Download [PORTABLE] Multivariate Statistical Machine Learning Methods For Genomic Prediction Pdf

The overfitting phenomenon happens when a statistical machine learning model learns very well about the noise as well as the signal that is present in the training data. On the other hand, an underfitted phenomenon occurs when only a few predictors are included in the statistical machine learning model that represents the complete structure of the data pattern poorly. This problem also arises when the training data set is too small and thus an underfitted model does a poor job of fitting the training data and unsatisfactorily predicts new data points. This chapter describes the importance of the trade-off between prediction accuracy and model interpretability, as well as the difference between explanatory and predictive modeling: Explanatory modeling minimizes bias, whereas predictive modeling seeks to minimize the combination of bias and estimation variance. We assess the importance and different methods of cross-validation as well as the importance and strategies of tuning that are key to the successful use of some statistical machine learning methods. We explain the most important metrics for evaluating the prediction performance for continuous, binary, categorical, and count response variables.

Download Multivariate Statistical Machine Learning Methods for Genomic Prediction pdf

On the other hand, an underfitted phenomenon occurs when few predictors are included in the statistical machine learning model, i.e., it is a very simple model that poorly represents the complete picture of the predominant data pattern. This problem also arises when the training data set is too small or not representative of the population data. An underfitted model does a poor job of fitting the training data and for this reason it is not expected to satisfactorily predict new data points. This implies that the predictions using unseen data are weak, since individuals are perceived as strangers unfamiliar with the training data set.

The paradox of overfitting is defined as complex models that contain more information about the training data, but less information about the testing data (future data we want to predict). In statistical machine learning, overfitting is a major issue and leads to some serious problems in research: (a) some relationships that seem statistically significant are only noise, (b) the complexity of the statistical machine learning model is very large for the amount of data provided, and (c) the model in general is not replicable and predicts poorly.

Since the main goal of developing and implementing statistical machine learning methods is to predict unseen data not used for training the statistical machine learning algorithm, researchers are mainly interested in minimizing the testing error (generalization error applicable to future samples) instead of minimizing the training error that is applicable to the observed data used for training the statistical machine learning algorithm.

Accuracy is the ability of a statistical machine learning model to make correct predictions and those models with more complexity (called flexible models) are better in terms of accuracy, while the simple, less complex models (called inflexible models) are less accurate but more interpretable. Interpretability indicates to what degree the model allows for human understanding of natural phenomena. For these reasons, when the goal of the study is prediction, flexible models should be used; however, when the goal of the study is inference, inflexible models are more appropriate because they more easily interpret the relationship between the response variables and the predictor variables. As the complexity of the statistical machine learning model increases, the bias is reduced and the variance increases. For this reason, when more parameters are included in the statistical machine learning model, the complexity of the model increases and the variance becomes the main concern while the bias steadily falls. For example, James et al. (2013) state that the linear regression model is a relatively inflexible method because it only generates linear functions, while the support vector machine method is one of the most flexible statistical machine learning methods.

The single hold-out set or validation set approach consists of randomly dividing the available data set into a training set and a validation or hold-out set (Fig. 4.3). The statistical machine learning model is trained with the training set while the hold-out set (testing set) is used to study how well that statistical machine learning model performs on unseen data. For example, 80% of the data can be used for training the model and the remaining 20% of the data for testing it. One weakness of the hold-out (validation) set approach is that it depends on just one training-testing split and its performance depends on how the data are split into the training and testing sets.

Schematic representation of the hold-out set approach. A set of observations are randomly split into a training set with individuals I40, I5, I82, among others, and into a testing set with observations I45, I88, among others. The statistical machine learning model is fitted on the training set and its performance is evaluated on the validation set (James et al. 2013)

It is important to point out that to reduce variability, we recommend implementing the k-fold CV multiple times, each time using different complementary subsets to form the folds; the validation results are combined (e.g., averaged) over the rounds (times) to give a better estimate of the statistical machine learning model predictive performance.

Learning curves (LC) are considered effective tools to monitor the performance of the employee exposed to a new task. LCs provide a mathematical representation of the learning process that takes place as the task is repeated. In statistical machine learning the LC is a line plot of learning (y-axis) over experience (x-axis). Learning curves are extensively used in statistical machine learning for algorithms that learn (their parameters) incrementally over time, such as deep neural networks. In general, there is considerable empirical evidence suggesting that five- or ten-fold cross-validation should be preferred to LOO.

It is important to highlight that when the data set is considerably large, it is better to randomly split it into three parts: a training set, a validation set (or tuning set), and a testing set. The training set and testing set are used as explained before, while the validation (tuning set) set is used to estimate the prediction error for model selection, which is the process of estimating the performance of different models in order to choose the best one, or to evaluate the chosen statistical machine learning model with a range of values of tuning hyperparameters to select the combination of hyperparameters with the best prediction performance and then use these hyperparameters (or best model) to evaluate the prediction performance in the testing set (Fig. 4.7). It is important to point out that Fig. 4.7 shows only one random split of the data in terms of the training, testing, and validation sets.

There are many ways of searching for the best hyperparameters. However, a general approach defines a set of candidate values for each hyperparameter. Each value of this set of candidate values is then applied with a resample of the training set of the chosen statistical machine learning method, where we aggregate all the hold-out predictions from which the best hyperparameters are chosen and refit the model with the entire set (Kuhn and Johnson 2013). A schematic representation of the tuning process proposed by Kuhn and Johnson (2013) is given in Fig. 4.9. It is important to highlight that this process should be performed correctly because when the same data are used for training and evaluating the prediction performance, the prediction performance obtained is extremely optimistic.

For example, suppose a breeder is interested in developing an algorithm to classify unseen plants as diseased or not diseased with an available training data set. The goal is to minimize the rate of misclassification or to maximize the percentage of cases correctly classified (PCCC). Also, assume that you are new to the world of statistical machine learning and that you only understand the k-nearest neighbor method. Since this algorithm depends only on the hyperparameter called the number of neighbors (k), the question is which value of k to choose in such a way that the prediction performance of this algorithm will be the best in the sample prediction of plants. To find the best value of the k hyperparameter, you must specify a range of values for k (for example, from 1 to 60 with increments of 1), then with a part of your training data set, called the training-inner (or tuning that corresponds to the training data in the inner loop) set, which is randomly selected. You proceed to evaluate the 60 values of k with the k-nearest neighbor method and evaluate the prediction performance in the remaining part of the training set (validation set). Next, you select the value of k from this range of values that best predicts (according, for example, to the PCCC) out-of-sample data (validation set) and use this value to perform the prediction of the unseen plants not used for training the model (testing set). This is a widely adopted practice that consists of searching for the parameter (usually through brute force loops) that yields the best performance over a validation set. However, the process illustrated here is very simple because the k-nearest neighbor model only depends on a unique hyperparameter; however, there are other statistical machine learning algorithms (for example, deep learning methods) where the tuning process is required for a considerable amount of hyperparameters. For this reason, we encourage caution when choosing the statistical machine learning algorithm, since the amount of work required for performing the tuning process depends on the chosen method. 041b061a72