TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 24, No. 4, August 2026, pp. 1187 1196 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v24i4.27683 1187 Methods of nding the maximum common transiti v e subgraph: experimental comparison Oleg Syche v, Anton Chupinin Department of Softw are for Automated Systems, F aculty of Electronics and Computer Engineering, V olgograd State T echnical Uni v ersity , V olgograd, Russia Article Inf o Article history: Recei v ed No v 20, 2025 Re vised Mar 19, 2026 Accepted Mar 29, 2026 K eyw ords: Backtracking Branch matching Directed ac yclic graph Dynamic programming Heuristic method Maximum common transiti v e graph ABSTRA CT The problem of nding a maximum common subgraph (MCS) in a graph has broad applications in practical domains. Ho we v er , certain sce narios require subgraphs with special properties, such as transiti vity , that must be k ept during b uilding the subgraph. W e formally dene the concept of a transiti v e subgraph, in v estig ate its properties. W e study four dif ferent algorithms for nding the max- imum common transiti v e subgraph (MCTS), compiled a list of tests aim at com- paring graphs after making v arious changes and e v aluated their accurac y and ef cienc y on a set of test cases. Benchmarking on 64 tests ranks the algorithms by scalability and accurac y: branch matching is the most scalable ( > 1000 v er - tices) and accurate (F1: 0.9907). MC S tree search is viable for graphs of up to 250 v ertices (F1: 0.9752). Backtracking is limited to < 30 v ertices (ac- curac y: 0.5625), and brute-force is only feasible for graphs with 10 v ertices, despite its high accurac y (0.9375). W e discuss the adv antages and disadv antages of each method, the test cases where each method demonstrates a non-optimal MCTS, identify the classes on which the methods w ork correctly and found that the branch matching method based on the longest common subsequence (LCS) algorithm performed the best. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ole g Syche v Department of Softw are for Automated Systems, F aculty of Electronics and Computer Engineering V olgograd State T echnical Uni v ersity 28 Lenina Prospect, 400005 V olgograd, Russia Email: oasyche v@gmail.com 1. INTR ODUCTION The problem of nding a maximum common subgraph (MCS) has a wide range of applications. The MCS algorithm ca n be a po werful tool in molecular analysis [1]-[3], chemistry [4], [5], reconguration plan- ning of modular robotic s ystems [6], structural health monitoring [7] tracking [8] and general graph similarity learning [9]-[11]. The problem of nding the MCS is nondeterministic polynomial-time hard (NP-hard) in the gener al case. Ho we v er , there are se v eral approximation methods that can nd not necessarily optimal b ut close to optimal solutions [12]. In addition, there are special classes of graphs for which the problem of nding the MCS can be solv ed in polynomial time [13]. F or e xample, for graphs with unique node labels, the MCS can be found unambiguously using the intersections of the v erte x and edge sets. F or trees, an algorithm based on dynamic programming is used, which relies on bipartite matching of the v ertices of the tw o trees compared [14]. J ournal homepage: http://telk omnika.uad.ac.id Evaluation Warning : The document was created with Spire.PDF for Python.
1188 ISSN: 1693-6930 There is also an approach using quadrati c unconstrained binary optimization (Q UBO) objecti v e functions for nding the MCS . P arallel [15] and genetic [16] approaches ha v e also been used for nding MCS. Additionally , machine learning (ML) is often used in the MCS search problem [17]-[20]. There are also approaches and attempts to inte grate lar ge language models (LLMs) into graph-related tasks [21]. There also e xist specialized MCS with additional properties [22]. One of the less-st ud i ed cases of the MCS problem is the common subgraph that preserv es transiti v e dependencies between nodes when deleting nodes that do not t the conditions of the common subgraph. Preserving transiti v e dependencies in a directed ac yclic graph (D A G) allo ws one to retain relationships l ik e “X precedes Y” e v en when an intermediate node cannot be included in the common subgraph. One of practical cases of preserving transiti vity when b uilding the MCS is the comparison of softw are repositories. Repositories that use branchi ng can be understood as D A G storing the transiti v e relationships of precedence between re visions. The pre vious studies ha v e demonstrated ef cienc y of using the longest common subsequences (LCS) algorithm in teaching