Add DIGER Semantic ID recommender

README.md

@@ -150,6 +150,7 @@ The table below lists the recommendation models/algorithms featured in Cornac. E
 
 | Year | Model and Paper | Type | Environment | Example |
 | :--: | --------------- | :--: | :---------: | :-----: |
+| 2026 | [Differentiable Semantic ID for Generative Recommendation (DIGER)](cornac/models/diger), [docs](https://cornac.readthedocs.io/en/stable/api_ref/models.html#module-cornac.models.diger.recom_diger), [paper](https://arxiv.org/abs/2601.19711) | Next-Item / Content-Based | [requirements](cornac/models/diger/requirements.txt), CPU / GPU | [quick-start](examples/diger_example.py) |
 | 2025 | [Generating Long Semantic IDs in Parallel for Recommendation (RPG)](cornac/models/rpg), [docs](https://cornac.readthedocs.io/en/stable/api_ref/models.html#module-cornac.models.rpg.recom_rpg), [paper](https://arxiv.org/abs/2506.05781) | Next-Item / Content-Based | [requirements](cornac/models/rpg/requirements.txt), CPU / GPU | [quick-start](examples/rpg_example.py)
 |      | [Diffusion-based Generative Recommendation Model (DiffGRM)](cornac/models/diffgrm), [docs](https://cornac.readthedocs.io/en/stable/api_ref/models.html#module-cornac.models.diffgrm.recom_diffgrm), [paper](https://arxiv.org/abs/2510.21805) | Next-Item / Content-Based | [requirements](cornac/models/diffgrm/requirements.txt), CPU / GPU | [quick-start](examples/diffgrm_example.py)
 | 2024 | [Comparative Aspects and Opinions Ranking for Recommendation Explanations (Companion)](cornac/models/companion), [docs](https://cornac.readthedocs.io/en/stable/api_ref/models.html#module-cornac.models.companion.recom_companion), [paper](https://lthoang.com/assets/publications/mlj24.pdf) | Hybrid / Sentiment / Explainable | CPU | [quick-start](examples/companion_example.py)

cornac/models/__init__.py

@@ -43,6 +43,7 @@
 from .cvae import CVAE
 from .cvaecf import CVAECF
 from .dmrl import DMRL
+from .diger import DIGER
 from .diffgrm import DiffGRM
 from .dnntsp import DNNTSP
 from .ease import EASE

