DOCSLIB.ORG

Cross-Validation for Detecting and Preventing Overfitting

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Cross-Validation for Detecting and Preventing Overfitting Cross-validation for detecting and preventing overfitting Note to other teachers and users of Andrew W. Moore these slides. Andrew would be delighted if you found this source material useful in giving your own lectures. Feel free to use Professor these slides verbatim, or to modify them to fit your own needs. PowerPoint originals are available. If you make use School of Computer Science of a significant portion of these slides in your own lecture, please include this message, or the following link to the Carnegie Mellon University source repository of Andrew’s tutorials: http://www.cs.cmu.edu/~awm/tutorials . www.cs.cmu.edu/~awm Comments and corrections gratefully received. [email protected] 412-268-7599 Copyright © Andrew W. Moore Slide 1 A Regression Problem y = f(x) + noise Can we learn f from this data? y x Let’s consider three methods… Copyright © Andrew W. Moore Slide 2 Linear Regression y x Copyright © Andrew W. Moore Slide 3 Linear Regression Univariate Linear regression with a constant term: X Y X= 3 y= 7 Originally 3 7 1 3 discussed in the previous Andrew 1 3 : : Lecture: “Neural Nets” : : x1=(3).. y1=7.. Copyright © Andrew W. Moore Slide 4 Linear Regression Univariate Linear regression with a constant term: X Y X= 3 y= 7 3 7 1 3 1 3 : : : : x =(3).. y =7.. Z= y1= 1 1 3 7 1 1 3 : : z1=(1,3).. y1=7.. zk=(1,xk) Copyright © Andrew W. Moore Slide 5 Linear Regression Univariate Linear regression with a constant term: X Y X= 3 y= 7 3 7 1 3 1 3 : : : : x =(3).. y =7.. Z= y1= 1 1 3 7 1 1 3 : : β=(ZTZ)-1(ZTy) z1=(1,3).. y1=7.. est y = β0+ β1 x zk=(1,xk) Copyright © Andrew W. Moore Slide 6 Quadratic Regression y x Copyright © Andrew W. Moore Slide 7 Quadratic Regression Much more about X Y X= 3 y= 7 this in the future Andrew Lecture: 3 7 1 3 “Favorite Regression 1 3 : : Algorithms” : : x1=(3,2).. y1=7.. 1 3 9 y= Z= 7 1 1 1 3 : β=(ZTZ)-1(ZTy) : z=(1 , x, x2 ) est 2 , y = β0+ β1 x+ β2 x Copyright © Andrew W. Moore Slide 8 Join-the-dots Also known as piecewise linear nonparametric regression if that makes you feel better y x Copyright © Andrew W. Moore Slide 9 Which is best? y y x x Why not choose the method with the best fit to the data? Copyright © Andrew W. Moore Slide 10 What do we really want? y y x x Why not choose the method with the best fit to the data? “How well are you going to predict future data drawn from the same distribution?” Copyright © Andrew W. Moore Slide 11 The test set method 1. Randomly choose 30% of the data to be in a test set 2. The remainder is a y training set x Copyright © Andrew W. Moore Slide 12 The test set method 1. Randomly choose 30% of the data to be in a test set 2. The remainder is a y training set 3. Perform your regression on the training x set (Linear regression example) Copyright © Andrew W. Moore Slide 13 The test set method 1. Randomly choose 30% of the data to be in a test set 2. The remainder is a y training set 3. Perform your regression on the training x set (Linear regression example) 4. Estimate your future performance with the test Mean Squared Error = 2.4 set Copyright © Andrew W. Moore Slide 14 The test set method 1. Randomly choose 30% of the data to be in a test set 2. The remainder is a y training set 3. Perform your regression on the training x set (Quadratic regression example) 4. Estimate your future performance with the test Mean Squared Error = 0.9 set Copyright © Andrew W. Moore Slide 15 The test set method 1. Randomly choose 30% of the data to be in a test set 2. The remainder is a y training set 3. Perform your regression on the training x set (Join the dots example) 4. Estimate your future performance with the test Mean Squared Error = 2.2 set Copyright © Andrew W. Moore Slide 16 The test set method Good news: •Very very simple •Can then simply choose the method with the best test-set score Bad news: •What’s the downside? Copyright © Andrew W. Moore Slide 17 The test set method Good news: •Very very simple •Can then simply choose the method with the best test-set score Bad news: •Wastes data: we get an