This paper highlights the problem that existing methods for generating personalized headlines based on users' past click data fail to account for irrelevant click noise in the clickstream, potentially generating headlines that do not match users' actual preferences. To address this issue, we propose a novel framework, PHG-DIF (Personalized Headline Generation framework via Denoising Fake Interests from Implicit Feedback). PHG-DIF removes clickstream noise through double filtering based on short dwell times and unusual click bursts, and dynamically models users' evolving and multifaceted interests through multi-level temporal fusion to achieve accurate user profiling. Furthermore, we present a new benchmark dataset, DT-PENS, consisting of click data from 1,000 users and approximately 10,000 annotated personalized headlines. Experimental results demonstrate that PHG-DIF significantly mitigates the negative impact of click noise and achieves state-of-the-art performance.
