Given the increase in the amount of accessible information on the Web, more attention has been drawn to information retrieval systems such as search engines. In web search, recency ranking refers to ranking documents by their relevance to the query, but also taking freshness into account. In this thesis, we propose two models for recency ranking. The first one is the query recency sensitivity model, and the second is a model to predict the publication time of documents. We extract temporal features from several sources and automatically construct ground truth datasets for both models. Furthermore, we integrate the models into an existing commercial search engine with a multi-stage ranking architecture. Our experiments demonstrate an improvement in the effectiveness of the commercial search engine.
keywords: recency ranking, web search, search engine, machine learning, gradient boosted decision trees, information retrieval, big data
Please see full thesis here.