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

<table> <tr> <th rowspan="2">Metric</th> <th colspan="2">Beauty</th> <th colspan="2">Sports</th> <th colspan="2">Toys</th> </tr> <tr> <th>Ours</th> <th>Paper</th> <th>Ours</th> <th>Paper</th> <th>Ours</th> <th>Paper</th> </tr> <tr> <td>Recall@5</td> <td>0.0392</td> <td>0.0454</td> <td>0.0233</td> <td>0.0264</td> <td>0.0396</td> <td>0.0521</td> </tr> <tr> <td>Recall@10</td> <td>0.0594</td> <td>0.0648</td> <td>0.0379</td> <td>0.0400</td> <td>0.0577</td> <td>0.0712</td> </tr> <tr> <td>NDCG@5</td> <td>0.0257</td> <td>0.0321</td> <td>0.0150</td> <td>0.0181</td> <td>0.0270</td> <td>0.0371</td> </tr> <tr> <td>NDCG@10</td> <td>0.0321</td> <td>0.0384</td> <td>0.0197</td> <td>0.0225</td> <td>0.0328</td> <td>0.0432</td> </tr> </table>

References

Recommender Systems with Generative Retrieval

Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation