These are the most intriguing AI research pieces published this year. It mixes artificial intelligence (AI) innovations with data science. It is ordered chronologically and contains a link to a longer article.

LIMIP: Lifelong Learning to solve Mixed Integer Programs

The Branch-and-Bound technique is commonly used to solve Mixed Integer Programs (MIPs). Recent interest has been drawn to Learning to copy quick approximations of the expert, strong branching heuristic due to its effectiveness in decreasing the running time for solving MIPs. Existing learning-to-branch approaches, however, presume that all training data is provided in a single training session. This assumption is frequently false, and if training data is delivered continuously over time, existing approaches are susceptible to catastrophic forgetting.

In this paper, the authors investigate the unknown Lifelong Learning to Branch to the Mixed Integer Programs paradigm. LIMIP uses a bipartite Graph Attention Network to transfer MIP instances to embedding spaces to minimise catastrophic forgetting. This rich embedding space prevents catastrophic forgetting by applying knowledge distillation and elastic weight consolidation, in which they acquire the parameters essential to retaining efficacy and are consequently safeguarded against considerable drift. The researchers assess LIMIP on several NP-hard tasks and prove that, when confronted with lifetime learning, LIMIP is up to 50 per cent more effective than existing baselines.

Task and Model Agnostic Adversarial Attack on Graph Neural Networks

Adversarial attacks on Graph Neural Networks (GNNs) expose inherent security flaws, restricting their use in safety-critical applications. Existing attack tactics, however, need either knowledge of the GNN model in use or of the predictive task being attacked. Is this information essential? For instance, we may utilise a graph for several downstream operations unknown to an attacker. Consequently, it is necessary to evaluate the sensitivity of GNNs to adversarial perturbations in a model- and task-independent setting.

In this paper, the authors investigate this issue and demonstrate that GNNs remain vulnerable even when the downstream job and model are unknown. The presented approach, TANDIS (Targeted Attack via Neighborhood DIStortion), indicates that the distortion of node neighbourhoods can significantly degrade prediction performance. Although neighbourhood distortion is an NP-hard problem, TANDIS designs a successful heuristic by combining Graph Isomorphism Network and deep Q-learning in a novel way. Extensive studies on real datasets and state-of-the-art models demonstrate that, on average, TANDIS is up to fifty per cent more effective than state-of-the-art procedures while being over one thousand times faster.

Global Counterfactual Explainer for Graph Neural Networks

Applications for graph neural networks (GNNs) include computational biology, natural language processing, and computer security. Since GNNs are black-box machine learning models, there is an increasing need to explain their predictions due to their rising popularity. Counterfactual reasoning, which aims to change the GNN prediction with minimal changes to the input graph, is one technique to overcome this issue. However, existing methods for GNN counterfactual explanation are limited to instance-specific local reasoning. This strategy has two key drawbacks: the inability to provide worldwide remedial policies and the overloading of human cognitive ability with too much data.

The researchers investigate the global explainability of GNNs through global counterfactual reasoning in this paper. Specifically, they wish to identify a limited collection of sample counterfactual graphs that explain all input graphs. To achieve this objective, the researchers suggest GCFExplainer, a novel method driven by vertex-reinforced random walks on an edit map of graphs with a greedy summary. Extensive experiments on real graph datasets demonstrate that the global explanation from GCFExplainer provides important high-level insights into the model's behaviour and achieves a 46.9% improvement in recourse coverage and a 9.5% reduction in recourse cost in comparison to the state-of-the-art local counterfactual explainers.