-
PlanE: Representation Learning over Planar Graphs
Authors:
Radoslav Dimitrov,
Zeyang Zhao,
Ralph Abboud,
İsmail İlkan Ceylan
Abstract:
Graph neural networks are prominent models for representation learning over graphs, where the idea is to iteratively compute representations of nodes of an input graph through a series of transformations in such a way that the learned graph function is isomorphism invariant on graphs, which makes the learned representations graph invariants. On the other hand, it is well-known that graph invariant…
▽ More
Graph neural networks are prominent models for representation learning over graphs, where the idea is to iteratively compute representations of nodes of an input graph through a series of transformations in such a way that the learned graph function is isomorphism invariant on graphs, which makes the learned representations graph invariants. On the other hand, it is well-known that graph invariants learned by these class of models are incomplete: there are pairs of non-isomorphic graphs which cannot be distinguished by standard graph neural networks. This is unsurprising given the computational difficulty of graph isomorphism testing on general graphs, but the situation begs to differ for special graph classes, for which efficient graph isomorphism testing algorithms are known, such as planar graphs. The goal of this work is to design architectures for efficiently learning complete invariants of planar graphs. Inspired by the classical planar graph isomorphism algorithm of Hopcroft and Tarjan, we propose PlanE as a framework for planar representation learning. PlanE includes architectures which can learn complete invariants over planar graphs while remaining practically scalable. We empirically validate the strong performance of the resulting model architectures on well-known planar graph benchmarks, achieving multiple state-of-the-art results.
△ Less
Submitted 24 October, 2023; v1 submitted 3 July, 2023;
originally announced July 2023.
-
Shortest Path Networks for Graph Property Prediction
Authors:
Ralph Abboud,
Radoslav Dimitrov,
İsmail İlkan Ceylan
Abstract:
Most graph neural network models rely on a particular message passing paradigm, where the idea is to iteratively propagate node representations of a graph to each node in the direct neighborhood. While very prominent, this paradigm leads to information propagation bottlenecks, as information is repeatedly compressed at intermediary node representations, which causes loss of information, making it…
▽ More
Most graph neural network models rely on a particular message passing paradigm, where the idea is to iteratively propagate node representations of a graph to each node in the direct neighborhood. While very prominent, this paradigm leads to information propagation bottlenecks, as information is repeatedly compressed at intermediary node representations, which causes loss of information, making it practically impossible to gather meaningful signals from distant nodes. To address this, we propose shortest path message passing neural networks, where the node representations of a graph are propagated to each node in the shortest path neighborhoods. In this setting, nodes can directly communicate between each other even if they are not neighbors, breaking the information bottleneck and hence leading to more adequately learned representations. Our framework generalizes message passing neural networks, resulting in a class of more expressive models, including some recent state-of-the-art models. We verify the capacity of a basic model of this framework on dedicated synthetic experiments, and on real-world graph classification and regression benchmarks, and obtain state-of-the art results.
△ Less
Submitted 20 January, 2023; v1 submitted 2 June, 2022;
originally announced June 2022.
-
On cohesive powers of linear orders
Authors:
Rumen Dimitrov,
Valentina Harizanov,
Andrey Morozov,
Paul Shafer,
Alexandra A. Soskova,
Stefan V. Vatev
Abstract:
Cohesive powers of computable structures are effective analogs of ultrapowers, where cohesive sets play the role of ultrafilters. Let $ω$, $ζ$, and $η$ denote the respective order-types of the natural numbers, the integers, and the rationals when thought of as linear orders. We investigate the cohesive powers of computable linear orders, with special emphasis on computable copies of $ω$. If…
▽ More
Cohesive powers of computable structures are effective analogs of ultrapowers, where cohesive sets play the role of ultrafilters. Let $ω$, $ζ$, and $η$ denote the respective order-types of the natural numbers, the integers, and the rationals when thought of as linear orders. We investigate the cohesive powers of computable linear orders, with special emphasis on computable copies of $ω$. If $\mathcal{L}$ is a computable copy of $ω$ that is computably isomorphic to the usual presentation of $ω$, then every cohesive power of $\mathcal{L}$ has order-type $ω+ ζη$. However, there are computable copies of $ω$, necessarily not computably isomorphic to the usual presentation, having cohesive powers not elementarily equivalent to $ω+ ζη$. For example, we show that there is a computable copy of $ω$ with a cohesive power of order-type $ω+ η$. Our most general result is that if $X \subseteq \mathbb{N} \setminus \{0\}$ is a Boolean combination of $Σ_2$ sets, thought of as a set of finite order-types, then there is a computable copy of $ω$ with a cohesive power of order-type $ω+ σ(X \cup \{ω+ ζη+ ω^*\})$, where $σ(X \cup \{ω+ ζη+ ω^*\})$ denotes the shuffle of the order-types in $X$ and the order-type $ω+ ζη+ ω^*$. Furthermore, if $X$ is finite and non-empty, then there is a computable copy of $ω$ with a cohesive power of order-type $ω+ σ(X)$.
△ Less
Submitted 22 February, 2023; v1 submitted 1 September, 2020;
originally announced September 2020.
-
On the Approximability of Weighted Model Integration on DNF Structures
Authors:
Ralph Abboud,
İsmail İlkan Ceylan,
Radoslav Dimitrov
Abstract:
Weighted model counting (WMC) consists of computing the weighted sum of all satisfying assignments of a propositional formula. WMC is well-known to be #P-hard for exact solving, but admits a fully polynomial randomized approximation scheme (FPRAS) when restricted to DNF structures. In this work, we study weighted model integration, a generalization of weighted model counting which involves real va…
▽ More
Weighted model counting (WMC) consists of computing the weighted sum of all satisfying assignments of a propositional formula. WMC is well-known to be #P-hard for exact solving, but admits a fully polynomial randomized approximation scheme (FPRAS) when restricted to DNF structures. In this work, we study weighted model integration, a generalization of weighted model counting which involves real variables in addition to propositional variables, and pose the following question: Does weighted model integration on DNF structures admit an FPRAS? Building on classical results from approximate volume computation and approximate weighted model counting, we show that weighted model integration on DNF structures can indeed be approximated for a class of weight functions. Our approximation algorithm is based on three subroutines, each of which can be a weak (i.e., approximate), or a strong (i.e., exact) oracle, and in all cases, comes along with accuracy guarantees. We experimentally verify our approach over randomly generated DNF instances of varying sizes, and show that our algorithm scales to large problem instances, involving up to 1K variables, which are currently out of reach for existing, general-purpose weighted model integration solvers.
△ Less
Submitted 13 July, 2020; v1 submitted 16 February, 2020;
originally announced February 2020.
-
Cohesive Powers of Linear Orders
Authors:
Rumen Dimitrov,
Valentina Harizanov,
Andrey Morozov,
Paul Shafer,
Alexandra Soskova,
Stefan Vatev
Abstract:
Cohesive powers of computable structures can be viewed as effective ultraproducts over effectively indecomposable sets called cohesive sets. We investigate the isomorphism types of cohesive powers $Π_{C}% \mathcal{L}$ for familiar computable linear orders $\mathcal{L}$. If $% \mathcal{L}$ is isomorphic to the ordered set of natural numbers $\mathbb{N}$ and has a computable successor function, then…
▽ More
Cohesive powers of computable structures can be viewed as effective ultraproducts over effectively indecomposable sets called cohesive sets. We investigate the isomorphism types of cohesive powers $Π_{C}% \mathcal{L}$ for familiar computable linear orders $\mathcal{L}$. If $% \mathcal{L}$ is isomorphic to the ordered set of natural numbers $\mathbb{N}$ and has a computable successor function, then $Π_{C}\mathcal{L}$ is isomorphic to $\mathbb{N}+\mathbb{Q}\times \mathbb{Z}.$ Here, $+$ stands for the sum and $\times $ for the lexicographical product of two orders. We construct computable linear orders $\mathcal{L}_{1}$ and $\mathcal{L}_{2}$ isomorphic to $\mathbb{N},$ both with noncomputable successor functions, such that $Π_{C}\mathcal{L}_{1}\mathbb{\ }$is isomorphic to $\mathbb{N}+% \mathbb{Q}\times \mathbb{Z}$, while $Π_{C}\mathcal{L}_{2}$ is not$.$ While cohesive powers preserve all $Π_{2}^{0}$ and $Σ_{2}^{0}$ sentences, we provide new examples of $Π_{3}^{0}$ sentences $Φ$ and computable structures $% \mathcal{M}$ such that $\mathcal{M}\vDash Φ$ while $Π_{C}\mathcal{M}% \vDash \urcorner Φ.$
△ Less
Submitted 15 January, 2019;
originally announced January 2019.
