In recent years, graph neural networks (GNNs) have emerged as a powerful neural architecture to learn vector representations of nodes and graphs in a supervised, end-to-end fashion. Up to now, GNNs have only been evaluated empirically-showing promising results. The following work investigates GNNs from a theoretical point of view and relates them to the 1-dimensional Weisfeiler-Leman graph isomorphism heuristic (1-WL). We show that GNNs have the same expressiveness as the 1-WL in terms of distinguishing non-isomorphic (sub-)graphs. Hence, both algorithms also have the same shortcomings. Based on this, we propose a generalization of GNNs, so-called k-dimensional GNNs (k-GNNs), which can take higher-order graph structures at multiple scales into account. These higher-order structures play an essential role in the characterization of social networks and molecule graphs. Our experimental evaluation confirms our theoretical findings as well as confirms that higher-order information is useful in the task of graph classification and regression.completing the equivalence. Since the power of the 1-WL has been completely characterized, see, e.g., (Arvind et al. 2015;Kiefer, Schweitzer, and Selman 2015), we can transfer these results to the case of GNNs, showing that both approaches have the same shortcomings.Going further, we leverage these theoretical relationships to propose a generalization of GNNs, called k-GNNs, which are neural architectures based on the k-dimensional WL algorithm (k-WL), which are strictly more powerful than GNNs. The key insight in these higher-dimensional variants is that they perform message passing directly between subgraph structures, rather than individual nodes. This higher-order form of message passing can capture structural information that is not visible at the node-level.Graph kernels based on the k-WL have been proposed in the past (Morris, Kersting, and Mutzel 2017). However, a key advantage of implementing higher-order message passing in GNNs-which we demonstrate here-is that we can design hierarchical variants of k-GNNs, which combine graph representations learned at different granularities in an end-to-end trainable framework. Concretely, in the presented hierarchical approach the initial messages in a k-GNN are based on the output of lower-dimensional k -GNN (with k < k), which allows the model to effectively capture graph structures of varying granularity. Many real-world graphs inherit a hierarchical structure-e.g., in a social network we must model both the ego-networks around individual nodes, as well as the coarse-grained relationships between entire communities, see, e.g., (Newman 2003)-and our experimental results demonstrate that these hierarchical k-GNNs are able to consistently outperform traditional GNNs on a variety of graph classification and regression tasks. Across twelve graph regression tasks from the QM9 benchmark, we find that our hierarchical model reduces the mean absolute error by 54.45% on average. For graph classification, we find that our hierarchical models...
We give a complexity theoretic classification of homomorphism problems for graphs and, more generally, relational structures obtained by restricting the left hand side structure in a homomorphism. For every class C of structures, let HOM(C,−) be the problem of deciding whether a given structure A ∈C has a homomorphism to a given (arbitrary) structure ß. We prove that, under some complexity theoretic assumption from parameterized complexity theory, HOM(C,−) is in polynomial time if and only if C has bounded tree width modulo homomorphic equivalence. Translated into the language of constraint satisfaction problems, our result yields a characterization of the tractable structural restrictions of constraint satisfaction problems. Translated into the language of database theory, it implies a characterization of the tractable instances of the evaluation problem for conjunctive queries over relational databases.
Relational joins are at the core of relational algebra, which in turn is the core of the standard database query language SQL. As their evaluation is expensive and very often dominated by the output size, it is an important task for database query optimisers to compute estimates on the size of joins and to find good execution plans for sequences of joins. We study these problems from a theoretical perspective, both in the worst-case model, and in an average-case model where the database is chosen according to a known probability distribution. In the former case, our first key observation is that the worst-case size of a query is characterised by the fractional edge cover number of its underlying hypergraph, a combinatorial parameter previously known to provide an upper bound. We complete the picture by proving a matching lower bound, and by showing that there exist queries for which the join-project plan suggested by the fractional edge cover approach may be substantially better than any join plan that does not use intermediate projections. On the other hand, we show that in the average-case model, every join-project plan can be turned into a plan containing no projections in such a way that the expected time to evaluate the plan increases only by a constant factor independent of the size of the database. Not surprisingly, the key combinatorial parameter in this context is the maximum density of the underlying hypergraph. We show how to make effective use of this parameter to eliminate the projections.
Many important combinatorial problems can be modeled as constraint satisfaction problems. Hence identifying polynomial-time solvable classes of constraint satisfaction problems has received a lot of attention. In this paper, we are interested in structural properties that can make the problem tractable. So far, the largest structural class that is known to be polynomial-time solvable is the class of bounded hypertree width instances introduced by Gottlob et al. [2002]. Here we identify a new class of polynomial-time solvable instances: those having bounded fractional edge cover number.Combining hypertree width and fractional edge cover number, we then introduce the notion of fractional hypertree width. We prove that constraint satisfaction problems with bounded fractional hypertree width can be solved in polynomial time (provided that a the tree decomposition is given in the input). Together with a recent approximation algorithm for finding such decompositions [Marx 2010a], it follows that bounded fractional hypertree width is now the most general known structural property that guarantees polynomialtime solvability.
Nowhere dense graph classes, introduced by Nešetřil and Ossona de Mendez [30], form a large variety of classes of "sparse graphs" including the class of planar graphs, actually all classes with excluded minors, and also bounded degree graphs and graph classes of bounded expansion.We show that deciding properties of graphs definable in first-order logic is fixed-parameter tractable on nowhere dense graph classes. At least for graph classes closed under taking subgraphs, this result is optimal: it was known before that for all classes C of graphs closed under taking subgraphs, if deciding first-order properties of graphs in C is fixed-parameter tractable, then C must be nowhere dense (under a reasonable complexity theoretic assumption).As a by-product, we give an algorithmic construction of sparse neighbourhood covers for nowhere dense graphs. This extends and improves previous constructions of neighbourhood covers for graph classes with excluded minors. At the same time, our construction is considerably simpler than those.Our proofs are based on a new game-theoretic characterisation of nowhere dense graphs that allows for a recursive version of localitybased algorithms on these classes. On the logical side, we prove a "rank-preserving" version of Gaifman's locality theorem.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.