Scala for Machine Learning

Description

Not a single day passes that we do not hear about big data in the news media, technical conferences, and even coffee shops. The ever-increasing amount of data collected in process monitoring, research, or simple human behavior becomes valuable only if you extract knowledge from it. Machine learning is the essential tool to mine data for knowledge. This book covers the what, why, and how of machine learning: • What are the objectives and the mathematical foundations of machine learning? • Why is Scala the ideal programming language to implement machine learning algorithms? • How can you apply machine learning to solve real-world problems? Throughout this book, machine learning algorithms are described with diagrams, mathematical formulations, and documented snippets of Scala code, allowing you to understand these key concepts in your own unique way. Cover Copyright Credits About the Author About the Reviewers www.PacktPub.com Customer Feedback Table of Contents Preface Chapter 1: Getting Started Mathematical notations for the curious Why machine learning? Classification Prediction Optimization Regression Why Scala? Scala as a functional language Abstraction Higher kinded types Functors Monads Scala as an object oriented language Scala as a scalable language Model categorization Taxonomy of machine learning algorithms Unsupervised learning Clustering Dimension reduction Supervised learning Generative models Discriminative models Semi-supervised learning Reinforcement learning Leveraging Java libraries Tools and frameworks Java Scala Eclipse Scala IDE IntelliJ IDEA Scala plugin Simple build tool Apache Commons Math Description Licensing Installation JFreeChart Description Licensing Installation Other libraries and frameworks Source code Convention Context bounds Presentation Primitives and implicits Immutability Let's kick the tires Writing a simple workflow Step 1 – scoping the problem Step 2 – loading data Step 3 – preprocessing data Step 4 – discovering patterns Step 5 – implementing the classifier Step 6 – evaluating the model Summary Chapter 2: Data Pipelines Modeling What is a model? Model versus design Selecting features Extracting features Defining a methodology Monadic data transformation Error handling Monads to the rescue mplicit models Explicit models Workflow computational model Supporting mathematical abstractions Step 1 – variable declaration Step 2 – model definition Step 3 – instantiation Composing mixins to build workflow Understanding the problem Defining modules Instantiating the workflow Modularizing Profiling data Immutable statistics Z-score and Gauss Assessing a model Validation Key quality metrics F-score for binomial classification F-score for multinomial classification Area under the curves Area under PRC Area under ROC Cross-validation One-fold cross-validation K-fold cross-validation Bias-variance decomposition Overfitting Summary Chapter 3: Data Preprocessing Time series in Scala Context bounds Types and operations Transpose operator Differential operator Lazy views Moving averages Simple moving average Weighted moving average Exponential moving average Fourier analysis Discrete Fourier transform (DFT) DFT-based filtering Detection of market cycles The discrete Kalman filter The state space estimation The transition equation The measurement equation The recursive algorithm Prediction Correction Kalman smoothing Fixed lag smoothing Experimentation Benefits and drawbacks Alternative preprocessing techniques Summary Chapter 4: Unsupervised Learning K-mean clustering K-means Measuring similarity Defining the algorithm Step 1 – Clusters configuration Step 2 – Clusters assignment Step 3 – Reconstruction error minimization Step 4 – Classification Curse of dimensionality Evaluation The results Tuning the number of clusters Validation Expectation-Maximization (EM) Gaussian mixture model EM overview Implementation Classification Testing Online EM Summary Chapter 5: Dimension Reduction Challenging model complexity The divergences The Kullback-Leibler divergence Overview Implementation Testing The mutual information Principal components analysis (PCA) Algorithm Implementation Test case Evaluation Extending PCA Validation Categorical features Performance Nonlinear models Kernel PCA Manifolds Summary Chapter 6: Naïve Bayes Classifiers Probabilistic graphical models Naïve Bayes classifiers Introducing the multinomial Naïve Bayes Formalism The frequentist perspective The predictive model The zero-Frequency problem Implementation Design Training Classification F1 Validation Features extraction Testing Multivariate Bernoulli classification Model Implementation Naïve Bayes and text mining Basics information retrieval Implementation Analyzing documents Extracting relative terms frequency Generating the features Testing Retrieving textual information Evaluating text mining classifier Pros and cons Summary Chapter 7: Sequential Data Models Markov decision processes The Markov property The first-order discrete Markov chain The hidden Markov model (HMM) Notation The lambda model Design Evaluation (CF-1) Alpha (forward pass) Beta (backward pass) Training (CF-2) Baum-Welch estimator (EM) Decoding (CF-3) The Viterbi algorithm Putting it all together Test case 1 – Training HMM as filtering technique Conditional random fields Introduction to CRF Linear chain CRF Regularized CRF and text analytics The feature functions model Design Implementation Configuring the CRF classifier Training the CRF model Applying the CRF model Tests The training convergence profile Impact of the size of the training set Impact of L2 regularization factor Comparing CRF and HMM Performance consideration Summary Chapter 8: Monte Carlo Inference The purpose of sampling Gaussian sampling Box-Muller transform Monte Carlo approximation Overview Implementation Bootstrapping with replacement Overview Resampling Implementation Pros and cons of bootstrap Markov Chain Monte Carlo (MCMC) Overview Metropolis-Hastings (MH) Implementation Test Summary Chapter 9: Regression and Regularization Linear regression Univariate linear regression Implementation Test case Ordinary least squares (OLS) regression Design Implementation Test case 1 – trending Test case 2 – features selection Regularization Ln roughness penalty Ridge regression Design Implementation Test case Numerical optimization Logistic regression Logistic function Design Training workflow Step 1 – configuring the optimizer Step 2 – computing the Jacobian matrix Step 3 – managing the convergence of optimizer Step 4 – defining the least squares problem Step 5 – minimizing the sum of square errors Test Classification Summary Chapter 10: Multi-Layer Perceptron Feed-forward neural networks (FFNN) The biological background Mathematical background The multilayer perceptron (MLP) Activation function Network topology Design Configuration Network components Network topology Input and hidden layers Output layer Synapses Connections Weights initialization Model Problem types (modes) Online versus batch training Training epoch Step 1 – input forward propagation Step 2 – error backpropagation Step 3 – exit condition Putting it all together Training and classification Regularization Model generation Fast Fisher-Yates shuffle Prediction Model fitness Evaluation Execution profile Impact of learning rate Impact of the momentum factor Impact of the number of hidden layers Test case Implementation Models evaluation Impact of hidden layers' architecture Benefits and limitations Summary Chapter 11: Deep Learning Sparse autoencoder Undercomplete autoencoder Deterministic autoencoder Categorization Feed-forward sparse, undercomplete autoencoder Sparsity updating equations Implementation Restricted Boltzmann Machines (RBMs) Boltzmann machine Binary restricted Boltzmann machines Conditional probabilities Sampling Log-likelihood gradient Contrastive divergence Configuration parameters Unsupervised learning Convolution neural networks Local receptive fields Weight sharing Convolution layers Sub-sampling layers Putting it all together Summary Chapter 12: Kernel Models and SVM Kernel functions Overview Common discriminative kernels Kernel monadic composition The support vector machine (SVM) The linear SVM The separable case (hard margin) The non-separable case (soft margin) The nonlinear SVM Max-margin classification The kernel trick Support vector classifier (SVC) The binary SVC Anomaly detection with one-class SVC Support vector regression (SVR) Overview SVR versus linear regression Performance considerations Summary Chapter 13: Evolutionary Computing Evolution The origin NP problems Evolutionary computing Genetic algorithms and machine learning Genetic algorithm components Encodings Value encoding Predicate encoding Solution encoding The encoding scheme Genetic operators Selection Crossover Mutation Fitness score Implementation Software design Key components Population Chromosomes Genes Selection Controlling population growth GA configuration Crossover Population Chromosomes Genes Mutation Population Chromosomes Genes Reproduction Solver GA for trading strategies Definition of trading strategies Trading operators The cost function Market signals Trading strategies Signal encoding Test case – Fall 2008 market crash Creating trading strategies Configuring the optimizer Finding the best trading strategy Tests Advantages and risks of genetic algorithms Summary Chapter 14: Multi-Armed Bandits K-armed bandit Exploration-exploitation trade-offs Expected cumulative regret Bayesian Bernoulli bandits Epsilon-greedy algorithm Thompson sampling Bandit context Prior/posterior beta distribution Implementation Simulated exploration and exploitation Upper bound confidence Confidence interval Implementation Summary Chapter 15: Reinforcement Learning Reinforcement learning Understanding the challenge A solution – Q-learning Terminology Concept Value of policy Bellman optimality equations Temporal difference for model-free learning Action-value iterative update Implementation Software design The states and actions The search space The policy and action-value The Q-learning components The Q-learning training Tail recursion to the rescue Validation The prediction Option trading using Q-learning Option property Option model Quantization Putting it all together Evaluation Pros and cons of reinforcement learning Learning classifier systems Introduction to LCS Combining learning and evolution Terminology Extended learning classifier systems XCS components Application to portfolio management XCS core data XCS rules Covering Example of implementation Benefits and limitations of learning classifier systems Summary Chapter 16: Parallelism in Scala and Akka Overview Scala Object creation Streams Memory on demand Design for reusing Streams memory Parallel collections Processing a parallel collection Benchmark framework Performance evaluation Scalability with Actors The Actor model Partitioning Beyond Actors – reactive programming Akka Master-workers Messages exchange Worker Actors The workflow controller The master Actor Master with routing Distributed discrete Fourier transform Limitations Futures Blocking on futures Future callbacks Putting it all together Summary Chapter 17 : Apache Spark MLlib Overview Apache Spark core Why Spark? Design principles In-memory persistency Laziness Transforms and actions Shared variables Experimenting with Spark Deploying Spark Using Spark shell MLlib library Overview Creating RDDs K-means using MLlib Tests Reusable ML pipelines Reusable ML transforms Encoding features Training the model Predictive model Training summary statistics Validating the model Grid search Apache Spark and ScalaTest Extending Spark Kullback-Leibler divergence Implementation Kullback-Leibler evaluator Streaming engine Why streaming? Batch and real-time processing Architecture overview Discretized streams Use case – continuous parsing Checkpointing Performance evaluation Tuning parameters Performance considerations Pros and cons Summary Appendix A: Basic Concepts Scala programming List of libraries and tools Code snippets format Best practices Encapsulation Class constructor template Companion objects versus case classes Enumerations versus case classes Overloading Design template for immutable classifiers Utility classes Data extraction Financial data sources Documents extraction DMatrix class Counter Monitor Mathematics Linear algebra QR decomposition LU factorization LDL decomposition Cholesky factorization Singular Value Decomposition (SVD) Eigenvalue decomposition Algebraic and numerical libraries First order predicate logic Jacobian and Hessian matrices Summary of optimization techniques Gradient descent methods Quasi-Newton algorithms Nonlinear least squares minimization Lagrange multipliers Overview dynamic programming Finances 101 Fundamental analysis Technical analysis Terminology Trading data Trading signal and strategy Price patterns Options trading Financial data sources Suggested online courses References Appendix B : References