-
Turing Degrees and Automorphism Groups of Substructure Lattices
Authors:
Rumen Dimitrov,
Valentina Harizanov,
Andrey Morozov
Abstract:
The study of automorphisms of computable and other structures connects computability theory with classical group theory. Among the noncomputable countable structures, computably enumerable structures are one of the most important objects of investigation in computable model theory. In this paper, we focus on the lattice structure of computably enumerable substructures of a given canonical computab…
▽ More
The study of automorphisms of computable and other structures connects computability theory with classical group theory. Among the noncomputable countable structures, computably enumerable structures are one of the most important objects of investigation in computable model theory. In this paper, we focus on the lattice structure of computably enumerable substructures of a given canonical computable structure. In particular, for a Turing degree $\mathbf{d}$, we investigate the groups of $\mathbf{d}$ -computable automorphisms of the lattice of $\mathbf{d}$-computably enumerable vector spaces, of the interval Boolean algebra $\mathcal{B}_{η}$ of the ordered set of rationals, and of the lattice of $\mathbf{d}$ -computably enumerable subalgebras of $\mathcal{B}_{η}$. For these groups we show that Turing reducibility can be used to substitute the group-theoretic embedding. We also prove that the Turing degree of the isomorphism types for these groups is the second Turing jump of $\mathbf{d}$ , $\mathbf{d^{\prime \prime }}$.
△ Less
Submitted 3 November, 2018;
originally announced November 2018.
-
A new formalism for calculation of the partition function of single stranded nucleic acids
Authors:
Roumen A. Dimitrov
Abstract:
A new formalism for calculation of the partition function of single stranded nucleic acids is presented. Secondary structures and the topology of structure elements are the level of resolution that is used. The folding model deals with matches, mismatches, symmetric and asymmetric interior loops, stacked pairs in loop and dangling end regions, multi-branched loops, bulges and single base stackin…
▽ More
A new formalism for calculation of the partition function of single stranded nucleic acids is presented. Secondary structures and the topology of structure elements are the level of resolution that is used. The folding model deals with matches, mismatches, symmetric and asymmetric interior loops, stacked pairs in loop and dangling end regions, multi-branched loops, bulges and single base stacking that might exist at duplex ends or at the ends of helices. Calculations on short and long sequences show, that for short oligonucleotides, a duplex formation often displays a two-state transition. However, for longer oligonucleotides, the thermodynamic properties of the single self-folding transition affects the transition nature of the duplex formation, resulting in a population of intermediate hairpin species in the solution. The role of intermediate hairpin species is analyzed in the case when a short oligonucleotides (molecular beacons) have to reliably identify and hybridize to accessible nucleotides within their targeted mRNA sequences. It is shown that the enhanced specificity of the molecular beacons is a result of their constrained conformational flexibility and the all-or-none mechanism of their hybridization to the target sequence.
△ Less
Submitted 5 October, 2005;
originally announced October 2005.
