The Evolution of Multi-modal Recommendation Algorithms for Short Videos

The rapid proliferation of short video platforms has necessitated the evolution of recommendation systems beyond traditional unimodal approaches. This survey comprehensively analyzes the advancements, challenges, and future directions of multi-modal recommendation algorithms tailored for short videos. Unlike conventional methods reliant on singular data sources (e.g., user logs or text), multi-modal systems integrate visual, audio, textual, and behavioral signals to address critical limitations such as data sparsity, cold starts, and dynamic user intent. We systematically categorize multi-modal fusion strategiesdata-level, feature-level, and decision-leveland highlight their applications in short video scenarios, including cross-modal alignment (e.g., contrastive learning), audio-visual synchronization, and behavior-driven personalization. Key challenges span data noise, computational inefficiency, interest drift, and privacy risks while emerging trends emphasize lightweight fusion, generative AI-driven content synthesis, and explainable recommendation mechanisms. By synthesizing cutting-edge frameworks like M3CSR and SVARM, this review underscores how multi-modal techniques enhance recommendation accuracy, user engagement, and platform competitiveness. Future research must prioritize efficient edge-compatible architectures, ethical AI governance, and interdisciplinary innovations to sustain the growth of short video ecosystems.