### Table 2: Shrinkage (shr), estimate (EB or EM) - pii, is within parenthesis, with the usage of the EB- and the EM-method.

2004

### Table 7: Performance of the EM algorithm when initialized by the Random, Marginal, and HAC methods on the digit6 dataset (average and standard deviation over 12 or more runs). Initialization Random Marginal HAC HAC HAC

1998

"... In PAGE 16: ... We examine HAC with N0 = N = 700 and with lower values N0 = 100; 300. Table7 summarizes the results. There are no clear winners.... In PAGE 16: ... Moreover, Random initialization tends to yield more small clusters. On average, for the models whose performance is recorded in Table7 , 1.5 clusters are supported by less than one case.... ..."

Cited by 66

### Table 1: Comparison of the IB-EM algorithm, 50 runs of EM with random starting points, and 50 runs of mean field EM from the same random starting points. Shown are train and test log-likelihood per instance for the best and 80th percentile of the random runs. Also shown is the percentile of the runs that are worse than the IB-EM results. Data sets shown include a Naive Bayes model for the Stock data set and the Digit data set; a 3 and 4 level hierarchical model for the Digit data set (DigH3 and DigH4); and an hierarchical model for the Yeast data set. For each model we show several cardinalities for the hidden variables, shown in the first column.

2005

"... In PAGE 20: ... The advantage of IB-EM is particularly pronounced for the more complex models with higher cardinalities. Table1 provides more details of these runs including train performance and comparison to 50 random mean field EM runs. We also compared the IB-EM method to the perturbation method of Elidan et al.... ..."

Cited by 9

### Table III gives the results of TRUST-TECH compared with other methods proposed in the literature like split and merge EM and k-means+EM [41]. MRS+EM indicates the multiple random starts experiment which includes the same number of starts as the number of valid directions in the case of the TRUST- TECH method. The multiple starts are made only in the promising regions. RS+EM is just a single random start and is given here to illustrate the lower bound on the performance of our method empirically. TABLE III COMPARISON OF TRUST-TECH-EM WITH OTHER METHODS Method Elliptical Iris

2007

Cited by 1

### Table 3. Common haplotypes and their frequencies obtained by the locus-based algorithm, MRH and the EM method. In the haplotypes, the alleles are encoded as 1=A, 2=C, 3=G, and 4=T.

2003

"... In PAGE 12: ...andomly selected autosome (i.e. chromosome 3). There are 10 blocks (i.e. regions with few recombinants) in the chromosome 3 data, of which one block consists of 16 marker loci and all the others have only 4-6 marker loci each. The test results of the locus-based algorithm, MRH and the EM algorithm are summarized in Table3 . (We obtained the results of the EM algorithm directly form the authors [5].... ..."

Cited by 11

### Table 6: Common haplotypes and their frequencies obtained by block-extension, ILP and the EM method. In haplotypes, the alleles are encoded as 1=A, 2=C, 3=G, and 4=T. Block Common haplotypes EM BE ILP

2005

"... In PAGE 26: ... Such frequency information can be used to estimate the likelihood of the haplotypes in a pedigree as described in Section 5. The common haplotypes, their frequencies and their total frequency in each block of chromosome 3 estimated by block-extension, ILP and the EM algorithm are summarized in Table6 . The majority of the common haplotypes identified by the three algorithms are the same.... ..."

Cited by 4