Researches on Combinations of Auxiliary Problems in ASO (Alternating Structure Optimization) Algorithm

Recently, a semi-supervised learning algorithm called ASO (Alternating Structure Optimization) has been proposed, which belongs to linear structural learning. It utilizes a number of auxiliary problems (APs) with unlabelled data and then extracts common structural parameter of APs to improve the performances of the target problems (TPs). How to select the appropriate APs is the keystone of ASO algorithm. This paper proposes another principle of APs selection: combinations. It determines optimal ratios between multi-combinations when proper total amounts of APs are given. Besides, we also analyze how to select appropriate total amounts. Both theoretical analysis and experimental results indicate that the principle of combinations is credible. Comparing with the principle of diversity that we have proposed, this principle immensely reduces the computational complexity. While the performances keep invariable.