Semi-supervised learning by DDD with a sharing base function

1
Semi-supervised learning by DDD with a sharing
base function

• - preliminary result on WDBC data

2
From the DDD formula
Consider all the Dirichlet distribution share a
common base function (similar to what
Dunson did),
Where
Affinity matrix
We choose
3
Semi-supervised learning (transductive way)
For those data , this is the
conditional posterior for .
By performing MCMC, we can get the histograms of
their full posteriors given the labeled data set.
4
Apply the DDD to one benchmark data set
WDBC 569 data with dimensionality 32
Randomly choose a portion of data as labeled data
and calculate the area under ROC (AUR) for each
trial. The number of labeled data 20, 40, 60,
80. For each case 20 random trials.
MCMC iteration 2000 Burn-in 500
5
(No Transcript)
6
Qiuhuas results
Note the accuracy is different from AUR.