A taxonomy of sequential pattern mining algorithms
ACM Computing Surveys (CSUR)
| Hi-index | 0.00 |
Web user modeling which studies users' diverse online navigation behavior and infers their interests has become crucial in many applications, such as Web recommendation, Web site evaluation and reorganization. There exists a wide range of Web user modeling approaches, including supervised learning approaches such as classification, and unsupervised learning approaches such as association rule mining and clustering. Most of existing approaches discover users' browsing patterns by either exploring the similarity between users' navigation paths, or inferring Web page associations based on their co-occurrence in users' visits. However, these discovered patterns are page-level patterns, such as an association rule pageA ⇒ pageB (support 50%, confidence 70%), implying these two pages are likely to be visited together. These patterns are too simple to represent users' diverse, complex navigation behavior, nor can they characterize users' underlying interests or information needs which lead to the specific navigation behavior. In this thesis, we present a Task-Oriented Web user modeling framework for studying Web users' navigation behavior from multiple sources of knowledge, including users' navigation data, Web site content data and other possible sources. In this framework, we introduce the notion of task, which is defined as a set of functionalities provided by a Web site, for meeting users' diverse information needs. These tasks are not directly visible, but implicitly lay underlying users' interactions with the site, and the site's content and structure information. Our framework consists of two task-oriented models, i.e. task learning from user navigation data, and task learning from Web site content information. In these two models, we use hidden variable models to automatically discover tasks from Web user navigation data and content data respectively, quantitatively characterize users' interests, discover page-level and task-level usage patterns, and conduct various types of user behavior analyses. Furthermore, usage patterns from these two task models are seamlessly integrated under the maximum entropy principle, allowing us to develop various types of Web applications.