This paper introduces PONTE, a human-in-the-loop framework for generating personalized and trustworthy natural language explanations from AI systems. It employs a closed-loop validation and adaptation process to ensure faithfulness, completeness, and stylistic alignment with user preferences, mitigating common issues associated with large language models in Explainable AI.

The paper introduces U-MASK, a novel user-adaptive spatio-temporal masking method that resolves the tension between immediacy, stability, and generalization in personalized mobile AI applications. It unifies short-term adaptation, long-horizon forecasting, and cold-start recommendations by formulating them as a conditional completion problem on partially observed spatio-temporal data, demonstrating significant improvements on real-world mobile datasets.