PhD Dissertation Proposal by LE Thi Phuong | Attribution Learning for Recommendation Systems

ABOUT THE TALK

Transformer-based models are now the dominant paradigm for recommendation, from classifiers that compare products based on their reviews to language models that generate recommendations from natural-language prompts. Their predictions, however, remain opaque, and attribution methods — the standard tool for explaining them — compute importance scores from a single input configuration, making them unstable under variations of the input that preserve its meaning.

This dissertation studies two such variations in recommendation: reversing the order of two compared entities and paraphrasing the instruction prompt describing a user's history. We show that in both cases, semantically equivalent inputs yield inconsistent attributions, undermining the reliability of the explanations. We propose two solutions: a post-hoc explanation framework for comparison classifiers that guarantees identical attributions across input orders by construction and a training-time regularization framework that aligns item-level attributions across prompt templates for sequential recommendation, improving consistency on unseen templates by up to 33% while preserving accuracy. Building on these results, ongoing and future work extends the program from consistency to correctness and utility: a training framework that aligns attributions with behaviorally grounded evidence and an attribution-ensemble approach that uses agreement across prompts to improve recommendation accuracy itself.

ABOUT THE SPEAKER

LE Thi Phuong is a PhD candidate at the School of Computing and Information Systems, Singapore Management University, supervised by Prof. Hady W. Lauw. Her research focuses on interpretability and trustworthy explanations for Transformer-based recommendation and comparison models.