Compressed Video Aggregator for Efficient Micro-Video Recommendation
A team of researchers has unveiled a new tool called the Compressed Video Aggregator (CVA), which focuses on enhancing micro-video recommendations by keeping video content separate from preference learning. This module uses a technique called masked mean pooling to turn static VFM frame embeddings into a semantic-consensus anchor, which is then converted into a smaller latent space. It undergoes further refinement through residual self-attention and feedforward layers, resulting in a single video embedding for recommendations. To reduce redundancy, titles help in selecting important frames using CLIP. Tests on the MicroLens and Short-Video datasets show notable improvements, along with significant cuts in training time and GPU memory, while also improving performance across all methods, including CVA.
Key facts
- CVA decouples video information from preference learning.
- It uses masked mean pooling to summarize frozen VFM frame embeddings.
- Key frames are re-selected based on CLIP using titles.
- Experiments on MicroLens and Short-Video datasets show gains.
- Training time and GPU memory are reduced by orders of magnitude.
- Re-selected frames improve all methods including CVA.
- The paper discusses erroneous title scenarios.
- CVA produces a single video embedding for the recommender.
Entities
—