✨ Official implementation of ADvisor, proposed in the ACL 2026 Industry Track paper:
Pre-Deployment Advertisement Ranking under Data Scarcity via Context-Aware Criteria Generation with VLMs
python main.py --use_cross_brand_for_featurespython main.pyThe expected directory structure is as follows:
data/
├── train_<brand_id>.csv
├── test_<brand_id>.csv
├── caption_dict.json
├── caption_embeddings.pkl
├── brand_embeddings.pkl
└── brand_descriptions.json
Each train_<brand_id>.csv and test_<brand_id>.csv file should contain advertisement-level information and target metric columns used for ranking.
Example:
train_brandA.csv
test_brandA.csv
train_brandB.csv
test_brandB.csvOptional files such as captions, caption embeddings, brand embeddings, and brand descriptions can be used to provide additional context for VLM-based criteria generation and scoring.