TRVR: A Trimmed Relevance Vector Regression Method
Banghua Yang, Zhizhen Zhang, Zhengyang Sun · 2007
A novel trimmed relevance vector regression method named TRVR is proposed to provide robust solution for regression. Firstly the likelihood function is redefined as the trimmed likelihood function over a trimmed subset. Then by maximizing the trimmed likelihood function within the relevance vector machine (RVM) framework, the model weights can be learnt. Simultaneously a re-weighted update strategy is utilized to update the subset iteratively until the optimized subset without outliers is obtained, which can lead to the robustness. Finally the experimental evidence has been gathered to show that this proposed method is very robust and effective.