hadoop and machine learning
DESCRIPTION
Slides for the talk by the Cloudera Data Science team on the state of machine learning and Hadoop at NIPS 2011.TRANSCRIPT
Machine Learning and HadoopPresent and FutureJosh Wills, Tom Pierce, and Jeff HammerbacherCloudera Data Science TeamDecember 17th, 2011
High Availability for Data Scientists
Copyright 2011 Cloudera Inc. All rights reserved
NIPS
Agenda
• Part 1: Industrial Machine Learning• Part 2: Machine Learning and Hadoop
• State of the World• Where Things Are Headed
• Part 3: Things Industry Needs From Academia
Copyright 2011 Cloudera Inc. All rights reserved
Industrial Machine Learning
Copyright 2011 Cloudera Inc. All rights reserved
Delta One: Model Evaluation
• ML Systems Are One Piece of a Complex System• Well-defined objective functions are the exception
• Multiple, often conflicting goals• Weights are fuzzy and shift with business priorities• Pareto optimization is the safest play
• Predictive Accuracy Is Only Useful Up to a Point• Examples
• Computational advertising• Friend recommendations on social networks
Copyright 2011 Cloudera Inc. All rights reserved
Delta Two: Systems Precede Algorithms
• Greenfield Projects Hardly Ever Happen• (and don’t usually launch)
• Industrial Computational Infrastructure• General-purpose• Cheap• Shared
• Constraints Drive Innovation• Vowpal Wabbit Hashing Trick• SETI @ Google
Copyright 2011 Cloudera Inc. All rights reserved
Delta Three: Workflow
Copyright 2011 Cloudera Inc. All rights reserved
Practice Over Theory Blog
Delta Three: Workflow
• Optimize the Overall Process• Model fitting is a small piece of the overall flow time• Parallelize everything
• Better Features > Better Models• Fast Model Deployment
• Common Feature Extraction Logic• Servable Models
• Validation as Sanity Checking• Deploy to a small subset of real data and evaluate
Copyright 2011 Cloudera Inc. All rights reserved
Agenda
• Part 1: Industrial Machine Learning• Part 2: Machine Learning and Hadoop
• State of the World• Where Things Are Headed
• Part 3: Things Industry Needs From Academia
Copyright 2011 Cloudera Inc. All rights reserved
Hadoop: It’s Where The Data Is
Copyright 2011 Cloudera Inc. All rights reserved
Hadoop Platform: Substrate
• Commodity servers• Open Compute
• Open source operating system• Linux
• Open source configuration management• Puppet• Chef
• Coordination service• ZooKeeper
Copyright 2011 Cloudera Inc. All rights reserved
Hadoop Platform: Storage
• Distributed schema-less storage• HDFS• Ceph
• Append-only storage formats and metadata• Avro• RCFile• HCatalog
• Mutable key-value storage and metadata• HBase
Copyright 2011 Cloudera Inc. All rights reserved
Hadoop Platform: Integration
• Tool Access• FUSE• JDBC• ODBC
• Data Ingestion• Flume• Sqoop
Copyright 2011 Cloudera Inc. All rights reserved
ML and Hadoop: The State of the World
Copyright 2011 Cloudera Inc. All rights reserved
Computation: Plain Old MapReduce
• Great for:• Data Preparation• Feature Engineering• Model Validation/Evaluation
• Works For Certain Model Fitting Problems• Recommendation Systems• Decision Trees (PLANET; Gradient Boosted Decision Trees)
• Not A Practical Option for Online Learning• Way More Detail from the KDD 2011 Talk
Copyright 2011 Cloudera Inc. All rights reserved
Tools for Data Preparation/Feature Engineering
• Languages/Environments• PigLatin• HiveQL• Need to deal with mismatch between offline/online feature
generation
• Java/Scala APIs• Crunch (Cloudera)• Scoobi (NICTA)• Cascading (Concurrent)• Jaql (IBM)
Copyright 2011 Cloudera Inc. All rights reserved
Apache Mahout
• The starting place for MapReduce-based machine learning algorithms• Not machine-learning-in-a-box• Custom tweaks/modifications are the rule
• A disparate collection of algorithms for:• Recommendations• Clustering• Classification• Frequent Itemset Mining
Copyright 2011 Cloudera Inc. All rights reserved
Apache Mahout (cont.)
• Best Library: Taste Recommender• Oldest project, most widely-deployed in production• SVD implementation is particularly active
• Good Libraries: Online SGD• Does not use MapReduce• Vowpal Rabbit + AllReduce is faster, has L-BFGS option
• Roll Your Own Instead: Naïve Bayes• Challenges
• “Secret sauce” effect• Delta between Mahout + the cutting edge in ML
Copyright 2011 Cloudera Inc. All rights reserved
More Machine Learning Interfaces for Hadoop
• Based on MapReduce• SystemML (IBM)• AllReduce (Vowpal Wabbit)
• No MapReduce• Spark
• R-Based Systems (Augment MapReduce with R)• Segue• RHIPE• RHadoop• Ricardo (IBM)
Copyright 2011 Cloudera Inc. All rights reserved
ML and Hadoop: Where Things are Headed
Copyright 2011 Cloudera Inc. All rights reserved
MRv2 and YARN
• Eliminates JobTracker bottleneck• Separate Resource Manager/Scheduler• Individual jobs have their own task masters
• Moves MapReduce into user-land• Enables Hadoop clusters to run all sorts of jobs
• MPI (Hamster; MAPREDUCE-2911)• Native BSP (Giraph)• Spark• AllReduce, GraphLab
Copyright 2011 Cloudera Inc. All rights reserved
Agenda
• Part 1: Industrial Machine Learning• Part 2: Machine Learning and Hadoop
• State of the World• Where Things Are Headed
• Part 3: Things Industry Needs From Academia
Copyright 2011 Cloudera Inc. All rights reserved
Machine Learning on Multivariate Time Series
• 1e5 writes/sec• Positive events are
relatively rare• Feature extraction
challenge• May not be clear what
the right time horizon is• Tight SLAs• Very high stakes
Copyright 2011 Cloudera Inc. All rights reserved
An Academic Language For Feature Engineering
• Feature extraction/selection is as important as model fitting• e.g., hierarchical feature representation, impact on training
time and experiment design, feature cost modeling, etc.
• Academic literature on this problem is sparse and dispersed across multiple fields• NIPS 2003• HCI, NLP, Information Retrieval, etc.
• We need a common language for talking about these problems across disciplines
Copyright 2011 Cloudera Inc. All rights reserved
A Broader Ontology For Model Selection
• Practical factors that enter into the “best” choice of model…• Data arrival rate• Data volume• Scoring latency• Model refresh time• Robustness/reliability
• …in addition to the standard predictive power/simplicity tradeoffs
Copyright 2011 Cloudera Inc. All rights reserved
Questions?Want A Job?
@josh_wills