부산대학교 도서관

In this work, we conduct a feature-aware comparison of approaches to Cumulative Citation Recommendation (CCR), a task that aims to filter and rank a stream of documents according to their relevance to entities in a knowledge base. We conducted experiments starting with a big feature set, identified a powerful subset and applied it to comparing classification and learning-to-rank algorithms. With few set of powerful features, we achieve better performance than the state-of-the-art. Surprisingly, our findings challenge the previously known preference of learning-to-rank over classification: in our study, the CCR performance of the classification approach outperforms that using learning-to-rank. This indicates that comparing two approaches is problematic due to the interplay between the approaches themselves and the feature sets one chooses to use.