programming and foreign languages [23], [24]. Git is an i mportant tool in modern softw are de v elopment, and l earning ho w to use v ersion control tool is a crucial step in training highly skilled information technology (IT) specialists . Numerous studies sho w that using specialized automatic tutors in student education can signicantly accelerate the learning process [25], [26]. But implementing an ef fecti v e tutor requires the ability to nd the maximum common transiti v e subgraph (MCTS) to identify misplaced, missing, and e xtraneous nodes (change sets) in the student’ s answer [27]. In this w ork, we propose a v ariant of the MCS problem that accounts for transiti v e relationships in D A Gs. W e dene the MCTS, compare se v eral methods for nding it, and analyze the comparison results. The paper is structured as follo ws. Section 2 pro vides the k e y denitions necessary to understand the subject. Section 3 describes the prospecti v e methods of nding the MCTS: brute-force search of all possible v ariants, an adaptation of dynamic programming method for nding the m aximum common subtree, a heuristic method based on matching each branch of graph G 1 with the closest branch of graph G 2 and calculating their LCS and backtracking method for unique node label. Section 4 outlines the core test cases and pro vides a rationale for the selected test cases. Section 5 presents the e xperimental results, and section 6 contains the study conclusions. 2. KEY DEFINITIONS In this section, the basic concepts and terminology used throughout the paper are introduced. Denition 1. A gr aph is denoted as G = ( V , E ) , wher e V is the set of vertices (nodes) and E V × V is the set of edg es (connections between vertices). Denition 2. A D A G is a dir ected gr aph that does not contain dir ected cycles. In other w ords, starting from a v erte x v and follo wing the direction of the arcs, it is impossible to return to the original v erte x v . Denition 3. Let G = ( V , E ) be a dir ected gr aph. The complete tr ansitive r emo val of the verte x v is t he oper ation that for all distinct vertices u 1 , . . . , u n V and w 1 , . . . , w m V satisfying: i, j : ( u i , v ) E ( v , w j ) E (1) pr oduces a gr aph G = ( V , E ) wher e: V = V \ { v } , E = ( E \ { ( u i , v ) , ( v , w j ) | i, j } ) { ( u i , w j ) | i, j } (2) That is, when a v erte x v is remo v ed, for each of its ancestors, an edge is created connecting it to each of v erte x v s descendants. Denition 4. A gr aph G 1 = ( V 1 , E 1 ) is called a tr ansitive subgr aph of G = ( V , E ) if ther e e xists a gr aph G 1 isomorphic to G 1 that can be obtained fr om G by applying a nite number (possibly zer o) of tr ansiti ve r emo vals of the verte x. Denition 5. A gr aph G is called the MCTS of gr aphs G 1 and G 2 if: (1) G is a tr ansitive subgr aph of both G 1 and G 2 ; (2) Ther e e xists no other common tr ansitive subgr aph G of G 1 and G 2 with | V ( G ) | > | V ( G ) | . TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1187–1196 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control 1189 Figure 1 sho ws the MCTS of the tw o graphs. Figure 1. MCTS 3. METHOD 3.1. Brute-f or ce sear ch The brute-force method for nding the MCTS is based on analyzing all possible mappings between the v ertices of tw o graphs G 1 = ( V 1 , E 1 ) and G 2 = ( V 2 , E 2 ) . Assuming | V 1 | | V 2 | , the algorithm e xamines all possible biject ions between subsets V 1 V 1 and V 2 V 2 , where | V 2 | = | V 1 | . At the same time, the graph structure is preserv ed: if v 1 , u 1 V 1 and v 2 , u 2 V 2 , then under the biject ion mapping v 1 to v 2 and u 1 to u 2 , the v erte x relations must be maintained: if u 1 is a descendant of v 1 , then u 2 must be a descendant of v 2 . The goal is to select the mapping with the maximum number of correct v erte x bijections. First, the method selects the k v ertices of G 1 and matches them to k v ertices of G 2 . T o search and v erify the number of v alid bijecti v e v erte x mappings for a complete graph matching, it is necessary to tra v erse all v ertices of both graphs using depth-rst search (DFS) or breadth-rst search (BFS) and to store all transiti v e connections for each graph before comparing them. Note that, for labeled graphs, the number of possible mappings can be reduced by requiring label equality for the corresponding v ertices, which can signicantly speed up the solution in practical applications. The brute-force method guarantees an e xact solution for nding the MCTS, as it e xhausti v ely enu- merates all possible subgraphs and selects the optimal one, b ut due to the comple xity of O ( n ! · 2 n ) , where n is the number of v ertices in the lar gest graph, its use is limited to small graphs. The k e y adv antage of e xhausti v e search is its absolute accurac y , though its e xponential comple xity limits practical applications. F or labeled graphs, optimization is possible through v erte x-label matching, which reduces the search space. Despite its computational limitations, this method remains the gold standard for v erifying approximation algorithms. 