Supervised Lazy Random Walk Classifier

Incorporating the k-nearest neighbor information and lazy random walks on graph, this paper presents a supervised classifier, namely supervised lazy random walk (SLRW) classifier. First, a partially labeled graph is built over the input data, where the edge weight represents the locally scaled pair wise similarity based on the k-nearest neighbors. And then the SLRW classifier is trained with lazy random walk technique so as to predict the labels of test data. This study brings random walk technology into classification problem without unlabeled training data and enriches its application into supervised learning area. Tests conducted over the real data sets demonstrate the effective robustness of our model to noise and comparisons to other classifiers indicate its excellent classification performance.