Researchers have long envisioned a Semantic Web, where unstructured Web content is replaced by documents with rich semantic annotations. Unfortunately, this vision has been hampered by the difficulty of acquiring semantic metadata for Web pages. This paper introduces a method for automatically "semantifying" structural page elements: using machine learning to train classifiers that can be applied in a post-hoc fashion. We focus on one popular class of semantic identifiers: those concerned with the structure―or information architecture―of a page. To determine the set of structural semantics to learn and to collect training data for the learning, we gather a large corpus of labeled page elements from a set of online workers. We discuss the results from this collection and demonstrate that our classifiers learn structural semantics in a general way.

This thesis describes how data-driven approaches to Web design problems can enable useful interactions for designers. It presents three machine learning applications which enable new interaction mechanisms for Web design: rapid retargeting between page designs, scalable design search, and gener- ative probabilistic model induction to support design interactions cast as probabilistic inference. It also presents a scalable architecture for efficient data-mining on Web designs, which supports these three applications.

One challenge in building a web design tool that attempts to leverage examples is gathering design alternatives and providing mappings between web page elements. We present a crowdsourcing interface to collect user-generated correspondences between two web pages. Our iterative refinement of the interface was guided by three main design principles: modularize the task, minimize user errors, and provide relevant information. As an initial experiment, we collected fifteen web pages with diverse style and layout, and deployed the interface on Amazon’s Mechanical Turk. Preliminary data analysis shows that Turkers take longer than experts and define fewer mappings in general. Further evaluation and experiments with different types of pages will identify directions for a web design tool that enables the use of any web page as a design template.

We present a novel technique for automatically retargeting content from one web page onto the layout of another. Web pages are decomposed into their perceptual hierarchical representations. We then use a structured-prediction algorithm to learn reasonable mappings between the perceptual trees. Using the mappings, we are able to merge the content of one page with the layout of another.

This thesis presents a volume-preserving analytic muscle model that can be embedded within a skin mesh to induce realistic, physics-based deformation during simulation. These volumetric muscles are layered on top of a dynamic framework of linear, segment-based muscles that drive the underlying skeletal structure. The result is an integrated system that supports realistic skin deformation along a specified target motion while requiring only minimal computational resources.