estimate of the We say the “test-set best method to apply to 30% less data estimator of performance •If we don’t have much data, our test-set has high might just be lucky or unlucky variance” Copyright © Andrew W. Moore Slide 18 LOOCV (Leave-one-out Cross Validation) For k=1 to R th 1. Let (xk,yk) be the k record y x Copyright © Andrew W. Moore Slide 19 LOOCV (Leave-one-out Cross Validation) For k=1 to R th 1. Let (xk,yk) be the k record 2. Temporarily remove (xk,yk) from the dataset y x Copyright © Andrew W. Moore Slide 20 LOOCV (Leave-one-out Cross Validation) For k=1 to R th 1. Let (xk,yk) be the k record 2. Temporarily remove (xk,yk) from the dataset 3. Train on the remaining R-1 y datapoints x Copyright © Andrew W. Moore Slide 21 LOOCV (Leave-one-out Cross Validation) For k=1 to R th 1. Let (xk,yk) be the k record 2. Temporarily remove (xk,yk) from the dataset 3. Train on the remaining R-1 y datapoints 4. Note your error (xk,yk) x Copyright © Andrew W. Moore Slide 22 LOOCV (Leave-one-out Cross Validation) For k=1 to R th 1. Let (xk,yk) be the k record 2. Temporarily remove (xk,yk) from the dataset 3. Train on the remaining R-1 y datapoints 4. Note your error (xk,yk) x When you’ve done all points, report the mean error. Copyright © Andrew W. Moore Slide 23 LOOCV (Leave-one-out Cross Validation) For k=1 to R 1. Let (xk,yk) be the kth record y y y 2. Temporarily remove (xk,yk) from the dataset x x x 3. Train on the remaining R-1 datapoints 4. Note your y y y error (xk,yk) When you’ve done all points, x x x report the mean error. MSELOOCV = 2.12 y y y x x x Copyright © Andrew W. Moore Slide 24 LOOCV for Quadratic Regression For k=1 to R 1. Let (xk,yk) be the kth record y y y 2. Temporarily remove (xk,yk) from the dataset x x x 3. Train on the remaining R-1 datapoints 4. Note your y y y error (xk,yk) When you’ve done all points, x x x report the mean error. MSELOOCV =0.962 y y y x x x Copyright © Andrew W. Moore Slide 25 LOOCV for Join The Dots For k=1 to R 1. Let (xk,yk) be the kth record y y y 2. Temporarily remove (xk,yk) from the dataset x x x 3. Train on the remaining R-1 datapoints 4. Note your y y y error (xk,yk) When you’ve done all points, x x x report the mean error. MSELOOCV =3.33 y y y x x x Copyright © Andrew W. Moore Slide 26 Which kind of Cross Validation? Downside Upside Test-set Variance: unreliable Cheap estimate of future performance Leave- Expensive. Doesn’t one-out Has some weird waste data behavior ..can we get the best of both worlds? Copyright © Andrew W. Moore Slide 27 k-fold Cross Randomly break the dataset into k partitions (in our example we’ll have k=3 Validation partitions colored Red Green and Blue) y x Copyright © Andrew W. Moore Slide 28 k-fold Cross Randomly break the dataset into k partitions (in our example we’ll have k=3 Validation partitions colored Red Green and Blue) For the red partition: Train on all the points not in the red partition. Find the test-set sum of errors on the red points. y x Copyright © Andrew W. Moore Slide 29 k-fold Cross Randomly break the dataset into k partitions (in our example we’ll have k=3 Validation partitions colored Red Green and Blue) For the red partition: Train on all the points not in the red partition. Find the test-set sum of errors on the red points. For the green partition: Train on all the points not in the green partition. y Find the test-set sum of errors on the green points. x Copyright © Andrew W. Moore Slide 30 k-fold Cross Randomly break the dataset into k partitions (in our example we’ll have k=3 Validation partitions colored Red Green and Blue) For the red partition: Train on all the points not in the red partition. Find the test-set sum of errors on the red points. For the green partition: Train on all the points not in the green partition. y Find the test-set sum of errors on the green points. For the blue partition: Train on all the points not in the blue partition. Find x the test-set sum of errors on the blue points. Copyright © Andrew W. Moore Slide 31 k-fold Cross Randomly break the dataset into k partitions (in our example we’ll have k=3 Validation partitions colored Red Green and Blue) For the red partition: Train on all the points not in the red partition.
