Learning to Rank in Vector Spaces and Social Networks

User query q, Web pages {v } (q, v) can be represented with a rich feature vector Text match score with title, anchor text, headings, bold text, body text,. .. , of v as a hypertext document Pagerank, topic-specific Pageranks, personalized Pageranks of v as a node in the Web graph Estimated location of user, commercial intent,. .. Must we guess the relative importance of these features? How to combine these into a single scoring function on (q, v) so as to induce a ranking on {v }? Here, the " query " is the surfer's contextual information More noisy than queries, which are noisy enough! Plus page and site contents A response is an ad to place, or a link to insert Must rank and select from a large pool of available ads or links (In this tutorial we will ignore issues of bidding and visibility pricing) The Web has only a few kinds of hyperlinks: same-host subdirectory, same-host superdirectory, same-host across-path, different-host same-domain, different-domain etc. Often differentiated by hardwired policy, e.g, HITS completely ignores same-host links Entity-relationship (ER) graphs are richer E.g. A personal information management (PIM) system has many node/entity types (person, organization, email, paper, conference, phone number) and edge/relation types (works-for, sent, received, authored, published-in) Ranking needs to exploit the richer type system Don't want to guess the relative importance of edge types (may be dependent on queries) Relevance feedback is well-explored in traditional IR User-assisted local modification of ranking function for vector-space models How to extend these to richer data representations that incorporate entities, relationship links, entity and relation types? Surprisingly unexplored area Training and evaluation scenarios Measurements to evaluate quality of ranking Label mismatch loss functions for ordinal regression Preference pair violations Area under (true positive, false positive) curve Average precision Rank-discounted reward for relevance Rank correlations What's useful vs. what's easy to learn