Multi-Graph Search: Methods & Applications

• Multi-Graph Search is a framework for coordinated search across multiple graph representations that leverages intra- and inter-graph information to optimize performance.
• It encompasses diverse formulations including motion planning, cross-graph representation learning, and matching strategies that ensure global consistency and efficient search.
• Advanced techniques such as partial solution tracking and auxiliary structures like meta-graphs enhance both computational efficiency and theoretical guarantees.

Multi-Graph Search (MGS) denotes a class of graph-centric search, inference, and optimization formulations in which computation is carried out across multiple graphs, multiple graph views, or multiple interacting graph-induced search structures rather than a single fixed graph. In the cited literature, the term is used in several distinct senses: as a search-based motion-planning algorithm that maintains multiple implicit graphs over a configuration space; as cross-graph representation learning and alignment in common latent spaces; as consistent correspondence search across graph collections; as continuous search over dynamic multi-relational graphs; and, in some papers, as an overloaded acronym for specialized optimizers. A plausible common denominator is coordinated search or alignment across several graph structures, with explicit mechanisms for consistency, merging, or information sharing (Mishani et al., 12 Feb 2026).

1. Scope, definitions, and recurring abstractions

One important formulation treats “multi-graph” data as a single node set with shared features $X \in \mathbb{R}^{n \times d}$ and multiple adjacency matrices $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$, where each $A^{(v)}$ represents a different relation or similarity notion on the same entities. In that setting, the central problem is to learn a representation $Z = h(X,\mathcal{A};\Theta)$ that simultaneously exploits intra-graph structure and inter-graph correlations, rather than averaging graphs or processing each graph independently (Jiang et al., 2019). A second formulation begins from a collection of separate graphs $$\{\mathcal{G}_1,\dots,\mathcal{G}_{N_{\mathcal{G}}}\}$$ and seeks globally consistent assignment matrices $\mathbf{X}_{lm}$ across all graph pairs, so that pairwise correspondences satisfy one-to-one constraints and cycle-consistency (Park et al., 2017). A third formulation, explicit in motion planning, defines MGS as a coordinated search over a set of implicit subgraphs $\mathcal{G}=\{G_1,\dots,G_m\}$ rooted at key states, with later merging of initially disconnected subgraphs through feasible transitions (Mishani et al., 12 Feb 2026).

Usage of MGS	Core object	Representative paper
Multi-graph motion planning	Multiple implicit search graphs in configuration space	(Mishani et al., 12 Feb 2026)
Multi-graph representation learning	Same nodes, multiple graph views	(Jiang et al., 2019)
Multiple graph matching	Pairwise assignments with global consistency	(Park et al., 2017)
Dynamic graph query processing	Continuous search on evolving multi-relational graphs	(Choudhury et al., 2013)
Continuous control planning	Layered directed graph with clustered states	(Kujanpää et al., 2022)

Across these formulations, several abstractions recur. One is the maintenance of partial solutions, such as partial correspondences, partial embeddings, or partial paths. Another is an explicit mechanism for coupling local structure with cross-graph information, whether by adversarial alignment, synchronization, similarity-based superedges, or subgraph merging. A third is the use of auxiliary structures—bidomains, SJ-Trees, tessellations, meta-graphs, or multiple OPEN/FOCAL queues—to control combinatorial growth while preserving search quality (Bai et al., 2020).

2. MGS as coordinated search over multiple implicit graphs

In high-dimensional robot motion planning, Multi-Graph Search is a search-based planner that generalizes classical unidirectional and bidirectional search to a multi-graph setting. The robot configuration space $\mathcal{C}\subseteq\mathbb{R}^d$ is discretized into an implicit graph, but MGS does not commit to a single search graph. Instead, it maintains several implicit subgraphs $G_i=(V_i,E_i)$, rooted at the start, the goal when available, and additional intermediate key states selected by a workspace-based attractor method (Mishani et al., 12 Feb 2026).

The anchor graph $G_1$ is rooted at $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$0 and uses focal search with an admissible heuristic $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$1, an inadmissible heuristic $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$2, an OPEN list ordered by $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$3, and a FOCAL list containing states satisfying $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$4. The remaining connect graphs $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$5 use a front-to-front connect heuristic,

$\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$6

so that each subgraph is driven toward the nearest frontier of another subgraph rather than toward a fixed terminal configuration (Mishani et al., 12 Feb 2026). When a collision-free local path exists between frontiers, MGS merges subgraphs, recomputes $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$7-values consistently with respect to the receiving root, and transfers OPEN, FOCAL, and CLOSED status accordingly.

This design yields precise reductions to classical search. If $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$8, MGS degenerates to standard single-root focal search. If $\mathcal{A} = \{A^{(1)},\dots,A^{(m)}\}$9 and the roots are start and goal, it behaves as a bounded-suboptimal bidirectional heuristic search. With more than two roots, it becomes a multi-directional search in which several landmark-rooted frontiers cooperate by connection and merging rather than by independent exploration (Mishani et al., 12 Feb 2026).

