Predictive Analytics with Hadoop
Upcoming SlideShare
Loading in...5
×
 

Predictive Analytics with Hadoop

on

  • 1,009 views

 

Statistics

Views

Total Views
1,009
Views on SlideShare
1,009
Embed Views
0

Actions

Likes
0
Downloads
70
Comments
1

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Microsoft PowerPoint

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel

11 of 1

  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
  • I want to run regression model on bigr.frame data with reference to slide 58. But I can not find the function bigr.lm
    It does not exist in any package
    is there any way to do that?
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment
  • http://www-01.ibm.com/software/data/infosphere/biginsights/

Predictive Analytics with Hadoop Predictive Analytics with Hadoop Presentation Transcript

  • Predictive Analytics with Hadoop Michael Hausenblas MapR Technologies
  • © 2014 MapR Technologies 2© 2014 MapR Technologies
  • © 2014 MapR Technologies 3 Agenda • Examples • Data-driven solutions • Obtaining big training data • Recommendation with Mahout and Solr • Operational considerations
  • © 2014 MapR Technologies 4 Recommendation is Everywhere Media and Advertising e-commerce Enterprise Sales • Recommend sales opportunities to partners • $40M revenue in year 1 • 1.5B records per day • Using MapR
  • © 2014 MapR Technologies 5 Classification is Everywhere IP address blacklisting Fraud Detection Customer 360 Scoring & Categorization • Customer 360 application • Each customer is scored and categorized based on all their activity • Data from hundreds of streams and databases • Using MapR • 600+ variables considered for every IP address • Billions of data points • Using MapR • Identify anomalous patterns indicating fraud, theft and criminal activity • Stop phishing attempts • Using MapR Fortune 100 Telco
  • © 2014 MapR Technologies 6 Data-Driven Solutions • Physics is „simple‟: f = ma; E=mc2 • Human behavior is much more complex – Which ad will they click? – Is a behavior fraudulent? Why? • Don‟t look for complex models that try to discover general rules – The size of the dataset is the most important factor – Simple models (n-gram, linear classifiers) with Big Data • A. Halevy, P. Norvig, and F. Pereira. The unreasonable effectiveness of data. IEEE Intelligent Systems, 24(2):8-12, 2009.
  • © 2014 MapR Technologies 7 The Algorithms Are Less Important
  • © 2014 MapR Technologies 8 Focus on the Data • Most algorithms come down to counting and simple math • Invest your time where you can make a difference – Getting more data can improve results by 2x • eg, add beacons everywhere to instrument user behavior – Tweaking an ML algorithm will yield a fraction of 1% • Data wrangling – Feature engineering – Moving data around – …
  • © 2014 MapR Technologies 9 Obtaining Big Training Data • Can‟t really rely on experts to label the data – Doesn‟t scale (not enough experts out there) – Too expensive • So how do you get the training data? – Crowdsourcing – Implicit feedback • “Obvious” features • User engagement
  • © 2014 MapR Technologies 10 Using Crowdsourcing for Annotation R. Snow, B. O'Connor, D. Jurafsky, and A. Ng. Cheap and fast – but is it good? Evaluating non-expert annotations for natural language tasks. EMNLP, 2008. Quantity: $2 for 7000 annotations (leveraging Amazon Mechanical Turk and a “flat world”) Quality: 4 non-experts = 1 expert
  • © 2014 MapR Technologies 11 Using “Obvious” Features for Annotation :) :(
  • © 2014 MapR Technologies 12 Leveraging Implicit Feedback • Users behavior provides valuable training data • Google adjusts search rankings based on engagement – Did the user click on the result? – Did the user come back to the search page within seconds? • Most recommendation algorithms are based solely on user activity – What products did they view/buy? – What ads did they click on? • T. Joachims, L. Granka, B. Pan, H. Hembrooke, F. Radlinski, and G. Gay. Evaluating the accuracy of implicit feedback from clicks and query reformulations in Web search. ACM TOIS, 25(2):1, 2007.
  • © 2014 MapR Technologies 13 Increasing Exploration We need to find ways to increase exploration to broaden our learning Exploration of the second page Can‟t learn much about results on page 2, 3 or 4!
