CS6220: DATA MINING TECHNIQUES

Transcription

1 CS6220: DATA MINING TECHNIQUES Matrix Data: Prediction Instructor: Yizhou Sun September 21, 2015

2 Announcements TA Monisha s office hour has changed to Thursdays 10-12pm, 462WVH (the same location) Team formation due this Sunday Homework 1 out by tomorrow. 2

3 Today s Schedule Course Project Introduction Linear Regression Model Decision Tree 3

4 Methods to Learn Matrix Data Text Data Set Data Sequence Data Time Series Graph & Network Images Classification Decision Tree; Naïve Bayes; Logistic Regression SVM; knn HMM Label Propagation* Neural Network Clustering K-means; hierarchical clustering; DBSCAN; Mixture Models; kernel k- means* PLSA SCAN*; Spectral Clustering* Frequent Pattern Mining Apriori; FP-growth GSP; PrefixSpan Prediction Linear Regression Autoregression Similarity Search Ranking DTW P-PageRank PageRank 4

5 How to learn these algorithms? Three levels When it is applicable? Input, output, strengths, weaknesses, time complexity How it works? Pseudo-code, work flows, major steps Can work out a toy problem by pen and paper Why it works? Intuition, philosophy, objective, derivation, proof 5

6 Matrix Data: Prediction Matrix Data Linear Regression Model Model Evaluation and Selection Summary 6

7 Example A matrix of n p: n data objects / points p attributes / dimensions x 11 x i1 x n1 x 1f x if x nf x 1p x ip x np 7

8 Numerical E.g., height, income Attribute Type Categorical / discrete E.g., Sex, Race 8

9 Categorical Attribute Types Nominal: categories, states, or names of things Hair_color = {auburn, black, blond, brown, grey, red, white} marital status, occupation, ID numbers, zip codes Binary Nominal attribute with only 2 states (0 and 1) Symmetric binary: both outcomes equally important e.g., gender Asymmetric binary: outcomes not equally important. e.g., medical test (positive vs. negative) Convention: assign 1 to most important outcome (e.g., HIV positive) Ordinal Values have a meaningful order (ranking) but magnitude between successive values is not known. Size = {small, medium, large}, grades, army rankings 9

10 Matrix Data: Prediction Matrix Data Linear Regression Model Model Evaluation and Selection Summary 10

11 Linear Regression Ordinary Least Square Regression Closed form solution Online updating Linear Regression with Probabilistic Interpretation 11

12 The Linear Regression Problem Any Attributes to Continuous Value: x y {age; major ; gender; race} GPA {income; credit score; profession} loan {college; major ; GPA} future income 12

13 Illustration 13

14 Formalization Data: n independent data objects y i, i = 1,, n x i = x i0, x i1, x i2,, x ip T, i = 1,, n Usually a constant factor is considered, say, x i0 = 1 Model: y: dependent variable x: explanatory variables β = β 0, β 1,, β p T : weight vector y = x T β = β 0 + x 1 β 1 + x 2 β x p β p 14

15 Model Construction A 2-step Process Use training data to find the best parameter β, denoted as β Model Usage Model Evaluation Use test data to select the best model Feature selection Apply the model to the unseen data: y = x T β 15

16 Least Square Estimation Cost function (Total Square Error): J β = i x i T β y i 2 Matrix form: J β = Xβ y T (Xβ y) 1, 1, 1, x 11 x i1 x n1 or Xβ y 2 x 1f x if x nf x 1p x ip x np X: n p + 1 matrix y 1 y i y n y: n 1 vector 16

17 Ordinary Least Squares (OLS) Goal: find β that minimizes J β J β = Xβ y T Xβ y = β T X T Xβ y T Xβ β T X T y + y T y Ordinary least squares Set first derivative of J β as 0 J β = 2βT X T X 2y T X = 0 β = X T X 1 X T y 17

18 Gradient Descent Minimize the cost function by moving down in the steepest direction 18

19 Gradient Descent Online Updating Move in the direction of steepest descend η = 0.1 in practice β (t+1) :=β (t) η J β β=β (t), Where J β = i x i T β y i 2 = i J i (β) J β = i When a new observation, i, comes in, only need to update: β (t+1) :=β (t) + 2η(y i x i T β (t) )x i J i β = i 2x i (x i T β y i ) If the prediction for object i is smaller than the real value, β should move forward to the direction of x i 19

20 Other Practical Issues What if X T X is not invertible? Add a small portion of identity matrix, λi, to it (ridge regression* ) What if some attributes are categorical? Set dummy variables E.g., x = 1, if sex = F; x = 0, if sex = M Nominal variable with multiple values? Create more dummy variables for one variable What if non-linear correlation exists? Transform features, say, x to x 2 20

21 Probabilistic Interpretation Review of normal distribution X~N μ, σ 2 f X = x = 1 x μ 2 2πσ 2 e 2σ 2 21

22 Probabilistic Interpretation Model: y i = x i T β + ε i ε i ~N(0, σ 2 ) y i x i, β~n(x i T β, σ 2 ) E y i x i Likelihood: = x i T β L β = i p y i x i, β) = i 1 2πσ 2 exp{ y i x i T β 2 2σ 2 } Maximum Likelihood Estimation find β that maximizes L β arg max L = arg min J, Equivalent to OLS! 22

23 Matrix Data: Prediction Matrix Data Linear Regression Model Model Evaluation and Selection Summary 23

24 Model Selection Problem Basic problem: how to choose between competing linear regression models Model too simple: underfit the data; poor predictions; high bias; low variance Model too complex: overfit the data; poor predictions; low bias; high variance Model just right: balance bias and variance to get good predictions 24

25 Bias: E( f x ) f(x) Bias and Variance True predictor f x : x T β Estimated predictor f x : x T β How far away is the expectation of the estimator to the true value? The smaller the better. Variance: Var f x = E[ f x E f x How variant is the estimator? The smaller the better. Reconsider the cost function J β = i x i T β y i 2 Can be considered as E[ f x f(x) ε 2 ] = bias 2 + variance + noise Note E ε = 0, Var ε = σ 2 2 ] 25

26 Bias-Variance Trade-off 26

27 Cross-Validation Partition the data into K folds Use K-1 fold as training, and 1 fold as testing Calculate the average accuracy best on K training-testing pairs Accuracy on validation/test dataset! Mean square error can again be used: i x i T β y i 2 /n 27

28 AIC & BIC* AIC and BIC can be used to test the quality of statistical models AIC (Akaike information criterion) AIC = 2k 2ln( L), where k is the number of parameters in the model and L is the likelihood under the estimated parameter BIC (Bayesian Information criterion) BIC = kln(n) 2ln( L), Where n is the number of objects 28

29 Stepwise Feature Selection Avoid brute-force selection 2 p Forward selection Starting with the best single feature Always add the feature that improves the performance best Stop if no feature will further improve the performance Backward elimination Start with the full model Always remove the feature that results in the best performance enhancement Stop if removing any feature will get worse performance 29

30 Matrix Data: Prediction Matrix Data Linear Regression Model Model Evaluation and Selection Summary 30

31 Summary What is matrix data? Attribute types Linear regression OLS Probabilistic interpretation Model Evaluation and Selection Bias-Variance Trade-off Mean square error Cross-validation, AIC, BIC, step-wise feature selection 31