Interactive video recommendation (IVR) plays a crucial role in enhancing user engagement and satisfaction in short-video platforms. Despite its benefits, IVR suffers from the filter bubble issue, primarily addressed through diversity-based strategies to mitigate overexposure effect. However, such strategies may lead to cognitive dissonance, which can negatively impact overall satisfaction when contents are recommended to the users that deviates from their expectations. Few studies in IVR have adequately considered the potential impact of cognitive dissonance on user experience and offered effective strategies. Therefore, we propose a cognitive dissonance-aware interactive video recommendation (CDIVR) model. Specifically, a cognitive reward model (CRM) is proposed to comprehensively estimate user satisfaction from three aspects: user preference, overexposure effect, and cognitive bias. Additionally, a denoised state tracker (DST) is also employed to the reduce noise interference and accurately track the changes in user states through cognitive state representation and adaptive denoising. Experiments on the KuaiRec and KuaiRand datasets demonstrate that the CDIVR model can effectively alleviate cognitive dissonance issue and outperform the baselines by 17.84% and 11.98% in cumulative user satisfaction, respectively. The code is available at: https://github.com/sana-mine/CDIVR-codes .