  • © 2014 MapR Technologies 14 Result Dithering • Dithering is used to re-order recommendation results – Re-ordering is done randomly • Dithering is guaranteed to make off-line performance worse • Dithering also has a near perfect record of making actual performance much better – “Made more difference than any other change”
  • © 2014 MapR Technologies 15 Simple Dithering Algorithm • Generate synthetic score from log rank plus Gaussian • Pick noise scale to provide desired level of mixing • Typically: • Oh… use floor(t/T) as seed so results don‟t change too often s = logr + N(0,e) e Î 0.4, 0.8[ ] Dr µrexpe
  • © 2014 MapR Technologies 16 Example:ε = 0.5 • Each line represents a recommendation of 8 items • The non-dithered recommendation would be 1, 2, …, 8 1 2 6 5 3 4 13 16 1 2 3 8 5 7 6 34 1 4 3 2 6 7 11 10 1 2 4 3 15 7 13 19 1 6 2 3 4 16 9 5 1 2 3 5 24 7 17 13 1 2 3 4 6 12 5 14 2 1 3 5 7 6 4 17 4 1 2 7 3 9 8 5 2 1 5 3 4 7 13 6 3 1 5 4 2 7 8 6 2 1 3 4 7 12 17 16
  • © 2014 MapR Technologies 17 Example:ε = log 2 = 0.69 • Each line represents a recommendation of 8 items • The non-dithered recommendation would be 1, 2, …, 8 1 2 8 3 9 15 7 6 1 8 14 15 3 2 22 10 1 3 8 2 10 5 7 4 1 2 10 7 3 8 6 14 1 5 33 15 2 9 11 29 1 2 7 3 5 4 19 6 1 3 5 23 9 7 4 2 2 4 11 8 3 1 44 9 2 3 1 4 6 7 8 33 3 4 1 2 10 11 15 14 11 1 2 4 5 7 3 14 1 8 7 3 22 11 2 33
  • © 2014 MapR Technologies 18© 2014 MapR Technologies Recommendations with Mahout and Solr
  • © 2014 MapR Technologies 19 What is Recommendation? The behavior of a crowd helps us understand what individuals will do…
  • © 2014 MapR Technologies 20 Batch and Real-Time • We can learn about the relationship between items every X hours – These relationships don‟t change often – People who buy Nikon D7100 cameras also buy a Nikon EN-EL15 battery • What to recommend to Bob has to be determined in real-time – Bob may be a new user with no history – Bob is shopping for a camera right now, but he was buying a baby bottle an hour ago • How do we do that? – Mahout for the heavy number crunching – Solr/Elasticsearch for the real-time recommendations
  • © 2014 MapR Technologies 21 Real-Time Recommender Architecture Co-occurrence (Mahout) Indexing (Solr) Index shardsItem metadata Web tier Search (Solr) Recent history for single user Recommendation Complete history (log files/tables) Note: All data lives in the cluster
  • © 2014 MapR Technologies 22 Recommendations Alice got an apple and a puppy Charles got a bicycle Alice Charles
  • © 2014 MapR Technologies 23 Recommendations Alice got an apple and a puppy Charles got a bicycle Bob got an apple Alice Bob Charles
  • © 2014 MapR Technologies 24 Recommendations ? What else would Bob like? Alice Bob Charles
  • © 2014 MapR Technologies 25 Log Files Alice Bob Charles Alice Bob Charles Alice
  • © 2014 MapR Technologies 26 History Matrix Alice Bob Charle s ✔ ✔ ✔ ✔ ✔ ✔ ✔
  • © 2014 MapR Technologies 27 Co-Occurrence Matrix: Items by Items Q: How do you tell which co-occurrences are useful? A: Let Mahout do the math… - 1 2 1 1 1 1 2 1 0 0 0 0
  • © 2014 MapR Technologies 29 Indicator Matrix: Anomalous Co-Occurrences ✔ ✔ Result: The marked row will be added to the indicator field in the item document…
  • © 2014 MapR Technologies 30 Indicator Matrix ✔ id: t4 title: puppy desc: The sweetest little puppy ever. keywords: puppy, dog, pet indicators: (t1) That one row from indicator matrix becomes the indicator field in the Solr document used to deploy the recommendation engine. Note: Data for the indicator field is added directly to meta-data for a document in Solr index. You don‟t need to create a separate index for the indicators.
  • © 2014 MapR Technologies 31 Internals of the Recommender Engine Q: What should we recommend if new user listened to 2122:Fats Domino & 303:Beatles? A: Search the index with “indicator_artists:2122 OR indicator_artists:303”
  • © 2014 MapR Technologies 32 Internals of the Recommender Engine 1710:Check Berry is the top recommendation
  • © 2014 MapR Technologies 33 Toolbox for Predictive Analytics with Hadoop • Mahout – Use it for Recommendations, Clustering, Math • Vowpal Wabbit – Use it for Classification (but it‟s harder to use) • SkyTree – Commercial (not open source) implementation of common algorithms
