Аннотация
In an increasingly fragmented digital media landscape, understanding the sustained acceptance of legacy media demands models that can effectively align heterogeneous platform representations and audience behaviors. Traditional techniques often struggle with poor generalizability and lack contextual awareness, particularly in cross-platform environments. To address these limitations, we propose MediaAdaptNet, a deep representation alignment framework that leverages a cross-platform attention mechanism to capture acceptance signals of legacy media across diverse online ecosystems. Our architecture includes two pivotal components: LEGEND (Latent Engagement Generative Encoder for Narrative Dynamics) and TRAILS (Temporal Reasoning for Acceptance Inference in Legacy Systems). LEGEND integrates multimodal content encoding, user trait embeddings, and graphical propagation to dynamically model audience acceptance behaviors. It employs topic-guided attention and structured representation projection for user behavior grouping to enable semantically grounded predictions over time. TRAILS complements LEGEND by introducing a graph-based temporal reasoning strategy that tracks cross-episode user-media interactions. It utilizes episodic memory, hierarchical segmentation, and structural alignment to infer stable, causal acceptance patterns while enhancing interpretability. Evaluated on multiple real-world datasets, MediaAdaptNet achieves substantial improvements, with a 12% increase in F1-score and an 18% boost in cross-domain retrieval R-precision over standard baselines. Ablation studies further validate the complementary strengths of LEGEND and TRAILS, demonstrating robust performance even under content noise and platform shifts. Our work contributes a scalable, interpretable, and empirically validated solution for legacy media acceptance discovery, advancing research at the intersection of digital media analytics, audience modeling, and deep learning.