Page 81 - 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

by another GNN model. In contrast to state‑of‑the‑art [6] Franco Scarselli, Marco Gori, Ah Chung Tsoi,
solutions based on costly optimization algorithms, the Markus Hagenbuchner, and Gabriele Monfardini.
proposed solution can be integrated into network con‑ “The graph neural network model”. In: IEEE Trans‑
trol and troubleshooting systems operating in real time. actions on Neural Networks 20.1 (2008), pp. 61–80.
We tested NetXplain over RouteNet, a GNN model that [7] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley,
predicts per‑source‑destination delays in computer net‑ Oriol Vinyals, and George E Dahl. “Neural message
works, and showed that our solution can produce an out‑ passing for quantum chemistry”. In: arXiv preprint
put equivalent to state‑of‑the‑art solutions with an exe‑ arXiv:1704.01212 (2017).
cution time more than 3 orders of magnitude faster in
[8] Peter Battaglia, Razvan Pascanu, Matthew Lai,
networks up to 24 nodes. Moreover, we discussed the
Danilo Jimenez Rezende, et al. “Interaction net‑
potential applications that can have this GNN‑based ex‑
works for learning about objects, relations and
plainability solution when applied to networking. As fu‑
physics”. In: Advances in neural information pro‑
ture work, it would be interesting to show experimentally
cessing systems (NIPS). 2016, pp. 4502–4510.
the potential applications of the proposed lightweight ex‑
plainability method to different networking use cases, [9] Marinka Zitnik, Monica Agrawal, and Jure
such as those described in Section 7, as well as making Leskovec. “Modeling polypharmacy side ef‑
a deep analysis on the knowledge extracted by NetXplain fects with graph convolutional networks”. In:
on different target GNN models. Bioinformatics 34.13 (2018), pp. i457–i466.
[10] Rex Ying et al. “Graph convolutional neural net‑
ACKNOWLEDGEMENT works for web‑scale recommender systems”. In:
Proceedings of the ACM SIGKDD International Con‑
This work has been supported by the Spanish MINECO ference on Knowledge Discovery & Data Mining.
under contract TEC2017‑90034‑C2‑1‑R (ALLIANCE) and 2018, pp. 974–983.
the Catalan Institution for Research and Advanced Stud‑ [11] Wenqi Fan et al. “Graph neural networks for social
ies (ICREA). recommendation”. In: The ACM World Wide Web
Conference (WWW). 2019, pp. 417–426.
REFERENCES [12] Krzysztof Rusek, José Suárez‑Varela, Albert
Mestres, Pere Barlet‑Ros, and Albert Cabellos‑
[1] David Silver, Aja Huang, Chris J Maddison, Arthur Aparicio. “Unveiling the potential of Graph Neural
Guez, Laurent Sifre, George Van Den Driessche, Ju‑ Networks for network modeling and optimiza‑
lian Schrittwieser, Ioannis Antonoglou, Veda Pan‑ tion in SDN”. In: Proceedings of the 2019 ACM
neershelvam, Marc Lanctot, et al. “Mastering the Symposium on SDN Research. 2019, pp. 140–151.
game of Go with deep neural networks and tree [13] Fabien Geyer and Georg Carle. “Learning and gen‑
search”. In: nature 529.7587 (2016), pp. 484–489.
erating distributed routing protocols using graph‑
[2] Oriol Vinyals, Igor Babuschkin, Wojciech M Czar‑ based deep learning”. In: Proceedings of the 2018
necki, Michaël Mathieu, Andrew Dudzik, Junyoung Workshop on Big Data Analytics and Machine
Chung, David H Choi, Richard Powell, Timo Ewalds, Learning for Data Communication Networks. 2018,
Petko Georgiev, et al. “Grandmaster level in Star‑ pp. 40–45.
Craft II using multi‑agent reinforcement learning”. [14] Hongzi Mao, Malte Schwarzkopf, Shaileshh Bojja
In: Nature 575.7782 (2019), pp. 350–354. Venkatakrishnan, Zili Meng, and Mohammad Al‑
[3] Zili Meng, Minhu Wang, Jiasong Bai, Mingwei Xu, izadeh. “Learning scheduling algorithms for data
Hongzi Mao, and Hongxin Hu. “Interpreting Deep processingclusters”.In:ProceedingsoftheACMSIG‑
Learning‑Based Networking Systems”. In: Proceed‑ COMM. 2019, pp. 270–288.
ings of ACM SIGCOMM. 2020, pp. 154–171.
[15] Paul Almasan, José Suárez‑Varela, Arnau Badia‑
[4] Wojciech Samek, Thomas Wiegand, and Klaus‑ Sampera, Krzysztof Rusek, Pere Barlet‑Ros, and Al‑
Robert Müller. “Explainable arti icial intelli‑ bert Cabellos‑Aparicio. “Deep reinforcement learn‑
gence: Understanding, visualizing and interpret‑ ing meets graph neural networks: Exploring a
ing deep learning models”. In: arXiv preprint routing optimization use case”. In: arXiv preprint
arXiv:1708.08296 (2017). arXiv:1910.07421 (2019).
[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, [16] Arnau Badia‑Sampera, José Suárez‑Varela, Paul Al‑
and Antonio Torralba. “Network dissection: Quan‑ masan, Krzysztof Rusek, Pere Barlet‑Ros, and Al‑
tifying interpretability of deep visual represen‑ bert Cabellos‑Aparicio. “Towards more realistic
tations”. In: Proceedings of the IEEE conference network models based on Graph Neural Networks”.
on computer vision and pattern recognition. 2017, In: Proceedings of the 15th International Conference
pp. 6541–6549. on emerging Networking EXperiments and Tech‑
nologies. 2019, pp. 14–16.

© International Telecommunication Union, 2021 65