ASER: Enhancing Sequential Recommendation with Sensory Attributes from Reviews
Researchers have rolled out a new approach called ASER, which stands for Attribute-based Sensory-Enhanced Representation. It aims to improve how items are represented in sequential recommendation systems by pulling in sensory details from product reviews. Using a large language model fine-tuned as a teacher, they extract structured sensory attribute-value pairs—like color: matte black or scent: vanilla—from the reviews. These attributes are then streamlined into a compact student transformer that generates fixed-dimensional sensory embeddings for each item. This method has been tested on five different Amazon categories and is integrated into standard systems like SASRec and BERT4Rec. You can find the research paper on arXiv under the reference 2603.02709.
Key facts
- ASER stands for Attribute-based Sensory-Enhanced Representation
- Uses a large language model fine-tuned as a teacher to extract sensory attribute-value pairs
- Extracted structures are distilled into a compact student transformer
- Produces fixed-dimensional sensory embeddings for each item
- Embeddings are incorporated into SASRec and BERT4Rec architectures
- Evaluated on five Amazon domains
- Paper available on arXiv: 2603.02709
- Sensory attributes include color and scent examples
Entities
Institutions
- arXiv