Hivemall: Scalable machine learning library for Apache Hive/Spark/Pig

1). Research Engineer, Treasure DataMakoto YUI @myui

2016/10/26 Hadoop Summit '16, Tokyo

2). Research Engineer, NTTTakashi Yamamuro @maropu

#hivemall

Plan of the talk

1. Introduction to Hivemall

2. Hivemall on Spark

2016/10/26 Hadoop Summit '16, Tokyo 2

2016/10/26 Hadoop Summit '16, Tokyo 3

Hivemall enters Apache Incubator on Sept 13, 2016 🎉

http://incubator.apache.org/projects/hivemall.html

@ApacheHivemall

• Makoto Yui <Treasure Data>• Takeshi Yamamuro <NTT>

Hivemall on Apache Spark• Daniel Dai <Hortonworks>

Hivemall on Apache Pig Apache Pig PMC member

• Tsuyoshi Ozawa <NTT>Apache Hadoop PMC member

• Kai Sasaki <Treasure Data>

2016/10/26 Hadoop Summit '16, Tokyo 4

Initial committers

Champion

Nominated Mentors

2016/10/26 Hadoop Summit '16, Tokyo 5

Project mentors

• Reynold Xin <Databricks, ASF member>Apache Spark PMC member

• Markus Weimer <Microsoft, ASF member> Apache REEF PMC member

• Xiangrui Meng <Databricks, ASF member>Apache Spark PMC member

• Roman Shaposhnik <Pivotal, ASF member>Apache Bigtop/Incubator PMC member

What is Apache Hivemall

Scalable machine learning library built as a collection of Hive UDFs

2016/10/26 Hadoop Summit '16, Tokyo

github.com/myui/hivemall

hivemall.incubator.apache.org

Hivemall’s Vision: ML on SQL

2016/10/26 Hadoop Summit '16, Tokyo 7

100+ lines

of code

Classification with Mahout

CREATE TABLE lr_model ASSELECT feature, -- reducers perform model averaging in parallel avg(weight) as weightFROM ( SELECT logress(features,label,..) as (feature,weight) FROM train) t -- map-only taskGROUP BY feature; -- shuffled to reducers

✓ Machine Learning made easy for SQL developers✓ Interactive and Stable APIs w/ SQL abstraction

This SQL query automatically runs in parallel on Hadoop

Hadoop HDFS

MapReduce(MRv1)

Hivemall

Apache YARN

Apache Tez DAG processing

Machine Learning

Query Processing

Parallel Data Processing Framework

Resource Management

Distributed File SystemCloud Storage

SparkSQL

Apache Spark

MESOS

Hive Pig

MLlib

Hivemall’s Technology Stack

2016/10/26 Hadoop Summit '16, Tokyo 8

Amazon S3

List of supported AlgorithmsClassification

✓ Perceptron✓ Passive Aggressive (PA, PA1, PA2)✓ Confidence Weighted (CW)✓ Adaptive Regularization of Weight

Vectors (AROW)✓ Soft Confidence Weighted (SCW)✓ AdaGrad+RDA✓ Factorization Machines✓ RandomForest Classification

2016/10/26 Hadoop Summit '16, Tokyo 9

Regression✓Logistic Regression (SGD)✓AdaGrad (logistic loss)✓AdaDELTA (logistic loss)✓PA Regression ✓AROW Regression✓Factorization Machines✓RandomForest Regression

SCW is a good first choiceTry RandomForest if SCW does not work

Logistic regression is good for getting a probability of a positive class

Factorization Machines is good where features are sparse and categorical ones

List of Algorithms for Recommendation

2016/10/26 Hadoop Summit '16, Tokyo 10

K-Nearest Neighbor✓ Minhash and b-Bit Minhash (LSH variant)

✓ Similarity Search on Vector Space (Euclid/Cosine/Jaccard/Angular)

