next up previous contents
Next: 4 Semi-Supervised Learning Up: 3 Instance Weighting Previous: 3.2 Covariate Shift   Contents

3.3 Change of Functional Relations

Both class imbalance and covariate shift simplify the difference between $ P_s(X, Y)$ and $ P_t(X, Y)$. It is still possible that $ P_t(X \vert Y)$ differs from $ P_s(X \vert Y)$ or $ P_t(Y \vert X)$ differs from $ P_s(Y \vert X)$.

Jiang and Zhai (2007a) considered the case when $ P_t(Y \vert X)$ differs from $ P_s(Y \vert X)$, and proposed a heuristic method to remove ``misleading'' training instances from the source domain, where $ P_s(y \vert x)$ is very different from $ P_t(y \vert x)$. To discover these ``misleading'' training instances, some labeled data from the target domain is needed. This method therefore is only suitable for supervised domain adaptation.

Jing Jiang 2008-03-06