Page 73 - 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

NETXPLAIN: REAL‑TIME EXPLAINABILITY OF
GRAPH NEURAL NETWORKS APPLIED TO NETWORKING

1
1
2
1
2
David Pujol‑Perich , José Suárez‑Varela , Shihan Xiao , Bo Wu , Albert Cabellos‑Aparicio , Pere Barlet‑Ros 1
1 2
Barcelona Neural Networking center, Universitat Politècnica de Catalunya., Network Technology Lab., Huawei
Technologies Co.,Ltd.
NOTE: Corresponding author: David Pujol‑Perich, david.pujol.perich@upc.edu
Abstract – Recent advancements in Deep Learning (DL) have revolutionized the way we can ef iciently tackle complex opti‑
mization problems. However, existing DL‑based solutions are often considered as black boxes with high inner complexity. As a
result, there is still certain skepticism among the networking industry about their practical viability to operate data networks.
In this context, explainability techniques have recently emerged to unveil why DL models make each decision. This paper fo‑
cuses on the explainability of Graph Neural Networks (GNNs) applied to networking. GNNs are a novel DL family with unique
properties to generalize over graphs. As a result, they have shown unprecedented performance to solve complex network
optimization problems. This paper presents NetXplain, a novel real‑time explainability solution that uses a GNN to interpret
the output produced by another GNN. In the evaluation, we apply the proposed explainability method to RouteNet, a GNN
model that predicts end‑to‑end QoS metrics in networks. We show that NetXplain operates more than 3 orders of magnitude
faster than state‑of‑the‑art explainability solutions when applied to networks up to 24 nodes, which makes it compatible with
real‑time applications; while demonstrating strong capabilities to generalize to network scenarios not seen during training.

Keywords – AI/ML for networks, explainability, graph neural networks
1. INTRODUCTION In this context, explainability solutions [4] have recently
emerged as practical tools to interpret systematically the
In recent years, Deep Learning (DL) has revolutionized
decisions produced by DL models. Particularly, these
the way we are able to solve a vast number of problems
recently proposed solutions analyze trained DL models
by inding meaningful patterns on large amounts of data.
from a black‑box perspective (i.e., they only analyze their
This acquired knowledge then enables us to make highly
inputs and outputs) and aim to discover which elements
accurate predictions, leading to systematically outper‑
mainly drive the output produced by these models. As a
forming state‑of‑the‑art solutions in many different prob‑
result, they can eventually determine what are the most
lems [1, 2]. However, in the ield of networking, DL‑based
critical input elements to reach the inal decisions. These
techniques still pose an important technological barrier
kinds of techniques have been intensely examined in the
to achieve market adoption. In general, Machine Learning
ield of computer vision, showing promising results [5].
(ML) solutions provide probabilistic performance guar‑
At the same time, the last few years have seen the explo‑
antees, which typically degrade as the data deviates from
sion of Graph Neural Networks (GNNs) [6], a new neural
the distribution observed during training. Moreover, neu‑
network family that has attracted large interest given its
ral networks have very complex internal architectures, of‑
numerous applications to different ields where the in‑
ten with thousands or even millions of parameters not in‑
formation is fundamentally represented as graphs (e.g.,
terpretable by humans. As a result, they are treated as
chemistry [7], physics [8], biology [9], information sci‑
black boxes [3]. This limits the viability of these solutions
ence [10, 11]). This newly introduced mechanism has
to be applied to networks, as these are critical infrastruc‑
proven, to date, to be the only DL technique capable of
tures where it is essential to deploy fully reliable solu‑
generalizing with high accuracy to graphs of different
tions. Otherwise, a potential miscon iguration could lead
sizes and structures not seen during the training phase.
to temporal service disruptions with serious economic
In this context, GNNs have shown good properties to be
damages for network operators.
applied in the ield of computer networks, as many key
In this vein, we do need mechanisms that can delimit the
components in network control and management prob‑
safe operational ranges of DL models. This makes it fun‑
damental to understand why and in what situations a DL‑ lems are fundamentally represented as graphs (e.g., topol‑
ogy, routing). Indeed, we have already witnessed some
based solution can fail. This can be achieved by producing
successful GNN‑based applications to network modeling
human‑readable interpretations of the decisions made by
and optimization [12, 13, 14, 15]. However, the fact that
these models (e.g., interpret a routing decision given a
we are not able to understand the inner architecture of
traf ic matrix and a network topology). This would not
GNNs presents nowadays a major barrier that may
only enable us to achieve more mature and reliable DL
solutions but also to enhance their performance by mak‑ hinder its adoption in real-world networks.
ing ad-hoc adjustments for a particular network scenario
(e.g., hyper-parameter tuning).
© International Telecommunication Union, 2021 57