Page 74 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

Explainability of GNNs has been recently explored in two In Section 5 we describe NetXplain, the proposed explain‑
main works. A irst work emerging from the ML commu‑ ability solution. Afterward, Section 6 presents an evalu‑
nity [10] analyzes a well‑known GNN model applied to a ation of the accuracy and cost of NetXplain with respect
chemistry problem to quantify the impact of the differ‑ to the state of the art. Finally, Section 7 presents a discus‑
ent input elements (atoms, bonds) on the inal model pre‑ sion on possible applications of the proposed explainabil‑
dictions (molecular properties). Likewise, the network‑ ity method, and Section 8 concludes the paper.
ing community has made a irst attempt to apply a simi‑
lar solution to several network optimization use cases 2. BACKGROUND
[3]. However, both solutions are based on costly iterative
optimization algorithms that need to be executed for each
sample on which we want to obtain interpretations. 2.1 Graph neural networks
Hence, they do not meet the requirements to make com‑
Graph neural networks are a novel neural network fam‑
prehensive analysis over large data sets and, more impor‑
ily designed to operate over graph‑structured data, by
tantly, to be used in real‑time applications. To address
capturing and modeling the inherent patterns in a graph.
these limitations, this paper proposes NetXplain, a novel
This has resulted in an unprecedented predictive power
real‑time explainability solution for GNNs. NetXplain in‑
in many applications where data is structured as graphs.
troduces a novel approach where we use a GNN that
Despite the several variants of GNNs introduced in recent
learns, from Tabula Rasa, how to interpret the outputs of years, in this paper we focus on Message‑Passing Neural
another GNN trained for a speci ic task (e.g., routing op‑
Networks (MPNN) [7], as they represent a generic GNN
timization). NetXplain produces human‑understandable
framework.
interpretations of GNNs comparable to those of state‑of‑
MPNN operates over a graph G, in which nodes ∈ are
the‑art solutions [10, 3]. However, it makes this at a much
characterized with some initial features . First, the

more limited cost. In our evaluation, we apply NetXplain 0
hidden‑state ℎ of each node ∈ are initialized us‑

to RouteNet [12], a GNN model that predicts the per‑path
ing their input node features . Once each element of

delay given a network snapshot as input (i.e., topology + 0
the graph has its hidden‑state ℎ initialized, they proceed

routing + traf ic matrix). To this end, we irst train NetX‑
to the message‑passing phase, which shall be repeated a
plain on a data set with samples produced by Metis [3].
given number of times . Fig. 1 illustrates how the mes‑
This training is done over a data set of limited size –5 to
sage passing phase works. In each iteration of the algo‑
10% of the original data set used in RouteNet [12]. Then,
rithm, every node receives a message from each of its
we validate the generalization power of our GNN‑based neighbors ∈ ( ). In MPNN, messages are generated
method, by applying it to network scenarios fundamen‑
using a function (·) computed with the hidden state of
tally different from those seen during training. The evalu‑
the neighbor node. Then, once every node has received
ation results reveal the feasibility to train NetXplain over
the messages from its immediate neighbors, these mes‑
a small dataset produced by costly explainability solutions
sages are combined with an aggregation function (·) pro‑
(e.g., [10, 3]), and be able to apply it over a wide variety
ducing a ixed‑size output (e.g., an element‑wise summa‑
of network scenarios. This eventually enables us to meet
tion).
the needed requirements to make a comprehensive anal‑
Finally, the algorithm reaches the update phase, in which
ysis of the safe operational range of GNN solutions at a
nodes use the aggregated information received from their
limited cost. In this context, we show that NetXplain far
neighbors to update their own hidden states via the up‑
outperforms state‑of‑the‑art algorithms in terms of com‑
date function (·).
putational cost, running more than 3 orders of magnitude
Formally, the message passing at a given iteration is de‑
faster on average than Metis [3] when applied to samples
ined as:
of three real‑world network topologies up to 24 nodes. (1)
, = (ℎ , ℎ , )

,

As an example, this explainability solution can be used
+1 = ({ | ∈ ( )}) (2)
as follows: given a GNN‑based solution and a network ,

scenario, NetXplain points to the particular network ele‑ ℎ +1 = (ℎ , +1 ) (3)

ments (e.g., devices, links, paths) that mostly affected the
In these equations, functions (·) and (·) can be com‑
output decisions of the GNN model. This can be help‑
puted through a universal function approximator, such as
ful for many different use cases, including: ( ) test &
neural networks (e.g., feed‑forward NN or recurrent NN).
troubleshooting of GNN‑based solutions, ( ) reverse en‑
After message passings, the hidden states of nodes typ‑
gineering, or ( ) improving network optimization solu‑
ically converge to some ixed values [6]. Thus, these i‑
tions.
nal hidden states pass through a readout function (·) that
The remainder of this paper is structured as follows. First,
computes the output of the GNN model. (·) automatically
Section 2 introduces the fundamental principles of GNNs
learns the mapping from hidden‑state representations to
and their application to networking. Then, Section 3
the output labels of the model :
presents the related work on explainability for GNNs.

̂ = (ℎ | ∈ ) (4)

58 © International Telecommunication Union, 2021