Mining Web Dynamics for Search

Abstract Billions of web users collectively contribute to a dynamic web that preserves how information sources and descriptions change over time. This dynamic process sheds light on the quality of web content, and even indicates the temporal properties of information needs expressed via queries. However, existing commercial search engines typically utilize one crawl of web content (the latest) without considering the complementary information concealed in web dynamics. As a result, the generated rankings may be biased due to the efficiency of knowledge on page or hyperlink evolution, and the time-sensitive facet within search quality, e.g., freshness, has to be neglected. While previous research efforts have been focused on exploring the temporal dimension in retrieval process, few of them showed consistent improvements on large-scale real-world archival web corpus with a broad time span.We investigate how to utilize the changes of web pages and hyperlinks to improve search quality, in terms of freshness and relevance of search results. Three applications that I have focused on are: (1) document representation, in which the anchortext (short descriptive text associated with hyperlinks) importance is estimated by considering its historical status; (2) web authority estimation, in which web freshness is quantified and utilized for controlling the authority propagation; and (3) learning to rank, in which freshness and relevance are optimized simultaneously in an adaptive way depending on query type. The contributions of this thesis are: (1) incorporate web dynamics information into critical components within search infrastructure in a principled way; and (2) empirically verify the proposed methods by conducting experiments based on (or depending on) a large-scale real-world archival web corpus, and demonstrated their superiority over existing state-of-the-art.