The theoretical guarantees are the principal distinction from multi-tree sampling methods. Under an admissible heuristic, the returned path cost is bounded by $A^{(v)}$0, where $A^{(v)}$1 is the optimal cost on the implicit discrete graph. Under a consistent heuristic, each state is expanded at most twice. The algorithm is complete, and the reported experiments on manipulation and mobile manipulation tasks show high success rates together with low variation in path cost, emphasizing deterministic and repeatable behavior rather than probabilistic coverage (Mishani et al., 12 Feb 2026).

3. MGS as multi-graph representation learning and alignment

A different research line uses MGS in a representational sense: learning one search or retrieval space from several graph views. Multiple Graph Adversarial Learning considers the case of one node set, one shared feature matrix, and multiple heterogeneous adjacency matrices. Its generator is a shared-parameter Kipf–Welling style GCN applied separately to each $A^{(v)}$2, producing embeddings $A^{(v)}$3. A discriminator then tries to identify the source graph of each embedding, while the generator tries to make embeddings indistinguishable across graphs. The resulting adversarial objective produces a structure-invariant common subspace, and the concatenated representation $A^{(v)}$4 is then used for semi-supervised prediction (Jiang et al., 2019). In the cited interpretation, this common subspace is directly relevant to multi-graph search because aligned embeddings permit unified nearest-neighbor, ranking, or retrieval operations across graph views.

The heterogeneous case is addressed by the Multi-Graph Meta-Transformer. MGMT assumes several graphs per entity, differing in topology, scale, and semantics, often without shared node identities. Graph Transformer encoders first map each graph to a common latent space. Task-relevant supernodes are then selected by attention, and a meta-graph is built by connecting supernodes across graphs when their latent embeddings are sufficiently similar. Additional Graph Transformer layers on the meta-graph perform joint reasoning over intra- and inter-graph structure. The supernodes and superedges provide built-in interpretability because they highlight influential substructures and cross-graph alignments (Moslemi et al., 30 Jan 2026).

A related architectural search perspective appears in meta-multigraph search on heterogeneous information networks. There, a meta-multigraph generalizes both meta-path and meta-graph by allowing multiple relation types between hyper-nodes. Search is performed over a super-net containing all candidate relation types, with architecture parameters

$A^{(v)}$5

controlling which relation types survive. The paper introduces Partial Message Meta-Multigraph search and Progressive Complex-to-Concise Meta-Multigraph search, the latter using depth-dependent thresholds so that shallow layers retain many relations while deeper layers become increasingly selective (Li et al., 2023). This suggests a broader MGS interpretation in which the searched objects are not only explicit graphs but also graph-structured message-passing templates.

4. MGS as correspondence search, matching, and common substructures

In graph matching, MGS denotes the search for globally consistent correspondences across many graphs. The multi-layer random walks synchronization framework represents each pair of graphs by a multi-layer association graph that preserves multiple attributes separately rather than collapsing them into a single descriptor. A distribution of random walkers is propagated on each pairwise structure, then reweighted per layer, normalized by Sinkhorn scaling, and synchronized across graph pairs by a global layer-confidence vector and permutation synchronization. The factorization

$A^{(v)}$6

encodes cycle-consistent pairwise assignments, and MatchEIG is used to project noisy pairwise estimates to a coherent low-rank structure (Park et al., 2017). In search terms, the algorithm does not solve pairwise matchings independently and repair them later; it injects global consistency constraints during the search itself.

The two-graph Maximum Common Subgraph problem provides another MGS template. GLSearch formulates MCS as branch-and-bound over partial node matchings between $A^{(v)}$7 and $A^{(v)}$8. The search state contains the current mapping, matched subgraphs, and bidomains, where each bidomain $A^{(v)}$9 groups nodes with identical connectivity signatures relative to the current partial solution. The upper bound

$Z = h(X,\mathcal{A};\Theta)$0

drives pruning, while a GNN-based Deep Q-Network replaces hand-crafted branching heuristics when selecting the next node pair to expand (Bai et al., 2020). Although the paper is restricted to two graphs, it explicitly provides a blueprint for multi-graph search: state representations for partial correspondences, constraint-aware partitions of the remaining search space, and a learned policy that prioritizes high-value expansions under a limited budget.

5. Dynamic, continuous, and hierarchical graph search

Continuous graph search on evolving data appears in the StreamWorks system and the closely related exact subgraph search algorithm for dynamic multi-relational graphs. In that setting, a data graph $Z = h(X,\mathcal{A};\Theta)$1 evolves by edge insertions, a query graph $Z = h(X,\mathcal{A};\Theta)$2 is fixed, and the problem is to report subgraphs $Z = h(X,\mathcal{A};\Theta)$3 with $Z = h(X,\mathcal{A};\Theta)$4 and temporal span $Z = h(X,\mathcal{A};\Theta)$5. Incremental evaluation is organized by the Subgraph Join Tree, a binary decomposition of the query graph in which each node stores partial matches of a query subgraph and internal nodes join child matches through a cut-subgraph key (Choudhury et al., 2013). StreamWorks exposes the same mechanism as a continuous-query system over dynamic multi-relational graphs and emphasizes real-time search, semantic filtering, and temporal windows for cyber-security, news, and social-media applications (Choudhury et al., 2013).

