This article needs additional citations for verification. (August 2017) (Learn how and when to remove this template message)
In statistics, overfitting is "the production of an analysis that corresponds too closely or exactly to a particular set of data, and may therefore fail to fit additional data or predict future observations reliably". An overfitted model is a statistical model that contains more parameters than can be justified by the data. The essence of overfitting is to have unknowingly extracted some of the residual variation (i.e. the noise) as if that variation represented underlying model structure.:45
Underfitting occurs when a statistical model cannot adequately capture the underlying structure of the data. An underfitted model is a model where some parameters or terms that would appear in a correctly specified model are missing. Underfitting would occur, for example, when fitting a linear model to non-linear data. Such a model will tend to have poor predictive performance.
Overfitting and underfitting can occur in machine learning, in particular. In machine learning, the phenomena are sometimes called "overtraining" and "undertraining".
The possibility of overfitting exists because the criterion used for selecting the model is not the same as the criterion used to judge the suitability of a model. For example, a model might be selected by maximizing its performance on some set of training data, and yet its suitability might be determined by its ability to perform well on unseen data; then overfitting occurs when a model begins to "memorize" training data rather than "learning" to generalize from a trend.
As an extreme example, if the number of parameters is the same as or greater than the number of observations, a simple model can perfectly predict the training data simply by memorizing the data in its entirety. (For an illustration, see Figure 2.) Such a model, though, will typically fail severely when making predictions.
The potential for overfitting depends not only on the number of parameters and data but also the conformability of the model structure with the data shape, and the magnitude of model error compared to the expected level of noise or error in the data. Even when the fitted model does not have an excessive number of parameters, it is to be expected that the fitted relationship will appear to perform less well on a new data set than on the data set used for fitting (a phenomenon sometimes known as shrinkage). In particular, the value of the coefficient of determination will shrink relative to the original data.
To lessen the chance of, or amount of, overfitting, several techniques are available (e.g. model comparison, cross-validation, regularization, early stopping, pruning, Bayesian priors, or dropout). The basis of some techniques is either (1) to explicitly penalize overly complex models or (2) to test the model's ability to generalize by evaluating its performance on a set of data not used for training, which is assumed to approximate the typical unseen data that a model will encounter.
This section needs expansion. You can help by adding to it. (October 2017)
In statistics, an inference is drawn from a statistical model, which has been selected via some procedure. Burnham & Anderson, in their much-cited text on model selection, argue that to avoid overfitting, we should adhere to the "Principle of Parsimony". The authors also state the following.:32-33
Overfitted models … are often free of bias in the parameter estimators, but have estimated (and actual) sampling variances that are needlessly large (the precision of the estimators is poor, relative to what could have been accomplished with a more parsimonious model). Spurious treatment effects tend to be identified, and spurious variables are included with overfitted models. … A best approximating model is achieved by properly balancing the errors of underfitting and overfitting.
Overfitting is more likely to be a serious concern when there is little theory is available to guide the analysis, in part because then there tend to be a large number of models to select from. The book Model Selection and Model Averaging (2008) puts it this way.
Given a data set, you can fit thousands of models at the push of a button, but how do you choose the best? With so many candidate models, overfitting is a real danger. Is the monkey who typed Hamlet actually a good writer?
In regression, overfitting occurs frequently. In the extreme case, if there are p variables in a linear regression with p data points, the fitted line will go exactly through every point. A recent study suggests that two observations per independent variable are sufficient for linear regression. For logistic regression or Cox proportional hazards models, there are a variety of rules of thumb (e.g. 5-9, 10 and 10-15 — the guideline of 10 observations per independent variable is known as the "one in ten rule"). In the process of regression model selection, the mean squared error of the random regression function can be split into random noise, approximation bias, and variance in the estimate of regression function, and bias–variance tradeoff is often used to overcome overfit models.
Usually a learning algorithm is trained using some set of "training data": exemplary situations for which the desired output is known. The goal is that the algorithm will also perform well on predicting the output when fed "validation data" that was not encountered during its training.
Overfitting is the use of models or procedures that violate Occam's razor, for example by including more adjustable parameters than are ultimately optimal, or by using a more complicated approach than is ultimately optimal. For an example where there are too many adjustable parameters, consider a dataset where training data for y can be adequately predicted by a linear function of two dependent variables. Such a function requires only three parameters (the intercept and two slopes). Replacing this simple function with a new, more complex quadratic function, or with a new, more complex linear function on more than two dependent variables, carries a risk: Occam's razor implies that any given complex function is a priori less probable than any given simple function. If the new, more complicated function is selected instead of the simple function, and if there was not a large enough gain in training-data fit to offset the complexity increase, then the new complex function "overfits" the data, and the complex overfitted function will likely perform worse than the simpler function on validation data outside the training dataset, even though the complex function performed as well, or perhaps even better, on the training dataset.
