Query-Conditioned Graph Propagation

• Query-conditioned graph propagation is a method that conditions the flow of graph information on query-specific vectors, unifying random walks, GNN feature propagation, and conditional inference.
• It utilizes unified mathematical frameworks and scalable algorithms, including randomized AGP and adaptive path filtering, to deliver rigorous error bounds and efficient query resolution.
• Applications range from graph clustering and knowledge graph reasoning to dynamic GNN preprocessing, achieving significant speedups and memory efficiency across varied graph tasks.

Query-conditioned graph propagation refers to a class of algorithms and frameworks in which the process of propagating information on a graph—such as features, labels, or query results—is explicitly conditioned on a user-specified query. The query may encode seeds, subgraphs, node subsets, or relation patterns of interest, and the resulting propagation is tailored to deliver results or representations strictly relevant to the query. This paradigm subsumes classical proximity queries (like random walks or Personalized PageRank), graph neural network feature smoothing, label propagation, localized query answering (e.g., model-driven graph queries restricted to subgraphs), adaptive GNN reasoning, and hybrid conditional inference in statistical network models. Modern research provides both unified mathematical foundations and efficient algorithms, including scalable randomized approaches with theoretical error guarantees, adaptive and learning-augmented path filtering, and fully incremental query engines.

1. Unified Mathematical Frameworks

At the heart of query-conditioned graph propagation are formal frameworks that express a wide range of operations—random walks, proximity scores, GNN feature propagation, and conditional inference—as specific instances of a generalized linear operator or probabilistic model, parameterized by a query object.

• Power Series Propagation: Let $G=(V,E)$ with adjacency $A$ and degree matrix $D$. Query-conditioned propagation is formulated as an infinite-series operator:

$$p = \sum_{i=0}^{\infty} w_i \left(D^{-a}AD^{-b}\right)^i x$$

where $x$ is the query vector (e.g., seed node, personalized teleport vector, input features), and $\{w_i\}$ is a weight sequence encoding the specific proximity or smoothing operator (e.g., $k$-step walk, PageRank, heat kernel, Katz, GNN feature propagation). Varying $x$, $\{w_i\}$, $a$, $b$ recovers a broad spectrum of tasks, all query-specific (Wang et al., 2021).

• Probabilistic Conditioning via Gaussian MRFs: For semi-supervised, feature, and hybrid tasks, a joint Markov random field over labels and features can be constructed. Observing (conditioning on) any subset (the "query") produces a conditional mean for the unknown variables:

$\hat Z_U = -\Omega_{UU}^{-1} \Omega_{UQ} Z_Q$

where $Q$ is the set of observed/query nodes, $U=V\setminus Q$, and $\Omega$ is a block-structured precision matrix encoding both within-node correlations and inter-node Laplacian smoothness (Jia et al., 2021).

• Localized Graph Query Languages and Engines: Query-conditioned propagation also generalizes to declarative query languages, where the user specifies patterns and relevant subgraphs. The propagation semantics are then defined in terms of completeness relative to these user-specified "local" subgraphs or nested query objects (Barkowsky et al., 2024).

2. Algorithmic Techniques

Algorithmic instantiations of query-conditioned propagation are diverse, with several central approaches:

• Approximate Graph Propagation (AGP): AGP offers a randomized "push-and-sample" algorithm that unifies and efficiently approximates a wide range of propagation operators. The key is to maintain, at each step, residue and reserve vectors, with neighbor updates handled via a combination of deterministic ("heavy-edge") pushes and randomized ("light-edge") sampling. For input query $x$, this yields a rigorous $(\epsilon,\delta)$-approximation to the desired query-conditioned propagation vector in output-sensitive time (Wang et al., 2021).
• Adaptive and Learned Propagation Paths: Techniques such as AdaProp incrementally construct the query-conditioned frontier in a knowledge graph, selecting a fixed budget of successors at each step using a learned, query-aware sampling distribution. The propagation path (which nodes and relations to visit at each layer) is actively filtered to maximize target coverage and semantic relevance while avoiding combinatorial explosion (Zhang et al., 2022).
• Dynamic and Incremental Propagation: For settings where the graph and/or queries change over time, incremental data structures (e.g., Localized RETE networks for subgraph matching, or amortized-update AGP-Dynamic algorithms) propagate only changes relevant to the queried subgraph or parameters, often with provable constant-factor time or memory overhead compared to global recomputation (Barkowsky et al., 2024, Zhao et al., 12 Sep 2025).

3. Theoretical Guarantees and Complexity

All principal methods provide explicit formal guarantees tailored to the query-conditioned setting.

