seminar

Conviction
conv(s->u) = 1/imp(s->~u)
it is generalized from improvement but taken negative case into consideration too.


Semi-Supervised Learning
using graph-based to connect nearest all instances in the datasets. Estimating the labels of unknown instances by energy propagation (or other functions) of connected instances.

Multiple-Instance Learning
we know the negative is all negative but for positive group is combined with at least one positive. An example of application is content-based image retrieval