Load more

Recommended publications
  • Harmless Overfitting: Using Denoising Autoencoders in Estimation of Distribution Algorithms
    Journal of Machine Learning Research 21 (2020) 1-31 Submitted 10/16; Revised 4/20; Published 4/20 Harmless Overfitting: Using Denoising Autoencoders in Estimation of Distribution Algorithms Malte Probst∗ [email protected] Honda Research Institute EU Carl-Legien-Str. 30, 63073 Offenbach, Germany Franz Rothlauf [email protected] Information Systems and Business Administration Johannes Gutenberg-Universit¨atMainz Jakob-Welder-Weg 9, 55128 Mainz, Germany Editor: Francis Bach Abstract Estimation of Distribution Algorithms (EDAs) are metaheuristics where learning a model and sampling new solutions replaces the variation operators recombination and mutation used in standard Genetic Algorithms. The choice of these models as well as the correspond- ing training processes are subject to the bias/variance tradeoff, also known as under- and overfitting: simple models cannot capture complex interactions between problem variables, whereas complex models are susceptible to modeling random noise. This paper suggests using Denoising Autoencoders (DAEs) as generative models within EDAs (DAE-EDA). The resulting DAE-EDA is able to model complex probability distributions. Furthermore, overfitting is less harmful, since DAEs overfit by learning the identity function. This over- fitting behavior introduces unbiased random noise into the samples, which is no major problem for the EDA but just leads to higher population diversity. As a result, DAE-EDA runs for more generations before convergence and searches promising parts of the solution space more thoroughly. We study the performance of DAE-EDA on several combinatorial single-objective optimization problems. In comparison to the Bayesian Optimization Al- gorithm, DAE-EDA requires a similar number of evaluations of the objective function but is much faster and can be parallelized efficiently, making it the preferred choice especially for large and difficult optimization problems.
    [Show full text]
  • More Perceptron
    More Perceptron Instructor: Wei Xu Some slides adapted from Dan Jurfasky, Brendan O’Connor and Marine Carpuat A Biological Neuron https://www.youtube.com/watch?v=6qS83wD29PY Perceptron (an artificial neuron) inputs Perceptron Algorithm • Very similar to logistic regression • Not exactly computing gradient (simpler) vs. weighted sum of features Online Learning • Rather than making a full pass through the data, compute gradient and update parameters after each training example. Stochastic Gradient Descent • Updates (or more specially, gradients) will be less accurate, but the overall effect is to move in the right direction • Often works well and converges faster than batch learning Online Learning • update parameters for each training example Initialize weight vector w = 0 Create features Loop for K iterations Loop for all training examples x_i, y_i … update_weights(w, x_i, y_i) Perceptron Algorithm • Very similar to logistic regression • Not exactly computing gradient Initialize weight vector w = 0 Loop for K iterations Loop For all training examples x_i if sign(w * x_i) != y_i Error-driven! w += y_i * x_i Error-driven Intuition • For a given example, makes a prediction, then checks to see if this prediction is correct. - If the prediction is correct, do nothing. - If the prediction is wrong, change its parameters so that it would do better on this example next time around. Error-driven Intuition Error-driven Intuition Error-driven Intuition Exercise Perceptron (vs. LR) • Only hyperparameter is maximum number of iterations (LR also needs learning rate) • Guaranteed to converge if the data is linearly separable (LR always converge) objective function is convex What if non linearly separable? • In real-world problems, this is nearly always the case.
