On the Trade-Off Between Iterative Classification and Collective Classification: First Experimental Results

Abstract. There have been two major approaches for classification of networked (linked) data. Local approaches (iterative classification) learn a model locally without considering unlabeled data and apply the model iteratively to classify unlabeled data. Global approaches (collective classification), on the other hand, exploit unlabeled data and the links occurring between labeled and unlabeled data for learning. Naturally, global approaches are computationally more demanding than local ones. Moreover, for large data sets, approximate inference has to be performed to make computations feasible. In the present work, we investigate the benefits of collective classification based on global probabilistic models over local approaches. Our experimental results show that global approaches do not always outperform local approaches with respect to the classification accuracy. More precisely, the results suggest that global approaches considerably outperform local approaches only for low ratios of labeled data.