Besides providing information, Web pages serve as the user interfaces of the Internet. The Internet has become indispensable in people’s daily life, so there is an increasing need to design Web pages with high-quality appearance. As user experience has been becoming more and more important recently, how to improve the visual experience of Web pages is of great importance and attracts the researchers from multiple disciplines such as human-computer interaction (HCI), industry design, cognitive psychology, etc. Due to lack of advanced Web mining and computer vision techniques for feature selection and sophisticated machine learning theory for model construction, the constructed Web appearance evaluation models in existing studies are usually naive. And the models are insufficient in the description of the relationships between features of Web appearance and human perception. To advance Web appearance research and construct generalized, automatically and more accurate evaluation models, this paper bridges the gap between the studies in the HCI and design communities and the methodologies in Web mining and machine learning. The main contributions of this thesis include: 1. We have analyzed the representation of human perception of Web appearance and proposed a new representation strategy. The perception on a page by a single person is represented by the person’s vote on a pre-defined set of basic ratings. However, the votes of a group of persons instead of a single person are meaningful for model learning. Current studies usually leverage a categorical label or a score to represent the perception of a group of persons of a particular page. This thesis proposes a new representation that utilize the distribution of use votes over basic labels, which can better capture the subjectivity nature of human perception on Web appearance. 2. We have proposed a general framework for the model construction for Web appearance evaluation. This framework considers three types of features: low-level features that are extracted directly from the source codes of Web pages, middle-level features that describe the structures of Web pages, and high-level features that describes the visual presentation of a page. These features cove the information of a Web page well. We have constructed classifying and scoring models for Web appearance. Existing evaluation models considers a limited number of features and the involved model learning algorithms are just simple data fitting a...
修改评论