Metrics and Why They Are Important

Metrics are the means by which performance of a machine learning model is measured. Establishing well-defined metrics is arguably the most important and frequently overlooked step in framing a machine learning problem.

The organizers of PLAsTiCC put significant effort in establishing metrics. They publishing a paper [1] on just that topic.

In the end they chose probabilistic classification as the metric, which means the model must assign a likelihood score for each of the class values. This differs from the more typical MAP (maximum a-priori) classification where the model would predict the most likely class. The reason given by the authors is that, in star research, lesser-probable classes might warrent investigation if that class is interesting. Specifically "an object with even a moderate probability of being of a very rare class could be worth the risk (of further investigation)".

In other words - probabilistic classification provides richer data to help guide further study of individual objects.

For our web application we need to make this tractable for the average user. To do so - we pose this as a betting game where the player has $100 and must place their bets on the N classes of star. They are free to go "all in" on one class, or spread the money around, but the bet must sum to 100. That is consistent with the approach defined by the metric (the value of the user's bet is their probabilistic prediction). This will be consistent with the output of the models from Trotta and Boone that were developed based on that metric.



  • (1)