Sep 9, 2018 · 1 min read
Baseline accuracy would be 88.7% by simply predicting 0 in every case, therefore accuracy increasing to ~90% is not much improvement. I feel accuracy is not the appropriate measure here, for instance if this is a marketing problem you need to apply cost/benefits of each marketing piece vs benefit for correctly identifying purchaser. In that case you can decide on how to balance precision/recall. An AUC of .59 is not much lift, which I think would be a more appropriate a metric than accuracy on a simple 50% threshold.
