From 33fe808985bc69552073ca5aed6dd466a44dec0b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Frank=20K=C3=B6hne?= Date: Mon, 2 Nov 2020 12:45:10 +0100 Subject: [PATCH] Clarify goal-setting, line 54. --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index ca04fc3..cdb5151 100644 --- a/README.md +++ b/README.md @@ -51,7 +51,8 @@ Relevant questions are: Not only can you use such insights to prioritize automation efforts, but you will also need to rely on them while implementing automated decisions with machine learning approaches: They become part of the *loss function*. -An example: In input management an AI classifies incoming business documents and forwards them to the responsible processes. The machine learning model to accomplish this willfro the history of manual classifications and try to classify them in the same way. Every machine learning needs a target key metric, which has to be optimized. The classic approach for classification problems of this kind is to optimize a kind of hit rate of correct to incorrect classifications. Although this works, it systematically wastes potential savings - in the end, the goal is incompletely defined. If we want to try to avoid expensive processes, this goal must become part of the target key metric, because in input management not all misclassifications are equally expensive. +An example: In input management an AI classifies incoming business documents and forwards them to the responsible processes. The machine learning model to accomplish this will learn from the history of manual classifications and try to classify them in the same way. +Every machine learning needs a target key metric, which has to be optimized. The classic approach for classification problems of this kind is to optimize a kind of _hit rate_ (accuracy) of correct to incorrect classifications. Although this works, it systematically wastes potential savings - in the end, the goal is incompletely defined. We not only want to approximate past decision making. We want to avoid expensive processes and errors. These goals must become part of the target key metric, because in input management not all misclassifications are equally expensive: * To wrongly categorize an incoming new customer contract as a termination is the biggest possible mistake. * Classifying a termination as an incoming new customer contract is a pity but has hardly any economic consequences.