In online continuous control, Continuous Monte Carlo Graph Search replaces the Monte Carlo Tree Search tree with a layered directed graph. At each planning time step, similar states are clustered into a limited number of stochastic action bandit nodes, each maintaining a Gaussian action policy and a Gaussian state distribution. Graph width grows when the replay buffer at a layer is large enough, and depth grows only when the final layer has accumulated sufficient evidence. The resulting planner shares action policies across similar states, reducing the branching explosion that affects tree-based methods in continuous state and action spaces (Kujanpää et al., 2022). This is not a multi-graph method in the explicit motion-planning sense of multiple independent subgraphs, but it is a graph-based generalization of tree search in which several state clusters jointly define the planning frontier.

A hierarchical version appears in multi-agent planning for thermalling gliders. There, a lower-level graph $Z = h(X,\mathcal{A};\Theta)$6 encodes visitation sequences for a single glider under a fixed allocation of interest points, and a uniform-cost graph search returns the optimal path under validity and curvature constraints. An upper-level allocation graph $Z = h(X,\mathcal{A};\Theta)$7 assigns interest points to gliders, and a Branch&Bound graph search uses lower-level weak costs as lower bounds on descendants in the allocation tree. The combined “multi level graph-search” is proven optimal and faster than brute-force allocation in simulations (Zaman et al., 2020). This suggests a hierarchical MGS pattern in which several coupled graph searches at different abstraction levels exchange bounds and feasibility information.

6. Acronym overload, adjacent methods, and conceptual boundaries

The literature also uses the acronym MGS for methods that are not literal searches over multiple explicit graphs. In zero-order neural architecture search, ZARTS introduces Multi-Graph Search as an optimizer over architecture parameters $Z = h(X,\mathcal{A};\Theta)$8. The method samples multiple candidate updates $Z = h(X,\mathcal{A};\Theta)$9, evaluates the resulting architecture graphs, and computes an importance-weighted update,

$$\{\mathcal{G}_1,\dots,\mathcal{G}_{N_{\mathcal{G}}}\}$$0

Within ZARTS, MGS is the zero-order method that best balances accuracy and speed, and it is presented as a robust gradient-free counterpart to DARTS rather than as a graph-search planner in the classical sense (Wang et al., 2021).

A different acronym collision appears in constrained simulation optimization, where MGS denotes min-max gradient search. That algorithm is a single-loop primal-dual method operating on the regularized Lagrangian $$\{\mathcal{G}_1,\dots,\mathcal{G}_{N_{\mathcal{G}}}\}$$1, with alternating projected descent-ascent updates and zeroth-order gradient estimators derived from noisy simulations. Its principal result is a finite-time convergence guarantee of $$\{\mathcal{G}_1,\dots,\mathcal{G}_{N_{\mathcal{G}}}\}$$2 to a stationary solution, which under mild conditions corresponds to a KKT point (Jin et al., 5 Jun 2026). This usage is terminologically important because it shares the acronym but not the multi-graph semantics of motion planning, matching, or dynamic graph querying.

Quantum search on bipartite multigraphs introduces yet another specialized meaning. That work studies marked-vertex search on connected bipartite multigraphs and on 2-tessellable graphs, the latter being exactly the line graphs of bipartite multigraphs. Using adapted Szegedy and staggered quantum walks together with the AGJK algorithm as a subroutine, it achieves a quadratic speedup in the number of queries, with complexity $$\{\mathcal{G}_1,\dots,\mathcal{G}_{N_{\mathcal{G}}}\}$$3 in terms of the hitting time of the underlying reversible Markov chain (Bezerra et al., 17 Apr 2025). Closely adjacent is Multi-Objective Graph Heuristic Search for robot co-design, where graph grammar derivations define the search space and a universal multi-objective heuristic function based on graph neural networks guides exploration toward high-quality Pareto fronts (Xu et al., 2021).

Taken together, these usages show that “Multi-Graph Search” is not a single canonical algorithmic object. The cited literature instead supports a more precise taxonomy. In some papers, MGS is literal coordinated search over multiple implicit graphs; in others it is search across graph views, graph correspondences, graph templates, or graph-parameterized action spaces; and in still others it is simply an acronym reused for a different optimization method. The main misconception, therefore, is to treat all occurrences of MGS as interchangeable. A more accurate reading is that the term names a family of graph-structured search ideas whose technical content depends on the domain-specific representation of state, constraint, and cross-graph coupling.