<p>As users increasingly turn to on-demand television platforms such as IPTV for leisure, effective recommendation systems that leverage users’ historical viewing sequences have become essential for enhancing user satisfaction and fostering long-term loyalty. However, existing sequential recommendation models exhibit notable limitations in the video-on-demand context. First, they often fail to adequately capture multi-dimensional temporal features, particularly watching duration, within user behavior sequences. Second, they typically overlook the distinct repeat and explore consumption patterns exhibited by users. Third, evaluation metrics tend to emphasize a single aspect – most commonly accuracy – while neglecting other dimensions such as novelty. To address these challenges, we propose TIER, a Temporal Information-Enhanced Repeat-aware sequential recommendation model. TIER innovatively incorporates rich temporal signals, including timestamps, time intervals, and viewing durations, and introduces a repeat-explore preference selection module to differentiate between habitual re-watching and novel content discovery behaviors. The architecture combines an attention mechanism tailored for modeling repeat behaviors with a custom-designed convolutional neural network – featuring both vertical and horizontal convolution – to effectively capture exploratory patterns. We evaluate TIER on a large-scale provincial IPTV dataset (more than 12 million interaction logs). The model achieves a 3.29% improvement in MRR@20 and comparable performance on novelty metrics. Comprehensive subgroup analyses, ablation studies, and case examinations further validate the model’s robustness and practical relevance. This work offers a novel perspective on integrating multi-dimensional temporal features into sequential recommendation and contributes a viable solution for balancing predictive accuracy with users’ demand for fresh content.</p>