• Approximation Bounds: AGP and its dynamic variants achieve error in pseudonorms (e.g., $\|p-\hat p\|_\infty\leq\epsilon$) for all nodes $v$ with $p(v)\geq\delta$, with high probability, under sublinear expected runtime:

$$\mathbb{E}[\text{time}] = O\left( L^2/\epsilon \cdot \sum_{i=1}^L \|Y_i T^i x\|_1 \right)$$

Output sensitivity is attained when most outputs are small, and memory usage is $O(n)$ (Wang et al., 2021, Zhao et al., 12 Sep 2025).

• Local Completeness: For query-localized RETE networks, completeness is defined with respect to the relevant subgraph $H_p$:

$$\forall m \in M_Q(H).\, m(V^Q) \cap H_p \neq \emptyset \implies m \in M$$

The marking mechanism, local navigation structures, and request-projection mechanisms guarantee that all and only matches relevant to $H_p$ are generated, with soundness and local-completeness formally proved (Barkowsky et al., 2024).

• Linear-Time Adaptive Propagation: AdaProp guarantees that the total propagated node set is $O(LK)$ when propagating up to depth $L$ with top-$K$ expansion at each step. Learning-based sampling is statistically justified by end-to-end optimization of a cross-entropy loss over the answer set, connecting the efficiency-accuracy tradeoff directly to query-conditioned path construction (Zhang et al., 2022).

4. Applications Across Domains

Query-conditioned graph propagation is foundational for numerous applications:

• Graph Neural Networks (GNNs): All linear decoupled GNNs (SGC, GDC, APPNP) can be realized as query-conditioned propagation by interpreting each feature column as a query vector, generalizing to massive-scale propagation for mini-batch GNN inference (Wang et al., 2021, Zhao et al., 12 Sep 2025).
• Knowledge Graph Reasoning: Relation-dependent GNN inference for answering $(e_q, r_q, ?)$ triples leverages adaptive, query-conditioned paths to restrict attention to semantically pertinent subgraphs, crucial for scalability and interpretability on large KGs (Zhang et al., 2022).
• Graph Query Systems: Declarative graph query engines, such as those based on RETE, benefit from user- or interaction-conditioned propagation engines to support incremental evaluation restricted to relevant subgraphs or patterns, minimizing global recomputation and memory footprint (Barkowsky et al., 2024).
• Node Proximity and Clustering: Local spectral methods (heat kernel PageRank, PPR) for community detection and local clustering are, in algorithmic terms, queries conditioning on seed nodes or regions of interest. Query-conditioned propagation yields near-optimal local cluster discovery and sweep quality (Wang et al., 2021).

5. Empirical Insights and Comparative Evaluation

Empirical studies validate the practical impact of query-conditioned propagation:

• AGP achieves $10\times$–$20\times$ speedup in GNN preprocessing on graphs with over a billion edges (e.g., Papers100M, Friendster), with negligible (subpercent) reduction in classification accuracy, and routinely halves peak memory compared to global methods (Wang et al., 2021).
• AGP-Dynamic attains up to $177\times$ faster update times in dynamic graphs over AGP-Static, with no loss of approximation quality, and query times within a factor $2\times$–$3\times$ of non-dynamic baselines for equivalent error tolerances (Zhao et al., 12 Sep 2025).
• AdaProp outperforms both full-neighborhood R-GCN/CompGCN and fixed-propagation (NBFNet, RED-GNN) on knowledge graph reasoning tasks, improving mean reciprocal rank and sharply increasing the "target-over-entities" ratio at increasing propagation depth (Zhang et al., 2022).
• Localized RETE—relative to global incremental query evaluation—demonstrates $5\times$–$20\times$ improvements in build and incremental update time, especially for queries and workloads concentrated in user-selected subgraphs or submodels (Barkowsky et al., 2024).

6. Practical Considerations and Extensions

Implementation and adaptation of query-conditioned graph propagation depend on application context:

• Parameter Tuning: The error threshold $\epsilon$, depth $L$, and path-budget $K$ mediate the statistical-accuracy vs. computational-efficiency frontier; increasing any (less stringent constraint) yields faster computation at the expense of approximation.
• Memory and Data Structures: All scalable methods optimize memory through reuse—AGP stores only two $n$-vectors in-core, localized RETE ensures constant-factor memory increase over standard RETE, and AdaProp's frontier is budgeted by $L\times K$.
• Integration into Larger Pipelines: AGP supports parallelization across multiple "query" feature vectors, enabling plug-in to GNN or metric learning pipelines; AdaProp supports end-to-end training of sampler and GNN, and RETE-based engines accommodate incremental updates driven by user or external edits.

A plausible implication is that future research will continue to diversify notions of "query"—including parameter-conditioned, temporally evolving, or functionally abstracted queries—demanding further generalization of efficient, robust query-conditioned propagation schemes.