This document discusses cost-sensitive multiclass classification using one-versus-one comparisons. It notes that traditional classification aims to minimize future errors, but some applications require treating different error types differently based on their costs. It then presents a tool that extends regular classification algorithms to handle cost-sensitive problems, and proposes a novel cost-sensitive classification method by coupling this tool with a one-versus-one algorithm. Promising experimental results are demonstrated for this new method.
Title: Cost-sensitive Multiclass Classification Using One-versus ...
1. Title: Cost-sensitive Multiclass Classification Using One-versus-oneComparisonsSpeaker: 台大資工 林軒田Classification is an important problem in machine learning. It can be usedin a variety of applications, such as separating apples, oranges, andbananas automatically. Traditionally, the regular classification setup aimsat minimizing the number of future mis-prediction errors. Nevertheless, insome applications, it is needed to treat different types of mis-predictionerrorsdifferently. For instance, a false-negative prediction for a spamclassification system only takes the user an extra secondto delete theemail, while a false-positive prediction can mean a huge loss when the emailactually carries important information. When recommending movies to asubscriber with preference ``romance over action over horror'', the cost ofmis-predicting a romance movie as a horror one should be significantlyhigher than the cost of mis-predicting the movie as an action one. Suchneeds can be formalized as the cost-sensitive classification setup, which isdrawing much research attention because of its many potential applications,including targeted marketing, fraud detection and web analysis.Because regular classification is a well-studied setup, there are many goodregular classification algorithms. In this talk, we first present a toolthat systematically extend those algorithms to deal with cost-sensitiveclassification problems. Then, by coupling the tool with the popularone-versus-one regular classification algorithm, we propose a simple andnovel cost-sensitive classification method. Finally, we demonstrate somestrong theoretical guarantees and some promising experimental results thatcome with our proposed method.The talk is self-contained and assumes no prior knowledge in machinelearning.<br />