Apache Spark Machine Learning
4 likes1,980 views
This document discusses machine learning techniques in Spark including classification, clustering, and collaborative filtering. It provides examples of building classification models with Spark including vectorizing data, training models, evaluating models, and making predictions. Clustering and collaborative filtering are also introduced. The document demonstrates collaborative filtering with Spark using alternating least squares to build a recommendation model from user ratings data.
![Ā®
Ā© 2014 MapR Technologies 9
Vectorizing Data
ā¢āÆ identify interesting features (those that contribute to the model)
ā¢āÆ assign features to dimensions
Example: vectorize an apple
Features: [size, color, weight]
Example: vectorize a text document
(Term Frequency Inverse Term Frequency)
Dictionary: [a, advance, after, ā¦, you, yourself, youth, zigzag]
[3.2, 16777184.0, 45.8][223,1,1,0,ā¦,12,10,6,1]](https://image.slidesharecdn.com/sparkmachinelearning-151215151615/85/Apache-Spark-Machine-Learning-9-320.jpg)
![Ā®
Ā© 2014 MapR Technologies 34
Test Model
val falsePositives =(testAndPredictionsJoinedRDD.filter{
case ((user, product), (ratingT, ratingP)) =>
(ratingT <= 1 && ratingP >=4)
})
falsePositives.take(2)
Array[((Int, Int), (Double, Double))] =
((3842,2858),(1.0,4.106488210964762)),
((6031,3194),(1.0,4.790778049100913))](https://image.slidesharecdn.com/sparkmachinelearning-151215151615/85/Apache-Spark-Machine-Learning-34-320.jpg)
Apache Spark Machine Learning
- 1. Ā® Ā© 2014 MapR Technologies 1 Ā® Ā© 2014 MapR Technologies Machine Learning with Spark Carol McDonald
- 2. Ā® Ā© 2014 MapR Technologies 2 Agenda ā¢āÆ Classification ā¢āÆ Clustering ā¢āÆ Collaborative Filtering with Spark ā¢āÆ Model training ā¢āÆ Alternating Least Squares ā¢āÆ The code
- 3. Ā® Ā© 2014 MapR Technologies 3 Three Categories of Techniques for Machine Learning classification Collaborative filtering (recommendation) clustering Groups similar items identifies category for item Recommend items
- 4. Ā® Ā© 2014 MapR Technologies 4 What is classification Form of ML that: ā¢āÆ identifies which category an item belongs to ā¢āÆ Uses supervised learning algorithms ā¢āÆ Data is labeled Examples: ā¢āÆ Spam Detection ā¢āÆ spam/non-spam ā¢āÆ Credit Card Fraud Detection ā¢āÆ fraud/non-fraud ā¢āÆ Sentiment analysis
- 5. Ā® Ā© 2014 MapR Technologies 5 Building and deploying a classifier model
- 6. Ā® Ā© 2014 MapR Technologies 6 If it walks/swims/quacks like a duck ā¦ Attributes, Features: ā¢āÆ If it walks ā¢āÆ If it swims ā¢āÆ If it quacks āWhen I see a bird that walks like a duck and swims like a duck and quacks like a duck, I call that bird a duck.ā classify something based on āifā conditions. Answer, Label: ā¢āÆ Duck ā¢āÆ Not duck
- 7. Ā® Ā© 2014 MapR Technologies 7 ā¦ then it must be a duck ducks not ducks walks quacks swims Label: ā¢āÆ Duck ā¢āÆ Not duck Features: ā¢āÆ walks ā¢āÆ swims ā¢āÆ quacks
- 8. Ā® Ā© 2014 MapR Technologies 8 Building and deploying a classifier model Reference Learning Spark Oreilly Book
- 9. Ā® Ā© 2014 MapR Technologies 9 Vectorizing Data ā¢āÆ identify interesting features (those that contribute to the model) ā¢āÆ assign features to dimensions Example: vectorize an apple Features: [size, color, weight] Example: vectorize a text document (Term Frequency Inverse Term Frequency) Dictionary: [a, advance, after, ā¦, you, yourself, youth, zigzag] [3.2, 16777184.0, 45.8][223,1,1,0,ā¦,12,10,6,1]
