Analyzing web layout structures using graph mining

Abstract

The layout of a web page commonly offers a limited variety of elements arranged in a number of ways, for example, in navigation panels, or as advertisements, text content, and images. Presumably, the layout of a web page will influence the way it is used, and this may or may not match the intentions of its designers. In this paper, we propose a novel graph mining algorithm and apply it to study the commercially important problem of how and what specific patterns and features of layout affect advertising click rates. Our proposed algorithm, MIGDAC (Mining Graph DAta for Classification), applies graph theory and an interestingness measure to discover interesting subgraphs that can allow one class to be both characterized and easily distinguished from other classes. We first extract the information as a block from the web pages and transform that information into sets of graphs. MIGDAC then uses an interestingness threshold and measure to extract a set of class-specific patterns from the frequent sub-graphs of each class. We then, calculate the weight of evidence to estimate whether the layout of the page will positively or negatively influence the advertisement click-rate on an unseen web page. The experiment is performed on a set of real web pages from a local web site. MIGDAC performs well, greatly improving the accuracy of traditional frequent graph mining algorithm.