Add TIGER model for Semantic ID recommendation

[hieuddo]

Description

This PR integrates TIGER (Recommender Systems with Generative Retrieval) into cornac, along with dataset utilities to support its content-embedding requirements.

Key Changes

• Added load_text() to cornac/datasets/amazon_review.py to parse product metadata (title, price, brand, categories, and description).
• Added unit tests for metadata parsing in tests/cornac/datasets/test_amazon_review.py.
• TIGER model implementation, using NextItemRecommender interface
  • Included default configurations (paischer_config.py and grid_config.py) matching the settings used in popular benchmarks, to replicate results from the original paper.
• Added examples/tiger_example.py with training pipeline on the Amazon Review (beauty) dataset

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).
• I have updated datasets/README.md (if you are adding a new dataset).

[hieuddo]

I've done some tuning with the three Amazon review datasets that are commonly used in Semantic ID papers.

The table below shows Recall@5 across three datasets, with GRID and PAISCHER configs, comparing to TIGER original paper and GRID benchmark.

Category	Defaults	GRID_CONFIG	PAISCHER_CONFIG	TIGER Paper	GRID published
beauty	0.0217	0.0340	0.0419	0.0454	0.0422
sports	0.0157	0.0188	0.0240	0.0264	0.0236
toys	0.0192	0.0298	0.0362 (+desc)	0.0521	0.0376

Setting: leave_last_out split (identical splits to original paper), mode="last", max_len=20, seed=123. Sentence-T5-base (768-d) item embeddings over TIGER-style Title/Price/Brand/Categories text.

Reproducing TIGER is quite tricky, especially for toys, because there are some references saying for beauty and sports, they use Title/Price/Brand/Categories textual information, while it's not entirely clear with toys (likely only Descriptions). (+desc) in the table means we use all Title/Price/Brand/Categories and descriptions.

Between the two configs, GRID prioritizes speed and sacrifices a bit of performance (uses Adam optimizer for faster convergence; original paper uses Adagrad); while PAISCHER tries to follow closely with the original paper to achieve the best results possible.

More detailed tuning, base encoder (ablation studies in GRID show that using rkmeans results in better performance than rqvae with their training style)

Category	Recipe	Tokenizer	R@5	N@5	R@10	N@10	R@20	N@20	Train(s)
beauty	default	rqvae	0.0217	0.0132	0.0386	0.0186	0.0591	0.0238	8072
beauty	grid	rkmeans	0.0340	0.0216	0.0564	0.0288	0.0885	0.0368	4772
beauty	paischer	rqvae	0.0419	0.0270	0.0634	0.0338	0.0953	0.0419	4341
beauty	paischer	rkmeans	0.0404	0.0259	0.0641	0.0335	0.1005	0.0427	9386
sports	default	rqvae	0.0157	0.0098	0.0262	0.0132	0.0442	0.0177	11075
sports	grid	rkmeans	0.0188	0.0117	0.0329	0.0162	0.0540	0.0215	5503
sports	paischer	rqvae	0.0230	0.0149	0.0373	0.0195	0.0596	0.0251	14102
sports	paischer	rkmeans	0.0240	0.0155	0.0394	0.0204	0.0635	0.0265	15917
toys	default	rqvae	0.0192	0.0115	0.0325	0.0158	0.0509	0.0204	6366
toys	grid	rkmeans	0.0298	0.0189	0.0506	0.0256	0.0780	0.0325	4270
toys	paischer	rqvae	0.0320	0.0205	0.0521	0.0269	0.0822	0.0345	9698
toys	paischer	rkmeans	0.0348	0.0222	0.0578	0.0295	0.0922	0.0382	9544
toys	paischer +desc	rkmeans	0.0362	0.0237	0.0583	0.0308	0.0915	0.0391	8609

A few other things to note:

• TIGER likely uses Sentence-T5-XXL embedding, but improvements are minimal (sometimes even decrease)
• beam_search doesn't show improvements over exact scoring
• beauty and sports are somewhat reproducible, but toys is not
• Using OpenAI's text-embedding-3-large embedding (provided by RPG) might improve the performance, but not a lot
• Well-tuned SASRec and BERT4Rec (within cornac) still outperform TIGER. Quick runs from Global temporal and leave-last-out splitting for next-item evaluation #708 (comment) already showed we have better results than SASRec and BERT4Rec's results from TIGER paper. I did some more detailed tuning for SASRec and BERT4Rec, and the results are much higher.

[hieuddo]

@qtuantruong @lthoang We do have all the tuning/experimental results from the notes above. Not sure if we should document it somewhere for reference?