3.2. Adapting the maximum common subtr ee sear ch method using dynamic pr ogramming to solv e the pr oblem of computing the maximum common transiti v e subtr ee. F or solving problems related to nding the maximum common subtree of tw o trees [14], there e xists a specic algorithm that emplo ys dynamic programming and a maximum-weight matching algorithm. Adapting this method to nd the MCTS could allo w solving a wide range of problems with a lo wer time comple xity that brute-force search, allo wing the analysis of lar ger graphs. The algorithm uses an approach that determines the v alue of a parent v erte x through the v alues of its child v ertices. Three main cases need to be considered: matching v ertices, matching one v erte x with the descendants of the other , and matching the second v erte x with the descendants of the rst. If a v erte x u from graph G 1 can be matched with a v erte x v from graph G 2 , then a set of pairs of v erte x matching from graph Methods of nding the maximum common tr ansitive subgr aph: e xperimental comparison (Ole g Syc he v) Evaluation Warning : The document was created with Spire.PDF for Python.
1190 ISSN: 1693-6930 G 1 to graph G 2 should be created for these v ertices, and the pair ( u, v ) should be added. Ne xt, it is necessary to iterate through all descendants of u and v and construct a weight matrix for each pair of descendants of the v ertices u and v . Then, using an algorithm to nd the ma x i mum weight matching, we determine the best pairs among the descendants. F or the cases of comparing v ertices with descendants, it is necessary to select a pair where the weight of the pair is maximal. Then, for the current pair of v ertices, it is necessary to choose the best result among the three presented. The complete algorithm is sho wn in Algorithm 1. Algorithm 1 Maximum common transiti v e subtree pr ocedur e M A X T R A N S I T S U B T R E E ( u, v , G 1 , G 2 , dp ) l ef t Case 1: match u and v center Case 2: skip v , recurse on children of u r ig ht Case 3: skip u , recurse on children of v if u = v then Only match if labels agree l ef t l ef t { ( u, v ) } Initialize matrix M f or all u C hil dr en ( G 1 , u ) do f or all v C hil dr en ( G 2 , v ) do subg r aph M A X T R A N S I T S U B T R E E ( u , v , G 1 , G 2 , dp ) Add subg r aph to M [ u , v ] end f or end f or match M A X W E I G H T M A T C H I N G ( M , dp ) Find best child pairing l ef t l ef t { match } end if f or all u C hil dr en ( G 1 , u ) do center max( center , M A X T R A N S I T S U B T R E E ( u , v , G 1 , G 2 , dp )) end f or f or all v C hil dr en ( G 2 , v ) do r ig ht max( r i g ht, M A X T R A N S I T S U B T R E E ( u, v , G 1 , G 2 , dp )) end f or dp [ u, v ] max( l ef t, center , r ig ht ) Choose lar gest set r etur n dp [ u, v ] end pr ocedur e Since the Hung arian algorithm (with time comple xity O ( n 3 ) ) is used to nd the maximum weight matching, the total time comple xity of the algorithm is O ( n 5 ) , where n is the number of v ertices in the graph. 3.3. Heuristic method based on matching branches of tw o graphs This heuristic method is based on the idea that dif ferent branches in a Git repository ha v e a di v erse set of changes. Each change can be vie wed as a label on the v ertices (commits). It follo ws that, when tw o graphs ha v e similar characteristics to a Git repository , it is necessary to match each branch of graph G 1 to the branch of graph G 2 that is most similar , such that each branch of G 1 is matched with only one branch from G 2 , and each branch of G 2 is matched with only one branch from G 1 . Subsequently , for each such pair , it is necessary to nd the LCS of v ertices, considering that some v ertices may contain multiple labels, and there may be cases of incomplete label correspondence for v ertices that we include in the LCS. T o this end, it will be necess ary to us e a modie d v ersion of the LCS algorithm s uch that, in addi tion to maximizing the length of the subsequence, it minimizes the number of mismatched labels. L [ i, j ] = 0 , if i = 0 or j = 0 , max( L [ i 1 , j ] , L [ i, j 1] , L [ i 1 , j 1] + v [ i ][ j ]) , if x i can compare y j , max( L [ i 1 , j ] , L [ i, j 1]) , otherwise. (3) Let B ( G ) denote the list of all branches in graph G . Then, the number of matched branch pairs between graphs G 1 and G 2 is equal to min {| B ( G 1 ) | , | B ( G 2 ) |} . The time comple xity of the LCS algorithm is O ( n · m ) , where n and m are the lengths of the sequences; in our case, this is the number of v ertices in the branch. Considering all of this, the o v erall comple xity of the heuristic method is O (min {| B ( G 1 ) | , | B ( G 2 ) |} · n · m ) , which is signicantly lo wer than the time comple xities of the algorithms that were described abo v e. