Graph neural networks (GNNs), similarly to other connectionist models, lack transparency in their decision-making. A number of sub-symbolic approaches, such as generating importance masks, have been developed to provide insights into the decision making process of such GNNs. These are first important steps on the way to model explainability, but leaving the interpretation of these sub-symbolic explanations to human analysts can be problematic since humans naturally rely on their background knowledge and therefore also their biases about the data and its domain. To overcome this problem we introduce a conceptual approach by suggesting model-level explanation rule extraction through a standard white-box learning method from the generated importance masks.
NeSy’21
application of AI models are more likely to succeed if the evaluation of these models is
focused more on the user’s needs [
While a variety of explainable models for graph neural networks are being developed [
In this position paper, we address this weakness of existing approaches by proposing a post-processing rule-based companion to such a sub-symbolic explainer method, with the conceptual schema shown in Figure 1. Thereby we want to complement the sub-symbolic instance-level explanations with model-level rules. By extracting and aggregating global rule-based explanations through a standard white-box machine learning method from the generated explainer subgraph, we reduce the amount of additional interpretation needed by the user and provide a model-level explanation, that captures explanations about the global behavior of a model by investigating what input patterns can lead to a specific prediction. As an example of such an approach, we developed a novel method known as SUBGREX. We take the output of the state-of-the-art explainer method as input as well as graph-specific attributes such as node distances and network motifs and use decision trees to generate rule-based explanations.
The goal is to generate systems that can provide model-level rule-based explanations, which don’t rely on the user’s understanding of the domain. We can achieve this by combining the results of a sub-symbolic explainer method with a white-box rule generator which satisfies the representation needs for human comprehensibility and reasoning. The rule-based explanation generation isn’t a stand-alone approach, but an add-on postprocessing method in order to enhance the explanations and make them more user-centric. After training a GNN, the GNNs decision making process is interpreted by identifying a sparse receptive field containing influential elements. Our post-processing approach consists of taking these initial symbolic explanations and lifting them to the level of rules. The proposed process can be seen in Algorithm 1 for a node classification task, where an edge mask and node feature mask are generated by a sub-symbolic explainer model , and subsequently rules for edge and node features are created by the white-box models and . The rules are created through a classification process, where the individual edges and features are assigned binary labels “influential" or “not-influential" based on their masking value. The white-box model is trained using the binary masks as the target variable. The generated rules then function as a model-level explanation for the respective classes of the original classification problem and furthermore, introduce a verbalization element that can make explanations more comprehensible for the user. An important part of generating an user-centric explanation is to make explanation customizable and include provenance, e.g. by including information about the domain knowledge utilized by the system to increase user-understandability and acceptability. Such domain knowledge can be included by extracting attributes from the subgraphs depending on the domain and classification task. An example of such inclusion of domain knowledge are graph-specific attributes, such as motifs mapped to the domain language (see Algorithm 1). In a more formalized setting, relational rules could also be included e.g. to represent domain-specific constraints.
To test this conceptual approach, we propose our SUBGREX model, which enables the
generation of model-level explanations. Since the state-of-the-art method outperforms
alternative baseline approaches by 43.0% in explanation accuracy [
The MUTAG dataset, which is comprised of molecule graphs labeled according to their
mutagenic efec [
We have proposed a conceptual vision for how to approach generating enhanced, more user-centric rule-based explanations from sub-symbolic instance-level explanations, which improve model-level understanding. We also report on initial experiments that demonstrate the validity of our method. Even with the rather simple SUBGREX method we show some surprisingly efective results in terms of meaningfulness of explanations and high sensitivity. In further research we plan to evaluate and compare the efectiveness of diferent white-box models including semantic web technologies such as inductive logic programming.
This work was supported by the German Federal Ministry of Economics and Technology (BMWI) in the project RAKI (no. 01MD19012D).