    [Show full text]
  • Measuring Overfitting and Mispecification in Nonlinear Models
    WP 11/25 Measuring overfitting and mispecification in nonlinear models Marcel Bilger Willard G. Manning August 2011 york.ac.uk/res/herc/hedgwp Measuring overfitting and mispecification in nonlinear models Marcel Bilger, Willard G. Manning The Harris School of Public Policy Studies, University of Chicago, USA Abstract We start by proposing a new measure of overfitting expressed on the untransformed scale of the dependent variable, which is generally the scale of interest to the analyst. We then show that with nonlinear models shrinkage due to overfitting gets confounded by shrinkage—or expansion— arising from model misspecification. Out-of-sample predictive calibration can in fact be expressed as in-sample calibration times 1 minus this new measure of overfitting. We finally argue that re-calibration should be performed on the scale of interest and provide both a simulation study and a real-data illustration based on health care expenditure data. JEL classification: C21, C52, C53, I11 Keywords: overfitting, shrinkage, misspecification, forecasting, health care expenditure 1. Introduction When fitting a model, it is well-known that we pick up part of the idiosyncratic char- acteristics of the data as well as the systematic relationship between a dependent and explanatory variables. This phenomenon is known as overfitting and generally occurs when a model is excessively complex relative to the amount of data available. Overfit- ting is a major threat to regression analysis in terms of both inference and prediction. When models greatly over-explain the data at hand they can even show relations which reflect chance only. Consequently, overfitting casts doubt on the true statistical signif- icance of the effects found by the analyst as well as the magnitude of the response.
    [Show full text]
  • Over-Fitting in Model Selection with Gaussian Process Regression
    Over-Fitting in Model Selection with Gaussian Process Regression Rekar O. Mohammed and Gavin C. Cawley University of East Anglia, Norwich, UK [email protected], [email protected] http://theoval.cmp.uea.ac.uk/ Abstract. Model selection in Gaussian Process Regression (GPR) seeks to determine the optimal values of the hyper-parameters governing the covariance function, which allows flexible customization of the GP to the problem at hand. An oft-overlooked issue that is often encountered in the model process is over-fitting the model selection criterion, typi- cally the marginal likelihood. The over-fitting in machine learning refers to the fitting of random noise present in the model selection criterion in addition to features improving the generalisation performance of the statistical model. In this paper, we construct several Gaussian process regression models for a range of high-dimensional datasets from the UCI machine learning repository. Afterwards, we compare both MSE on the test dataset and the negative log marginal likelihood (nlZ), used as the model selection criteria, to find whether the problem of overfitting in model selection also affects GPR. We found that the squared exponential covariance function with Automatic Relevance Determination (SEard) is better than other kernels including squared exponential covariance func- tion with isotropic distance measure (SEiso) according to the nLZ, but it is clearly not the best according to MSE on the test data, and this is an indication of over-fitting problem in model selection. Keywords: Gaussian process · Regression · Covariance function · Model selection · Over-fitting 1 Introduction Supervised learning tasks can be divided into two main types, namely classifi- cation and regression problems.
    [Show full text]
  • On Overfitting and Asymptotic Bias in Batch Reinforcement Learning With
    Journal of Artificial Intelligence Research 65 (2019) 1-30 Submitted 06/2018; published 05/2019 On Overfitting and Asymptotic Bias in Batch Reinforcement Learning with Partial Observability Vincent François-Lavet [email protected] Guillaume Rabusseau [email protected] Joelle Pineau [email protected] School of Computer Science, McGill University University Street 3480, Montreal, QC, H3A 2A7, Canada Damien Ernst [email protected] Raphael Fonteneau [email protected] Montefiore Institute, University of Liege Allée de la découverte 10, 4000 Liège, Belgium Abstract This paper provides an analysis of the tradeoff between asymptotic bias (suboptimality with unlimited data) and overfitting (additional suboptimality due to limited data) in the context of reinforcement learning with partial observability. Our theoretical analysis formally characterizes that while potentially increasing the asymptotic bias, a smaller state representation decreases the risk of overfitting. This analysis relies on expressing the quality of a state representation by bounding L1 error terms of the associated belief states. Theoretical results are empirically illustrated when the state representation is a truncated history of observations, both on synthetic POMDPs and on a large-scale POMDP in the context of smartgrids, with real-world data. Finally, similarly to known results in the fully observable setting, we also briefly discuss and empirically illustrate how using function approximators and adapting the discount factor may enhance the tradeoff between asymptotic bias and overfitting in the partially observable context. 1. Introduction This paper studies sequential decision-making problems that may be modeled as Markov Decision Processes (MDP) but for which the state is partially observable.