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1187–1196 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control 1191 Some v ertices may belong to se v eral branches at once, causing a single v erte x to participate multi ple times in the LCS search operations ( comple xity O ( n 2 ) ), which implies that the algorithm’ s comple xity is O ( n 2 · k ) , where n is the number of v ertices and k is the number of branches. 3.4. Backtracking method tr ee f or unique node label It is w orth noting that primarily for educational purposes, to teach students ho w to w ork with v ersion control system (VCS), repositories are used where dif fs in commits are unique. These repositories can be vie wed as trees with unique node labels, where each v erte x is assigned e xactly one label, i mplying that for e v ery v erte x v G 1 , there is a unique corresponding v erte x u G 2 . F or each v erte x of each tree, we compute the list of all its ancestors. Then, we perform a DFS algorithm of the rst tree G 1 , proceeding from the root to the lea v es. During this tra v ersal, for each v erte x we e xamine the list of its ancestors in the second tree G 2 and the corresponding branches in G 1 .There is no need to consider branches all of whos e v ertices belong to another branch; formally , we disre g ard an y branch B such that B ̸ = B with V ( B ) V ( B ) . T w o branches are conicting if the y contain a pair of conicting v ertices. V ertices u and v are conict- ing if u is not an ancestor of v in G 1 , whil e u is an ancestor of v in G 2 . F ormally , ( u, v ) are conicting u ̸⪯ G 1 v u G 2 v where G denotes the ancestor relation in tree G . During the backtracking phas e once we ha v e obta ined the list of acti v e branches containing v for each v erte x v , we handle the case where v has tw o or more direct children. If deg + ( v ) 2 , (i.e., v has direct children u 1 , u 2 , . . . , u n with n 2 ), we must select a set of no n- conicting branches each containing one of the children u 1 , u 2 , . . . , u n and attempt to mer ge them into a single branch, which is added to the global branch list. F or this algorithm, the solution is the lar gest branch by cardinality . The time comple xity of the backtracking method is O ( n !) , where n is the number of branches in the graph. 4. TEST SAMPLE When creating the tests necessary to v erify the correctness of the methods for nding the m aximum transiti v e subgraph, we based on the properties of possible practical usage: comparing tw o Git repositories of the same project when learning to use Git. The number of v ertices for the majority of test graphs is 7-12 v ertices (up to 20 v ertices), which is suf cient for testing basic graph transformation cases. W e considered dif ferent types of changes in the graph (which are interpreted as errors). One class of errors dif ferent placement of v erte x labels (i.e., a label being remo v ed from a v erte x, an e xtra label being added to a v erte x, or the label mo v ed to another v erte x). The latter case can be interpreted as a combination of tw o operations: de leting the label from one v erte x and adding it to another . F or the errors of this type, a linear graph (also called a bamboo) is suf cient. The other kind of dif ferences is the dif ferences in the order of v ertices. Examples include deleting an leaf v erte x, deleting a v erte x from the middle of the graph (which has descendant v ertices), adding a v erte x to the middle, adding a v erte x to the end and changing the v erte x placement. There are also more specic changes, such as mer ging se v eral v ertices into one (remo ving a v ert e x and adding all its labels to another v erte x) and splitting a single v erte x i nto a fe w v ertices. F or those changes, directed trees with 2–3 branches are used. This is important for testing the methods’ ability to correctly identify changes in the branching structure. It is also important to v erify if the method correctly handles the mer ging of multiple branches into one and accurately