Matrix Completion✓ Matrix Factorization✓ Factorization Machines (regression)

each_top_k function of Hivemall is useful for recommending top-k items

2016/10/26 Hadoop Summit '16, Tokyo 11

student class score1 b 702 a 803 a 904 b 505 a 706 b 60

Top-k query processing

student class score3 a 902 a 801 b 706 b 60

List top-2 students for each class

2016/10/26 Hadoop Summit '16, Tokyo 12

student class score1 b 702 a 803 a 904 b 505 a 706 b 60

Top-k query processing

List top-2 students for each class

SELECT * FROM ( SELECT *, rank() over (partition by class order by score desc) as rank FROM table) tWHERE rank <= 2

2016/10/26 Hadoop Summit '16, Tokyo 13

student class score1 b 702 a 803 a 904 b 505 a 706 b 60

Top-k query processing

List top-2 students for each classSELECT each_top_k( 2, class, score, class, student ) as (rank, score, class, student)FROM ( SELECT * FROM table DISTRIBUTE BY class SORT BY class) t

2016/10/26 Hadoop Summit '16, Tokyo 14

Top-k query processing by RANK OVER()

partition by class

Node 1

Sort by class, score

rank over()

rank >= 2

2016/10/26 Hadoop Summit '16, Tokyo 15

Top-k query processing by EACH_TOP_K

distributed by class

Node 1

Sort by class

each_top_k

OUTPUT only K items

2016/10/26 Hadoop Summit '16, Tokyo 16

Comparison between RANK and EACH_TOP_K

distributed by class

Sort by class

each_top_k

Sort by class, score

rank over()

rank >= 2

SORTING IS HEAVY

NEED TO PROCESS ALL

OUTPUT only K items

Each_top_k is very efficient where the number of class is large

Bounded Priority Queueis utilized

Performance reported by TD customer

2016/10/26 Hadoop Summit '16, Tokyo 17

•1,000 students in each class•20 million classes

RANK over() query does not finishes in 24 hours EACH_TOP_K finishes in 2 hours

Refer for detailhttps://speakerdeck.com/kaky0922/hivemall-meetup-20160908

Other Supported Algorithms

2016/10/26 Hadoop Summit '16, Tokyo 18

Anomaly Detection✓ Local Outlier Factor (LoF)

Feature Engineering✓Feature Hashing✓Feature Scaling (normalization, z-score)

✓ TF-IDF vectorizer✓ Polynomial Expansion (Feature Pairing)

✓ Amplifier

NLP✓Basic Englist text Tokenizer ✓Japanese Tokenizer (Kuromoji)

Industry use cases of Hivemall

CTR prediction of Ad click logs• Algorithm: Logistic regression• Freakout Inc., Smartnews, and more

Gender prediction of Ad click logs• Algorithm: Classification• Scaleout Inc.

2016/10/26 Hadoop Summit '16, Tokyo 19

http://www.slideshare.net/eventdotsjp/hivemall

Industry use cases of Hivemall

Item/User recommendation• Algorithm: Recommendation• Wish.com, GMO pepabo

2016/10/26 Hadoop Summit '16, Tokyo 20

Problem: Recommendation using hot-item is hard in hand-crafted product market because each creator sells few single items (will soon become out-of-stock)

minne.com

Industry use cases of Hivemall

Value prediction of Real estates• Algorithm: Regression• Livesense

2016/10/26 Hadoop Summit '16, Tokyo 21

Industry use cases of Hivemall

User score calculation• Algrorithm: Regression• Klout

2016/10/26 Hadoop Summit '16, Tokyo 22

bit.ly/klout-hivemall

Influencer marketing

klout.com

2016/10/26 Hadoop Summit '16, Tokyo 23

Efficient algorithm for finding change point and outliers from timeseries data

2016/10/26 Hadoop Summit '16, Tokyo 24

J. Takeuchi and K. Yamanishi, “A Unifying Framework for Detecting Outliers and Change Points from Time Series,” IEEE transactions on Knowledge and Data Engineering, pp.482-492, 2006.

Anomaly/Change-point Detection by ChangeFinder

Efficient algorithm for finding change point and outliers from timeseries data

2016/10/26 Hadoop Summit '16, Tokyo 25

Anomaly/Change-point Detection by ChangeFinder

J. Takeuchi and K. Yamanishi, “A Unifying Framework for Detecting Outliers and Change Points from Time Series,” IEEE transactions on Knowledge and Data Engineering, pp.482-492, 2006.

