Figure 2
From: Identifying protein complexes directly from high-throughput TAP data with Markov random fields
MRF on simulated data. We tested two sets of simulated data: (1) N = 500, K = 11 and (2) N = 3000, K = 500 and the false positive rate φ is set to 0.005 and the false negative rates ν is 0.2 or 0.5. With ν = 0.2 (2(a), 2(b)), MRF can recover the true clustering with the minimum negative log-likelihood which is taken on for 11 clusters. Notice that any more clusters do not reduce the cost any further; additional clusters simply remain empty. For ν = 0.5, the accuracy is worse and needs more empty clusters to reach convergence. In 2(c) and 2(d) the convergence rate fluctuates more.