processes other changes when the branches are mer ged. W e also made comple x tests for combinations of the graph changes described abo v e. T able 1 sho ws the main parameters that were combined to create dif ferent tests and their v alues; T able 2 sho ws a part of the test sample using the short names of parameters. T able 1. T est parameters T est parameter Short name V alues Graph type type D A G, directed tree Number of v ertices v ertices 1, 3-8, 12-15, 16-20 Number of branches br anches 1, 2-3 Number of v erte x labels labels 0, 1, 2 or more Number of changes dif f 0, 1, 2 or more Methods of nding the maximum common tr ansitive subgr aph: e xperimental comparison (Ole g Syc he v) Evaluation Warning : The document was created with Spire.PDF for Python.
1192 ISSN: 1693-6930 T able 2. T est cases No. Name G 1 type G 1 v ertices G 1 labels G 1 branches G 2 type G 2 v ertices G 2 labels G 2 branches dif f. 1 Equal trees T ree 10 1-2 1 T ree 10 1-2 1 0 2 Equal D A Gs D A G 14 1 3 D A G 14 1 3 0 3 V erte x deleted in middle T ree 7 1 3 T ree 6 3 1 1 4 Leaf deleted T ree 7 1 3 T ree 6 3 1 1 ... 20 Branch deleted T ree 9 1 3 T ree 7 2 1 2 21 Re v ersed branch T ree 6 1 1 T ree 6 1 6 5 22 D A G branch deleted D A G 11 1 3 T ree 9 2 1 2 ... 35 Label added T ree 10 1 2-3 T ree 10 1 1 1 36 Label missing T ree 10 1 2-3 T ree 10 1 1 1 ... 50 One v erte x split into se v eral T ree 6 1-4 1 T ree 7 1-2 1 2 ... 63 Comple x test 5 T ree 8 2-3 2 T ree 8 2-3 2 4 64 Comple x test 6 T ree 8 1-2 2 T ree 10 2-3 2 4 This test suite co v ers most of the error cases that arise when students w ork independently to master v ersion control systems and learn repository-editing skills using the pro vided tools. 5. RESUL TS This section presents the results of testing the listed algorithms on the listed set of tests that s imulate the results of performing tasks for students to w ork with control systems. A comparison of the methods with the performance of the DeepSeek-V3.2 model is also carried out, for which a prompt w as specically engi- neered to nd MCTS. Metrics such as precision, recall, and F1 will be used to e v aluate the accura c y of MCTS determination. F or the condence interv al of the pass rate metric, the W ilson score interv al with a 95% con- dence le v el is used. F or this purpose, the follo wing v ertices are dened: true positi v e (TP) a pair of v ertices matched by the algorithm that is present in the reference set. F alse positi v e (FP) a pair of v ertices predicted by the algorithm b ut absent in the reference set. F als e ne g ati v e (FN) a pair of v ertices from the reference set not found by the algorithm. T rue ne g ati v e (TN) an y other pair of v ertices (not predicted by the algorithm and not included in the reference set). All tests were conducted on a computer with an Apple M2 processor and 16 GB of random access memory (RAM). The DeepSeek model successfully passed 37 tests (pass rate 0.58). Ho we v er , it frequently made errors when matching graph nodes, indicating certain dif culties in correctly interpreting the structure of graph data or their interconnections. The brute-force method demonstrated a hi gh pass rate (0.94). All f ailed tests had e xcessi v ely long e x ecution times, which necessitated the termination of these tests. Ho we v er , it scales poorly with the graph size. F or instance, on a graph with just 11 v ertices, this method can tak e about 2 minutes to nd an answer , while other methods solv e it in under one second. Furthermore, there is a si gnicant issue with memory consumption. F or a graph of 12 v ertices, the method already requires more than 5 GB of RAM. The backtracking method for tree-based unique node labeling sho wed the lo west pass rate among all algorithms tested (0.56). The primary set of tests where this algorithm incorrectl y identies the maximum transiti v e subgraph in v olv ed graphs with v ertices containing multiple labels and graphs with repeated labels. Additionally , all tests in which the graphs were D A Gs indicate that this algorithm is unsuitable for checking repositories with multiple branch mer ges. The method demonstrates acceptable runtime performance for graphs with a number of v ertices n < 30 . Ho we v er , as n increases be yond this threshold and the graph’ s branching f actor gro ws, the e x ecution time scales accordingly . The ne xt algorithm in terms of pass rate (0.83) is the method adapting dynamic programming for nding the maximum transiti v e subtree. The primary class of problems where this method made errors in v olv es comparing graphs after remo ving a common parent of multiple branches. In this case, the