Page 78 - 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

5.3 Generalization power of NetXplain Algorithm 1: Architecture of the NetXplain’s ex‑

By analyzing the training process of NetXplain, we iden‑ plainability GNN applied to RouteNet
tify the generation of the data set as the most computa‑ input : x , x ,ℛ

tionally costly task, even when considering that the num‑ output: ℎ , ℎ , ℎ ,

ber of selected samples of is only a small portion of the begin
original data set . // Initialize states of paths and links
Note that our proposal aims to learn how to explain po‑ foreach ∈ ℛ do ℎ ← [ , 0 … , 0] ;
0

tentially any sample that our target GNN could face dur‑ foreach ∈ do ℎ ← [ , 0 … , 0] ;
0
ing operation. This motivates our choice of using a GNN
to explain a target GNN. To the best of our knowledge, for = 1 to do
// Message passing from links to paths
GNNs are the only technique that offers high generaliza‑ foreach ∈ ℛ do
tion power over graph‑structured data. As a result, once −1
trained, the GNN explainability model generalizes to net‑ = {ℎ | ∈ }

−1
work scenarios not present in its training data set . This ℎ ← (ℎ , )

means that NetXplain’s GNN can be trained over a small end
data set to make predictions of the critical connections // Message passing from paths to links
from the perspective of the target GNN and, once trained, foreach ∈ do

it can predict in one step these critical connections over ← ∑ ∶ ∈ ℎ

arbitrary network scenarios (e.g., topologies of variable ℎ ← (ℎ −1 , )

size and structure). All this while offering an accuracy end
comparable to state‑of‑the‑art costly solutions. end
// Readout function
6. EVALUATION foreach ∈ ℛ do
foreach ∈ do
In this section, we irst evaluate the accuracy of the pre‑ ← (ℎ | ℎ )

,

dictions made by NetXplain with respect to the state‑of‑ ← ( q )
the‑art solutions (Metis [3]). Second, we quantify the end , ,
speed‑up when using NetXplain compared to Metis. In end
our experiments, we train an explainability model that
makes interpretations over RouteNet [12], a GNN model end
used to make QoS inference in networks, previously intro‑ ′
duced in more detail in Section 2.2. this subset needs only ≈ 5% of samples randomly ex‑
All these experiments are evaluated over the same data tracted from the original data set (i.e., approximately
sets used in RouteNet [12], which are publicly available 15k samples) to ensure that NetXplain learns properly.
at [18]. Afterward, we generate with Metis the inal explainabil‑
ity data set , as described in Section 5.2. In this process
′
6.1 Generating the explainability model Metis maps each of the selected samples ∈ to its cor‑
responding mask , using as a target GNN the RouteNet

First, we need to generate the explainability data set and model previouslytrainedonsamples ofNSFNet. Notethat
de ine an architecture for the explainability GNN model: Metis [3] is an iterative optimization algorithm. Hence,
we limit it to run 2,000 iterations per sample, after ob‑
6.1.1 Explainability data set serving this was suf icient to ensure convergence.
Finally, to train our NetXplain model, we make a random
To train a NetXplain explainability model for RouteNet we split of the explainability data set (80%, 10%, and 10%)
irst need to generate the explainability data set (Sec‑ to produce the training, validation, and test data sets re‑
tion 5.1). In this case, we generate this data set using spectively.
Metis [3].
To this end, we irst train RouteNet as the target GNN 6.1.2 Architecture of the explainability GNN
model, using 300k samples simulated in the NSFNet net‑
work, including scenarios with various routing con igura‑ As previously mentioned in Section 5.2, we use for the ex‑
tions and traf ic matrices [18]. plainability GNN a similar architecture to the target GNN,
Before generating the explainability data set , we ran‑ RouteNet [12] in this case. The only change introduced
′
domly sample a subset ⊆ from the original data with respect to the original formulation of RouteNet is in
sets [18]. Note that the different experiments made in the readout function. Algorithm 1 provides a detailed de‑
′
this section use different subsets to generate the ex‑ scription of the NetXplain’s explainability GNN when ap‑
plainability data sets , inally used to train the NetX‑ plied to RouteNet (see scheme of Fig. 2). In this case, the
plain’s GNN models. This is then speci ied in the respec‑ readout function outputs a weight , for each link‑path
tive sections. In general, our experimentation shows that connection ( , ). To this end, we concatenate the corre‑
62 © International Telecommunication Union, 2021