MINCUT POOLING IN GRAPH NEURAL NETWORKS
Filippo Maria Bianchi
∗
Daniele Grattarola
†
Cesare Alippi
‡
ABSTRACT
The advance of node pooling operations in a Graph Neural Network (GNN) has
lagged behind the feverish design of new graph convolution techniques, and pool-
ing remains an important and challenging endeavor for the design of deep archi-
tectures. In this paper, we propose a pooling operation for GNNs that implements
a differentiable unsupervised loss based on the minCUT optimization objective.
First, we validate the effectiveness of the proposed loss function by clustering
nodes in citation networks and through visualization examples, such as image seg-
mentation. Then, we show how the proposed pooling layer can be used to build a
deep GNN architecture for graph classification.
1 INTRODUCTION
Graph Neural Networks (GNNs) learn how to represent entities and how to combine them, according
to arbitrary relationships that are given as part of the task inputs. State of the art GNNs extend the
convolution operation from regular domains, such as images or time series, to data with arbitrary
topologies and unordered structures described by graphs (Bronstein et al., 2017; Battaglia et al.,
2018; Wu et al., 2019).
A fundamental component in deep learning is the pooling operation, which replaces the output of
convolutions with local summaries of nearby points and is implemented by low-pass operations,
such as maximum or average (Lee et al., 2016). An effective synergy is attained by alternating the
convolution operation, which extrapolates local patterns irrespective of the specific location on the
signal, and pooling, which lets the ensuing convolutions capture broader patterns. Pooling allows
to learn more abstract representations in deeper layers of the network, by discarding information
superfluous for the final inference task and keeps complexity under control avoiding an exponential
growth of intermediate features. However, traditional pooling techniques, which subsample a regular
grid and replace its elements by summary statistics, are not suitable for graphs.
The development of pooling strategies for GNNs has lagged behind the design of new and more
efficient graph convolutions. The reason is mainly the difficulty in defining a coarsened version of
the original graph that supports the pooled signal. A na
¨
ıve pooling in GNNs averages all nodes
features (Li et al., 2015), but has limited flexibility since no further convolutions can be applied
afterwards. Since all nodes are aggregated at once, it is not possible to extract local summaries that
capture heterogeneous behaviors and local structures on the graphs. A solution is to pre-compute
coarsened versions of the original graph and then fit the data to these deterministic structures (Bruna
et al., 2013). This aggregation accounts for the graph connectivity, but ignores the supervised task
objective as well as node the features that evolve during training.
In this paper, we propose a differentiable pooling operation implemented as a neural network layer,
which can be seamlessly combined with other layers that perform message-passing operations such
as graph convolutions (see Fig. 1). The parameters in the pooling layer are learned by combining a
supervised loss with a regularization term, which optimizes a continuous relaxation of the normal-
ized minCUT objective. The proposed minCUTpool yields partitions that i) cluster together nodes
which are similar and strongly connected on the graph, and ii) are optimal in terms of supervised
∗
NORCE Norwegian Research Center (Tromsø, NO), fibi@norceresearch.no
†
Universit
`
a della Svizzera italiana (Lugano, CH), grattd@usi.ch
‡
Universit
`
a della Svizzera italiana (Lugano, CH) and Politecnico di Milano (Milano, IT)
Work in progress.
1
arXiv:1907.00481v1 [cs.LG] 30 Jun 2019