correct solution w ould be to match corr esponding v ertices in the compared graphs, b ut the method instead proposes discarding smaller branches and retaining only the lar gest branch in the matching. Performance testi ng re v ealed that the TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1187–1196 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control 1193 proposed method maintains f ast e x ecution on graphs with the number of v ertices n > 100 , which is comparable to the scale of real-w orld repositories. The heuristic method based on matching each branch of the graph performed best on this test set, achie ving a pass rate of 0.98. It successfully handled tas k s in v olving graphs without branch mer ges, reposito- ries with mer ges, graphs with repeated label s, and graphs with unique label sets. It demonstrated high speed in e x ecuting test cases. The method only performed poorly in a test where tw o or more branches of a graph had identical label sets b ut dif fered in the order of v ert e x arrangement, leading to incorrect matching of cor - responding branches across dif ferent graphs and f ailure to identify the MCTS. This method pro v ed to be the most performance, maintaining f ast e x ecution on graphs with n > 1000 v ertices. All methods are in one w ay or another suited for a specic class of graphs on which the y correctly locate the MCTS. When the MCTS search is incorrect, the methods return not the maximum, b ut one of the common transiti v e subgraphs. It is possible to use se v eral methods at once to search for the MCTS and choose the lar gest output . The T able 3 presents a comprehensi v e comparison of these algorithms. T able 4 and Figure 2 demonstrate the program’ s e x ecution time as a function of the number of v ertices in the graph. Figure 3(a) presents an e xample of tw o graphs on which the backtracking method performs an incorrect search; Figure 3(b) sho ws an e xample of tw o graphs where the adapted dynamic programming method conducts an incorrect search; Figure 3(c) illustrates an e xample of tw o graphs where the heuristic matching-based method f ails to perform a correct search. T able 3. T est results No Method P assed F ailed P ass rate C I 95% Precision Recall F1-score 1 Bruteforce 60 4 0.9375 0.8500–0.9754 1.0000 0.8843 0.9386 2 MCS tree search 53 11 0.8281 0.7178–0.9012 0.9952 0.9560 0.9752 3 Branch matching 63 1 0.9844 0.9167–0.9972 0.9953 0.9861 0.9907 4 Backtracking 36 28 0.5625 0.4409–0.6771 0.9922 0.8796 0.9325 5 deepseek-v3.2 37 27 0.5781 0.4561–0.6913 0.8986 0.8819 0.8902 T able 4. Algorithm performance by problem size (seconds) Number of v ertices Method 9 10 11 25 30 32 34 50 100 200 250 1000 5000 Branch 0.20 0.20 0.20 0.22 0.23 0.22 0.22 0.23 0.22 0.26 0.29 0.64 8.00 MCS tree search 0.03 0.03 0.03 0.07 0.10 0.11 0.10 0.35 6.00 120.0 360.0 Backtracking 0.03 0.03 0.03 0.64 8.00 24.00 105.0 Bruteforce 2.00 20.00 99.00 Figure 2. Performance comparison of four algorithms Methods of nding the maximum common tr ansitive subgr aph: e xperimental comparison (Ole g Syc he v) Evaluation Warning : The document was created with Spire.PDF for Python.
1194 ISSN: 1693-6930 (a) (b) (c) Figure 3. Graphs for which the methods f ail: (a) backtracking, (b) MCS tree search, and (c) branch matching 6. CONCLUSION Finding MCTS is necessary for nding changes in transiti v e D A Gs, which can be used, to v erify students’ w orks when learning to use v ersion control systems lik e Git. W e considered four methods of nd- ing MCTS and designed a test sample to study their ef cienc y when dif ferent changes to graphs were made. Experimental results conrmed the theoretical e xpectation that the w orst-performing methods w ould be the LLM-based method (0.58) and the backtracking method (0.56). The brute-force method achie v ed perfect ac- curac y , b u t only on small graphs ( n < 10) . The best o v erall performance w as demonstrated b y the branch matching method (0.98), which f ailed only on graphs containing highly simi lar branches. The MCS tree search method yielded a solid result (0.83), although it underperformed relati v e to initial e xpectations. Although none of the methods solv ed all test cases, we identied distinct graph classes where each method tends to f ail. That enables dynamic selection or combination of methods based on graph structure for e xample, by multiple xing approaches or running se v eral in parallel and select ing the best result using a “bag of e xperts” strate gy to fur - ther impro v e o v erall accurac y . The studied methods will be used in de v eloping an intelligent tutoring system for teaching v ersion