  • © 2014 MapR Technologies 34 Operational Considerations • Snapshot your raw data before training – Reproducible process – MapR snapshots • Leverage NFS for real-time data ingestion – Train the model on today‟s data – Learning schedule independent from ingestion schedule • Look for HA, data protection, disaster recovery – Predictive analytics increases revenue or reduces cost – Quickly becomes a must-have, not a nice-to-have
  • © 2014 MapR Technologies 35 Q&AEngage with us! mhausenblas@mapr.com +353 86 0215164 @MapR_EMEA maprtech +MaprTechnologies maprtech MapR Technologies
  • Analyzing Big Data with Open Source R and Hadoop Steven Sit IBM
  • Analyzing big data with open source R and Hadoop Steven Sit IBM Silicon Valley Laboratory
  • 39 Outline • R on Hadoop • Rationale and requirements • Various approaches • Using R and “Big R” packages • Data exploration, statistical analysis and Visualization • Scale out R with data partitioning • Distributed machine learning • Demo
  • 40 Rationale and requirements Productivity • Use natural R syntax to access massive data in Hadoop • Leverage existing packages and libraries Platform for analytics • Common input and output data formats across analytic algorithms Customizability & Extensibility • Easily customize existing algorithms • Support for both conventional data mining and large scale ML needs Scalability & Performance • Scale to massively parallel clusters • Analyze terabytes and petabytes of data
  • 41 Streams ETL Tools SQOOP Flume NFSRest API Files NoSQL DB Hadoop Business Intelligence Tools ODBC JDBC, Utilities Warehouses, Marts, MDM Analytic Runtimes Analytical Tools SQL engines DDL Database Tools PIG, Java, etc Exploration Tools Search Indexes Models Sources Common Patterns for Hadoop and Analytics
  • 42 Quick Tour of R • R is an interpreted language • Open-source implementation of the S language (1976) • Best suited for statistical analysis and modeling • Data exploration and manipulation • Descriptive statistics • Predictive analytics and machine learning • Visualization • +++ • Can produce “publication quality graphics” • Emerging as a competitor to proprietary platforms
  • 43 R is hot!... but • Quite lean as far as software goes • Free! • State of the art algorithms • Statistical researchers often provide their methods as R packages • New techniques available without delay • Commercial packages usually behind the curve • 4700+ packages as of today • Active and vibrant user community • Universities are teaching R • IT shops are adopting R • Companies are integrating R into their products • R jobs and demand for R skills on the rise Unfortunately R is not built for Big Data Working with large datasets is limited by RAM
  • 44 R and Big Data: Various approaches• R + Hadoop streaming • Use R as a scripting language • Write map/reduce jobs in R • Or Python for mappers, and R for the reducers • R + open-source packages for Hadoop • RHadoop • RMR2: Write low-level map/reduce with API in R • RHBase and RHDFS: Simple R API over Hadoop capabilities • RHipe • R + SparkR + Hadoop • R frontend to Spark in memory structure and data from Hadoop • R + High Level Languages • JaQL<->R “bridge” • R interfaces over non-R engines • ScalaR from Revolution R • Oracle’s ORE, Netezza’s NZR, Teradata’s teradataR, +++
  • 45 RHadoop Example Given a table describing flight information for every airline spanning several decades: “Carrier”, “Year”, “Month”, “Day”, “DepDelay”, … What’s the mean departure delay (“DepDelay”) for each airline for each month?
  • 46 RHadoop Example (contd.) Map Reduce key = c(Carrier, Year, Mo nth) value = DepDelay key = c(Carrier, Year, Mo nth) value = vector(DepDelay) key = NULL value = vector (columns in a line) key = c(Carrier, Year, Mo nth) value = mean(DepDelay) Shuffle
  • 47 RHadoop Example (contd.) csvtextinputformat = function(line) keyval(NULL, unlist(strsplit(line, ","))) deptdelay = function (input, output) { mapreduce(input = input, output = output, textinputformat = csvtextinputformat, map = function(k, fields) { # Skip header lines and bad records: if (!(identical(fields[[1]], "Year")) & length(fields) == 29) { deptDelay <- fields[[16]] # Skip records where departure dalay is "NA": if (!(identical(deptDelay, "NA"))) { # field[9] is carrier, field[1] is year, field[2] is month: keyval(c(fields[[9]], fields[[1]], fields[[2]]), deptDelay)}}}, reduce = function(keySplit, vv) { keyval(keySplit[[2]], c(keySplit[[3]], length(vv), keySplit[[1]], mean(as.numeric(vv))))})} from.dfs(deptdelay("/data/airline/1987.csv", "/dept-delay-month")) Source: http://blog.revolutionanalytics.com/2012/03/r-and-hadoop-step-by-step-tutorials.html
