R语言igraph包rewire函数niter参数实际含义及取值建议

Great question! Let's break this down clearly, since the niter parameter is a key detail that makes a big difference in how your rewired networks turn out.

What does niter actually mean? #

First, let's recap what keeping_degseq() does: it uses the configuration model to rewire edges while preserving every node's degree (so the degree sequence stays identical to the original network).

The niter parameter sets the total number of edge swap attempts the function will make. Here's the catch: not every attempt succeeds. Each attempt works like this:

• Randomly pick two existing edges, say u-v and x-y
• Try to swap them to form u-x and v-y (or u-y and v-x)
• If those two new edges don't already exist (no duplicate edges or self-loops), the swap is kept; if not, the attempt is discarded.

So when you run rewire(keeping_degseq(niter = 20)), the function makes 20 swap attempts—only some will actually change the network. With niter = 100, there are way more chances to successfully swap edges, so the resulting network will be much more "randomized" compared to the original small-world network g. That's why you see differences in metrics like betweenness centrality: the 20-iteration network is still pretty close to the original small-world structure, while the 100-iteration one is closer to a random network with the same degree sequence.

How to choose niter for Bootstrap? #

The goal of Bootstrap here is usually to generate a set of random networks that match your original network's degree sequence, so you can test if your original network's properties are statistically different from random expectations. To do this well, you need niter large enough to let the network "converge" to a stable random state—but not so large that you waste computation time.

Here are practical steps to pick the right value:

• Test for convergence: Generate networks with increasing niter values (e.g., 10, 50, 100, 500, 1000). For each set of networks, calculate the mean and standard deviation of the metric you care about (like betweenness centrality or clustering coefficient). When increasing niter no longer causes a noticeable shift in these stats, you've hit a converged value.
• Use an empirical rule of thumb: A common starting point is to set niter to 5-10 times the total number of edges in your network. For your example network g: 10 nodes, each with degree 3, so total edges are (10*3)/2 = 15. 5-10 times that is 75-150, which explains why niter=100 gives a more randomized network than niter=20 (which is less than 2x the edge count).
• Balance computation cost: If you're working with a huge network, very large niter values will slow things down. Start with a moderate value, check if your metrics stabilize, and adjust up or down as needed.
• Validate consistency: Generate 10-20 networks with your chosen niter and check if their metric distributions are consistent. If they are, your niter is sufficient; if not, you need to increase it.

For your specific Bootstrap use case, niter=100 is already a better choice than 20, but you might want to test 150 or 200 to see if your metrics stabilize further.

内容的提问来源于stack exchange，提问作者FilipeTeixeira