Predicting user behavior on video streaming by using watch-time duration analysis

• Introduce CBCB with sequential (CBCB-S) and revert (CBCB-R) behaviours. • Leverage watch-time duration with user history for short-term prediction. • Outperform VideoReach and UVCAN on Precision, Recall, F1, and Accuracy. • CBCB-R achieves Recall 1.000 and F1-Score 0.990 on JAWWY logs. • Decision Tree is the strongest conventional baseline across datasets. Predicting future user actions is essential for enhancing video recommendation systems. Machine learning-based recommendation models analyze user behavior on video streaming platforms to deliver personalized content suggestions. However, existing approaches often rely solely on user-item interaction history or optimize video-level watch-time, overlooking key implicit factors such as watch-time duration and video intrinsic characteristics, which significantly influence user engagement and content preference. To address these gaps, this study introduces the Content-Based Captivation Behavior (CBCB) framework, a novel two-step approach that enhances short-term behavior prediction in video streaming platforms. First, the User Sequential Captivation Behavior (CBCB-S) component analyzes historical viewing patterns to track engagement trends. Next, the User Revert Captivation Behavior (CBCB-R) component identifies instances where users return to previously viewed content, refining personalized recommendations. By integrating watch-time duration and historical behavioral patterns, CBCB provides a more nuanced understanding of user engagement and captures sequential and revert behaviors at the user level—enabling more precise personalized recommendations than video-level watch-time-based approaches. The proposed method is implemented using machine learning algorithms and evaluated on a real-world JAWWY (Saudi Telecom Company) dataset. Experimental results demonstrate that CBCB achieves high precision, recall, F1-score, and overall accuracy for personalized recommendations. These findings highlight the importance of modeling sequential and revert user behaviors to enhance content personalization, recommendation relevance, and user satisfaction.