scores_metrics
Machine Learning in Climate Sciences University of Tübingen
Jakob Schloer
-
Quantile score
-
Cross validation error
- Leave one out error
-
Correlation coefficients
-
Wasserstein metric
Comparing Time series:
-
Correlations
-
Eucledian Distance
-
Dynamic Time Warping: DTW finds an optimal match between two sequences of feature vectors which allows for stretched and compressed sections of the sequence.
-
Mutual Information: Entropy based metric, introduced by Shannon. Applied to time series, there are quite a few papers by now. For example (this)[https://arxiv.org/abs/0904.4753].
-
iSAX: The final one I want to flag is the so-called “Motif Discovery” and the related (iSAX)[http://www.cs.ucr.edu/~eamonn/iSAX.pdf] representation of time series (by Eamon Keogh), which is very scalable.
The errors are on the same scale as the data, i.e. different data sets cannot be compared
Mean absolute error (MAE)
$$ MAE = mean(|y_i - \hat{y}i|) = \frac{1}{N} \sum{i=1}^{N} |y_i - \hat{y}_i| $$
Minimizing MAE leads to prediction of median.
Mean square error (MSE)
Strongly penalizes large wrong predictions
Root mean square error (RMSE)
Minimizing RMSE leads to prediction of mean
Percentage errors are unit free and therefore allow comparison of different data sets.
Mean absolute percentage error (MAPE)
Problems:
-
cannot be used if there are zero values
-
puts more weight on negative errors
Symmetric absolute percentage error (sMAPE) Used to overcome the problems of MAPE
Alternative to percentage errors when comparing different datasets
Mean absolute scaled error (MASE)
-
scale invariance
-
symmetric
-
less than one if it arises from a better forecast than the average naïve forecast and conversely it is greater than one if the forecast is worse than the average naïve forecast
https://otexts.com/fpp2/accuracy.html https://scikit-learn.org/stable/modules/model_evaluation.html#
Check if a hypothesis is correct. Often referred to as explained variance scores.
Coefficient of determination ((R^2)) Describes the variance (of y) which is explained by the model prediction.
F-test The F-value expresses how much of the model has improved compared to the mean (null hypothesis) given the variance of the model and data. The F-test is obtained by
where (Y) is the set of data points, (\hat{Y} = \left[ \hat{y}_{i} \right]) with (i = 1,.., N) are a set of predicted points.
Chi-square test
Leave-one out error (LOOE)
Pearson correlation Pearson correlation measures how two continuous signals co-vary over time. The linear relationship between these signals are given from -1 (anticorrelated) to 0 (incorrelated) to 1 (perfecly correlated).
The Pearson correlation coefficient for two random variables (X_1) and (X_2) is:
For time-series on can calculate a
-
global correlation coefficient: a single value
-
local correlation coefficient: determine correlation in a rolling window over time
Caution:
-
outliers can skew the correlation
-
assuming the data is homoscedatic, i.e. constant variances
Time Lagged Cross Correlation (TLCC) TLCC is a measure of similarity of two series as a function of displacement. It captures directionality between two signals, i.e. leader-follower relationship. Idea: Similar to convolution of two signals, i.e. shifting one signal with respect to the other while repeatedly calculating the correlation.
0.5
0.5
Windowed time lagged cross correlations (WTLCC) are an extension of TLCC where local correlations coefficients are computed for each lag-time which is then plotted as a matrix.
Granger causality
Dynamic Time Wrapping (DTW)
-
saliency maps: highlights which changes in the input would most affect the output.
-
heat maps: highlights which inputs are most important for the prediction
Saliency maps