[hieuddo]

Description

Integrates DIGER (Differentiable Semantic ID for Generative Recommendation) into Cornac through the NextItemRecommender interface.

• Adds joint semantic-ID codebook and generator training.
• Includes Sentence-T5 and OpenAI feature configurations, example usage, and semantic-ID evaluation.
• Documents retained-checkpoint results alongside TIGER/RPG references, with evaluation differences stated.

DIGER configurations were selected from completed six-candidate learning-rate searches per feature setting. Broader model, tokenizer, and decoding tuning was not exhaustive.

Quick results and comparisons with TIGER/RPG:

Model / recipe	Features	Recall@5	NDCG@5	Recall@10	NDCG@10	Recall@20	NDCG@20
DIGER	ST5	0.041378 ± 0.001322	0.027401 ± 0.000342	0.063021 ± 0.001201	0.034355 ± 0.000252	0.093294 ± 0.003556	0.042015 ± 0.000854
TIGER / PAISCHER_CONFIG, RQVAE	ST5-base	0.0419	0.0270	0.0634	0.0338	0.0953	0.0419
DIGER	OpenAI	0.044135 ± 0.001538	0.028961 ± 0.000630	0.069236 ± 0.001704	0.037026 ± 0.000674	0.105338 ± 0.001701	0.046100 ± 0.000367
TIGER / PAISCHER_CONFIG, rkmeans	OpenAI	0.0405	0.0265	0.0678	0.0353	0.1045	0.0445
RPG / RPG_BEAUTY_CONFIG, graph	OpenAI	0.0533	0.0372	0.0789	0.0454	0.1113	0.0536

Related Issues

Checklist:

• I have added tests.
• I have updated the documentation accordingly.
• I have updated README.md (if you are adding a new model).
• I have updated examples/README.md (if you are adding a new example).