1
$\begingroup$

I am using graphs to represent some structured data. In my case, I have a time series of undirected unweighted graphs with the same topology (i.e. isomorphic graphs with same number of nodes and edges, and same connections). The only thing keeps changing over time are (some of) the node labels. In my case, the node labels are discrete (categorical labels). I want to use some kind of measurement to describe the similarity (or dis-similarity) between two given isomorphic graphs with different node labels. Note that this is fundamentally different from comparing two sequences in the sense that the local similarity at each node not only depends on the matchings of the node label, but also the matchings of the labels of the 1st, 2nd, ... nearest neighbors.

I am looking into graph kernels. In theory, there should be a significant speed-up given the graph topology is fixed, since one can precompute an original kernel and reuse it with different label distributions for different isomorphic graphs. However, I found that all common graph kernels assign the majority of the weight to topological features, which means I will always get very high similarity scores between two isomorphic graphs, even with completely different sets of node labels, which I would expect the similarity to be 0 in this case.

Can anyone suggest a good algorithm that can measure similarity between isomorphic graphs with different node labels? I have been looking for literature related to this, but could not find any. Any suggestion is appreciated.

Edit: note that there can be multiple isomorphisms. For example, as depicted below, $G_1$ and $G_2$ have different node labels at node 1, 3, 4, but they are completely identical, so they should have the highest similarity. This means that any similarity measurement that compares node labels by indices is incorrect.

enter image description here

$\endgroup$
11
  • 1
    $\begingroup$ You can maybe use a formula of the type : $\sum_{v \in V} \sum_{k = 1...n}|N_k(v)| \alpha_k \phi(v)$ where $\phi(v)$ is equal to $1$ if the labels do not match on $v$, $N_k(v)$ are the $k$-nearest neighbors, $\alpha_k$ is a decreasing function of $k$, and $n$ is fixed to make the computation tractable. $\endgroup$
    – caduk
    Oct 8 at 9:48
  • 2
    $\begingroup$ A better one could be $\sum_{v \in V} \sum_{k = 1...n}\sum_{v'\in V}s_k(v') \alpha_k \phi(v)$ where $s_k(v')$ is the number of $k$-paths from $v$ to $v'$. $\endgroup$
    – caduk
    Oct 8 at 9:57
  • 1
    $\begingroup$ What would it change? To compute $\phi$, you just need to compare the labels, it does not matter how they are encoded. $\endgroup$
    – caduk
    Oct 8 at 11:10
  • 1
    $\begingroup$ As I understand your problem, your graphs always have an isomorphism and you only want to compute the similarity of graphs on exact isomorphisms (no costs on structural dissimilarity). The first step is probably to compute the automorphism group. If the group is relativey small, then you can compute the dissimilarity on all elements of the group (using for example the dissimilarity I proposed, which is a linear one). If it's bigger, then your problem reduce to optimizing hamming distance on a permutation group (assuming your dissimilarity is linear), and it is NP-hard $\endgroup$
    – caduk
    Oct 13 at 8:14
  • 1
    $\begingroup$ If I understand right, you have a sort of automorphism group of a toroidal grid, with only the translational symmetry. You may look at the cyclic hamming distance for string (here on 2d), there are interesting results using fourier transform (cs.stackexchange.com/questions/131994/…), which would give a $n^2log(n)|\Sigma|$ complexity I believe. $\endgroup$
    – caduk
    Oct 13 at 9:48
1
$\begingroup$

This paper might be what you're looking for. You can phrase the algorithm as an eigenvalue problem. I remember implementing it, but it's kind of confusing until you look up what the different matrix products mean. Good luck!

$\endgroup$
2
  • 1
    $\begingroup$ Thanks for the paper suggestion. Yes the paper is confusing and I failed to understand how it is related to my question. Would you like to explain it in detail? Btw I have started a bounty. $\endgroup$
    – Shaun Han
    Oct 12 at 22:06
  • 1
    $\begingroup$ The central idea is this: assign a score of local similarity to each node and edge in graph 1 to all nodes and edges in graph 2. Propagate the scores in graph 1 according to the edges in graph 2. If you do this iteratively, eventually the node with the most similar neighborhood in graph 1 will have the highest score for the corresponding node in graph 2. The point of the paper is proving this and rephrasing the question as an eigenvalue problem (Eq 5). The eigenvalue is the overall similarity score, while the elements of the eigenvector are the scores for pairs of nodes from graph 1 x graph 2. $\endgroup$
    – MSha
    Nov 5 at 22:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.