How to pick the best threshold_match_probability through ROC Curve?

[yangg1224]

I am using ground truth table to plot the ROC curve, using
linker.roc_chart_from_labels_table("combined_GT")

Here is the ROC curve:

The screenshot show the top left corner point.

linker = DuckDBLinker([AccessOAP,OACIS])
linker.load_model("model\saved_model_link.json")
results = linker.predict(threshold_match_probability=0.95)

Please help how to pick the best threshold_match_probability through the ROC curve...
results = linker.predict(threshold_match_probability= ? )

[RobinL]

There isn't really a single answer to this question I'm afraid - it really depends on your users preference for false positives vs false negatives. In general, there's no reason to assume it would be the top left point (although it's often thereabouts)

[yangg1224]

Thanks for your reply, I wonder does the Match_Probability in the screenshot (0.0115%) is the value of threshold_match_probability, if I pick the top left corner point?