  • 48 What is “Big R” • Explore, visualize, transform, and model big data using familiar R syntax and paradigm • Scale out R with MR programming • Partitioning of large data • Parallel cluster execution of R code • Distributed Machine Learning • A scalable statistics engine that provides canned algorithms, and an ability to author new ones, all via R R Clients Scalabl e Machin e Learnin g Data Sources Embedde d R Executio n IBM R Packages IBM R Packages Pull data (summaries) to R client Or, push R functions right on the data 1 2 3
  • Let‟s mix it up a little ... ... with a demo interspersed with slides.
  • 50 Big R Demo Data • Publicly available “airline” data • 22 years of actual arrival/departure information • Every scheduled flight in the US • 1987 onwards • From U.S. Department of Transportation • Cleansed version • http://stat-computing.org/dataexpo/2009/the-data.html
  • 51 Airline data description Year 1987-2008 Month 1-12 DayofMonth 1-31 DayOfWeek 1 (Monday) - 7 (Sunday) DepTime actual departure time (local, hhmm) CRSDepTime scheduled departure time (local, hhmm) ArrTime actual arrival time (local, hhmm) CRSArrTime scheduled arrival time (local, hhmm) UniqueCarrier unique carrier code FlightNum flight number TailNum plane tail number ActualElapsedTime in minutes CRSElapsedTime in minutes AirTime in minutes ArrDelay arrival delay, in minutes
  • Airline data description (contd.) Origin origin IATA airport code Dest destination IATA airport code Distance in miles TaxiIn taxi in time, in minutes TaxiOut taxi out time in minutes Cancelled was the flight cancelled? CancellationCode reason for cancellation (A = carrier, B = weather, C = NAS, D = security) Diverted 1 = yes, 0 = no CarrierDelay in minutes WeatherDelay in minutes NASDelay in minutes SecurityDelay in minutes LateAircraftDelay in minutes
  • 53 Explore, visualize, transform, model Hadoop data with R • Represent Big Data objects as R datatypes • R's programming syntax and paradigm • Data stays in HDFS • R classes (e.g. bigr.data.frame) as proxies • No low-level map/reduce constructs • No underlying scripting languages 1 # Construct a bigr.frame to access large data set air <- bigr.frame(dataPath="airline_demo.csv", …) # How many flights were flown by United or Delta? length(UniqueCarrier[UniqueCarrier %in% c("UA", "DL")]) # Filter all flights that were delayed by 15+ minutes at departure or arrival. airSubset <- air[air$Cancelled == 0 & (air$DepDelay >= 15 | air$ArrDelay >= 15), c("UniqueCarrier", "Origin", "Dest", "DepDelay", "ArrDelay")] # For these filtered flights, compute key statistics (# of flights, # average flying distance and flying time), grouped by airline summary(count(UniqueCarrier) + mean(Distance) + mean(CRSElapsedTime) ~ UniqueCarrier, dataset = airSubset) Bigr.boxplot(air$Distance ~ air$UniqueCarrier, …)
  • 54 Scale out R in Hadoop • Support parallel / partitioned execution of R • Work around R’s memory limitations • Execute R snippets on chunks of data • Partitioned by key, by #rows, via sampling, … • Follows R’s “apply” model • Parallelized seamlessly Map/Reduce engine 2 # Filter the airline data on United and Hawaiian bf <- air[air$UniqueCarrier %in% c("HA", "UA"),] # Build one decision-tree model per airliner models <- groupApply(data = bf, groupingColumns = list(bf$UniqueCarrier), rfunction = function(df) { library(rpart) predcols <- c('ArrDelay', 'DepDelay', 'DepTime'', 'Distance') return (rpart(ArrDelay ~ ., df[,predcols]))}) # Pull the model for HA to the client modelHA <- bigr.pull(models$HA) # Visualize the model prettyTree(modelHA)
  • Big R API: Core Classes & Methods Connection handling bigr.connect() and bigr.disconnect() is.bigr.connected() bigr.frame Modeled after R’s data.frame Proxy for tabular data bigr.vector Modeled after R’s vector datatype Proxy for a column bigr.list (Loosely) modeled after R’s list Proxy for collections of serialized R objects • Basic exploration • head(), tail() • dim(), length(), nrow( ), ncol() • str(), summary() • Selection and Projections • [ • $ • Arithmetic and Logical operators • +, -, /, - • &, |, ! • ifelse() • String and Math functions • Lots of these • Other relational operators • table() • unique() • merge() • summary() • sort() • na.omit(), na.exclude() • Data movement • as.data.frame, as.bigr.fram e • bigr.persist • Sampling • bigr.sample() • bigr.random() • Visualizations • built into R packages (E.g. ggplot2)
