Subset selection based fusion for biomedical information retrieval tasks.

Area of Science:

• Biomedical Informatics
• Information Retrieval
• Computer Science

Background:

• Biomedical information retrieval systems face challenges in effectiveness and efficiency.
• Data fusion techniques can improve retrieval performance but require optimal system selection.
• Existing methods for selecting retrieval systems for fusion may not be optimal.

Purpose of the Study:

• To propose and evaluate novel ranking-based methods for selecting optimal subsets of retrieval systems for data fusion.
• To enhance the effectiveness and efficiency of biomedical information retrieval.

Main Methods:

• Proposed three ranking-based subset selection methods: Sequential Forward Search (SFS), Diversity & Performance (D&P), and Performance & Diversity (P&D).
• Applied these methods in conjunction with Reciprocal Rank Fusion (RRF).
• Conducted experiments on four TREC medical datasets using 62-125 candidate retrieval systems, selecting up to 15 for fusion.

Main Results:

• The proposed subset selection methods significantly improved retrieval performance.
• Fusion of selected systems using RRF resulted in performance improvements of 10% to over 60% compared to the best individual system.
• The methods outperformed state-of-the-art technology by a considerable margin.

Conclusions:

• The proposed subset selection approach provides a practical and cost-efficient solution for biomedical information retrieval.
• Achieved substantial performance gains while reducing computational overhead.
• Offers a significant advancement over existing methods for optimizing retrieval system fusion.