cornac/models/diger/README.md

@@ -0,0 +1,116 @@
+# DIGER
+
+Cornac implementation of **DIGER** (Differentiable Semantic ID for Generative
+Recommendation, Fu et al., SIGIR 2026,
+[arXiv:2601.19711](https://arxiv.org/abs/2601.19711)). A residual quantizer turns
+item content embeddings into semantic IDs; a T5 generator predicts the next
+item's ID. Frequency-gated Gumbel assignments let recommendation gradients
+update the codebooks alongside the generator.
+
+## Requirements
+
+```bash
+pip install -r cornac/models/diger/requirements.txt
+```
+
+## Usage
+
+Run the [Amazon Beauty example](../../../examples/diger_example.py) with
+precomputed Sentence-T5 embeddings:
+
+```bash
+python examples/diger_example.py --features beauty-st5.npz
+```
+
+The NPZ archive must contain `features` (a finite float32 matrix) and `ids`
+(raw item IDs aligned with its rows), covering every item in the evaluation
+splits. Add `--embedding openai` for precomputed OpenAI embeddings; no API key is
+needed. Without `--features`, the example generates Sentence-T5 embeddings from
+Cornac's item texts (requires `sentence-transformers`). These regenerated
+features may differ from those used for the results below.
+
+Supply embeddings through `FeatureModality` in the Python API:
+
+```python
+from cornac.data import FeatureModality
+from cornac.eval_methods import NextItemEvaluation
+from cornac.models import DIGER
+from cornac.models.diger import BEAUTY_ST5_CONFIG
+
+eval_method = NextItemEvaluation.from_splits(
+    train_data=train,
+    val_data=val,
+    test_data=test,
+    mode="last",
+    item_feature=FeatureModality(features=item_embeddings, ids=item_ids),
+)
+model = DIGER(**{**BEAUTY_ST5_CONFIG, "device": "auto", "seed": 2025})
+```
+
+`BEAUTY_ST5_CONFIG` and `BEAUTY_OPENAI_CONFIG` provide the selected Beauty
+training recipe. Validation NDCG@10 is evaluated every epoch; training stops
+after 15 non-improving epochs and restores the paired generator, tokenizer and
+catalog.
+
+## Implementation notes
+
+- Training uses every history/next-item prefix; validation uses each session's
+  final target.
+- A fresh fit pretrains a feature-width-matched RQ autoencoder (20 warmup epochs,
+  residual Lloyd initialization, 100 reconstruction/VQ epochs), then freezes its
+  encoder. This initialization adapts the released recipe. A bundle from
+  `training.pretrain_tokenizer` can be reused through `initializer` with the same
+  ordered features and tokenizer options.
+- Joint training updates the codebooks and generator together; full-catalog
+  Sinkhorn assignments refresh each epoch. A deterministic suffix resolves
+  semantic-ID collisions. Unconstrained beam search scores valid full addresses;
+  set `n_beams` at least as large as the largest evaluation cutoff and no larger
+  than the codebook vocabulary.
+
+## Results
+
+The DIGER rows report CPU inference through the Cornac wrapper using six retained
+checkpoints on Amazon Reviews 2014 Beauty: 12,068 items, 22,363 test queries,
+beam 20 and history limit 210. Training was not rerun through this wrapper.
+ST5 uses 768-dimensional Sentence-T5 features; OpenAI uses 3072-dimensional
+`text-embedding-3-large` features. Values are mean ± sample SD across joint-training
+seeds 2025/2026/2027 with a fixed seed-2024 initializer; metrics are fractions.
+
+For reference, the table also includes Beauty results from the
+[TIGER](../tiger/README.md#best-tiger-configs-vs-tuned-sasrec-beauty) and
+[RPG](../rpg/README.md#graph-versus-exact-scoring) documentation. DIGER rows
+report three-seed mean ± sample SD; reference rows report a single run (seed 123).
+
+| 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](../tiger/README.md#best-tiger-configs-vs-tuned-sasrec-beauty) | 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](../tiger/README.md#best-tiger-configs-vs-tuned-sasrec-beauty) | OpenAI | 0.0405 | 0.0265 | 0.0678 | 0.0353 | 0.1045 | 0.0445 |
+| [RPG / `RPG_BEAUTY_CONFIG`, graph](../rpg/README.md#graph-versus-exact-scoring) | OpenAI | 0.0533 | 0.0372 | 0.0789 | 0.0454 | 0.1113 | 0.0536 |
+
+TIGER uses history length 20 and constrained beam 50; RPG uses history length
+50 and graph beam 20, versus DIGER's history limit 210 and unconstrained beam
+20. All use Amazon Reviews 2014 Beauty, but identical filtering, feature arrays
+and evaluation populations have not been established. These are reference
+results rather than a controlled comparison.
+
+For each DIGER feature setting, all six learning-rate candidates were evaluated,
+and the configuration with the highest validation NDCG@10 was selected. This
+limited search did not exhaustively explore model, tokenizer or decoding
+settings.
+
+The retained DIGER features, initializers and checkpoints are not distributed
+with the package; the presets alone do not reproduce this evaluation.
+
+## Semantic ID evaluation
+
+Set `semantic_id_metrics=True` on `NextItemEvaluation`, or add
+`--semantic-id-evaluation` to the example, to report SID and CCE Recall/NDCG
+alongside ordinary item metrics from one inference. DIGER's complete semantic
+IDs include a deterministic collision suffix, so each collision class contains
+one item and all three metric views coincide in the reported runs.
+
+Invalid and repeated generated IDs are compacted for semantic-ID metrics. The
+evaluator warns about short beams without expanding them. See the
+[Semantic ID guide](../../../docs/source/api_ref/eval_methods.rst).

cornac/models/diger/__init__.py

@@ -0,0 +1,21 @@
+# Copyright 2026 The Cornac Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""DIGER next-item recommendation."""
+
+from .recom_diger import DIGER
+from .diger_config import BEAUTY_ST5_CONFIG, BEAUTY_OPENAI_CONFIG
+
+__all__ = ["DIGER", "BEAUTY_ST5_CONFIG", "BEAUTY_OPENAI_CONFIG"]