metrics Module

Metrics that can be used to evaluate the performance of learners.

author:

Nitin Madnani (nmadnani@ets.org)

author:

Michael Heilman (mheilman@ets.org)

author:

Dan Blanchard (dblanchard@ets.org)

organization:

ETS

skll.metrics.correlation(y_true, y_pred, corr_type='pearson')[source]

Calculate given correlation type between y_true and y_pred.

y_pred can be multi-dimensional. If y_pred is 1-dimensional, it may either contain probabilities, most-likely classification labels, or regressor predictions. In that case, we simply return the correlation between y_true and y_pred. If y_pred is multi-dimensional, it contains probabilties for multiple classes in which case, we infer the most likely labels and then compute the correlation between those and y_true.

Parameters:

  • y_true (numpy.ndarray) – The true/actual/gold labels for the data.

  • y_pred (numpy.ndarray) – The predicted/observed labels for the data.

  • corr_type (str, default="pearson") – Which type of correlation to compute. Possible choices are “pearson”, “spearman”, and “kendall_tau”.

Returns:

correlation value if well-defined, else 0.0

Return type:

float

skll.metrics.f1_score_least_frequent(y_true, y_pred)[source]

Calculate F1 score of the least frequent label/class.

Parameters:

  • y_true (numpy.ndarray) – The true/actual/gold labels for the data.

  • y_pred (numpy.ndarray) – The predicted/observed labels for the data.

Returns:

F1 score of the least frequent label.

Return type:

float

skll.metrics.kappa(y_true, y_pred, weights=None, allow_off_by_one=False)[source]

Calculate the kappa inter-rater agreement.

The agreement is calculated between the gold standard and the predicted ratings. Potential values range from -1 (representing complete disagreement) to 1 (representing complete agreement). A kappa value of 0 is expected if all agreement is due to chance.

In the course of calculating kappa, all items in y_true and y_pred will first be converted to floats and then rounded to integers.

It is assumed that y_true and y_pred contain the complete range of possible ratings.

This function contains a combination of code from yorchopolis’s kappa-stats and Ben Hamner’s Metrics projects on Github.

Parameters:

  • y_true (numpy.ndarray) – The true/actual/gold labels for the data.

  • y_pred (numpy.ndarray) – The predicted/observed labels for the data.

  • weights (Optional[Union[str, numpy.ndarray]], default=None) – Specifies the weight matrix for the calculation. Possible values are: None (unweighted-kappa), "quadratic" (quadratically weighted kappa), "linear" (linearly weighted kappa), and a two-dimensional numpy array (a custom matrix of weights). Each weight in this array corresponds to the \(w_{ij}\) values in the Wikipedia description of how to calculate weighted Cohen’s kappa.

  • allow_off_by_one (bool, default=False) – If true, ratings that are off by one are counted as equal, and all other differences are reduced by one. For example, 1 and 2 will be considered to be equal, whereas 1 and 3 will have a difference of 1 for when building the weights matrix.

Returns:

The weighted or unweighted kappa score.

Return type:

float

Raises:
skll.metrics.register_custom_metric(custom_metric_path, custom_metric_name)[source]

Import, load, and register the custom metric function from the given path.

Parameters:

  • custom_metric_path (skll.types.PathOrStr) – The path to a custom metric.

  • custom_metric_name (str) – The name of the custom metric function to load. This function must take only two array-like arguments: the true labels and the predictions, in that order.

Raises:

  • ValueError – If the custom metric path does not end in ‘.py’.

  • NameError – If the name of the custom metric file conflicts with an already existing attribute in skll.metrics or if the custom metric name conflicts with a scikit-learn or SKLL metric.

skll.metrics.use_score_func(func_name, y_true, y_pred)[source]

Call the given scoring function.

This takes care of handling keyword arguments that were pre-specified when creating the scorer. This applies any sign-flipping that was specified by make_scorer() when the scorer was created.

Parameters:

  • func_name (str) – The name of the objective function to use.

  • y_true (numpy.ndarray) – The true/actual/gold labels for the data.

  • y_pred (numpy.ndarray) – The predicted/observed labels for the data.

Returns:

The scored result from the given scorer.

Return type:

float