Stefan Duffner

In pursuit of the hidden features of GNN’s internal representations

Abstract

We consider the problem of explaining Graph Neural Networks (GNNs). While most attempts aim at explaining the final decision of the model, we focus on the hidden layers to examine what the GNN actually captures and shed light on the hidden features built by the GNN. To that end, we first extract activation rules that identify sets of exceptionally co-activated neurons when classifying graphs in the same category. These rules define internal representations having a strong impact in the classification process. Then - this is the goal of the current paper - we interpret these rules by identifying a graph that is fully embedded in the related subspace identified by the rule. The graph search is based on a Monte Carlo Tree Search directed by a proximity measure between the graph embedding and the internal representation of the rule, as well as a realism factor that constrains the distribution of the labels of the graph to be similar to that observed on the dataset. Experiments including 6 real-world datasets and 3 baselines demonstrate that our method DISCERN generates realistic graphs of high quality which allows providing new insights into the respective GNN models.

Continue reading

On GNN explainability with activation rules

Abstract

GNNs are powerful models based on node representation learning that perform particularly well in many machine learning problems related to graphs. The major obstacle to the deployment of GNNs is mostly a problem of societal acceptability and trustworthiness, properties which require making explicit the internal functioning of such models. Here, we propose to mine activation rules in the hidden layers to understand how the GNNs perceive the world. The problem is not to discover activation rules that are individually highly discriminating for an output of the model. Instead, the challenge is to provide a small set of rules that cover all input graphs. To this end, we introduce the subjective activation pattern domain. We define an effective and principled algorithm to enumerate activations rules in each hidden layer. The proposed approach for quantifying the interest of these rules is rooted in information theory and is able to account for background knowledge on the input graph data. The activation rules can then be redescribed thanks to pattern languages involving interpretable features. We show that the activation rules provide insights on the characteristics used by the GNN to classify the graphs. Especially, this allows to identify the hidden features built by the GNN through its different layers. Also, these rules can subsequently be used for explaining GNN decisions. Experiments on both synthetic and real-life datasets show highly competitive performance, with up to 200% improvement in fidelity on explaining graph classification over the SOTA methods.

Continue reading

What does my GNN really capture? On exploring internal GNN representations

By Luca Veyrin-Forrer, Ataollah Kamal, Stefan Duffner, Marc Plantevit, Céline Robardet

2022-07-23

In Proceedings of the 31st international joint conference on artificial intelligence (IJCAI’22)

Abstract

GNNs are efficient for classifying graphs but their internal workings is opaque which limits their field of application. Existing methods for explaining GNN focus on disclosing the relationships between input graphs and the model’s decision. In contrary, the method we propose isolates internal features, hidden in the network layers, which are automatically identified by the GNN to classify graphs. We show that this method makes it possible to know the parts of the input graphs used by GNN with much less bias than the SOTA methods and therefore to provide confidence in the decision process.

Continue reading

Qu’est-ce que mon GNN capture vraiment ? Exploration des représentations internes d’un GNN

By Luca Veyrin-Forrer, Ataollah Kamal, Stefan Duffner, Marc Plantevit, Céline Robardet

2022-03-24

In Extraction et gestion des connaissances, EGC 2022, blois, france, 24 au 28 janvier 2022

Abstract

While existing GNN’s explanation methods explain the decision by studying the output layer, we propose a method that analyzes the hidden layers to identify the neurons that are co-activated for a class. We associate to them a graph.

Continue reading