Web page and image semi-supervised classification with heterogeneous information fusion

Abstract

Web data, such as web pages and web images, can be naturally partitioned into multiple heterogeneous attribute sets. Concretely speaking, web pages consist of hyperlink and contents, and web images consist of the textual and visual information. In this paper, we propose a new multi-view semi-supervised learning method, named local co-training, for web page and image classification. Local co-training employs local linear models to represent data points on each view (i.e. one attribute set), and iteratively refines them using unlabelled data with co-training strategy. In each iteration, only a part of local models that we call dominant local models needs to be incrementally updated. The method is thus efficient and fit for the learning of large-scale web data. In addition, we introduce a new measurement based on both the confidence and the disagreement to describe which unlabelled examples are ‘good’ for the enrichment of training sets. Local co-training builds a bridge between two dominant types of semi-supervised methods: graph-based methods and co-training. Experiments on web page and web image datasets demonstrate that local co-training can effectively improve the classification performance by exploiting multiple attribute sets and unlabelled data.