FSMRank: feature selection algorithm for learning to rank.

Area of Science:

• Machine Learning
• Information Retrieval
• Data Mining

Background:

• Learning to rank (LTR) is a rapidly developing field.
• Feature selection is crucial for improving the performance of LTR models.
• Existing methods often treat feature selection and ranking separately.

Purpose of the Study:

• To investigate feature selection for learning to rank.
• To propose a joint optimization framework for feature selection and ranking.
• To enhance the performance and efficiency of LTR algorithms.

Main Methods:

• Developed a joint convex optimization formulation for simultaneous feature selection and ranking error minimization.
• Utilized Nesterov's accelerated gradient algorithm for efficient optimization.
• Derived a generalization bound using Rademacher complexities.

Main Results:

• The proposed method achieved significant ranking performance gains over feature selection baselines.
• Demonstrated competitive performance against state-of-the-art learning-to-rank algorithms.
• The framework flexibly incorporates various feature importance and similarity measures.

Conclusions:

• Joint feature selection and ranking optimization is effective for LTR.
• The proposed accelerated gradient algorithm offers fast convergence.
• The method provides a robust and high-performing solution for learning to rank tasks.