Argumentation meets matrix factorization: A dual perspective for explainable recommendations
Jinfeng Zhong, Elsa Negre · Applied Soft Computing · 2025
Factorization-based models gained prominence during the Netflix Challenge (2007) and have since demonstrated strong performance in predicting user ratings. However, their limited interpretability often hinders users from understanding the rationale behind recommendations. In contrast, argumentation-based methods offer a different perspective: they model human-like reasoning by structuring information as arguments and counterarguments. They excel in explainability but typically fall short in accuracy To address this trade-off, we propose a novel framework, Context-Aware Feature-Attribution Through Argumentation (CA-FATA), which combines the predictive power of matrix factorization with the interpretability of argumentation frameworks. In CA-FATA, each user–item interaction is modeled using an argumentation framework. Items’ features are represented as arguments, and users’ ratings determine the arguments’ strengths. Additionally, the model incorporates users’ contextual information (e.g., time, location) to further improve predictive performance. Empirical evaluations on real-world datasets show that CA-FATA excels in predictive accuracy and interpretability. It outperforms existing argumentation-based methods and achieves comparable results with state-of-the-art context-free and context-aware models. CA-FATA also supports multiple explanation formats, including template-based explanations, interactive feedback, and contrastive reasoning. Furthermore, it alleviates the cold-start problem by clustering users based on feature preferences.