    [Show full text]
  • Overfitting Princeton University COS 495 Instructor: Yingyu Liang Review: Machine Learning Basics Math Formulation
    Machine Learning Basics Lecture 6: Overfitting Princeton University COS 495 Instructor: Yingyu Liang Review: machine learning basics Math formulation • Given training data 푥푖, 푦푖 : 1 ≤ 푖 ≤ 푛 i.i.d. from distribution 퐷 1 • Find 푦 = 푓(푥) ∈ 퓗 that minimizes 퐿෠ 푓 = σ푛 푙(푓, 푥 , 푦 ) 푛 푖=1 푖 푖 • s.t. the expected loss is small 퐿 푓 = 피 푥,푦 ~퐷[푙(푓, 푥, 푦)] Machine learning 1-2-3 • Collect data and extract features • Build model: choose hypothesis class 퓗 and loss function 푙 • Optimization: minimize the empirical loss Machine learning 1-2-3 Feature mapping Maximum Likelihood • Collect data and extract features • Build model: choose hypothesis class 퓗 and loss function 푙 • Optimization: minimize the empirical loss Occam’s razor Gradient descent; convex optimization Overfitting Linear vs nonlinear models 2 푥1 2 푥2 푥1 2푥1푥2 푦 = sign(푤푇휙(푥) + 푏) 2푐푥1 푥2 2푐푥2 푐 Polynomial kernel Linear vs nonlinear models • Linear model: 푓 푥 = 푎0 + 푎1푥 2 3 푀 • Nonlinear model: 푓 푥 = 푎0 + 푎1푥 + 푎2푥 + 푎3푥 + … + 푎푀 푥 • Linear model ⊆ Nonlinear model (since can always set 푎푖 = 0 (푖 > 1)) • Looks like nonlinear model can always achieve same/smaller error • Why one use Occam’s razor (choose a smaller hypothesis class)? Example: regression using polynomial curve 푡 = sin 2휋푥 + 휖 Figure from Machine Learning and Pattern Recognition, Bishop Example: regression using polynomial curve 푡 = sin 2휋푥 + 휖 Regression using polynomial of degree M Figure from Machine Learning and Pattern Recognition, Bishop Example: regression using polynomial curve 푡 = sin 2휋푥 + 휖 Figure from Machine Learning
    [Show full text]
  • Word2bits - Quantized Word Vectors
    Word2Bits - Quantized Word Vectors Maximilian Lam [email protected] Abstract Word vectors require significant amounts of memory and storage, posing issues to resource limited devices like mobile phones and GPUs. We show that high quality quantized word vectors using 1-2 bits per parameter can be learned by introducing a quantization function into Word2Vec. We furthermore show that training with the quantization function acts as a regularizer. We train word vectors on English Wikipedia (2017) and evaluate them on standard word similarity and analogy tasks and on question answering (SQuAD). Our quantized word vectors not only take 8-16x less space than full precision (32 bit) word vectors but also outperform them on word similarity tasks and question answering. 1 Introduction Word vectors are extensively used in deep learning models for natural language processing. Each word vector is typically represented as a 300-500 dimensional vector, with each parameter being 32 bits. As there are millions of words, word vectors may take up to 3-6 GB of memory/storage – a massive amount relative to other portions of a deep learning model[25]. These requirements pose issues to memory/storage limited devices like mobile phones and GPUs. Furthermore, word vectors are often re-trained on application specific data for better performance in application specific domains[27]. This motivates directly learning high quality compact word representations rather than adding an extra layer of compression on top of pretrained word vectors which may be computational expensive and degrade accuracy. Recent trends indicate that deep learning models can reach a high accuracy even while training in the presence of significant noise and perturbation[5, 6, 9, 28, 32].
