This study examines user engagement with online video content using a multi-task learning approach. In this study, we combine viewing histories, basic user attributes, and content datasets from several public sources to predict both the proportion of a video watched and whether a user skips a video. The two tasks are learned jointly, using a shared representation with separate outputs for regression and classification. Several common multi-task architectures are evaluated and compared under the same experimental setup. Techniques like Multi-Gate Mixture-of-Experts (MMoE), and Progressive Layered Extraction (PLE), and cross stick network were employed. Results of this study on a held-out test set show that watch ratio can be predicted with reasonable accuracy, while skip prediction remains challenging and only marginally better than random guessing. Differences between model architectures are small, suggesting that data size and label definition might have a stronger influenceon performance than model choice. These findings highlight the difficulty of modeling discreteengagement outcomes from noisy behavioral data and point to the importance of careful labelconstruction in future work. Especially, this study highlights the challenges of prediction of skip prediction due to likely reason of subjectively setting the threshold.