The expressive power and effectiveness of large language models (LLMs) is going to increasingly push intelligent agents towards sub-symbolic models for natural language processing (NLP) tasks in human–agent interaction. However, LLMs are characterised by a performance vs. transparency trade-off that hinders their applicability to such sensitive scenarios. This is the main reason behind many approaches focusing on local post-hoc explanations, recently proposed by the XAI community in the NLP realm. However, to the best of our knowledge, a thorough comparison among available explainability techniques is currently missing, as well as approaches for constructing global post-hoc explanations leveraging the local information. This is why we propose a novel framework for comparing state-of-the-art local post-hoc explanation mechanisms and for extracting logic programs surrogating LLMs. Our experiments—over a wide variety of text classification tasks—show how most local post-hoc explainers are loosely correlated, highlighting substantial discrepancies in their results. By relying on the proposed novel framework, we also show how it is possible to extract faithful and efficient global explanations for the original LLM over multiple tasks, enabling explainable and resource-friendly AI techniques.@en

Although popular and effective, large language models (LLM) are characterised by a performance vs. transparency trade-off that hinders their applicability to sensitive scenarios. This is the main reason behind many approaches focusing on local post-hoc explanations recently proposed by the XAI community. However, to the best of our knowledge, a thorough comparison among available explainability techniques is currently missing, mainly for the lack of a general metric to measure their benefits. We compare state-of-the-art local post-hoc explanation mechanisms for models trained over moral value classification tasks based on a measure of correlation. By relying on a novel framework for comparing global impact scores, our experiments show how most local post-hoc explainers are loosely correlated, and highlight huge discrepancies in their results—their “quarrel” about explanations. Finally, we compare the impact scores distribution obtained from each local post-hoc explainer with human-made dictionaries, and point out that there is no correlation between explanation outputs and the concepts humans consider as salient.

@en

Background and objective:This paper focuses on nutritional recommendation systems (RS), i.e. AI-powered automatic systems providing users with suggestions about what to eat to pursue their weight/body shape goals. A trade-off among (potentially) conflictual requirements must be taken into account when designing these kinds of systems, there including: (i) adherence to experts’ prescriptions, (ii) adherence to users’ tastes and preferences, (iii) explainability of the whole recommendation process. Accordingly, in this paper we propose a novel approach to the engineering of nutritional RS, combining machine learning and symbolic knowledge extraction to profile users—hence harmonising the aforementioned requirements. MethodsOur contribution focuses on the data processing workflow. Stemming from neural networks (NN) trained to predict user preferences, we use CART Breiman et al.(1984) to extract symbolic rules in Prolog Körner et al.(2022) form, and we combine them with expert prescriptions brought in similar form. We can then query the resulting symbolic knowledge base via logic solvers, to draw explainable recommendations. ResultsExperiments are performed involving a publicly available dataset of 45,723 recipes, plus 12 synthetic datasets about as many imaginary users, and 6 experts’ prescriptions. Fully-connected 4-layered NN are trained on those datasets, reaching ∼86% test-set accuracy, on average. Extracted rules, in turn, have ∼80% fidelity w.r.t. those NN. The resulting recommendation system has a test-set precision of ∼74%. The symbolic approach makes it possible to devise how the system draws recommendations. ConclusionsThanks to our approach, intelligent agents may learn users’ preferences from data, convert them into symbolic form, and extend them with experts’ goal-directed prescriptions. The resulting recommendations are then simultaneously acceptable for the end user and adequate under a nutritional perspective, while the whole process of recommendation generation is made explainable.

@en

Building on prior works on explanation negotiation protocols, this paper proposes a general-purpose protocol for multi-agent systems where recommender agents may need to provide explanations for their recommendations. The protocol specifies the roles and responsibilities of the explainee and the explainer agent and the types of information that should be exchanged between them to ensure a clear and effective explanation. However, it does not prescribe any particular sort of recommendation or explanation, hence remaining agnostic w.r.t. such notions. Novelty lays in the extended support for both ordinary and contrastive explanations, as well as for the situation where no explanation is needed as none is requested by the explainee. Accordingly, we formally present and analyse the protocol, motivating its design and discussing its generality. We also discuss the reification of the protocol into a re-usable software library, namely PyXMas, which is meant to support developers willing to build explainable MAS leveraging our protocol. Finally, we discuss how custom notions of recommendation and explanation can be easily plugged into PyXMas.

@en

In the context of Information-Centric Networking, Interest Flooding Attacks (IFAs) represent a new and dangerous sort of distributed denial of service. Since existing proposals targeting IFAs mainly focus on local information, in this paper we propose GNN4IFA as the first mechanism exploiting complex non-local knowledge for IFA detection by leveraging Graph Neural Networks (GNNs) handling the overall network topology.In order to test GNN4IFA, we collect SPOTIFAI, a novel dataset filling the current lack of available IFA datasets by covering a variety of IFA setups, including ~40 heterogeneous scenarios over three network topologies. We show that GNN4IFA performs well on all tested topologies and setups, reaching over 99% detection rate along with a negligible false positive rate and small computational costs. Overall, GNN4IFA overcomes state-of-the-art detection mechanisms both in terms of raw detection and flexibility, and - unlike all previous solutions in the literature - also enables the transfer of its detection on network topologies different from the one used in its design phase.

@en