    [Show full text]
  • On the Use of Entropy to Improve Model Selection Criteria
    entropy Article On the Use of Entropy to Improve Model Selection Criteria Andrea Murari 1, Emmanuele Peluso 2, Francesco Cianfrani 2, Pasquale Gaudio 2 and Michele Lungaroni 2,* 1 Consorzio RFX (CNR, ENEA, INFN, Universita’ di Padova, Acciaierie Venete SpA), 35127 Padova, Italy; [email protected] 2 Department of Industrial Engineering, University of Rome “Tor Vergata”, 00133 Roma, Italy; [email protected] (E.P.); [email protected] (F.C.); [email protected] (P.G.) * Correspondence: [email protected]; Tel.: +39-(0)6-7259-7196 Received: 21 January 2019; Accepted: 10 April 2019; Published: 12 April 2019 Abstract: The most widely used forms of model selection criteria, the Bayesian Information Criterion (BIC) and the Akaike Information Criterion (AIC), are expressed in terms of synthetic indicators of the residual distribution: the variance and the mean-squared error of the residuals respectively. In many applications in science, the noise affecting the data can be expected to have a Gaussian distribution. Therefore, at the same level of variance and mean-squared error, models, whose residuals are more uniformly distributed, should be favoured. The degree of uniformity of the residuals can be quantified by the Shannon entropy. Including the Shannon entropy in the BIC and AIC expressions improves significantly these criteria. The better performances have been demonstrated empirically with a series of simulations for various classes of functions and for different levels and statistics of the noise. In presence of outliers, a better treatment of the errors, using the Geodesic Distance, has proved essential. Keywords: Model Selection Criteria; Bayesian Information Criterion (BIC); Akaike Information Criterion (AIC); Shannon Entropy; Geodesic Distance 1.
    [Show full text]
  • Piecewise Regression Analysis Through Information Criteria Using Mathematical Programming
    Piecewise Regression Analysis through Information Criteria using Mathematical Programming Ioannis Gkioulekas, Lazaros G. Papageorgiou ∗ Centre for Process Systems Engineering, Department of Chemical Engineering, UCL (University College London), Torrington Place, London WC1E 7JE, UK ∗Corresponding author. Tel: +44 20 7679 2563 E-mail addresses: [email protected] (I. Gkioulekas), [email protected] (L.G. Papageorgiou) Abstract: Regression is a predictive analysis tool that examines the relationship be- tween independent and dependent variables. The goal of this analysis is to fit a mathe- matical function that describes how the value of the response changes when the values of the predictors vary. The simplest form of regression is linear regression which in the case multiple regression, tries to explain the data by simply fitting a hyperplane minimising the absolute error of the fitting. Piecewise regression analysis partitions the data into multiple regions and a regression function is fitted to each one. Such an approach is the OPLRA (Optimal Piecewise Linear Regression Analysis) model (Yang et al., 2016) which is a mathematical programming approach that optimally partitions the data into multiple regions and fits a linear regression functions minimising the Mean Absolute Error between prediction and truth. However, using many regions to describe the data can lead to overfitting and bad results. In this work an extension of the OPLRA model is proposed that deals with the problem of selecting the optimal number of regions as well as overfitting. To achieve this result, information criteria such as the Akaike and the Bayesian are used that reward predictive accuracy and penalise model complexity.
