Learning to Rank with Additive Ensembles of Regression Trees
Speaker(s) : Nicola Tonellotto (Univ. Pise)
Ranking query results according to a relevance criterion is a fundamental problem in Information Retrieval. Nowadays, an emerging research area named Learning-to-Rank has shown that effective solutions to the ranking problem can leverage machine learning techniques. The ranking process is particularly challenging for Web search engines, which, besides the demanding requirements for result pages of high quality in response to user queries, have also to deal with efficiency constraints, which are not so common in other ranking-based applications. Indeed, two of the most effective LtR-based rankers are based on additive ensembles of regression trees, namely Gradient-Boosted Regression Trees, and Lambda-MART. Due to the thousands of trees to be traversed at scoring time for each document, these rankers are also the most expensive in terms of computational time, thus impacting on response time and throughput of query processing. In this seminar we will introduce the Learning to Rank framework from the Information Retrieval perspective. Moreover, we will illustrate and discuss a novel algorithm to efficiently score documents by using a machine-learned ranking function modeled by an additive ensemble of regression trees.