2016/10/26 Hadoop Summit '16, Tokyo 26

• T. Ide and K. Inoue, "Knowledge Discovery from Heterogeneous Dynamic Systems using Change-Point Correlations", Proc. SDM, 2005T.

• T. Ide and K. Tsuda, "Change-point detection using Krylov subspace learning", Proc. SDM, 2007.

Change-point detection by Singular Spectrum Transformation

2016/10/26 Hadoop Summit '16, Tokyo 27

Evaluation Metrics

2016/10/26 Hadoop Summit '16, Tokyo 28

Feature Engineering – Feature Binning

Maps quantitative variables to fixed number of bins based on quantiles/distribution

Map Ages into 3 bins

2016/10/26 Hadoop Summit '16, Tokyo 29

Feature Selection – Signal Noise Ratio

2016/10/26 Hadoop Summit '16, Tokyo 30

Feature Selection – Chi-Square

2016/10/26 Hadoop Summit '16, Tokyo 31

Feature Transformation – Onehot encoding

Maps a categorical variable to a unique number starting from 1

Spark 2.0 support XGBoost Integration Field-aware Factorization Machines Generalized Linear Model

• Optimizer framework including ADAM• L1/L2 regularization

2016/10/26 Hadoop Summit '16, Tokyo 32

Other new features to come

Copyright©2016 NTT corp. All Rights Reserved.

Hivemall onTakeshi Yamamuro @ NTT

Hadoop Summit 2016Tokyo, Japan

34Copyright©2016 NTT corp. All Rights Reserved.

Who am I?

35Copyright©2016 NTT corp. All Rights Reserved.

What’s Spark?• 1. Unified Engine

• support end-to-end APs, e.g., MLlib and Streaming

• 2. High-level APIs• easy-to-use, rich optimization

• 3. Integrate broadly• storages, libraries, ...

36Copyright©2016 NTT corp. All Rights Reserved.

• Hivemall wrapper for Spark• Wrapper implementations for DataFrame/SQL• + some utilities for easy-to-use in Spark

• The wrapper makes you...• run most of Hivemall functions in Spark• try examples easily in your laptop• improve some function performance in Spark

What’s Hivemall on Spark?

37Copyright©2016 NTT corp. All Rights Reserved.

• Hivemall already has many fascinating ML algorithms and useful utilities

• High barriers to add newer algorithms in MLlib

Why’s Hivemall on Spark?

https://cwiki.apache.org/confluence/display/SPARK/Contributing+to+Spark

38Copyright©2016 NTT corp. All Rights Reserved.

• Most of Hivemall functions supported in Spark v1.6 and v2.0

Current Status

- For Spark v2.0$ git clone https://github.com/myui/hivemall$ cd hivemall$ mvn package –Pspark-2.0 –DskipTests$ ls target/*spark*target/hivemall-spark-2.0_2.11-XXX-with-dependencies.jar...

39Copyright©2016 NTT corp. All Rights Reserved.

• Most of Hivemall functions supported in Spark v1.6 and v2.0

Current Status

- For Spark v2.0

$ git clone https://github.com/myui/hivemall$ cd hivemall$ mvn package –Pspark-2.0 –DskipTests$ ls target/*spark*target/hivemall-spark-2.0_2.11-XXX-with-dependencies.jar...

-Pspark-1.6 for Spark v1.6

40Copyright©2016 NTT corp. All Rights Reserved.

• 1. Download a Spark binary• 2. Fetch training and test data• 3. Load these data in Spark• 4. Build a model• 5. Do predictions

Running an Example

41Copyright©2016 NTT corp. All Rights Reserved.

1. Download a Spark binary Running an Example

http://spark.apache.org/downloads.html

42Copyright©2016 NTT corp. All Rights Reserved.

• E2006 tfidf regression dataset• http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/

datasets/regression.html#E2006-tfidf

2. Fetch training and test dataRunning an Example

$ wget http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression/E2006.train.bz2

$ wget http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression/E2006.test.bz2

43Copyright©2016 NTT corp. All Rights Reserved.