- 10. Ā® Ā© 2014 MapR Technologies 10 Build Term Frequency Feature vectors // examples of spam val spam = sc.textFile("spam.txt") // examples of not spam val normal = sc.textFile("normal.txtā) // Create a HashingTF map email text to vectors of features val tf = new HashingTF(numFeatures = 10000) // Each email each word is mapped to one feature. val spamFeatures = spam .map(email => tf.transform(email.split(" "))) val normalFeatures = normal .map(email => tf.transform(email.split(" "))) Reference Learning Spark Oreilly Book
- 11. Ā® Ā© 2014 MapR Technologies 11 Building and deploying a classifier model
- 12. Ā® Ā© 2014 MapR Technologies 12 Build Model val trainingData = positiveExamples.union(negativeExamples) trainingData.cache() // Cache for iterative algorithm. // Run Logistic Regression using the SGD algorithm. val model = new LogisticRegressionWithSGD() .run(trainingData) Reference Learning Spark Oreilly Book
- 13. Ā® Ā© 2014 MapR Technologies 13 Building and deploying a classifier model Reference Learning Spark Oreilly Book
- 14. Ā® Ā© 2014 MapR Technologies 14 Model Evaluation // Test on a positive example (spam) Vector posTest = tf.transform(Arrays.asList( "O M G GET cheap stuff by sending money to...".split(" "))); // negative test not spam Vector negTest = tf.transform(Arrays.asList( "Hi Dad, I started studying Spark the other ...".split(" "))); System.out.println("Prediction for positive: " + model.predict(posTest)); System.out.println("Prediction for negative: " + model.predict(negTest));
- 15. Ā® Ā© 2014 MapR Technologies 15 Three Categories of Techniques for Machine Learning classification Collaborative filtering (recommendation) clustering
- 16. Ā® Ā© 2014 MapR Technologies 16 Clustering ā¢āÆ Clustering is the unsupervised learning task that involves grouping objects into clusters of high similarity āāÆ Search results grouping āāÆ grouping of customers by similar habits āāÆ Anomaly detection ā¢āÆ data traffic āāÆ Text categorization
- 17. Ā® Ā© 2014 MapR Technologies 17 What is Clustering? Clustering = (unsupervised) task of grouping similar objects MLlib K-means algorithm for clustering 1.āÆ randomly initialize centers of clusters 2.āÆ Assign all points to the closest cluster center 3.āÆ Change cluster centers to be in the middle of its points 4.āÆ Repeat until convergence
- 18. Ā® Ā© 2014 MapR Technologies 18 What is Clustering? Clustering = (unsupervised) task of grouping similar objects
- 19. Ā® Ā© 2014 MapR Technologies 19 Examples of ML Algorithms machine learning supervised unsupervised ā¢āÆ Classification ā¢āÆ NaĆÆve Bayes ā¢āÆ SVM ā¢āÆ Random Decision Forests ā¢āÆ Regression ā¢āÆ Linear ā¢āÆ logistic ā¢āÆ Clustering ā¢āÆ K-means ā¢āÆ Dimensionality reduction ā¢āÆ Principal Component Analysis ā¢āÆ SVD
- 20. Ā® Ā© 2014 MapR Technologies 20 ML Algorithms http://scikit-learn.org/stable/tutorial/machine_learning_map/index.html
- 21. Ā® Ā© 2014 MapR Technologies 21 Three Categories of Techniques for Machine Learning classification Collaborative filtering (recommendation) clustering
- 22. Ā® Ā© 2014 MapR Technologies 22 Collaborative Filtering with Spark ā¢āÆ Recommend Items āāÆ (filtering) ā¢āÆ Based on User preferences data āāÆ (collaborative)
- 23. Ā® Ā© 2014 MapR Technologies 23 Train a Model to Make Predictions New Data Model Predictions Training Data ModelAlgorithm Ted and Carol like Movie B and C Bob likes Movie B, What might he like ? Bob likes Movie B, Predict C
- 24. Ā® Ā© 2014 MapR Technologies 24 Alternating Least Squares ā¢āÆ approximates sparse user item rating matrix āāÆ as product of two dense matrices, User and Item factor matrices āāÆ tries to learn the hidden features of each user and item āāÆ algorithm alternatively fixes one factor matrix and solves for the other ?
- 25. Ā® Ā© 2014 MapR Technologies 25 ML Cross Validation Process Data Model Training/ Building Test Model Predictions Test Set Train Test loop Training Set
- 26. Ā® Ā© 2014 MapR Technologies 26 Ratings Data
- 27. Ā® Ā© 2014 MapR Technologies 27 Parse Input // parse input UserID::MovieID::Rating def parseRating(str: String): Rating= { val fields = str.split("::") Rating(fields(0).toInt, fields(1).toInt, fields(2).toDouble) } // create an RDD of Ratings objects val ratingsRDD = ratingText.map(parseRating).cache()
- 28. Ā® Ā© 2014 MapR Technologies 28 Build Model Data Build Model Test Set Training Set split ratings RDD into training data RDD (80%) and test data RDD (20%) build a user product matrix model
- 29. Ā® Ā© 2014 MapR Technologies 29 Create Model // Randomly split ratings RDD into training data RDD (80%) and test data RDD (20%) val splits = ratingsRDD.randomSplit(Array(0.8, 0.2), 0L) val trainingRatingsRDD = splits(0).cache() val testRatingsRDD = splits(1).cache() // build a ALS user product matrix model with rank=20, iterations=10 val model = (new ALS().setRank(20).setIterations(10) .run(trainingRatingsRDD))
- 30. Ā® Ā© 2014 MapR Technologies 30 Get predictions // get predicted ratings to compare to test ratings val testUserProductRDD = testRatingsRDD.map { case Rating(user, product, rating) => (user, product) } // call model.predict with test Userid, MovieId input data val predictionsForTestRDD = model.predict(testUserProductRDD) User, Movie Test Data Model Predicted Ratings
- 31. Ā® Ā© 2014 MapR Technologies 31 Compare predictions to Tests Join predicted ratings to test ratings in order to compare ((user, product),test rating) ((user, product), predicted rating) ((user, product),(test rating, predicted rating)) Key, Value Key, Value Key, Value
- 32. Ā® Ā© 2014 MapR Technologies 32 Test Model // prepare predictions for comparison val predictionsKeyedByUserProductRDD = predictionsForTestRDD.map{ case Rating(user, product, rating) => ((user, product), rating) } // prepare test for comparison val testKeyedByUserProductRDD = testRatingsRDD.map{ case Rating(user, product, rating) => ((user, product), rating) } //Join the test with predictions val testAndPredictionsJoinedRDD = testKeyedByUserProductRDD .join(predictionsKeyedByUserProductRDD)
- 33. Ā® Ā© 2014 MapR Technologies 33 Compare predictions to Tests Find False positives: Where test rating <= 1 and predicted rating >= 4 ((user, product),(test rating, predicted rating)) Key, Value
- 34. Ā® Ā© 2014 MapR Technologies 34 Test Model val falsePositives =(testAndPredictionsJoinedRDD.filter{ case ((user, product), (ratingT, ratingP)) => (ratingT <= 1 && ratingP >=4) }) falsePositives.take(2) Array[((Int, Int), (Double, Double))] = ((3842,2858),(1.0,4.106488210964762)), ((6031,3194),(1.0,4.790778049100913))
- 35. Ā® Ā© 2014 MapR Technologies 35 Test Model Mean Absolute Error //Evaluate the model using Mean Absolute Error (MAE) between test and predictions val meanAbsoluteError = testAndPredictionsJoinedRDD.map { case ((user, product), (testRating, predRating)) => val err = (testRating - predRating) Math.abs(err) }.mean() meanAbsoluteError: Double = 0.7244940545944053
- 36. Ā® Ā© 2014 MapR Technologies 36 Get Predictions for new user val newRatingsRDD=sc.parallelize(Array(Rating(0,260,4),Rating(0,1,3)) // union val unionRatingsRDD = ratingsRDD.union(newRatingsRDD) // build a ALS user product matrix model val model = (new ALS().setRank(20).setIterations(10) .run(unionRatingsRDD)) // get 5 recs for userid 0 val topRecsForUser = model.recommendProducts(0, 5)
- 37. Ā® Ā© 2014 MapR Technologies 37 Soon to Come ā¢āÆ Spark On Demand Training āāÆ https://www.mapr.com/services/mapr-academy/ ā¢āÆ Blogs and Tutorials: āāÆ Movie Recommendations with Collaborative Filtering āāÆ Spark Streaming
- 38. Ā® Ā© 2014 MapR Technologies 38 Machine Learning Blog ā¢āÆ https://www.mapr.com/blog/parallel-and-iterative-processing- machine-learning-recommendations-spark
- 39. Ā® Ā© 2014 MapR Technologies 39 Spark on MapR ā¢āÆ Certified Spark Distribution ā¢āÆ Fully supported and packaged by MapR in partnership with Databricks ā¢āÆ YARN integration āāÆ Spark can then allocate resources from cluster when needed
- 40. Ā® Ā© 2014 MapR Technologies 40 References ā¢āÆ Spark Online course: learn.mapr.com ā¢āÆ Spark web site: http://spark.apache.org/ ā¢āÆ https://databricks.com/ ā¢āÆ Spark on MapR: āāÆ http://www.mapr.com/products/apache-spark ā¢āÆ Spark SQL and DataFrame Guide ā¢āÆ Apache Spark vs. MapReduce ā Whiteboard Walkthrough ā¢āÆ Learning Spark - O'Reilly Book ā¢āÆ Apache Spark
- 41. Ā® Ā© 2014 MapR Technologies 41 Q&A @mapr maprtech Engage with us! MapR maprtech mapr-technologies