New JAIR Papers Cover Federated Learning, Negative Campaigning, Targeted Explanations

The latest issue of the Journal of Artificial Intelligence Research (JAIR) released three papers that address distinct challenges in AI today.

Formal Logic Inference Guided Uncertainty Quantification for Personalized Federated Learning Authors Guocheng He, Ziyan An, and Meiyi Ma describe LogiCP, a federated learning system that combines Signal Temporal Logic (STL) to identify temporal patterns across distributed sensors with Conformal Prediction (CP) to generate distribution‑free prediction intervals. The approach clusters clients into semantically coherent groups, reducing intra‑cluster heterogeneity while allowing dynamic client assignment at runtime without retraining. Evaluations run on traffic and temperature datasets show improved scalability and personalization compared with Bayesian Neural Networks and clustering‑based methods. This work builds on prior research into BNNs and offers practical benefits for smart grid forecasting and climate impact predictions. ¹

Efficiently Negative: Complexity and Approximations of Targeted Negative Campaigning Avishai Zagoury, Orgad Keller, Avinatan Hassidim, and Noam Hazon present a theoretical model where election outcomes can be altered by demoting non‑favored candidates through negative campaigning. They prove that finding such strategies is easy for Plurality and Veto rules but hard for t‑approval and Borda; they also provide approximation algorithms—a min(t,m‑t)-factor scheme for all t∈{1,…,m‑1} and a 3‑factor algorithm for Borda—alongside inapproximability results when different demotion operations carry varying prices. These findings relate to political science literature showing campaign effectiveness depends on candidate type and demographics, suggesting the need for careful implementation to avoid bias. ¹

Towards Verified and Targeted Explanations through Formal Methods Hanchen David Wang, Diego Lopez, Preston Robinette, Ipek Oguz, Taylor Johnson, and Meiyi Ma introduce ViTaX, a formal explainable AI framework that generates targeted semi‑counterfactual explanations with mathematical guarantees about decision boundaries. Unlike existing XAI methods that rely on heuristic attribution or untargeted robustness checks, ViTaX focuses on high‑risk misclassifications—for example confusing a “Stop” sign with a speed limit—and provides stakeholders confidence that neural network behavior aligns with human expectations while maintaining rigorous proofs. The motivation stems from incidents like the fatal Tesla crash of March 2018 involving Autopilot features, underscoring the necessity of trustworthy explanations in safety‑critical systems. ¹