    [Show full text]
  • Improving Generalization in Reinforcement Learning with Mixture Regularization
    Improving Generalization in Reinforcement Learning with Mixture Regularization Kaixin Wang1 Bingyi Kang1 Jie Shao2 Jiashi Feng1 1National University of Singapore 2ByteDance AI Lab {kaixin.wang, kang}@u.nus.edu, [email protected], [email protected] Abstract Deep reinforcement learning (RL) agents trained in a limited set of environments tend to suffer overfitting and fail to generalize to unseen testing environments. To improve their generalizability, data augmentation approaches (e.g. cutout and random convolution) are previously explored to increase the data diversity. How- ever, we find these approaches only locally perturb the observations regardless of the training environments, showing limited effectiveness on enhancing the data diversity and the generalization performance. In this work, we introduce a simple approach, named mixreg, which trains agents on a mixture of observa- tions from different training environments and imposes linearity constraints on the observation interpolations and the supervision (e.g. associated reward) inter- polations. Mixreg increases the data diversity more effectively and helps learn smoother policies. We verify its effectiveness on improving generalization by conducting extensive experiments on the large-scale Procgen benchmark. Results show mixreg outperforms the well-established baselines on unseen testing envi- ronments by a large margin. Mixreg is simple, effective and general. It can be applied to both policy-based and value-based RL algorithms. Code is available at https://github.com/kaixin96/mixreg. 1 Introduction Deep Reinforcement Learning (RL) has brought significant progress in learning policies to tackle various challenging tasks, such as board games like Go [19, 21], Chess and Shogi [20], video games like Atari [15, 1] and StarCraft [27], and robotics control tasks [14].
    [Show full text]
  • Practical Issues in Machine Learning Overfitting Overfitting and Model Selection and Model Selection
    Practical Issues in Machine Learning Overfitting and Model selection Aarti Singh Machine Learning 10-701/15-781 Feb 3, 2010 True vs. Empirical Risk True Risk : Target performance measure Classification – Probability of misclassification Regression – Mean Squared Error Also known as “Generalization Error” – performance on a random test point (X,Y) True vs. Empirical Risk True Risk : Target performance measure Classification – Probability of misclassification Regression – Mean Squared Error Also known as “Generalization Error” – performance on a random test point (X,Y) Empirical Risk : Performance on training data Classification – Proportion of misclassified examples Regression – Average Squared Error Overfitting Is the following predictor a good one? What is its empirical risk? (performance on training data) zero ! What about true risk? > zero Will predict very poorly on new random test point, Large generalization error ! Overfitting If we allow very complicated predictors, we could overfit the training data. Examples: Classification (0-NN classifier, decision tree with one sample/leaf) Football player ? No Yes Weight Weight Height Height Overfitting If we allow very complicated predictors, we could overfit the training data. Examples: Regression (Polynomial of order k – degree up to k-1) code online 1.5 1.4 k=1 k=2 1.2 1 1 0.8 0.6 0.5 0.4 0.2 0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.4 5 0 k=3 1.2 k=7 -5 1 -10 0.8 -15 0.6 -20 -25 0.4 -30 0.2 -35 0 -40 -0.2 -45 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Effect of Model Complexity If we allow very complicated predictors, we could overfit the training data.
    [Show full text]
  • Piecewise Regression Through the Akaike Information Criterion Using Mathematical Programming ?
    Piecewise Regression through the Akaike Information Criterion using Mathematical Programming ? Ioannis Gkioulekas, Lazaros G. Papageorgiou Centre for Process Systems Engineering, Department of Chemical Engineering, UCL (University College London), Torrington Place, London WC1E 7JE, UK Corresponding author: [email protected] Abstract: In machine learning, regression analysis is a tool for predicting the output variables from a set of known independent variables. Through regression analysis, a function that captures the relationship between the variables is fitted to the data. Many methods from literature tackle this problem with various degrees of difficulty. Some simple methods include linear regression and least squares, while some are more complicated such as support vector regression. Piecewise or segmented regression is a method of analysis that partitions the independent variables into intervals and a function is fitted to each interval. In this work, the Optimal Piecewise Linear Regression Analysis (OPLRA) model is used from literature to tackle the problem of segmented analysis. This model is a mathematical programming approach that is formulated as a mixed integer linear programming problem that optimally partitions the data into multiple regions and calculates the regression coefficients, while minimising the Mean Absolute Error of the fitting. However, the number of regions is a known priori. For this work, an extension of the model is proposed that can optimally decide on the number of regions using information criteria. Specifically, the Akaike Information Criterion is used and the objective is to minimise its value. By using the criterion, the model no longer needs a heuristic approach to decide on the number of regions and it also deals with the problem of overfitting and model complexity.
    [Show full text]