3. Load data in Spark Running an Example

$ <SPARK_HOME>/bin/spark-shell -conf spark.jars=hivemall-spark-2.0_2.11-XXX-with-dependencies.jar

// Create DataFrame from the bzip’d libsvm-formatted filescala> val trainDf = sqlContext.sparkSession.read.format("libsvm”) .load(“E2006.train.bz2")

scala> trainDf.printSchemaroot |-- label: double (nullable = true) |-- features: vector (nullable = true)

44Copyright©2016 NTT corp. All Rights Reserved.

3. Load data in Spark Running an Example

0.000357499151147113 6066:0.00079327062196048 6069:0.000311377727123504 6070:0.000306754934580457 6071:0.000276992485786437 6072:0.00039663531098024 6074:0.00039663531098024 6075:0.00032548335…

trainDfPartition1 Partition2 Partition3 PartitionN

…

…

…

Load in parallel becausebzip2 is splittable

45Copyright©2016 NTT corp. All Rights Reserved.

4. Build a model - DataFrame Running an Example

scala> paste:val modelDf = trainDf.train_logregr($"features", $"label") .groupBy("feature”) .agg("weight" -> "avg”)

46Copyright©2016 NTT corp. All Rights Reserved.

4. Build a model - SQL Running an Example

scala> trainDf.createOrReplaceTempView("TrainTable")scala> paste:val modelDf = sql(""" | SELECT feature, AVG(weight) AS weight | FROM ( | SELECT train_logregr(features, label) | AS (feature, weight) | FROM TrainTable | ) | GROUP BY feature """.stripMargin)

47Copyright©2016 NTT corp. All Rights Reserved.

5. Do predictions - DataFrame Running an Example

scala> paste:val df = testDf.select(rowid(), $"features") .explode_vector($"features") .cache

# Do predictionsdf.join(modelDf, df("feature") === model("feature"), "LEFT_OUTER") .groupBy("rowid") .avg(sigmoid(sum($"weight" * $"value")))

48Copyright©2016 NTT corp. All Rights Reserved.

5. Do predictions - SQL Running an Example

scala> modelDf.createOrReplaceTempView(”ModelTable")scala> df.createOrReplaceTempView(”TestTable”)scala> paste:sql(""" | SELECT rowid, sigmoid(value * weight) AS predicted | FROM TrainTable t | LEFT OUTER JOIN ModelTable m | ON t.feature = m.feature | GROUP BY rowid """.stripMargin)

49Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.1) Compute a sigmoid function

Improve Some Functions in Spark

scala> val sigmoidFunc = (d: Double) => 1.0 / (1.0 + Math.exp(-d))scala> val sparkUdf = functions.udf(sigmoidFunc)scala> df.select(sparkUdf($“value”))

50Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.1) Compute a sigmoid function

Improve Some Functions in Spark

scala> val hiveUdf = HivemallOps.sigmoidscala> df.select(hiveUdf($“value”))

51Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.1) Compute a sigmoid function

Improve Some Functions in Spark

52Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.2) Compute top-k for each key group

Improve Some Functions in Spark

scala> paste:df.withColumn( “rank”, rank().over(Window.partitionBy($"key").orderBy($"score".desc)) .where($"rank" <= topK)

53Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.2) Compute top-k for each key group

• Fixed the overhead issue for each_top_k• See pr#353: “Implement EachTopK as a generator expression“ in Spark

Improve Some Functions in Spark

scala> df.each_top_k(topK, “key”, “score”, “value”)

54Copyright©2016 NTT corp. All Rights Reserved.

• Spark has overheads to call Hive UD*Fs• Hivemall heavily depends on them

• ex.2) Compute top-k for each key group

Improve Some Functions in Spark

~4 times faster than rank!!

55Copyright©2016 NTT corp. All Rights Reserved.

• supports under development• fast implementation of the gradient tree boosting• widely used in Kaggle competitions

• This integration will make you...• load built models and predict in parallel• build multiple models in parallel for cross-validation

3rd–Party Library Integration

Conclusion and TakeawayHivemall provides a collection of machine learning algorithms as Hive UDFs/UDTFs

2016/10/26 Hadoop Summit '16, Tokyo 56

We welcome your contributions to Apache Hivemall

2016/10/26 Hadoop Summit '16, Tokyo 57

Any feature request or questions?