Fusion Evaluation¶
Fusion results aren’t easily evaluated like other machine learning methods, as there is not usually labeled training data to measure accuracy or other common metrics.
However, we do have available the features used to match the records. As we want the matches to be as accurate as possible, evaluating how these values compare for the matched records can give an good indication of how reasonable the fusion is.
We can also look at how well donated target variables retain their distributions. Recall, a main goal of data fusion is to donate information and treat it as though it were measured. Since the datasets being fused represent the same population, the donated data should resemble as closely as possible the original data that was measured. Comparing the target variable distributions pre and post fusion can give an indication of the fusion’s effectiveness.