Mining Models for Non-Visual Web Transactions

Web transactions (e.g. buying a CD player on the Web) typically involve a number of steps spanning several pages. This task gets strenuous when the Web is accessed non-visually (e.g. when the user is a visually handicapped individual). But usually one needs to browse only a small fragment of a Web page in a transactional step such as a form fill-out, selecting an item from search results, etc. Identifying and presenting such segments from a Web page can overcome the information overload problem when accessing the Web using non-visual interaction modalities like speech. Based on the aforementioned observation I have developed a transactional model that delivers only the “relevant” page fragments at each transactional step, thereby reducing the information overload. In my prior research, such transactional models were learned using a supervised learning approach. Supervised learning requires manually labeled training examples, consequences being that it is nether scalable nor flexible for creating personalized transaction models for arbitrary Web sites. In my PhD research, I am exploring automatic mining of transaction models from unlabeled transactional sequences. My approach is centered on leveraging contextual information of a hyper-link, geometric segmentation of a Web page, and clustering to mine the models. The approach will be scalable but more importantly end-users (with visual disabilities) will be able to create their own “personalized” transactional models for Web sites that they need to use on a regular basis for doing online transactions.