  • Big R API: Core Classes & Methods (contd.) • Summarization and aggregation • summary(), very powerful when used with “formula” notation • max(), min(), ran ge() • Mean, variance and standard deviation • mean() • var() • sd() • Correlation • cov() and cor() • Hadoop options • bigr.get.server.option() • bigr.set.server.option() • bigr.debug(T) • Useful for servicing • Will print out internal debugging output • Catalog access • bigr.listfs() • bigr.listTables(), bigr.lis tColumns() • groupApply() • Primary function for embedded execution • Can return tables or objects • Run “help(groupApply)” inside R for extensive documentation • Examining R execution logs • bigr.logs() • Other *Apply functions • rowApply(), for running R on batches of rows • tableApply(), for running R on entire dataset
  • 57 Example: What is the average scheduled flight time, actual gate-to-gate time, and actual airtime for each city pair per year? mapper.year.market.enroute_time = function(key, val) { # Skip header lines, cancellations, and diversions: if ( !identical(as.character(val['Year']), 'Year') & identical(as.numeric(val['Cancelled']), 0) & identical(as.numeric(val['Diverted']), 0) ) { # We don't care about direction of travel, so construct 'market' # with airports ordered alphabetically # (e.g, LAX to JFK becomes 'JFK-LAX' if (val['Origin'] < val['Dest']) market = paste(val['Origin'], val['Dest'], sep='-') else market = paste(val['Dest'], val['Origin'], sep='-') # key consists of year, market output.key = c(val['Year'], market) # output gate-to-gate elapsed times (CRS and actual) + time in air output.val = c(val['CRSElapsedTime'], val['ActualElapsedTime'], val['AirTime']) return( keyval(output.key, output.val) ) } } reducer.year.market.enroute_time = function(key, val.list) { # val.list is a list of row vectors # a data.frame is a list of column vectors # plyr's ldply() is the easiest way to convert IMHO if ( require(plyr) ) val.df = ldply(val.list, as.numeric) else { # this is as close as my deficient *apply skills can come w/o plyr val.list = lapply(val.list, as.numeric) val.df = data.frame( do.call(rbind, val.list) ) } colnames(val.df) = c('actual','crs','air') output.key = key output.val = c( nrow(val.df), mean(val.df$actual, na.rm=T), mean(val.df$crs, na.rm=T), mean(val.df$air, na.rm=T) ) return( keyval(output.key, output.val) ) } mr.year.market.enroute_time = function (input, output) { mapreduce(input = input, output = output, input.format = asa.csvtextinputformat, map = mapper.year.market.enroute_time, reduce = reducer.year.market.enroute_time, backend.parameters = list( hadoop = list(D = "mapred.reduce.tasks=10") ), verbose=T) } hdfs.output.path = file.path(hdfs.output.root, 'enroute-time') results = mr.year.market.enroute_time(hdfs.input.path, hdfs.output.path) results.df = from.dfs(results, to.data.frame=T) colnames(results.df) = c('year', 'market', 'flights', 'scheduled', 'actual', 'in.air') save(results.df, file="out/enroute.time.RData") RHadoop Implementation (>35 lines of code) Equivalent Big R Implementation (4 lines of code) air <- bigr.frame(dataPath = "airline.csv", dataSource = “DEL", na.string="NA") air$City1 <- ifelse(air$Origin < air$Dest, air$Origin, air$Dest) air$City2 <- ifelse(air$Origin >= air$Dest, air$Origin, air$Dest) summary(count(UniqueCarrier) + mean(ActualElapsedTime) + mean(CRSElapsedTime) + mean(AirTime) ~ Year + City1 + City2 , dataset = air[air$Cancelled == 0 & air$Diverted == 0,])
  • 58 Use Cases • Where Big R works well • When data can be partitioned cleanly ... • ... and each partition fits in the memory of a server node • In other words: • size of entire data can be bigger than cluster memory • size of individual partitions limited to node memory (due to R) • Real-world customer scenarios we’ve seen: • Model data from individual tractors • Build time-series anomaly detection on IP address pairs • Build models on each customer’s behavior • And where it doesn’t ... • When building one monolithic model on the entire data • without sampling • Without use some form of blending such as ensembles
