R中用transitiveweights建模赋值有向ERGM遇MCMC异常求助

背景 #

使用R的ergm.count包对84节点、75条加权边的赋值有向网络建模，邻接矩阵来自本地文件mat.csv。运行代码时CPU负载极高，进程长时间停滞，最终终止并抛出MCMC不混合的错误。

运行代码 #

library(ergm.count)
mat <- as.matrix(read.csv("./mat.csv",header=TRUE)[,-1])
net<-as.network(mat , directed=TRUE, matrix.type="a", 
                      ignore.eval=FALSE, names.eval="cnts"
                     )
m1<-ergm(net~ transitiveweights("min", "max", "min"),
         reference=~Poisson,
         response="cnts",
         control=control.ergm(parallel=6, parallel.type="PSOCK",
                              main.method="MCMLE")
)

现象与报错信息 #

1. 运行期间CPU核心占用拉满
2. 初始CD-MCMLE阶段收敛后，MCMLE阶段停滞3小时，最终终止输出：

> source("/root/transfer_analysis/test.R", encoding = "UTF-8")
Evaluating network in model.
Initializing unconstrained Metropolis-Hastings proposal: ‘ergm.count:MH_DiscTNT’.
Initializing model...
Model initialized.
Using initial method 'CD'.
Fitting initial model.
Starting contrastive divergence estimation via CD-MCMLE:
Iteration 1 of at most 60:
Convergence test P-value:1.5e-06
The log-likelihood improved by 0.1578.
Iteration 2 of at most 60:
Convergence test P-value:2.6e-02
The log-likelihood improved by 0.02119.
Iteration 3 of at most 60:
Convergence test P-value:4.6e-01
The log-likelihood improved by 0.002026.
Iteration 4 of at most 60:
Convergence test P-value:1e-01
The log-likelihood improved by 0.01116.
Iteration 5 of at most 60:
Convergence test P-value:1.4e-01
The log-likelihood improved by 0.007586.
Iteration 6 of at most 60:
Convergence test P-value:1e+00
Convergence detected. Stopping.
The log-likelihood improved by < 0.0001.
Finished CD.
Starting Monte Carlo maximum likelihood estimation (MCMLE):
Density guard set to 10000 from an initial count of 75 edges.

Iteration 1 of at most 60 with parameter:
transitiveweights.min.max.min 
                    -1.483437 
Starting unconstrained MCMC...
Back from unconstrained MCMC.
Error in ergm.MCMLE(init, s, s.obs, control = control, verbose = verbose,  : 
  Unconstrained MCMC sampling did not mix at all. Optimization cannot continue.
Additional warning message:
In ergm_MCMC_sample(s, control, theta = mcmc.init, verbose = max(verbose -  :
  Unable to reach target effective size in iterations alotted.

环境信息 #

• R版本：4.3.2
• ergm版本：4.5.0

解决方案建议 #

1. 调整MCMC控制参数 #

默认MCMC迭代配置可能不匹配稀疏加权网络，尝试增加燃烧期、采样间隔与样本量：

control=control.ergm(
  parallel=6, 
  parallel.type="PSOCK",
  main.method="MCMLE",
  MCMC.burnin = 10000,  # 延长燃烧期，让MCMC稳定
  MCMC.interval = 100,   # 增加采样间隔，减少自相关性
  MCMC.samplesize = 5000 # 提升样本量，保证有效样本数
)

2. 更换MCMC提议机制 #

默认的MH_DiscTNT提议可能不适合当前稀疏网络，改用Poisson类型提议：

control=control.ergm(
  parallel=6, 
  parallel.type="PSOCK",
  main.method="MCMLE",
  MCMC.prop = "ergm.count:MH_Poisson"
)

3. 简化模型基础结构 #

先拟合包含基础边项的模型，验证流程可行性后再加入高阶项：

m1<-ergm(net~ edges + transitiveweights("min", "max", "min"),
         reference=~Poisson,
         response="cnts",
         control=control.ergm(parallel=6, parallel.type="PSOCK",
                              main.method="MCMLE")
)

4. 预处理网络数据 #

• 检查权重是否存在极端值，对权重做标准化（如缩放至0-1区间）后再建模
• 确认网络稀疏性：84节点理论最大边数为6972，实际仅75条，极端稀疏可能导致MCMC难以探索样本空间

5. 优化并行配置 #

过多并行核数可能导致内存竞争，尝试减少核数并确保R有足够内存分配：

control=control.ergm(
  parallel=3, 
  parallel.type="PSOCK",
  main.method="MCMLE"
)

内容的提问来源于stack exchange，提问作者b39b332d