Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
“Data Science” IRDS: Data Mining Process Charles Sutton University of Edinburgh • Our working definition! • Data science is the study of the computational principles, methods, and systems for extracting knowledge from data.! • A relatively new term. A lot of current hype…! • “If you have to put ‘science’ in the name…”! • Component areas have a long history • • • • machine learning! databases! statistics! optimization • • • • natural language processing! computer vision ! speech processing! applications to science, business, health…. • Difficult to find another term for this intersection (many figures used from Murphy. Machine Learning: A Probabilistic Perspective.) 1 The term “data mining” Data mining is the analysis of (often large) observational data sets to find unsuspected relationships and to summarise the data in novel ways that are both understandable and useful to the data owner. — Hand, Mannila, Smyth, 2001 2 The term “data mining” Data mining is the analysis of (often large) observational data sets to find unsuspected relationships and to summarise the data in novel ways that are both understandable and useful to the data owner. — Hand, Mannila, Smyth, 2001 not collected for the purpose of your analysis 3 4 The term “data mining” Data mining is the analysis of (often large) observational data sets to find unsuspected relationships and to summarise the data in novel ways that are both understandable and useful to the data owner. — Hand, Mannila, Smyth, 2001 Many “easy” patterns already known e.g., pregnant example from association rule mining 5 The term “data mining” Data mining is the analysis of (often large) observational data sets to find unsuspected relationships and to summarise the data in novel ways that are both understandable and useful to the data owner. — Hand, Mannila, Smyth, 2001 Tradeoff between • predictive performance • human interpretability Ex: neural networks vs decision trees 6 What problem am I trying to solve? Before I get too far ahead of myself… 7 8 Problem Types • Visualization • Prediction: Learn a map x —> y • Prediction Examples • • Classification: Predict categorical value • Regression: Predict a real value • Others • Collaborative filtering • Learning to rank • Structured prediction supervised learning • Description • Clustering • Dimensionality reduction • Density estimation • Finding patterns unsupervised learning • Association rule mining • Detecting anomalies / outliers • Classification • Advertising • Ex: Given the text of an online advertisement and a search engine query, predict whether a user will click on the ad • Document classification • Ex: Spam filtering • Object detection • Ex: Given an image patch, dose it contain a face? Regression • Predict the final vote in an election (or referendum) from polls • Predict the temperature tomorrow given the previous few days Sometimes augmented with other structure / information • Structured prediction • Spatial data, Time series data • Ex: Predicting coding regions in DNA • Collaborative filtering (Amazon, Netflix) • Semi-supervised learning 9 Description Examples • • • • Clustering • Assign data into groups with high intra-group similarity • (like classification, except without examples of “correct” group assignments) • Ex: Cluster users into groups, based on behaviour • Social network analysis • Autoclass system (Cheeseman et al. 1988) discovered a new type of star, Dimensionality reduction • Eigenfaces • Topic modelling Discovering graph structure • Ex: Transcription networks • Ex: JamBayes for Seattle traffic jams Association rule mining • Market basket data • Computer security 11 10 2 Dimensionality reduction General problem 0 −2 4 2 0 −2 −4 −8 −6 −4 −2 0 Application to Images Data Basis 12 2 4 6 8 Phrase as a machine learning problem Data Analysis Process Roadmap Prepare, clean data More data Collect data Better features Select features Train model In the next few weeks, we’ll talk about Different learning algorithm Evaluate model Performance good enough? Decide how to improve model Yes Inspired by Wagstaff, 2012. “Machine Learning that Matters” For another more industrial process, see CRISP-DM. No • Visualization • Feature extraction • Evaluation and debugging But to talk about these, we still need to understand representation behind the algorithms No Deploy model Performance still good? Monitor deployed model Yes 13 14 Two Representation Problems