  • 59 Large scale analytics in Hadoop • Some workloads are not logically partitionable, or partitions are still large • SystemML - a scalable engine running natively over Hadoop: • Deliver pre-built ML algorithms • Regression, Classification, Clustering, etc. • Ability to author new algorithms 3 # Build a model on the entire data set model <- bigr.lm(ArrDelay ~ ., df) # Or, build several models, partitioned by airline models <- groupApply( input = air, groupBy = air$UniqueCarrier, function(df) { # Linear regression model <- bigr.lm(ArrDelay ~ ., df) return (model) })
  • 60 Association Rules Association • Apriori and pattern growth for frequent itemsets • sequence miner Clustering K-Means Data Mining Dimension Reduction • Non-negative Matrix Factorizations • Principal Component Analysis (large n, small p) • Singular Value Decompositions Time Series Analysis Granger modeling Predictive Analytics Regression Linear Regression for large, sparse datasets. Generalized Linear Models. Classification Linear Logistic Regression Trust Region Method Linear SVMs Modified Finite Newton Method Random decision trees for classification & regression Ranking PageRank of a directed graph HITS Hubs and Authorities Optimization Conjugate Gradient for Sparse Linear Systems Parallel Optimization for sparse linear models Stochastic Gradient Descent Outlier Detection Recursive Binning and reprojection for distance based outlier detection Univariate Scale, nominal, ordinal variables Scale: min, max, range, mean, variance, moments (kurtosis, skewness), order statistics (median, quantile, iqm), outliers Categorical: mode, histograms Bivariate Scale/categorical Eta, ftest, grouped mean/variance/weight Categorical/categorical cramer‟s V, pearson ChiSquare, spearman Data Exploration Large scale analytics (systemML) modules Recommender Systems Matrix Completion algorithms Meta Learning Ensemble Learning Cross Validation
  • 61 Example - Recommendation Systems with Collaborative Filtering ratings people W H Kfactors moviesK factors people 1 1 0.10 1 2 0.30 : : : 1 1 0.10 1 2 0.30 1 3 0.22 1 4 1.24 : : : : : : movies Analyzing both similarity of people and products
  • 62 Example: Topic Evolution in Social Media with Clustering tokens documents 1 1 0.10 1 2 0.30 1 3 0.22 1 4 1.24 : : : : : : W H Ktopics wordsK topics documents 1 1 0.10 1 2 0.30 : : :
  • 63 Example: Gaussian Non-negative Matrix Factorization package gnmf; import java.io.IOException; import java.net.URISyntaxException; import org.apache.hadoop.fs.FileSystem; import org.apache.hadoop.fs.Path; import org.apache.hadoop.mapred.JobConf; public class MatrixGNMF { public static void main(String[] args) throws IOException, URISyntaxException { if(args.length < 10) { System.out.println("missing parameters"); System.out.println("expected parameters: [directory of v] [directory of w] [directory of h] " + "[k] [num mappers] [num reducers] [replication] [working directory] " + "[final directory of w] [final directory of h]"); System.exit(1); } String vDir = args[0]; String wDir = args[1]; String hDir = args[2]; int k = Integer.parseInt(args[3]); int numMappers = Integer.parseInt(args[4]); int numReducers = Integer.parseInt(args[5]); int replication = Integer.parseInt(args[6]); String outputDir = args[7]; String wFinalDir = args[8]; String hFinalDir = args[9]; JobConf mainJob = new JobConf(MatrixGNMF.class); String vDirectory; String wDirectory; String hDirectory; FileSystem.get(mainJob).delete(new Path(outputDir)); vDirectory = vDir; hDirectory = hDir; wDirectory = wDir; String workingDirectory; String resultDirectoryX; String resultDirectoryY; long start = System.currentTimeMillis(); System.gc(); System.out.println("starting calculation"); System.out.print("calculating X = WT * V... "); workingDirectory = UpdateWHStep1.runJob(numMappers, numReducers, replication, UpdateWHStep1.UPDATE_TYPE_H, vDirectory, wDirectory, outputDir, k); resultDirectoryX = UpdateWHStep2.runJob(numMappers, numReducers, replication, workingDirectory, outputDir); FileSystem.get(mainJob).delete(new Path(workingDirectory)); System.out.println("done"); System.out.print("calculating Y = WT * W * H... "); workingDirectory = UpdateWHStep3.runJob(numMappers, numReducers, replication, wDirectory, outputDir); resultDirectoryY = UpdateWHStep4.runJob(numMappers, replication, workingDirectory, UpdateWHStep4.UPDATE_TYPE_H, hDirectory, outputDir); FileSystem.get(mainJob).delete(new Path(workingDirectory)); System.out.println("done"); System.out.print("calculating H = H .