Generalized Median Graph via Iterative Alternate Minimizations
Pith reviewed 2026-05-25 15:57 UTC · model grok-4.3
The pith
Block coordinate descent approximates generalized median graphs by alternating node and edge edit optimizations.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central claim is that an alternating minimization procedure, which optimizes node edits in one block and edge edits in the next within a coordinate descent loop, produces an efficient approximate generalized median graph for labeled graphs.
What carries the argument
Block coordinate descent that alternates between optimizing node edit operations and edge edit operations.
If this is right
- The method works for graphs whose nodes and edges both carry labels.
- It yields a practical computational procedure for an NP-hard problem.
- Experiments across multiple datasets confirm that the procedure runs efficiently.
Where Pith is reading between the lines
- Different initialization graphs could be tested to see how much the final median cost varies with starting point.
- The alternation pattern might be applied to other graph prototype problems that separate node and edge contributions.
- The same block structure could be combined with local search heuristics to escape poorer local solutions.
Load-bearing premise
That alternating between node and edge optimization blocks produces a result close to the true generalized median instead of a poor local minimum that depends heavily on initialization.
What would settle it
On small collections of graphs where an exact generalized median can be computed by exhaustive search, measure whether the alternating method returns a graph whose total edit cost to the set is substantially higher than the exact minimum.
Figures
read the original abstract
Computing a graph prototype may constitute a core element for clustering or classification tasks. However, its computation is an NP-Hard problem, even for simple classes of graphs. In this paper, we propose an efficient approach based on block coordinate descent to compute a generalized median graph from a set of graphs. This approach relies on a clear definition of the optimization process and handles labeling on both edges and nodes. This iterative process optimizes the edit operations to perform on a graph alternatively on nodes and edges. Several experiments on different datasets show the efficiency of our approach.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims to provide an efficient block coordinate descent algorithm for computing a generalized median graph from a set of (labeled) graphs. The method iteratively alternates between optimizing node edit operations and edge edit operations on a candidate prototype; the problem is acknowledged to be NP-hard, and the approach is evaluated empirically on several datasets to demonstrate practical efficiency.
Significance. A reliable, scalable heuristic for the generalized median graph would be useful for downstream graph clustering and classification tasks. The alternating-minimization framing is a natural coordinate-descent idea, but its value hinges on whether the attained local solutions are sufficiently close to the global median; without supporting analysis this remains an open empirical question.
major comments (2)
- [Abstract] Abstract: the central claim that the block coordinate descent procedure 'computes a generalized median graph' is not accompanied by any convergence analysis, approximation bound, or comparison against exact or branch-and-bound solvers, despite the explicit statement that the underlying problem is NP-hard.
- [Abstract] Abstract (and method description): because the edit-distance objective couples node and edge labels, an alternation that optimizes nodes then edges (or vice versa) can miss configurations that are jointly optimal; no argument or experiment is supplied showing that the attained cost is close to the true median or that the gap is bounded independently of initialization.
Simulated Author's Rebuttal
We thank the referee for the detailed and constructive report. We agree that the current wording in the abstract overstates the capabilities of the proposed block coordinate descent heuristic for this NP-hard problem. In the revised manuscript we will tone down the claims, explicitly label the method as a heuristic, and add a limitations paragraph discussing the absence of convergence guarantees and approximation bounds.
read point-by-point responses
-
Referee: [Abstract] Abstract: the central claim that the block coordinate descent procedure 'computes a generalized median graph' is not accompanied by any convergence analysis, approximation bound, or comparison against exact or branch-and-bound solvers, despite the explicit statement that the underlying problem is NP-hard.
Authors: We accept the criticism. The manuscript will be revised to replace the phrase 'computes a generalized median graph' with 'approximates a generalized median graph via block coordinate descent' in the abstract and introduction. A new paragraph will be added in the discussion section acknowledging that no convergence analysis or approximation guarantees are provided and that the method is a practical heuristic whose quality is assessed only empirically. Direct comparisons against exact solvers are infeasible on the larger datasets used in the experiments; on the smallest synthetic instances we can add a brief note that the heuristic matches the optimum found by exhaustive search. revision: yes
-
Referee: [Abstract] Abstract (and method description): because the edit-distance objective couples node and edge labels, an alternation that optimizes nodes then edges (or vice versa) can miss configurations that are jointly optimal; no argument or experiment is supplied showing that the attained cost is close to the true median or that the gap is bounded independently of initialization.
Authors: We agree that the node-edge coupling can cause the alternating procedure to converge to a local rather than global optimum. The revision will explicitly state in Section 3 that the algorithm finds a coordinate-wise local minimum and will report results from multiple random initializations on each dataset to illustrate variability. We do not possess a theoretical bound on the optimality gap that holds independently of initialization, and deriving such a bound appears difficult given the NP-hardness of the underlying problem; the manuscript will therefore not claim any such guarantee. revision: yes
Circularity Check
No circularity: algorithmic heuristic for NP-hard problem with no self-referential reduction
full rationale
The paper presents a block coordinate descent algorithm that alternates optimization of node and edge edit operations to approximate the generalized median graph. No equations, parameters, or claims reduce the output to a fitted input, self-definition, or self-citation chain. The method is explicitly positioned as an efficient heuristic for an acknowledged NP-hard problem, without any derivation that equates the attained local optimum to the global median by construction. The central procedure is a direct iterative minimization whose correctness is evaluated empirically rather than proven via tautological steps.
