3Prediction definition set of inputs: X1, X2, …, Xpthe output: YWe want to analyze the relationship between these variables (interpretation)We want to estimate output based on inputs (prediction)
4Prediction same concept in different literatures Machine learning: supervised learningFinance: forecastingPolitics: predictionEstimation theory: function approximationStatisticsMLEconomyX1, X2, …, XnPredictorsFeaturesIndependent variablesYResponseClassDependent variables
5Regression Why?Well-performed and accurate in both Interpretation and PredictionStrong fundamental in math, statistics and computationMany modern and advanced methods are based on Regression, even they are variant of regressionNew methods are still invented for regression:Nobel prize are still given to investigations in regression, Hot topicCould be formulated as optimization problem:that’s the reason I choose it for this class, it’s more related to subject of class than any other methods I’ve known for prediction
6Regression classification Linear RegressionLeast squareBest sub-sut Selection, Regression with feature selectionStepwise RegressionShrinkage regularization for Regression:Ridge RegressionLasso RegressionNon-Linear RegressionNumerical Data fittingANNDiscrete regressionLogistic Regression
7Before proceeding with regression Let’s investigate on some statistical property of ESTIMATION
8Estimating the parameter assume that we have iid (identically independent distributed) samples X1, ,Xn with unknown distribution.Estimating p.d.f of them is too hard in many situations, Instead of that, We want to estimate a parameter θ .is estimation of θ, it is function of X1, ,Xn .
9Bias-Variance dilemma Definition 1 : The bias of an estimator is. If it is 0, the estimator is said to be unbiased.Definition 2 : The mean squared error (MSE) of an estimator isAn interesting equation:What does it really mean?
10[Image from “More on Regularization and (Generalized) Ridge Operators”, Takane,(2007)]
11Test and training error as a function of model complexity. [ Image from “The Elements of Statistical Learning”,Second Edition, Hastie et al. (2008)]
12Linear Regression Model Set of training data :Linear Regression model:Real-valued coefficients β need to be estimated
13Linear Regression Least square Most popular estimation methodMinimize the Residual Sum of Squares:How do we minimize it?
14Linear Regression Least square Let’s rewrite last formula in this form:Quadratic function (not a point here but we shall use this property later)Differentiating respect to β and set it to zero:Unique Solution: ;Under which assumptions we could obtain unique solution?
15Linear Regression Least square, Assumptions X should be full-rank, hence is p.d and invertible, unique solution could be obtainedIn another word, features vectors should be linearly independent or uncorrelatedWhat will be happened to β if X would be non-full-rank matrix or some features would be highly correlated?
16Linear Regression Least square, flaws Low bias but High variance:and one could estimate Var(y) by:It’s hard to find meaning-full relation if we have too many features.What would you recommend to solve these problems?
17Linear Regression Improvements Model Selection (Feature Selection):Best-Subset Selection (Branch and Leap , Furnival (1974))Step-wise Selection (Greedy approach, sub-optimal but preferred)mRMR (using mutual information criterion for selection)Shrinkage Methods: impose constraint on βRidge RegressionLasso Regression
18Ridge RegressionWhen you have a problem want to be solved in statistics, There is always a Russian statistician waiting for you to solve it. (Be careful! just in statistics I guarantee , they will betray you in any other situations)Andrey Nikolayevich Tychonoff provides a Tikhonov (!!!) regularization for ill-posed problems , Also known as Ridge Regression in statistics.
19Ridge Regression first attempt Remember this?:Tychonoff added a term to avoid singularity and changed above formula to this:Now, the inverse could be computed even ifIs not of full-rank, Also β is still linear function of y.Every thing start from above formula but now we have better point of view than Tychonoff, let’s take a look!
20Ridge Regression better motivation To avoid high variance of β we just impose a constraint on it, our problem is now an optimization problem with constraints.
21Even better representation: using lagrangian form Or again even better Even better representation: using lagrangian form Or again even better! in matrix representation form, we could differentiate this formula and set it to zero Could you guess the solution?Could you find a relation between β and βridge when inputs are orthonormal?
22LASSOLeast Absolute Selection and Shrinkage Operator
23LASSO We impose L1-norm constraint on our regression No close form exists, it’s non-linear function of yHow could you solve above problem?(hint: ask Mr.Iranmehr!)
24LASSO Why?First attempt for usage of L1-norm, show significant results in signal processing, denoising[Chen et al. (1998)]Base method for LAR (new and novel method for regression, not covered here)[Efron et al. (2004)]Good for Sparse model selection where p>N[Donoho (2006b)]
25REFERENCES“The Elements of Statistical Learning”, Second Edition, Hastie et al. , 2008“More on Regularization and (Generalized) Ridge Operators”, Takane, 2007“Bias, Variance and MSE of Estimators”, Guy Lebanon, 2004“Least Squares Optimization with L1-Norm Regularization”, Mark Schmidt, 2005“Regularization: Ridge Regression and the LASSO”, Tibshirani, 2006