* X ./ Y... "); workingDirectory = UpdateWHStep5.runJob(numMappers, numReducers, replication, hDirectory, resultDirectoryX, resultDirectoryY, hFinalDir, k); System.out.println("done"); FileSystem.get(mainJob).delete(new Path(resultDirectoryX)); FileSystem.get(mainJob).delete(new Path(resultDirectoryY)); System.out.print("storing back H... "); FileSystem.get(mainJob).delete(new Path(hDirectory)); hDirectory = workingDirectory; System.out.println("done"); System.out.print("calculating X = V * HT... "); workingDirectory = UpdateWHStep1.runJob(numMappers, numReducers, replication, UpdateWHStep1.UPDATE_TYPE_W, vDirectory, hDirectory, outputDir, k); resultDirectoryX = UpdateWHStep2.runJob(numMappers, numReducers, replication, workingDirectory, outputDir); FileSystem.get(mainJob).delete(new Path(workingDirectory)); System.out.println("done"); System.out.print("calculating Y = W * H * HT... "); workingDirectory = UpdateWHStep3.runJob(numMappers, numReducers, replication, hDirectory, outputDir); resultDirectoryY = UpdateWHStep4.runJob(numMappers, replication, workingDirectory, UpdateWHStep4.UPDATE_TYPE_W, wDirectory, outputDir); FileSystem.get(mainJob).delete(new Path(workingDirectory)); System.out.println("done"); System.out.print("calculating W = W .* X ./ Y... "); workingDirectory = UpdateWHStep5.runJob(numMappers, numReducers, replication, wDirectory, resultDirectoryX, resultDirectoryY, wFinalDir, k); System.out.println("done"); FileSystem.get(mainJob).delete(new Path(resultDirectoryX)); FileSystem.get(mainJob).delete(new Path(resultDirectoryY)); System.out.print("storing back W... "); FileSystem.get(mainJob).delete(new Path(wDirectory)); wDirectory = workingDirectory; System.out.println("done"); package gnmf; import gnmf.io.MatrixObject; import gnmf.io.MatrixVector; import gnmf.io.TaggedIndex; import java.io.IOException; import java.util.Iterator; import org.apache.hadoop.fs.Path; import org.apache.hadoop.mapred.FileInputFormat; import org.apache.hadoop.mapred.FileOutputFormat; import org.apache.hadoop.mapred.JobClient; import org.apache.hadoop.mapred.JobConf; import org.apache.hadoop.mapred.MapReduceBase; import org.apache.hadoop.mapred.Mapper; import org.apache.hadoop.mapred.OutputCollector; import org.apache.hadoop.mapred.Reducer; import org.apache.hadoop.mapred.Reporter; import org.apache.hadoop.mapred.SequenceFileInputFormat; import org.apache.hadoop.mapred.SequenceFileOutputFormat; public class UpdateWHStep2 { static class UpdateWHStep2Mapper extends MapReduceBase implements Mapper<TaggedIndex, MatrixVector, TaggedIndex, MatrixVector> { @Override public void map(TaggedIndex key, MatrixVector value, OutputCollector<TaggedIndex, MatrixVector> out, Reporter reporter) throws IOException { out.collect(key, value); } } static class UpdateWHStep2Reducer extends MapReduceBase implements Reducer<TaggedIndex, MatrixVector, TaggedIndex, MatrixObject> { @Override public void reduce(TaggedIndex key, Iterator<MatrixVector> values, OutputCollector<TaggedIndex, MatrixObject> out, Reporter reporter) throws IOException { MatrixVector result = null; while(values.hasNext()) { MatrixVector current = values.next(); if(result == null) { result = current.getCopy(); } else { result.addVector(current); } } if(result != null) { out.collect(new TaggedIndex(key.getIndex(), TaggedIndex.TYPE_VECTOR_X), new MatrixObject(result)); } } } public static String runJob(int numMappers, int numReducers, int replication, String inputDir, String outputDir) throws IOException { String workingDirectory = outputDir + System.currentTimeMillis() + "-UpdateWHStep2/"; JobConf job = new JobConf(UpdateWHStep2.class); job.setJobName("MatrixGNMFUpdateWHStep2"); job.setInputFormat(SequenceFileInputFormat.class); FileInputFormat.setInputPaths(job, new Path(inputDir)); job.setOutputFormat(SequenceFileOutputFormat.class); FileOutputFormat.setOutputPath(job, new