Transcript Hands On Mahout - Mammoth Scale Machine

Hands on!
Speakers: Ted Dunning, Robin Anil
OSCON 2011, Portland
About Us
 Ted Dunning:
Chief Application Architect at MapR
Committer and PMC Member at Apache Mahout
Previously: MusicMatch (Yahoo! Music), Veoh recommendation, ID Analytics
 Robin Anil:
Software Engineer at Google
Committer and PMC Member at Apache Mahout
Previously: Yahoo! (Display ads), Minekey recommendation
Agenda
 Intro to Mahout (5 mins)
 Overview of Algorithms in Mahout (10 mins)
 Hands on Mahout!
- Clustering (30 mins)
- Classification (30 mins)
- Advanced topics with Q&A (15 mins)
Mission
To build a scalable machine learning library
Scale!
 Scale to large datasets
- Hadoop MapReduce implementations that scales linearly with data.
- Fast sequential algorithms whose runtime doesn’t depend on the size of the data
- Goal: To be as fast as possible for any algorithm
 Scalable to support your business case
- Apache Software License 2
 Scalable community
- Vibrant, responsive and diverse
- Come to the mailing list and find out more
Current state of ML libraries
 Lack community
 Lack scalability
 Lack documentations and examples
 Lack Apache licensing
 Are not well tested
 Are Research oriented
 Not built over existing production quality libraries
 Lack “Deployability”
Algorithms and Applications
Clustering
 Call it fuzzy grouping based on a notion of similarity
Mahout Clustering
 Plenty of Algorithms: K-Means,
Fuzzy K-Means, Mean Shift,
Canopy, Dirichlet
 Group similar looking objects
 Notion of similarity: Distance measure:
- Euclidean
- Cosine
- Tanimoto
- Manhattan
Classification
 Predicting the type of a new object based on its features
 The types are predetermined
Dog
Cat
Mahout Classification
 Plenty of algorithms
- Naïve Bayes
- Complementary Naïve Bayes
- Random Forests
- Logistic Regression (SGD)
- Support Vector Machines (patch ready)
 Learn a model from a manually classified data
 Predict the class of a new object based on its
features and the learned model
Part 1 - Clustering
Understanding data - Vectors
Y
X=5, Y=3
(5, 3)
 The vector denoted by point (5, 3) is simply
Array([5, 3]) or HashMap([0 => 5], [1 => 3])
X
Representing Vectors – The basics
 Now think 3, 4, 5, ….. n-dimensional
 Think of a document as a bag of words.
“she sells sea shells on the sea shore”
 Now map them to integers
she => 0
sells => 1
sea => 2
and so on
 The resulting vector [1.0, 1.0, 2.0, … ]
Vectors
 Imagine one dimension for each word.
 Each dimension is also called a feature
 Two techniques
- Dictionary Based
- Randomizer Based
Clustering Reuters dataset
Step 1 – Convert dataset into a Hadoop Sequence File
 http://www.daviddlewis.com/resources/testcollections/reuters21578/reuters21578.tar.gz
 Download (8.2 MB) and extract the SGML files.
- $ mkdir -p mahout-work/reuters-sgm
- $ cd mahout-work/reuters-sgm && tar xzf ../reuters21578.tar.gz && cd ..
&& cd ..
