercise, you should try to see why in expectation, the map-reduce below outputs cosine similarities.
[1] Bosagh-Zadeh, Reza and Carlsson, Gunnar (2013), Dimension Independent Matrix Square using MapReduce, arXiv:1304.1467 http://arxiv.org/abs/1304.1467
[2] Bosagh-Zadeh, Reza and Goel, Ashish (2012), Dimension Independent Similarity Computation, arXiv:1206.2082 http://arxiv.org/abs/1206.2082
# Testing
Tests for all invocations included.
Added L1 and L2 norm computation to MultivariateStatisticalSummary since it was needed. Added tests for both of them.
Author: Reza Zadeh <rizlar@gmail.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#1778 from rezazadeh/dimsumv2 and squashes the following commits:
404c64c [Reza Zadeh] Merge remote-tracking branch 'upstream/master' into dimsumv2
4eb71c6 [Reza Zadeh] Add excludes for normL1 and normL2
ee8bd65 [Reza Zadeh] Merge remote-tracking branch 'upstream/master' into dimsumv2
976ddd4 [Reza Zadeh] Broadcast colMags. Avoid div by zero.
3467cff [Reza Zadeh] Merge remote-tracking branch 'upstream/master' into dimsumv2
aea0247 [Reza Zadeh] Allow large thresholds to promote sparsity
9fe17c0 [Xiangrui Meng] organize imports
2196ba5 [Xiangrui Meng] Merge branch 'rezazadeh-dimsumv2' into dimsumv2
254ca08 [Reza Zadeh] Merge remote-tracking branch 'upstream/master' into dimsumv2
f2947e4 [Xiangrui Meng] some optimization
3c4cf41 [Xiangrui Meng] Merge branch 'master' into rezazadeh-dimsumv2
0e4eda4 [Reza Zadeh] Use partition index for RNG
251bb9c [Reza Zadeh] Documentation
25e9d0d [Reza Zadeh] Line length for style
fb296f6 [Reza Zadeh] renamed to normL1 and normL2
3764983 [Reza Zadeh] Documentation
e9c6791 [Reza Zadeh] New interface and documentation
613f261 [Reza Zadeh] Column magnitude summary
75a0b51 [Reza Zadeh] Use Ints instead of Longs in the shuffle
0f12ade [Reza Zadeh] Style changes
eb1dc20 [Reza Zadeh] Use Double.PositiveInfinity instead of Double.Max
f56a882 [Reza Zadeh] Remove changes to MultivariateOnlineSummarizer
dbc55ba [Reza Zadeh] Make colMagnitudes a method in RowMatrix
41e8ece [Reza Zadeh] style changes
139c8e1 [Reza Zadeh] Syntax changes
029aa9c [Reza Zadeh] javadoc and new test
75edb25 [Reza Zadeh] All tests passing!
05e59b8 [Reza Zadeh] Add test
502ce52 [Reza Zadeh] new interface
654c4fb [Reza Zadeh] default methods
3726ca9 [Reza Zadeh] Remove MatrixAlgebra
6bebabb [Reza Zadeh] remove changes to MatrixSuite
5b8cd7d [Reza Zadeh] Initial files
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
This PR for [SPARK-3614](https://issues.apache.org/jira/browse/SPARK-3614) adds functionality for filtering out terms which do not appear in at least a minimum number of documents.
This is implemented using a minimumOccurence parameter (default 0). When terms' document frequencies are less than minimumOccurence, their IDFs are set to 0, just like when the DF is 0. As a result, the TF-IDFs for the terms are found to be 0, as if the terms were not present in the documents.
This PR makes the following changes:
* Add a minimumOccurence parameter to the IDF and DocumentFrequencyAggregator classes.
* Create a parameter-less constructor for IDF with a default minimumOccurence value of 0 to remain backwards-compatibility with the original IDF API.
