Searching and Ranking in Entity-Relation Graphs

Ranking results to queries over information spaces where the data is represented in a directed graph has been the focus of much recent research work. In this report, we study three ranking paradigms addressing di erent issues in this regard. Entity-Relation(ER) graphs have entities as nodes, with relations between them being represented as edges. The PageRank system for ranking relies on the link-structure of the graph alone, and does not incorporate the query into search results explicitly. The HubRank system developed at IITB presented a viable way to dynamically personalize PageRank at query-time on ER graphs, by utilizing clever hubset selection strategies and early termination bounds. In this work, we rst look at extending the HubRank system to allow for personalization based on recent user-browsing history. Queries with hard predicates, such as Find papers written before 1990 near cryptography require the ranking to be carried out only on a subset of the graph (papers before 1990). We present a 'reverse push' algorithm to handle such queries, and evaluate its performance. Finally, we shift to looking for novel ways to combine link-structure information with local text-matches on nodes on a graph. The problem is rst formulated as an optimization problem for a single query. We then propose methods to extend it to learn to rank from multiple queries at once, and discuss issues faced in the same.