Data Feature Vectors Two Representation Problems Models/Patterns Image Classification o o o Feature Extraction X w x Learning x y x x x x x o o o o Feature Vectors Models/Patterns x2 o o o o o o Feature Extraction o x1 x Document Clustering X w x Learning x y x x x x x o o o o o o o o x1 x Document Clustering Feature Extraction Learning X o o o o o o o o o o Feature Extraction o o o o o o o o oo o o o o o o o o Association Rule Mining Transaction Database Data Image Classification x2 Learning X o o o o o o o o o o o o o o o o o o oo o o o o o o o o Association Rule Mining Feature Extraction Learning X if Diapers = 1 then Beer = 1 if Ham = 1 then Pineapple = 1 Transaction Database Feature Extraction Learning X … if Diapers = 1 then Beer = 1 if Ham = 1 then Pineapple = 1 … 1. Given input, what goes in the feature vector? 15-1 15-2 Two Representation Problems Data Feature Vectors Two Representation Problems Models/Patterns Image Classification x2 o o o Feature Extraction X w x Learning x y x x o x x o o o 1. What features to use o o o o x1 x x 2. What is the space of possible models Document Clustering Feature Extraction o o o o o o o o o o Learning X o o o o o o oo o o oo o o o o o o Association Rule Mining Transaction Database Feature Extraction if Diapers = 1 then Beer = 1 Learning if Ham = 1 then Pineapple = 1 X … • In this course, we discuss features. • Model —> IAML, PMR, MLPR • But: To pick features, must understand model. • So: Whirlwind tour of models, leaving out learning algorithms 2. What is the set of possible models? 15-3 16 Linear regression Nonlinear regression Let x 2 Rd denote the feature vector. Trying to predict y 2 R Simplest choice a linear function. Define parameters w 2 R d X ŷ = f (x, w) = w> x = w j xj What if we want to learn a nonlinear function? 2 > Trick: Define new features, e.g., for scalar x, define(x) = (1, x, x ) d ŷ = f (x, w) = w> (x) j=1 this is still linear in w (to keep notation simple assume that always xd = 1 ) Given a data set x(1) . . . x(N ) , y (1) , . . . , y (N ) find the best parameters N ⇣ ⌘2 X min y (i) w> x(i) w i=1 which can be solved easily (but I won’t say how) To find parameters, the minimisation problem is now 2.5 2 min w 1.5 1 17 i=1 y (i) w > (i) (x ) ⌘2 exactly the same form as before (because x is fixed) so still just as easy 0.5 0 −2 N ⇣ X −1 0 1 2 degree 2 15 10 5 0 −5 −10 3 18 0 5 10 15 20 Logistic regression K-Nearest Neighbour (a classification method, despite the name) simple method for classification or regression Define a distance function between feature vectors D(x, x0 ) x2 Linear regression was easy. Can we do linear classification too? o x x Define a discriminant function f (x, w) = w> x x x Then predict using ( 1 if f (x, w) y= 0 otherwise o o w o o o o o o x x To classify a new feature vector x o 1. Look through your training set. Find the K closest points. Call them NK (x) o (this is memory-based learning.) x1 x x 2. Return the majority vote. 3. If you want a probability, take the proportion 0 ! p(y = c|x) = yields linear decision boundary 1 K Can get class probabilities from this idea, using logistic regression: p(y = 1|x) = 1 1 + exp{ w> x} X (the running time of this algorithm is terrible. See IAML for better indexing.) (to show decision boundaries same, compute log odds log p(y = 1|x) p(y = 0|x) 19 20 K-Nearest Neighbour data K-means clustering train 5 Decision boundaries can be highly nonlinear 4 The bigger the K, the smoother the boundary Initialize cluster centroids randomly: µ1 . . . µK This is nonparametric: the complexity of the boundary varies depending on the amount of training data Repeat 2 1 0 −1 −2 −1 0 1 2 predicted label, K=5 5 4 4 3 3 2 2 1 1 0 0 −2 −3 −1 c1 c2 c3 −2 −1 0 Assign to closest cluster ! 5 −1 For each data point i 3 predicted label, K=1 k=1 To split the data into k clusters, iterate: ! 3 −2 −3 I{y 0 = c} (y 0 ,x0 )2NK (x) 1 2 3 −2 −3 21 ki arg max D(µk , x(i) ) k2{1...K} ! Move cluster centroids (average all points currently in cluster) k=5 ! ! ! c1 c2 c3 −2 For all k in 1,2,…K PN (i) i=1 x I{ki = k} µk PN i=1 I{ki = k} Until converged −1 0 1 2 Results in “convex” clusters 3 22 Summary • Different types of model structures 1. Linear boundaries (for classification and regression) 2. Nonlinear boundaries (but linear in a set of features) 3. “Wavy” boundaries (nonparametric, piecewise linear) 4. Convex boundaries (with respect to Euclidean distance) • This will affect feature construction, soon. 23