Path(workingDirectory)); job.setNumMapTasks(numMappers); job.setMapperClass(UpdateWHStep2Mapper.class); job.setMapOutputKeyClass(TaggedIndex.class); job.setMapOutputValueClass(MatrixVector.class); package gnmf; import gnmf.io.MatrixCell; import gnmf.io.MatrixFormats; import gnmf.io.MatrixObject; import gnmf.io.MatrixVector; import gnmf.io.TaggedIndex; import java.io.IOException; import java.util.Iterator; import org.apache.hadoop.filecache.DistributedCache; import org.apache.hadoop.fs.FileSystem; import org.apache.hadoop.fs.Path; import org.apache.hadoop.mapred.FileInputFormat; import org.apache.hadoop.mapred.FileOutputFormat; import org.apache.hadoop.mapred.JobClient; import org.apache.hadoop.mapred.JobConf; import org.apache.hadoop.mapred.MapReduceBase; import org.apache.hadoop.mapred.Mapper; import org.apache.hadoop.mapred.OutputCollector; import org.apache.hadoop.mapred.Reducer; import org.apache.hadoop.mapred.Reporter; import org.apache.hadoop.mapred.SequenceFileInputFormat; import org.apache.hadoop.mapred.SequenceFileOutputFormat; public class UpdateWHStep1 { public static final int UPDATE_TYPE_H = 0; public static final int UPDATE_TYPE_W = 1; static class UpdateWHStep1Mapper extends MapReduceBase implements Mapper<TaggedIndex, MatrixObject, TaggedIndex, MatrixObject> { private int updateType; @Override public void map(TaggedIndex key, MatrixObject value, OutputCollector<TaggedIndex, MatrixObject> out, Reporter reporter) throws IOException { if(updateType == UPDATE_TYPE_W && key.getType() == TaggedIndex.TYPE_CELL) { MatrixCell current = (MatrixCell) value.getObject(); out.collect(new TaggedIndex(current.getColumn(), TaggedIndex.TYPE_CELL), new MatrixObject(new MatrixCell(key.getIndex(), current.getValue()))); } else { out.collect(key, value); } } @Override public void configure(JobConf job) { updateType = job.getInt("gnmf.updateType", 0); } } static class UpdateWHStep1Reducer extends MapReduceBase implements Reducer<TaggedIndex, MatrixObject, TaggedIndex, MatrixVector> { private double[] baseVector = null; private int vectorSizeK; @Override public void reduce(TaggedIndex key, Iterator<MatrixObject> values, OutputCollector<TaggedIndex, MatrixVector> out, Reporter reporter) throws IOException { if(key.getType() == TaggedIndex.TYPE_VECTOR) { if(!values.hasNext()) throw new RuntimeException("expected vector"); MatrixFormats current = values.next().getObject(); if(!(current instanceof MatrixVector)) throw new RuntimeException("expected vector"); baseVector = ((MatrixVector) current).getValues(); } else { while(values.hasNext()) { MatrixCell current = (MatrixCell) values.next().getObject(); if(baseVector == null) { out.collect(new TaggedIndex(current.getColumn(), TaggedIndex.TYPE_VECTOR), new MatrixVector(vectorSizeK)); } else { if(baseVector.length == 0) throw new RuntimeException("base vector is corrupted"); MatrixVector resultingVector = new MatrixVector(baseVector); resultingVector.multiplyWithScalar(current.getValue()); if(resultingVector.getValues().length == 0) throw new RuntimeException("multiplying with scalar failed"); out.collect(new TaggedIndex(current.getColumn(), TaggedIndex.TYPE_VECTOR), resultingVector); } } baseVector = null; } } @Override public void configure(JobConf job) { vectorSizeK = job.getInt("dml.matrix.gnmf.k", 0); if(vectorSizeK == 0) Java Implementation (>1500 lines of code) Equivalent Big R - SystemML Implementation (12 lines of code) # Perform matrix operations, say non-negative factorization # V ~~ WH V <- bigr.matrix("V.mtx”); # initial matrix on HDFS W <- bigr.matrix(nrow=nrow(V), ncols=k); # initialize starting points H <- bigr.matrix(nrow=k, ncols=ncols(V)); for (i in 1:numiterations) { H <- H * (t(W) %*% V / t(W) %*% W %*% H); W <- W * (V %*% t(H) / W %*% H %*% t(H)); }
  • 64 Summary - Popular R Analytics on Hadoop • R is the preferred language for Data Scientists • Many approaches exist to enable R on Hadoop with pros and cons • “Big R” approach: • Data exploration, statistical analysis and Visualization with natural R syntax • Scale out R with data partitioning • Support for standard R tools and existing packages and libraries • SystemML engine for Distributed Machine Learning that provides canned algorithms, and an ability to author new ones, all via R-like syntax Data Sources Hive Tables HBase Tables Files R Runtime Hadoop R Clients Distributed ML Runtime
  • 65 65 www.ibm.com/software/data/infosphere/hadoop