Balancing Precision and Recall with Selective Search
Oral Defence Date:
Thursday, May 3, 2018 - 10:00
Asst. Prof. Anagha Kulkarni, Assoc. Prof. Kaz Okada, and Professor Arno Puder
Balancing precision and recall is a long-standing problem in Information Retrieval field. We tackle this problem with Selective Search, which divides the document collection into small shards and passes the user query to only a few selected shards. In contrast to the idea of Selective Search, is Exhaustive Search which passes the user query to the whole collection. Selective Search has shown higher performance than Exhaustive Search on efficiency and precision. However, recall with Selective Search has been lower because it misses some relevant documents. We address this problem in our research by improving the accuracy of shard selection process. We propose three new shard ranking approaches that provide statistically significant improvements in both precision and recall. Furthermore, we integrate three cutoff estimators that select how many of the top-ranked shards are searched for the query. A thorough evaluation with multiple large datasets demonstrates that the proposed shard ranking and cutoff estimators facilitate Selective Search that outperforms Exhaustive Search in both precision and recall.
Information Retrieval, Selective Search, Learning to Rank