Bayesian-Based Instance Weighting Techniques for Instance-Based Learners

摘要

Instance-Based learners such as the kNN algorithm classify a new instance based on the k most similar instances. Usually these instances have equal weights or votes. Some systems assign them weights that are inversely proportional to their distance from the new instance. In this work, we present several Bayesian-based instance weighting technique that are more suitable for noisy data sets. We use the Naive Bayesian probability that an instance truly belongs to its class or does not belong to another class, to calculate its weight. Our empirical study shows that these weighting techniques make the kNN algorithm far less sensitive to noisy training data.