* Sets the IDFs to 0 for terms which DFs are less than minimumOccurence
* Add tests to the Spark IDFSuite and Java JavaTfIdfSuite test suites
* Updated the MLLib Feature Extraction programming guide to describe the new feature
Author: RJ Nowling <rnowling@gmail.com>
Closes#2494 from rnowling/spark-3614-idf-filter and squashes the following commits:
0aa3c63 [RJ Nowling] Fix identation
e6523a8 [RJ Nowling] Remove unnecessary toDouble's from IDFSuite
bfa82ec [RJ Nowling] Add space after if
30d20b3 [RJ Nowling] Add spaces around equals signs
9013447 [RJ Nowling] Add space before division operator
79978fc [RJ Nowling] Remove unnecessary semi-colon
40fd70c [RJ Nowling] Change minimumOccurence to minDocFreq in code and docs
47850ab [RJ Nowling] Changed minimumOccurence to Int from Long
9fb4093 [RJ Nowling] Remove unnecessary lines from IDF class docs
1fc09d8 [RJ Nowling] Add backwards-compatible constructor to DocumentFrequencyAggregator
1801fd2 [RJ Nowling] Fix style errors in IDF.scala
6897252 [RJ Nowling] Preface minimumOccurence members with val to make them final and immutable
a200bab [RJ Nowling] Remove unnecessary else statement
4b974f5 [RJ Nowling] Remove accidentally-added import from testing
c0cc643 [RJ Nowling] Add minimumOccurence filtering to IDF
Add warnings to KMeans, GeneralizedLinearAlgorithm, and computeSVD when called with input data that is not cached. KMeans is implemented iteratively, and I believe that GeneralizedLinearAlgorithm’s current optimizers are iterative and its future optimizers are also likely to be iterative. RowMatrix’s computeSVD is iterative against an RDD when run in DistARPACK mode. ALS and DecisionTree are iterative as well, but they implement RDD caching internally so do not require a warning.
I added a warning to GeneralizedLinearAlgorithm rather than inside its optimizers, where the iteration actually occurs, because internally GeneralizedLinearAlgorithm maps its input data to an uncached RDD before passing it to an optimizer. (In other words, the warning would be printed for every GeneralizedLinearAlgorithm run, regardless of whether its input is cached, if the warning were in GradientDescent or other optimizer.) I assume that use of an uncached RDD by GeneralizedLinearAlgorithm is intentional, and that the mapping there (adding label, intercepts and scaling) is a lightweight operation. Arguably a user calling an optimizer such as GradientDescent will be knowledgable enough to cache their data without needing a log warning, so lack of a warning in the optimizers may be ok.
Some of the documentation examples making use of these iterative algorithms did not cache their training RDDs (while others did). I updated the examples to always cache. I also fixed some (unrelated) minor errors in the documentation examples.
Author: Aaron Staple <aaron.staple@gmail.com>
Closes#2347 from staple/SPARK-1484 and squashes the following commits:
bd49701 [Aaron Staple] Address review comments.
ab2d4a4 [Aaron Staple] Disable warnings on python code path.
a7a0f99 [Aaron Staple] Change code comments per review comments.
7cca1dc [Aaron Staple] Change warning message text.
c77e939 [Aaron Staple] [SPARK-1484][MLLIB] Warn when running an iterative algorithm on uncached data.
3b6c511 [Aaron Staple] Minor doc example fixes.
Currently, we serialize the data between JVM and Python case by case manually, this cannot scale to support so many APIs in MLlib.
This patch will try to address this problem by serialize the data using pickle protocol, using Pyrolite library to serialize/deserialize in JVM. Pickle protocol can be easily extended to support customized class.
All the modules are refactored to use this protocol.