control, which is important for training programmers and IT specialists. Finding MCTS between the correct solution reposi tory and the student’ s repository will allo w detection of missing, e xtraneous, misplaced and wrong commits, which can be sho w to the student with appropriate messages. That will allo w training with feedback without human in the loop, which signicantly increases the number of solv ed tasks and the resulting skills compared to manual v erication. FUNDING INFORMA TION The study w as carried out wi th the support of the Center for Digital Scientic and Educational Projects and De v elopments in the Field of Industrial Articial Intelligence (C2RED-AI) of V olgograd State T echnical Uni v ersity , created as part of the implementation of top-le v el educational programs in the eld of articial intelligence (Agreement No. 70-2025-000756). A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the C o nt rib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1187–1196 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control 1195 Name of A uthor C M So V a F o I R D O E V i Su P Fu Ole g Syche v Anton Chupinin C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data supporting the ndings of this study are openly a v ailable in the GitHub Project Repository at https://github .com/v akhe w1900/master -dis/tree/main REFERENCES [1] M. Stroet et al. , “OfraMP: a fragment-based tool to f acilitate the parametrization of lar ge molecules, J ournal of Computer -Aided Molecular Design , v ol. 37, no. 8, pp. 357–371, 2023, doi: 10.1007/s10822-023-00511-7. [2] L. Schietg at et al. , Automated detection of toxicophores and prediction of mutagenicity using PMCSFG algorithm, Molecular Informatics , v ol. 42, no. 3, p. 2200232, 2023, doi: 10.1002/minf.202200232. [3] N. P arisutham, “Ho w do centrality measures help to predict similarity patterns in molecular chemical structural graphs?” Articial Intellig ence Chemistry , v ol. 1, no. 2, p. 100007, 2023, doi: 10.1016/j.aichem.2023.100007. [4] T .-L. Phan et al. , “Reaction rebalancing: a no v el approach to curating reaction databases, J ournal of Cheminf ormatics , v ol. 16, no. 1, p. 82, 2024, doi: 10.1186/s13321-024-00875-4. [5] Y . Chang et al. , “High-entrop y allo y electrocatalysts screened using machine learning informed by quantum-inspired similarity analysis, Matter , v ol. 7, no. 11, pp. 4099–4113, 2024, doi: 10.1016/j.matt.2024.10.001. [6] C. Zhang, L. Zhou, and Y . Li, “P areto optimal reconguration planning and distrib uted parallel motion control of mobile modular robots, IEEE T r ansactions on Industrial Electr onics , v ol. 71, no. 8, pp. 9255–9264, 2024, doi: 10.1109/TIE.2023.3321997. [7] J. Goslig a, D. Hester , K. W orden, and A. Bunce, “On population-based structural health monitoring for bridges, Mec hanical Systems and Signal Pr ocessing , v ol. 173, p. 108919, 2022, doi: 10.1016/j.ymssp.2022.108919. [8] U. Ahmed, G. Sri v asta v a, Y . Djenouri, and J. C.-W . Lin, “Kno wledge graph based trajectory outlier detection in sustainable smart cities, Sustainable Cities and Society , v ol. 78, p. 103580, 2022, doi: 10.1016/j.scs.2021.103580. [9] Z. Lan, B. Hong, Y . Ma, and F . Ma, “More interpretable graph similarity computation via maximum common subgraph inference, IEEE T r ansactions on Knowledg e and Data Engineering , v ol. 36, no. 11, pp. 6588–6599, 2024, doi: 10.1109/TKDE.2024.3387044. [10] Z. Liu, Y . Chen, N. Liu, J. He, and D. Li, “Graph2re gion: Ef cient graph similarity learning with structure and scale restoration, IEEE T r ansactions on Knowledg e and Data Engineering , v ol. 37, no. 12, pp. 7213–7225, 2025, doi: 10.1109/TKDE.2025.3617461. [11] C . Zou, G. Lu, L. Du, X. Zeng, and S. Lin, “Graph similarity learning for cross-le v el interactions, Information Pr ocessing & Mana g ement , v ol. 62, no. 1, p. 103932, 2025, doi: 10.1016/j.ipm.2024.103932. [12] L. Cardone and S. Quer , “The multi-maxi mum and quasi-maxi mum common subgra ph problem, Computation , v ol. 11, no. 4, p. 69, 2023, doi: 10.3390/computation11040069. [13] D. J. Aldous, “On the lar gest common subtree of random leaf-labeled binary trees, SIAM J ournal on Discr ete Mathematics , v ol. 36, no. 1, pp. 299–314, 2022, doi: 10.1137/20M1347504. [14] A. Gupta and N. Nishimura, “Finding lar gest subtrees and smallest supertrees, Algorithmica , v ol. 21, no. 2, pp. 183–210, 1998, doi: 10.1007/PL00009212. [15] V . V asilchik o v , “Recursi v e-parallel algorithm for solving the maximum common subgraph problem, A utomatic Contr ol and Com- puter Sciences , v ol. 58, no. 7, pp. 827–835, 2024, doi: 10.3103/S0146411624700287. [16] C. V alenti, A