Cross Validation#

Cross Validation (CV) is a technique for assessing the generalization performance of a model using data it has never seen before. The validation score gives us a sense for how well the model will perform in the real world. In addition, it allows the user to identify problems such as underfitting, overfitting, and selection bias which are discussed in the last section.


Cross validation Metrics are used to score the predictions made by an Estimator with respect to their known ground-truth labels. There are different metrics for different types of problems as shown in the table below.

Note: All metrics follow the schema that higher scores are better - thus, common loss functions such as Mean Squared Error and RMSE are given as their negative to conform to this schema.

Classification and Anomaly Detection#

Metric Range
Accuracy [0, 1]
F Beta [0, 1]
Informedness [-1, 1]
MCC [-1, 1]


Metric Range
Mean Absolute Error [-∞, 0]
Mean Squared Error [-∞, 0]
Median Absolute Error [-∞, 0]
R Squared [-∞, 1]
RMSE [-∞, 0]
SMAPE [-100, 0]


Metric Range
Completeness [0, 1]
Homogeneity [0, 1]
Rand Index [-1, 1]
V Measure [0, 1]

To return a validation score from a Metric pass the predictions and labels to the score() method like in the example below.

use Rubix\ML\CrossValidation\Metrics\Accuracy;

$metric = new Accuracy();

$score = $metric->score($predictions, $labels);



Metrics can be used stand-alone or they can be used within a Validator object as the scoring function. Validators automate the cross validation process by training and testing a learner on different subsets of a master dataset. The way in which subsets are chosen depends on the algorithm employed under the hood. Most validators implement the Parallel interface which allows multiple tests to be run at the same time on multiple CPU cores.

Validator Test Coverage Parallel
Hold Out Partial
K Fold Full
Leave P Out Full
Monte Carlo Partial

For example, the K Fold validator automatically selects one of k folds of the dataset to use as a validation set and then uses the rest of the folds to train the learner. It will do this until the learner is trained and tested on every sample in the dataset at least once. The final score is then an average of the k validation scores returned by each test. To begin, pass an untrained Learner, a Labeled dataset, and the chosen validation metric to the validator's test() method.

use Rubix\ML\CrossValidation\KFold;
use Rubix\ML\Datasets\Labeled;
use Rubix\ML\CrossValidation\Metrics\FBeta;

$validator = new KFold(10);

$dataset = new Labeled($samples, $labels);

$score = $validator->test($estimator, $dataset, new FBeta());



Cross validation Reports give you a deeper sense for how well a particular model performs with fine-grained information. The generate() method takes a set of predictions and their corresponding ground-truth labels and returns an associative array (i.e. dictionary or map) filled with useful statistics.

Classification and Anomaly Detection#



For example, the Error Analysis report computes a variety of regression metrics.

use Rubix\ML\CrossValidation\Reports\ErrorAnalysis;

$report = new ErrorAnalysis();

$result = $report->generate($predictions, $labels);

array(18) {
  ["mean_absolute_error"]=> float(0.8)
  ["median_absolute_error"]=> float(1)
  ["mean_squared_error"]=> float(1)
  ["mean_absolute_percentage_error"]=> float(14.020774976657)
  ["rms_error"]=> float(1)
  ["mean_squared_log_error"]=> float(0.019107097505647)
  ["r_squared"]=> float(0.99589305515627)
  ["error_mean"]=> float(-0.2)
  ["error_midrange"]=> float(-0.5)
  ["error_median"]=> float(0)
  ["error_variance"]=> float(0.96)
  ["error_mad"]=> float(1)
  ["error_iqr"]=> float(2)
  ["error_skewness"]=> float(-0.22963966338592)
  ["error_kurtosis"]=> float(-1.0520833333333)
  ["error_min"]=> int(-2)
  ["error_max"]=> int(1)
  ["cardinality"]=> int(10)

Common Problems#

Here are some common problems that cross validation can help you identify.


A poorly performing model can sometimes be explained as underfiting the training data - a condition in which the learner is unable to capture the underlying pattern or trend given the model constraints. Underfitting mostly occurs when a simple model is chosen to represent data that is complex and non-linear. Adding more features can help with underfitting, however if the problem is too severe, a more flexible learner can be chosen for the task instead.


When a model performs well on training data but poorly during cross validation it could be that the model has overfit the training data. Overfitting occurs when the model conforms too closely to the training set data and therefore fails to generalize well to new data or make predictions reliably. Some degree of overfitting can occur with any model, but more flexible models are more prone to overfitting due to their ability to memorize individual samples. Most learners employ strategies such as regularization, early stopping, and/or pruning to control overfitting. Adding more samples to the dataset can also help.

Selection Bias#

When a model performs well on certain samples but poorly on others it could be that the learner was trained with a dataset that exhibits selection bias. Selection bias is the bias introduced when a population is disproportionally represented in a dataset. For example, if a learner is trained to classify pictures of cats and dogs with mostly (say 90%) images of cats, it will likely have difficulty in the real world where cats and dogs are more equally represented.