Machine Learning for Language Technology (2015) – DRAFT July 2015 Detailed Outline: Last updated: Thu 30 July 2015 Legend: D2014=Daume’ III, Hal (2014). A course in Machine Learning (online version 0.9) W2011=Witten, Ian and Frank, Eibe and Hall, Mark (2011). Data Mining. Practical Machine Learning Tools and Techniques. Morgan Kafmann, 3rd Edition. Lect 1 2 Topics Online: N/A In class: Opening Lecture Introduction to the course Scalable Learning What is machine learning? o It draws from exploratory statistics, inferential statistics, information theory, pattern recognition, linear algebra, calculus, etc. o Its basic characteristics: Training data & test data Generalization (hypothesis space, overfitting, underfitting) Evaluation etc. Online: Preliminaries Exploratory Statistics o variables: numeric/nominal/categorical o raw data and feature representation; o sampling, mean, variance, standard deviation, outliers, noise, etc. o graphs: how to read a histogram, scatter plot, etc. o feature standardization/normalization; o etc. Reading - D2014: 8-10 - W2011: Ch 1 - W2011: Ch2; Ch10; Ch11: 407-410; Ch17: 559562; Lab: Weka - Data Exploration and Preprocessing <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect02_LAB_Assignment.pdf> 3 Required concepts: Concepts, attributes, instances Understanding the arff format (cf .csv) The iris dataset (3 classes) Standard deviation Standardization Normalization Data Visualization – Graphs (W p. 562) Online: Decision Trees (1) Learning model - D2014: 1016; 60-62; Machine Learning for Language Technology (2015) – DRAFT July 2015 Loss function - W2011: 562565; Lab: Weka – Decision Tree (1): Reading the output <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect03_LAB_Assignment.pdf> 4 Required concepts: Crossvalidation Accuracy, P/R, F-measure Confusion matrix English Past Tense dataset (several classes, description: http://coltekin.net/cagri/ml08/lab3.html) Online: Decision Trees (2) Inductive Bias Models, parameters, hyper-parameters ID3(=J48), C4.5, etc. Pruning - D2014: 1623; 51-58; - W2011: 487494; 567; 575577; Digression: Information theory: entropy, surprisal Digression: Math: logarithms Lab: Weka – Decision Tree (2): Feature Selection and Reduction <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect04_LAB_Assignment.pdf> 5 Required concepts: Entropy Pruning Online: Perceptron (1) Numerical features Perceptron learning Convergence & separability - D2014: 3746; - W2011: 314316; 574-575; Math Digression: Dot product Lab: Weka – Perceptron (1): Discretizing numeric attributes <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect05_LAB_Assignment.pdf> 6 Required concepts: discretization Online: Perceptron (2) Voted and Averaged perceptron Limitations Digression: the “kernel trick” (weka p. 229-230) Lab: Weka – Perceptron (2): Parameter tuning <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect06_LAB_Assignment.pdf> Required concepts: - D2014: 4650; 60-62; - W2011: 147156; 577-578; Machine Learning for Language Technology (2015) – DRAFT July 2015 7 Kernel Online: Practical Issues (1) The importance of good features; Evaluating Model performance (Roc curves, etc.) - D2014: 5160; - W2011: 172177; 580-581 Lab: Weka – Practical Issues (1): Evaluating Model performance <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect07_LAB_Assignment.pdf> Required concepts: ROC curves 8 Online: Practical Issues (2): Hypothesis testing & statistically significance Lab: Weka – Practical Issues (2): Testing with Paired TTest <http://stp.lingfil.uu.se/~santinim/ml/2015/labs/Lect08_LAB_Assignment.pdf> 9 Required concepts: Paired t-test Online: Beyond binary classification Learning with unbalanced data Multiclass classification Ranking Digression: k-statistic 10 11 12 Lab: Bringing all together Online: Statistical learning (1) Theory of probability (quick repetition) Density estimation Statistical estimation (MLE etc) Statistical inference Lab: Weka – ??? Online: Statistical learning (2) Naïve bayes Prediction Lab: Weka – ??? Online: Statistical learning (3) Conditional models (logistic regression) priors Lab: Weka – ??? D2014: 63-65; W2011: 505515;