genetic approach to the maximum common subgraph problem, in Pr oceedings of the 20th International Confer ence on Computer Systems and tec hnolo gies , 2019, pp. 98–104, doi: 10.1145/3345252.3345272. [17] S. Quer , T . Madeo, A. Calabrese, G. Squillero, and E. Carraro, “Node embedding and cosine similarity for ef cient maximum common subgraph disco v ery , Applied Sciences , v ol. 15, no. 16, p. 8920, 2025, doi: 10.3390/app15168920. [18] I. Ro y , S. Chakrabarti, and A. De, “Maximum common subgraph guided graph retrie v al: late and early interaction netw orks, Advances in Neur al Information Pr ocessing Systems , v ol. 35, pp. 32 112–32 126, 2022. [19] K. S. Y o w , N. Liao, S. Luo, and R. Cheng, “Machine learning for subgraph e xtraction: Methods, applications and challenges, Pr oceedings of the VLDB Endowment , v ol. 16, no. 12, pp. 3864–3867, 2023, doi: 10.14778/3611540.3611571. [20] Z. Y an, C. Ding, L. Ma, L. Cao, and H. Y ou, “Rotated graph similarity computation via graph transformer netw orks, Neur ocom- puting , v ol. 658, p. 131474, 2025, doi: 10.1016/j.neucom.2025.131474. [21] X. Ren, J. T ang, D. Y in, N. Cha wla, and C. Huang, A surv e y of lar ge language models for graphs, in Pr oceedings of the 30th A CM SIGKDD Confer ence on Knowledg e Disco very and Data Mining , 2024, pp. 6616–6626, doi: 10.1145/3637528.3671460. Methods of nding the maximum common tr ansitive subgr aph: e xperimental comparison (Ole g Syc he v) Evaluation Warning : The document was created with Spire.PDF for Python.
1196 ISSN: 1693-6930 [22] E. de Gastines and A. Knippel, “F ormulations for the maximum common edge subgraph problem, Discr ete Applied Mathematics , v ol. 346, pp. 115–130, 2024, doi: 10.1016/j.dam.2023.11.044. [23] O. Syche v , “Questions for teaching phrase b uilding with automatic feedback, Softwar e Impact s , v ol. 15, p. 100461, Mar . 2023, doi: 10.1016/j.simpa.2022.100461. [24] E. No v ozhenina, O. Syche v , O. T opork o v a, and O. Evtushenk o, “T eaching english w ord order with correctwriting softw are, in Computational Science and Its Applications ICCSA 2021 . Springer International Publishing, 2021, p. 681–692, doi: 10.1007/978- 3-030-86970-0 47. [25] A. W . F . K ouam, “The ef fecti v eness of intelligent tutoring systems in supporting st udents with v arying le v els of programming e xperience, Disco ver Education , v ol. 3, no. 1, p. 278, 2024, doi: 10.1007/s44217-024-00385-3. [26] A. Lieb and T . Goel, “Student interaction with Ne wtBot: An LLM-as-tutor Chatbot for Secondary Ph ysics Education, in Extended Abstr acts of the CHI Confer ence on Human F actor s in Computing Systems , ser . CHI EA ’24. Ne w Y ork, NY , USA: Association for Computing Machinery , 2024, doi: 10.1145/3613905.3647957. [27] O. Syche v and D. Mamonto v , Automatic error detection and hint generation in the teaching of formal languages syntax using correctwriting question type for moodle lms, in 2018 3r d Russian-P acic Confer ence on Computer T ec hnolo gy and Applications (RPC) , 2018, pp. 1–4, doi: 10.1109/RPC.2018.8482125. BIOGRAPHIES OF A UTHORS Oleg Syche v recei v ed the M.Sc. de gree in Computer Science and Engineering from V olgograd State T echnical Uni v ersity , R ussia. He is a Associate Professor of Softw are Engineer - ing Department, V olgograd State T echnical Uni v ersity , V olgograd, Russia. He is currently w orking on the design and implementation of intelligent tutoring systems capable of presenting w ork ed e x- amples, determining semantic errors, and pro viding e xplanatory feedback on wh y an answer is in- correct—specically detai ling what subject-domain rules are brok en.His research i nterests in v olv es de v eloping systems to ask follo w-up questions to stimulate student thinking and automatically clas- sify generated questions to minimize direct human in v olv ement in the creation of assessments, all based on a single subject-domain model. His professional e xperience spans 23 years, including 22 years in scientic and pedagogical w ork. He can be contacted at email: oasyche v@gmail.com. Anton Chupinin recei v ed the B.Eng. de gree in Softw are Engineering from V olgograd State T echnical Uni v ersity (VSTU), V olgograd, Russia, in 2024. He is currently a graduate student of the M.Eng. de gree in the same eld. He is also a back end de v eloper in the V olg ablob compan y . His research interests in v olv es de v eloping intelligent tut oring systems for teaching computer science. He can be contacted at email: antchupinin@gmail.com. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1187–1196 Evaluation Warning : The document was created with Spire.PDF for Python.