Learning to Efficiently Rank on Big Data

Ranking in response to user queries is a central prob-lem in information retrieval, data mining, and machine learning. In the era of“Big data”, traditional effectiveness-centric ranking techniques tend to get more and more costly (requiring additional hardware and energy costs) to sustain reasonable ranking speed on large data. The mentality of combating big data by throwing in more hardware/machines will quickly become highly expensive since data is growing at an extremely fast rate oblivious to any cost concerns from us. “Learning to efficiently rank ” [5, 7, 6] offers a cost-effective solution to ranking on large data (e.g., billions of documents). That is, it addresses a critically important question – whether it is possible to improve ranking effectiveness on large data without incurring (too much) additional cost? At the basic level, the “Learning to efficiently rank” framework explores speed-effectiveness tradeoffs in the context of building highly effective, and very fast rank-ing models for large-scale document collections. A key observation is that there is an inherent tradeoff between the speed and effectiveness of ranking models – models that are highly effective typically are slow (due to poten-tially using a larger number of costly features and com-plex model structures), and models that are fast typically are not as effective. We would like to design and learn new ranking models that can break through the trade-off barrier and be simultaneously fast and effective when applied to large data (e.g., billions of documents). In view of the inherent tradeoffs, this is quite challenging, as most of the previous ranking models focus on a fixed operating point along the space of effectiveness vs speed. Copyright is held by the author/owner(s)