Viewer engagement in videos has consistently low due to excessive advertisements during or before playback, long startup delays, pauses, blurring, poor resolution, and repetitive, intrusive ads. Although many well-established algorithms and techniques claim to improve user engagement, there is no universal solution that addresses all user needs and adapts to varying network conditions. This study elaborates methods for assessing user Quality of Experience (QoE) while watching ad-supported videos by analyzing viewer facial expressions, video metadata, network conditions, and questionnaire feedback. Participants in this subjective study allow facial and screen capture, watch a series of videos with embedded ads, and provide feedback after each session. The collected comments and concerns are then analyzed to understand how advertisements affect engagement, with the goal of making more informed ad placement decisions for future video streaming. Our proposed method, AELIX (Adaptive Ensemble for Imbalance Classification using eXplainable artificial intelligence), based on experimental data and analysis, can infer user QoE and offer advertisers insights on placing ads without diminishing viewer engagement. AELIX achieves a no accuracy of 0.95 and 0.994 with utilization of advanced metrics. And it surpasses our benchmark, ITU-T P.1203, by 55.3%.