Axiom & Free-Parameter Ledger
Reference graph
Works this paper leans on
-
[1]
In: International Conference on Pattern Recognition
Bougleux, S., Ga¨ uz` ere, B., Brun, L.: Graph edit distance as a quadratic program. In: International Conference on Pattern Recognition. pp. 1701– 1706 (2016). https://doi.org/10.1109/ICPR.2016.7899881 9 Letter (HIGH) Dataset TS 1st phase 2nd phase pt % SM t(SM) % GM t(GM) % TS t(TS) 10% mBipartite mBipartite 0.023 76.42 0.325 82.82 0.325 83.01 5.275 mBi...
-
[2]
Pattern Recognition Letters 1(4), 245–253 (1983)
Bunke, H., Allermann, G.: Inexact graph matching for structural pattern recognition. Pattern Recognition Letters 1(4), 245–253 (1983). https://doi.org/10.1016/0167-8655(83)90033-8
-
[3]
Pattern Recognition72, 266–284 (2017)
Chaieb, R., Kalti, K., Luqman, M.M., Coustaty, M., Ogier, J.M., Amara, N.E.B.: Fuzzy generalized median graphs computation: Application to content-based document retrieval. Pattern Recognition72, 266–284 (2017). https://doi.org/10.1016/j.patcog.2017.07.030
-
[4]
In: International Conference on Pattern Recognition Applications and Methods
Daller, ´E., Bougleux, S., Ga¨ uz` ere, B., Brun, L.: Approximate graph edit distance by several local searches in parallel. In: International Conference on Pattern Recognition Applications and Methods. pp. 149–158 (2018). https://doi.org/10.5220/0006599901490158
-
[5]
Application to Graph-based Classification and Clustering
Ferrer, M.: Theory and Algorithms on the Median Graph. Application to Graph-based Classification and Clustering. Ph.D. thesis, Universitat Aut` onoma de Barcelona (2008),http://hdl.handle.net/10803/5788
work page 2008
-
[6]
In: Graph Embedding for Pattern Analysis, pp
Ferrer, M., Bardaj´ ı, I., Valveny, E., Karatzas, D., Bunke, H.: Median graph computation by means of graph embedding into vector spaces. In: Graph Embedding for Pattern Analysis, pp. 45–71. Springer New York (2013). https://doi.org/10.1007/978-1-4614-4457-2 3 10
-
[7]
Computer Vision and Image Understanding 115(7), 919– 928 (2011)
Ferrer, M., Karatzas, D., Valveny, E., Bardaji, I., Bunke, H.: A generic framework for median graph computation based on a recursive embed- ding approach. Computer Vision and Image Understanding 115(7), 919– 928 (2011). https://doi.org/10.1016/j.cviu.2010.12.010
-
[8]
Pattern Recognition 43(4), 1642–1655 (2010)
Ferrer, M., Valveny, E., Serratosa, F., Riesen, K., Bunke, H.: Generalized median graph computation by means of graph embed- ding in vector spaces. Pattern Recognition 43(4), 1642–1655 (2010). https://doi.org/10.1016/j.patcog.2009.10.013
-
[9]
Soft Computing 10(1), 47–53 (2006)
Hlaoui, A., Wang, S.: Median graph computation for graph clustering. Soft Computing 10(1), 47–53 (2006). https://doi.org/10.1007/s00500-005-0464- 1
-
[10]
Jiang, X., Munger, A., Bunke, H.: On median graphs: properties, algo- rithms, and applications. IEEE Trans. on Pattern Analysis and Machine Intelligence 23(10), 1144–1151 (2001). https://doi.org/10.1109/34.954604
-
[11]
Pattern Recognition Letters (2018)
Moreno-Garc´ ıa, C.F., Serratosa, F., Jiang, X.: Correspondence edit dis- tance to obtain a set of weighted means of graph correspondences. Pattern Recognition Letters (2018). https://doi.org/10.1016/j.patrec.2018.08.027
-
[12]
Journal of Combinatorial Opti- mization 17(1), 21–44 (2009)
Mukherjee, L., Singh, V., Peng, J., Xu, J., Zeitz, M.J., Berezney, R.: Gen- eralized median graphs and applications. Journal of Combinatorial Opti- mization 17(1), 21–44 (2009). https://doi.org/10.1007/s10878-008-9184-7
-
[13]
European Journal of Operational Research 254(2), 371– 384 (2016)
Musmanno, L.M., Ribeiro, C.C.: Heuristics for the generalized median graph problem. European Journal of Operational Research 254(2), 371– 384 (2016). https://doi.org/10.1016/j.ejor.2016.03.048
-
[14]
Scalable Funding of Bitcoin Micropayment Channel Networks
Nienk¨ otter, A., Jiang, X.: Improved prototype embedding based general- ized median computation by means of refined reconstruction methods. In: Structural, Syntactic, and Statistical Pattern Recognition. pp. 107–117. Springer International Publishing (2016). https://doi.org/10.1007/978-3- 319-49055-7 10
-
[15]
In: Structural, Syn- tactic, and Statistical Pattern Recognition
Rebagliati, N., Sol´ e-Ribalta, A., Pelillo, M., Serratosa, F.: On the relation between the common labelling and the median graph. In: Structural, Syn- tactic, and Statistical Pattern Recognition. pp. 107–115. Springer Berlin Heidelberg (2012). https://doi.org/10.1007/978-3-642-34166-3 12
-
[16]
Advances in Computer Vision and Pattern Recognition, Springer (2015)
Riesen, K.: Structural Pattern Recognition with Graph Edit Distance. Advances in Computer Vision and Pattern Recognition, Springer (2015). https://doi.org/10.1007/978-3-319-27252-8 11
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.