TIGER

This is an Unofficial Pytorch Implementation for the paper:

Recommender Systems with Generative Retrieval

Model Architecture

Data Preprocess

Step 1: Decompress the downloaded 5-core reviews and metadata from Amazon Review 2014, which are in the format reviews_Beauty_5.json.gz and meta_Beauty.json.gz. Use the command provided in the TIGER/data/process.ipynb file to perform the decompression.

Step 2: Use main.py in TIGER/rqvae folder to train a rqvae model using semantic embeddings obtained in Step 1.

Step 3: Use generate_code.py in TIGER/rqvae folder to select the best model to generate discrete code for semantic embeddings in Step 1 and padding at the last position to resolve duplicate codes.

Train a T5 encoder-decoder model

Use main.py in TIGER/model folder to train a T5 encoder-decoder model with semantic code.

Experimental Results

Metric	Beauty		Sports		Toys
	Ours	Paper	Ours	Paper	Ours	Paper
Recall@5	0.0392	0.0454	0.0233	0.0264	0.0396	0.0521
Recall@10	0.0594	0.0648	0.0379	0.0400	0.0577	0.0712
NDCG@5	0.0257	0.0321	0.0150	0.0181	0.0270	0.0371
NDCG@10	0.0321	0.0384	0.0197	0.0225	0.0328	0.0432

References

Recommender Systems with Generative Retrieval

Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation