Skip to content

proposedOutline.md

Latest commit

 

History

History
200 lines (140 loc) · 8.3 KB

proposedOutline.md

File metadata and controls

200 lines (140 loc) · 8.3 KB

##A Programmer's Guide to Machine Intelligence with PredictionIO

####Proposed outline The book would be a short (~250 page or so) book on using PredictionIO for data mining/machine learning focusing on real world applications. It will cover 3 three general data mining themes: recommendation systems, clustering systems, and classifiers. The book does not intend to offer a comprehensive discussion of all the algorithms in these areas. And it will not fully explain the theory of the algorithms. It will give sufficient information so readers will generally know how the algorithms work. The goal is a practical one---for readers to be able to use PredictionIO on their own data to solve problems. After a few introductory chapters, each of the three themes is presented in 2 or 3 chapters. The first chapter provides an general description of the theme (for example, recommendation systems) and presents a step-by-step example of using PredictionIO to solve a representative problem in the area. The next chapter presents several common recipes/algorithms and works through a complete example. The final chapter of each section presents a larger real-world application, where we use a cluster of Amazon EC2 instances. This final chapter of the trio may also describe issues in deploying such systems and may end with one or more short case studies of how these systems are used. An alternative presentation is to present a simple cluster example early but reserve most of the discussion of using clusters to a final chapter or two of the book.

#####Preface - TBD

####Part 1: Introduction

#####Chapter 1: Getting Started w/ PredictionIO - rough draft written ######estimated page count: 20 Readers learn what PredictionIO is, how to install it on their local machine, and how to build and run a simple recommendation system. This is really a "hit the ground running" chapter -- meaning that without much explanation the readers are led through building a simple recommender. At the end of the chapter readers will be able to build their own, simple, recommendation systems.

Topics to be covered:

  • a description of what a recommendation system is

  • installing PredictionIO on a machine (both on a VM and cloud based)

  • starting the server

  • building and training a simple music recommendation system

    • creating an app / generating an access key
    • importing data
    • getting and modifying an engine template
    • building
    • training
    • deploying
    • getting recommendations

  • reader challenge - build a recommendation system using a different dataset?

#####Chapter 2: What is PredictionIO ######estimated page count: 15 Readers learn about the components of PredictionIO. Why those components are important. And why they might consider using PredictionIO for their proejcts. There is no hands-on component to this chapter.

Topics to be covered:

  • HBase

  • ElasticSearch

  • Spark

    • Spark core
    • save some details (RDD for ex.,) for later
    • or maybe bite the bullet and give a few page intro to Spark. -

  • Zookeeper - synchronization

  • General Architecture

  • Why PredictionIO?

####Part 2: Recommendation Systems

#####Chapter 3: a closer look at recommendation systems. - rough draft written ######estimated page count: 20 Readers examine what they built in chapter 1 in more detail. They have their first peak at the engine.json file. They learn the basics of matrix factorization. They use a moderate sized dataset: the 20 million rating movieLens dataset. After reading this chapter readers will minimally know what the default recommendation algorithm is.

Topics to be covered:

  • a look at engine.json and the 'als' entry

  • a look at how people might make recommendations based on features like 'country' or 'female vocalist'

  • matrix factorization

  • the relationship of matrices P, Q, and R and how we can determine one of we have the other 2.

  • a brief mention of stochastic gradient descent.

  • a look at alternating least squares (not all the details but enough to understand the algorithm)

  • using a moderate sized dataset: the 20 million rating movieLens.

    • importing data in 'batch'

  • the effect of using a different number of latent features on accuracy.

  • mention matrix factorization and the Netflix competition. (meaning it is a reasonable algorithm to use)

  • perhaps a look at Engine.scala or reserve that for another chapter.

Chapter 4. Other recommendation systems. (not great name)

Readers will continue their examination of recommendation systems. They will look at systems that use implicit data and look at other recommendation like templates. Readers will get their first experience editing Scala code.

Topics to be covered:

  • implicit ratings
  • altering the template to use implicit ratings
  • sample dataset TBD
  • Product Ranking
  • E-Commerce Recommendation
  • Lead Scoring
  • Similar Product
  • practical considerations (not sure what I would say here yet)

####Part 3: Classification Systems

#####Chapter 5: Introducing Classification ######estimated page count: 20 Readers will learn about the fundamentals of classification systems and their uses (for ex., predicting the interest of customers). Readers will implement a simple Naive Bayes classifier.

List of topics to be covered in the chapter:

  • What are classification systems?
  • a quick hands-on intro of classifiers
  • dataset - maybe Pima Indian - or some other dataset
  • editing DataSource.scala to set parameters and class.
  • how do we measure performance of classifiers? (Evaluation.scala?)
  • (n.b. not entirely clear on evaluation in prediction.io)
  • Why use PredictionIO for classification?

#####Chapter 6: Classifying text using Naive Bayes ######estimated page count: 20 Readers will learn how to handle unstructured data including how to extract features from text. They will learn the workflow of a typical classification task (training & testing)

List of topics to be covered in the chapter:

  • types of data
  • preprocessing text
  • building a naive Bayes classifier
  • unning the classifier on X(20newsGroups data? ) and evaluating the results

#####Chapter 7: Other classification algorithms ######Estimated Page Count 20 List of topics to be covered in the chapter:

  • decision trees
  • Random Forest
  • linear regression

#####Chapter 8: Classifying Big Data ######Estimated Page Count 25 Readers will learn common pitfalls of building classifiers using big data. They will learn how to run a classifier on a big data set (Wikipedia articles) on an Amazon EC2 cluster.

List of topics that will be covered in the chapter:

  • How the time and space complexity of analyzing big data grows and why PredictionIO is a good choice.
  • Common Pitfalls and how to work incrementally
  • step-by-step walking though of using a Naive Bayes classifier on the Wikipedia Article Database.
  • CASE STUDY TBD

####Part 4: Clustering Systems

#####Chapter 9: Jumping into Clustering ######estimated page count: 25 Readers will learn about clustering systems and their real world applications. They will learn how to build a simple k-means clustering system using PredictionIO.

List of topics to be covered in the chapter:

  • fundamentals of clustering
  • how we measure similarity
  • hands-on intro to clustering
  • the k-means clustering algorithm
  • clustering a very simple example.
  • a closer look at k-means
  • working with a moderate sized dataset.

#####Chapter 10: Clustering and unstructured data ######estimated page count: 20 I am not totally sure on this chapter. I am thinking of explaining and using Latent Dirichlet Allocation (LDA) on text. Or we might look at streaming k-means.

List of topics to be covered in the chapter:

  • working with unstructured data
  • clustering something like the 20 news group dataset
  • tentative: clusetering people on Twitter

#####Chapter 11: TENTATIVE: Hierarchical Clustering a ######estimated page count: 20 TBD

List of topics to be covered in the chapter:

  • a closer look at k-means
  • working with unstructured data - turning text into vectors
  • clustering something like the 20 news group dataset
  • possibly a discussion of canopy clustering and a hands-on exercise
  • beyond k-means
  • tentative: clusetering people on Twitter

#####Chapter 12: Clustering big data on a cluster ######estimated page count: 20

Part 5: Deploying Production Level Systems

not sure here. Probably do this in collaboration with several people.