Accuracy¶
Importance of Accurate Training Data¶
When training a machine learning model, care must be taken to ensure that the dataset being used is a quality one. Indeed, the concept of "garbage-in, garbage-out" is highly applicable to models trained on poor a low-quality dataset - a model is only as good as the data it was trained on.
Consider a hypothetical machine learning model which is able to achieve perfect accuracy on any dataset it's given, without overfitting. Now consider what would happen if such a model were trained on two different datasets of vastly different quality - for example, electronic energies calculated using the Local Density Approximation 1 versus electronic energies calculated with a Coupled Cluster 2 technique. Despite performing perfectly, the model would only ever be as accurate as the underlying data used to train it.
Common Accuracy Metrics¶
Assessing model performance is a topic which has received significant attention, and there is as a result no shortage of methods by-which a model's performance can be evaluated. We present a few more-common metrics below.
Regression¶
- The Root Mean Squared Error is the square root of the average of the squared difference between the training data and the model's predictions.
- The Mean Absolute Error is the average of the absolute difference between the training data and the model's predictions.
Classification¶
- The Confusion Matrix can be used to derive a plethora of error metrics 3, including the true positive rate, the false positive rate, precision, recall, and many others.