Abstract:This paper presents a two-stage ranking algorithm based on meta-learning. The first stage aims at
directly optimizing NDCG (Normalized Discounted Cumulative Gain). In this stage, the algorithm uses
the one-slack SVM as the framework, and defines an optimization objective related to NDCG. In order
to solve the problem of exponential size of constraints in the optimization, a cutting plane algorithm is
introduced. For the sub-problem of finding the most violated constraint, we address it by Quick Sort.
The second stage is a refinement of the first one. In this phase, a non-convex sigmoid loss function
is proposed which guarantees the result is a local maximum of the first stage. Empirical studies on
benchmark collection justify the effectiveness of the proposed algorithm.
Fan CHENG;Xufa WANG. RANK-META: A Two-Stage Ranking Algorithm Based on Meta-Learning[J]. , 2011, 8(16): 4317-4326.
Fan CHENG;Xufa WANG. RANK-META: A Two-Stage Ranking Algorithm Based on Meta-Learning. , 2011, 8(16): 4317-4326.