 Extract content from SGML to text file
- $ bin/mahout org.apache.lucene.benchmark.utils.ExtractReuters mahoutwork/reuters-sgm mahout-work/reuters-out
Step 1 – Convert dataset into a Hadoop Sequence File
 Use seqdirectory tool to convert text file into a Hadoop Sequence File
- $ bin/mahout seqdirectory \
-i mahout-work/reuters-out \
-o mahout-work/reuters-out-seqdir \
-c UTF-8 -chunk 5
Hadoop Sequence File
 Sequence of Records, where each record is a <Key, Value> pair
- <Key1, Value1>
- <Key2, Value2>
- …
- …
- …
- <Keyn, Valuen>
 Key and Value needs to be of class org.apache.hadoop.io.Text
- Key = Record name or File name or unique identifier
- Value = Content as UTF-8 encoded string
 TIP: Dump data from your database directly into Hadoop Sequence Files (see next slide)
Writing to Sequence Files
Configuration conf = new Configuration();
FileSystem fs = FileSystem.get(conf);
Path path = new Path("testdata/part-00000");
SequenceFile.Writer writer = new SequenceFile.Writer(
fs, conf, path, Text.class, Text.class);
for (int i = 0; i < MAX_DOCS; i++)
writer.append(new Text(documents(i).Id()),
new Text(documents(i).Content()));
}
writer.close();
Generate Vectors from Sequence Files
 Steps
1. Compute Dictionary
2. Assign integers for words
3. Compute feature weights
4. Create vector for each document using word-integer mapping and feature-weight
Or
 Simply run $ bin/mahout seq2sparse
Generate Vectors from Sequence Files
 $ bin/mahout seq2sparse \
-i mahout-work/reuters-out-seqdir/ \
-o mahout-work/reuters-out-seqdir-sparse-kmeans
 Important options
- Ngrams
- Lucene Analyzer for tokenizing
- Feature Pruning
- Min support
- Max Document Frequency
- Min LLR (for ngrams)
- Weighting Method
- TF v/s TFIDF
- lp-Norm
- Log normalize length
Start K-Means clustering
 $ bin/mahout kmeans \
-i mahout-work/reuters-out-seqdir-sparse-kmeans/tfidf-vectors/ \
-c mahout-work/reuters-kmeans-clusters \
-o mahout-work/reuters-kmeans \
-dm org.apache.mahout.distance.CosineDistanceMeasure –cd 0.1 \
-x 10 -k 20 –ow
 Things to watch out for
- Number of iterations
- Convergence delta
- Distance Measure
- Creating assignments
K-Means clustering
c2
c1
c3
K-Means clustering
c2
c1
c3
K-Means clustering
c2
c2
c1
c1
c3
c3
K-Means clustering
c2
c1
c3
Inspect clusters
 $ bin/mahout clusterdump \
-s mahout-work/reuters-kmeans/clusters-9 \
-d mahout-work/reuters-out-seqdir-sparse-kmeans/dictionary.file-0 \
-dt sequencefile -b 100 -n 20
Typical output
:VL-21438{n=518 c=[0.56:0.019, 00:0.154, 00.03:0.018, 00.18:0.018, …
Top Terms:
iran
=>
3.1861672217321213
strike
=>
2.567886952727918
iranian
=>
2.133417966282966
union
=>
2.116033937940266
said
=>
2.101773806290277
workers
=>
2.066259451354332
gulf
=>
1.9501374918521601
had
=>
1.6077752463145605
he
=>
1.5355078004962228
FAQs
 How to get rid of useless words
 How to see documents to cluster assignments
 How to choose appropriate weighting
 How to run this on a cluster
 How to scale
 How to choose k
 How to improve similarity measurement
FAQs
 How to get rid of useless words
- Increase minSupport and or decrease dfPercent
- Use StopwordsAnalyzer
 How to see documents to cluster assignments
- Run clustering process at the end of centroid generation using –cl
 How to choose appropriate weighting
- If its long text, go with tfidf. Use normalization if documents different
in length
 How to run this on a cluster
- Set HADOOP_CONF directory to point to your hadoop cluster conf directory
 How to scale
- Use small value of k to partially cluster data and then do full
clustering on each cluster.
FAQs
 How to choose k
- Figure out based on the data you have. Trial and error
- Or use Canopy Clustering and distance threshold to figure it out
- Or use Spectral clustering
 How to improve Similarity Measurement
- Not all features are equal
- Small weight difference for certain types creates a large semantic
difference
- Use WeightedDistanceMeasure
- Or write a custom DistanceMeasure
Interesting problems
 Cluster users talking about OSCON’11 and cluster them based on what they are tweeting
- Can you suggest people to network with.
 Use user generate tags that people have given for musicians and cluster them
- Use the cluster to pre-populate suggest-box to autocomplete tags when users type
 Cluster movies based on abstract and description and show related movies.
- Note: How it can augment recommendations or collaborative filtering algorithms.
More clustering algorithms
 Canopy
 Fuzzy K-Means
 Mean Shift
 Dirichlet process clustering
 Spectral clustering.
Part 2 - Classification
Preliminaries
 Code is available from github:
- [email protected]:tdunning/Chapter-16.git
 EC2 instances available
 Thumb drives also available
 Email to [email protected]
 Twitter @ted_dunning
A Quick Review
 What is classification?
- goes-ins: predictors
- goes-outs: target variable
 What is classifiable data?
- continuous, categorical, word-like, text-like
- uniform schema
 How do we convert from classifiable data to feature vector?