Known issues: There will be some performance regression (both CPU and memory, the serialized data increased)
Author: Davies Liu <davies.liu@gmail.com>
Closes#2378 from davies/pickle_mllib and squashes the following commits:
dffbba2 [Davies Liu] Merge branch 'master' of github.com:apache/spark into pickle_mllib
810f97f [Davies Liu] fix equal of matrix
032cd62 [Davies Liu] add more type check and conversion for user_product
bd738ab [Davies Liu] address comments
e431377 [Davies Liu] fix cache of rdd, refactor
19d0967 [Davies Liu] refactor Picklers
2511e76 [Davies Liu] cleanup
1fccf1a [Davies Liu] address comments
a2cc855 [Davies Liu] fix tests
9ceff73 [Davies Liu] test size of serialized Rating
44e0551 [Davies Liu] fix cache
a379a81 [Davies Liu] fix pickle array in python2.7
df625c7 [Davies Liu] Merge commit '154d141' into pickle_mllib
154d141 [Davies Liu] fix autobatchedpickler
44736d7 [Davies Liu] speed up pickling array in Python 2.7
e1d1bfc [Davies Liu] refactor
708dc02 [Davies Liu] fix tests
9dcfb63 [Davies Liu] fix style
88034f0 [Davies Liu] rafactor, address comments
46a501e [Davies Liu] choose batch size automatically
df19464 [Davies Liu] memorize the module and class name during pickleing
f3506c5 [Davies Liu] Merge branch 'master' into pickle_mllib
722dd96 [Davies Liu] cleanup _common.py
0ee1525 [Davies Liu] remove outdated tests
b02e34f [Davies Liu] remove _common.py
84c721d [Davies Liu] Merge branch 'master' into pickle_mllib
4d7963e [Davies Liu] remove muanlly serialization
6d26b03 [Davies Liu] fix tests
c383544 [Davies Liu] classification
f2a0856 [Davies Liu] mllib/regression
d9f691f [Davies Liu] mllib/util
cccb8b1 [Davies Liu] mllib/tree
8fe166a [Davies Liu] Merge branch 'pickle' into pickle_mllib
aa2287e [Davies Liu] random
f1544c4 [Davies Liu] refactor clustering
52d1350 [Davies Liu] use new protocol in mllib/stat
b30ef35 [Davies Liu] use pickle to serialize data for mllib/recommendation
f44f771 [Davies Liu] enable tests about array
3908f5c [Davies Liu] Merge branch 'master' into pickle
c77c87b [Davies Liu] cleanup debugging code
60e4e2f [Davies Liu] support unpickle array.array for Python 2.6
Local `SparseMatrix` support added in Compressed Column Storage (CCS) format in addition to Level-2 and Level-3 BLAS operations such as dgemv and dgemm respectively.
BLAS doesn't support sparse matrix operations, therefore support for `SparseMatrix`-`DenseMatrix` multiplication and `SparseMatrix`-`DenseVector` implementations have been added. I will post performance comparisons in the comments momentarily.
Author: Burak <brkyvz@gmail.com>
Closes#2294 from brkyvz/SPARK-3418 and squashes the following commits:
88814ed [Burak] Hopefully fixed MiMa this time
47e49d5 [Burak] really fixed MiMa issue
f0bae57 [Burak] [SPARK-3418] Fixed MiMa compatibility issues (excluded from check)
4b7dbec [Burak] 9/17 comments addressed
7af2f83 [Burak] sealed traits Vector and Matrix
d3a8a16 [Burak] [SPARK-3418] Squashed missing alpha bug.
421045f [Burak] [SPARK-3418] New code review comments addressed
f35a161 [Burak] [SPARK-3418] Code review comments addressed and multiplication further optimized
2508577 [Burak] [SPARK-3418] Fixed one more style issue
d16e8a0 [Burak] [SPARK-3418] Fixed style issues and added documentation for methods
204a3f7 [Burak] [SPARK-3418] Fixed failing Matrix unit test
6025297 [Burak] [SPARK-3418] Fixed Scala-style errors
dc7be71 [Burak] [SPARK-3418][MLlib] Matrix unit tests expanded with indexing and updating
d2d5851 [Burak] [SPARK-3418][MLlib] Sparse Matrix support and additional native BLAS operations added
Added minInstancesPerNode, minInfoGain params to:
* DecisionTreeRunner.scala example
* Python API (tree.py)
Also:
* Fixed typo in tree suite test "do not choose split that does not satisfy min instance per node requirements"
* small style fixes
CC: mengxr
Author: qiping.lqp <qiping.lqp@alibaba-inc.com>
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Author: chouqin <liqiping1991@gmail.com>
Closes#2349 from jkbradley/chouqin-dt-preprune and squashes the following commits:
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
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
0278a11 [chouqin] remove `noSplit` and set `Predict` private to tree
d593ec7 [chouqin] fix docs and change minInstancesPerNode to 1
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
Updating this to reflect the newest SVD via ARPACK
Author: Reza Zadeh <rizlar@gmail.com>
Closes#2389 from rezazadeh/irmdocs and squashes the following commits:
7fa1313 [Reza Zadeh] Update svd docs
715da25 [Reza Zadeh] Updated computeSVD documentation IndexedRowMatrix