When comparing different types of models, complexity cannot be measured solely by counting how many parameters exist in each model; the expressivity of each parameter must be considered as well. For example, it is nontrivial to directly compare the complexity of a neural net (which can track curvilinear relationships) with m parameters to a regression model with n parameters.
Overfitting is especially likely in cases where learning was performed too long or where training examples are rare, causing the learner to adjust to very specific random features of the training data, that have no causal relation to the target function. In this process of overfitting, the performance on the training examples still increases while the performance on unseen data becomes worse.
As a simple example, consider a database of retail purchases that includes the item bought, the purchaser, and the date and time of purchase. It's easy to construct a model that will fit the training set perfectly by using the date and time of purchase to predict the other attributes; but this model will not generalize at all to new data, because those past times will never occur again.
Generally, a learning algorithm is said to overfit relative to a simpler one if it is more accurate in fitting known data (hindsight) but less accurate in predicting new data (foresight). One can intuitively understand overfitting from the fact that information from all past experience can be divided into two groups: information that is relevant for the future and irrelevant information ("noise"). Everything else being equal, the more difficult a criterion is to predict (i.e., the higher its uncertainty), the more noise exists in past information that needs to be ignored. The problem is determining which part to ignore. A learning algorithm that can reduce the chance of fitting noise is called robust.
The most obvious consequence of overfitting is poor performance on the validation dataset. Other negative consequences include:
- A function that is overfitted is likely to request more information about each item in the validation dataset than does the optimal function; gathering this additional unneeded data can be expensive or error-prone, especially if each individual piece of information must be gathered by human observation and manual data-entry.
- A more complex, overfitted function is likely to be less portable than a simple one. At one extreme, a one-variable linear regression is so portable that, if necessary, it could even be done by hand. At the other extreme are models that can be reproduced only by exactly duplicating the original modeler's entire setup, making reuse or scientific reproduction difficult.
Underfitting occurs when a statistical model or machine learning algorithm cannot adequately capture the underlying structure of the data. It occurs when the model or algorithm does not fit the data enough. Underfitting occurs if the model or algorithm shows low variance but high bias (to contrast the opposite, overfitting from high variance and low bias). It is often a result of an excessively simple model.
Burnham & Anderson state the following.:32
… an underfitted model would ignore some important replicable (i.e., conceptually replicable in most other samples) structure in the data and thus fail to identify effects that were actually supported by the data. In this case, bias in the parameter estimators is often substantial, and the sampling variance is underestimated, both factors resulting in poor confidence interval coverage. Underfitted models tend to miss important treatment effects in experimental settings.
- Leinweber, D. J. (2007). "Stupid Data Miner Tricks". The Journal of Investing. 16: 15–22. doi:10.3905/joi.2007.681820.
- Tetko, I. V.; Livingstone, D. J.; Luik, A. I. (1995). "Neural network studies. 1. Comparison of Overfitting and Overtraining" (PDF). J. Chem. Inf. Comput. Sci. 35 (5): 826–833. doi:10.1021/ci00027a006.
- Definition of "overfitting" at OxfordDictionaries.com: this definition is specifically for Statistics.
- Everitt B.S., Skrondal A. (2010), Cambridge Dictionary of Statistics, Cambridge University Press.
- Burnham, K. P.; Anderson, D. R. (2002), Model Selection and Multimodel Inference (2nd ed.), Springer-Verlag. (This has over 38000 citations on Google Scholar.)
- Claeskens, G.; Hjort, N.L. (2008), Model Selection and Model Averaging, Cambridge University Press.
- Harrell, F. E., Jr. (2001), Regression Modeling Strategies, Springer.
- Martha K. Smith (2014-06-13). "Overfitting". University of Texas at Austin. Retrieved 2016-07-31.
- Austin, P. C.; Steyerberg, E. W. (2015). "The number of subjects per variable required in linear regression analyses". Journal of Clinical Epidemiology. 68 (6): 627–636. doi:10.1016/j.jclinepi.2014.12.014.
- Vittinghoff, E.; McCulloch, C. E. (2007). "Relaxing the Rule of Ten Events per Variable in Logistic and Cox Regression". American Journal of Epidemiology. 165 (6): 710–718. doi:10.1093/aje/kwk052.
- Draper, Norman R.; Smith, Harry (1998). Applied Regression Analysis (3rd ed.). Wiley. ISBN 978-0471170822.
- Jim Frost (2015-09-03). "The Danger of Overfitting Regression Models". Retrieved 2016-07-31.
- Hawkins, Douglas M. (2004), "The problem of overfitting", Journal of Chemical Information and Modeling, 44.1: 1-12.
- Cai, Eric (2014-03-20). "Machine Learning Lesson of the Day – Overfitting and Underfitting". StatBlogs.
- Overfitting: when accuracy measure goes wrong - an introductory video tutorial.
- The Problem of Overfitting Data
- CSE546: Linear Regression Bias / Variance Tradeoff