GROCLM: LLM for Grocery Category Recommendation in E-Commerce
A team of researchers has introduced GROCLM, an advanced language model designed for recommending grocery categories in e-commerce, tackling the complexities of cyclical buying habits and varied user intentions. This model employs a two-stage LoRA-based training approach to embed rebuying signals into its parameters, surpassing traditional prompt-based methods. Additionally, a trie-based constrained decoding technique guarantees valid outputs within a specified category range. Testing on both proprietary production data and a public benchmark demonstrates its consistent superiority over robust baseline models. The findings are documented in arXiv:2607.24764.
Key facts
- GROCLM is a fine-tuned language model for grocery category recommendation.
- It uses a two-stage LoRA-based training strategy.
- The model encodes cyclical purchasing patterns into model parameters.
- A trie-based constrained decoding mechanism ensures valid outputs.
- Experiments were conducted on proprietary production data and a public benchmark.
- GROCLM consistently outperforms strong baselines.
- The paper is available on arXiv with ID 2607.24764.
- The approach targets category-level recommendation as a structured alternative to item-level methods.
Entities
Institutions
- arXiv