Greedy and Randomized Feature Selection for Web Search Ranking

Modern search engines have to be fast to satisfy users, so there are hard back-end latency requirements. The set of features useful for search ranking functions, though, continues to grow, making feature computation a latency bottleneck. As a result, not all available features can be used for ranking, and in fact, much of the time only a small percentage of these features can be used. Thus, it is crucial to have a feature selection mechanism that can find a subset of features that both meets latency requirements and achieves high relevance. To this end, we explore different feature selection methods using boosted regression trees, including both greedy approaches (i.e., selecting the features with the highest relative influence as computed by boosted trees, discounting importance by feature similarity) and randomized approaches (i.e., best-only genetic algorithm, a proposed more efficient randomized method with feature-importance-based backward elimination). We evaluate and compare these approaches using two data sets, one from a commercial Wikipedia search engine and the other from a commercial Web search engine. The experimental results show that the greedy approach that selects top features with the highest relative influence performs close to the full-feature model, and the randomized feature selection with feature-importance-based backward elimination outperforms all other randomized and greedy approaches, especially on the Wikipedia data.