ScopeRec

Confidence-budgeted adaptation for cold-start generative recommendation.

ScopeRec-C keeps a semantic-ID recommender frozen and makes new items reachable through a content-conditioned suffix distribution. A generated-prefix gate and an explicit probability budget control how much the existing ranking changes.

View the full-resolution architecture | Architecture and implementation details

• No additional trainable weights: adapt through content and probability allocation, with fixed backbone and existing semantic IDs.
• Controlled output scope: mix only after the model actually generates a selected prefix; no correct-target routing or oracle input in recommendation.
• Reversible and reproducible: version-checked artifacts, explicit edit disable, executed GenRecEdit function parity, and audited saved predictions.
• Measured tradeoffs: cold-item gains, warm-item costs, random seeds and confidence intervals are reported together.

Amazon Reviews 2023 Software, released-code-compatible GenRecEdit protocol, 9,386 test requests including 2,226 cold targets. Three training seeds share the same features, SID map and evaluation requests.

ScopeRec-C gains 1.887 percentage points of exact cold Recall@20 over TIGER while retaining 97.56% of warm NDCG@20.

See final results and uncertainty. A separate, locked 2020 temporal confirmation supports the same coverage/retention analysis; its numbers are not pooled with this benchmark. Only final comparisons are public, not grids or intermediate runs.

1. Encode item text and assign fixed semantic IDs; train the base T5 once.
2. Index the active cold items by SID prefix and retain their content vectors.
3. Build a query from the most recent historical item; compute cold-item similarity.
4. After the generated prefix selects a branch, combine its original suffix distribution with a confidence-budgeted content distribution.
5. Continue natural Beam Search and resolve full item codes; edits can be disabled.

The current default is ScopeRec-C, distinct from the original learned low-rank patch. The base distribution and its prefix probabilities stay fixed; the mass budget bounds old-path probability change, not Recall/NDCG degradation.

Technical architecture

Install a suitable PyTorch build, then:

python -m pip install -e '.[ml,dev]'
python -m pytest -q
python -m scoperec results verify
python -m scoperec results figures

Tests and figure rendering work without GPUs, datasets or checkpoints. The included final JSON preserves unrounded measurements and individual seeds.

To reproduce the main experiment on one GPU:

export SCOPE_PYTHON="$(command -v python)"
export SCOPE_DEVICE=cuda:0
bash scripts/reproduce_official.sh all

After the model/data artifacts exist:

python -m scoperec recommend --device cuda:0 --request 0 --top-k 10
python -m scoperec recommend --method tiger --device cuda:0 --request 0
python -m scoperec recommend --disable-edit --device cuda:0 --request 0

The demo uses a fixed benchmark catalog, not live products. Released-protocol invalid paths are flagged rather than removed to inflate metrics. Weights and datasets are not bundled. Detailed setup, stages and temporal comparison are in REPRODUCE.md.

The public tree contains source, tests, configs, CI, the architecture illustration and two final comparison charts in PNG and SVG. Experiment logs, validation grids, datasets, weights and upstream source remain local and are excluded from the staged GitHub snapshot.

Project code is licensed under Apache-2.0, as selected in unihsy/ScopeRec. Data, model and dependency licenses remain separate from the project code.