What does a trained machine learning model measure using a test dataset?

A trained machine learning model evaluates its ability to generalize to new data using a test dataset. The test dataset consists of examples that the model has not encountered during the training phase. Therefore, it provides an unbiased assessment of how well the model can predict outcomes for unseen data, which is crucial for understanding its performance in real-world applications.

Generalization refers to the model’s capability to perform well not just on the training data but also on new, unseen examples. This is significant because a model that fits the training data too closely may not perform adequately on other datasets, a phenomenon known as overfitting.

In contrast, accuracy on seen data pertains to how well the model performs on the data it was trained on, which does not provide insights into its application to new situations. Computational efficiency and training time are metrics related to performance and resource usage during model building, rather than its predictive capabilities on new data. Thus, focusing on how well the model generalizes is essential for evaluating its effectiveness in practical scenarios.