Data Flow
Not quite so
simple
Classifiable Data
 Continuous
- A number that represents a quantity, not an id
- Blood pressure, stock price, latitude, mass
 Categorical
- One of a known, small set (color, shape)
 Word-like
- One of a possibly unknown, possibly large set
 Text-like
- Many word-like things, usually unordered
But that isn’t quite there
 Learning algorithms need feature vectors
- Have to convert from data to vector
 Can assign one location per feature
- or category
- or word
 Can assign one or more locations with hashing
- scary
- but safe on average
Data Flow
The pipeline
Classifiable Data
Vectors
Instance and Target Variable
Instance and Target Variable
Hashed Encoding
What about collisions?
Let’s write some code
(cue relaxing background music)
Generating new features
 Sometimes the existing features are difficult to use
 Restating the geometry using new reference points may help
 Automatic reference points using k-means can be better than manual references
K-means using target
K-means features
More code!
(cue relaxing background music)
Integration Issues
 Feature extraction is ideal for map-reduce
- Side data adds some complexity
 Clustering works great with map-reduce
- Cluster centroids to HDFS
 Model training works better sequentially
- Need centroids in normal files
 Model deployment shouldn’t depend on HDFS
Parallel Stochastic Gradient Descent
Model
I
n
p
u
t
Train
sub
model
Average
models
Variational Dirichlet Assignment
Model
I
n
p
u
t
Gather
sufficient
statistics
Update
model
Old tricks, new dogs
 Mapper
- Assign point to cluster
Read from local disk
from distributed cache
- Emit cluster id, (1, point)
 Combiner and reducer
Read from
HDFS to local disk by
distributed cache
- Sum counts, weighted sum of points
- Emit cluster id, (n, sum/n)
 Output to HDFS
Written by
map-reduce
Old tricks, new dogs
 Mapper
Read
from
NFS
- Assign point to cluster
- Emit cluster id, 1, point
 Combiner and reducer
- Sum counts, weighted sum of points
- Emit cluster id, n, sum/n
 Output to HDFS
MapR FS
Written by
map-reduce
Modeling architecture
Side-data
Now via NFS
I
n
p
u
t
Feature
extraction
and
down
sampling
Data
join
Map-reduce
Sequential
SGD
Learning
More in Mahout
Topic modeling
 Grouping similar or co-occurring features into a topic
- Topic “Lol Cat”:
- Cat
- Meow
- Purr
- Haz
- Cheeseburger
- Lol
Mahout Topic Modeling
 Algorithm: Latent Dirichlet Allocation
- Input a set of documents
- Output top K prominent topics and the
features in each topic
Recommendations
 Predict what the user likes based on
- His/Her historical behavior
- Aggregate behavior of people similar to him
Mahout Recommenders
 Different types of recommenders
- User based
- Item based
 Full framework for storage, online
online and offline computation of recommendations
 Like clustering, there is a notion of similarity in users or items
- Cosine, Tanimoto, Pearson and LLR
Frequent Pattern Mining
 Find interesting groups of items based on how they co-occur in a dataset
Mahout Parallel FPGrowth
 Identify the most commonly
occurring patterns from
- Sales Transactions
buy “Milk, eggs and bread”
- Query Logs
ipad -> apple, tablet, iphone
- Spam Detection
Yahoo! http://www.slideshare.net/hadoopusergroup/mail-antispam
Get Started
 http://mahout.apache.org
 [email protected] - Developer mailing list
 [email protected] - User mailing list
 Check out the documentations and wiki for quickstart
 http://svn.apache.org/repos/asf/mahout/trunk/ Browse Code
 Send me email!
- [email protected]
- [email protected]
- [email protected]
 Try out MapR!
- www.mapr.com
Resources
 “Mahout in Action” Owen, Anil, Dunning, Friedman
http://www.manning.com/owen
 “Taming Text” Ingersoll, Morton, Farris
http://www.manning.com/ingersoll
 “Introducing Apache Mahout”
http://www.ibm.com/developerworks/java/library/j-mahout/
Thanks to
 Apache Foundation
 Mahout Committers
 Google Summer of Code Organizers
 And Students
 OSCON
 Open source!
References
 news.google.com
 Cat http://www.flickr.com/photos/gattou/3178745634/
 Dog http://www.flickr.com/photos/30800139@N04/3879737638/
 Milk Eggs Bread http://www.flickr.com/photos/nauright/4792775946/
 Amazon Recommendations
 twitter