GroupRec — RecSys 2026 demo submission¶
Patrik Dokoupil, Ludovico Boratto, and Ladislav Peska. GroupRec: A Unified Toolkit for Reproducible and Inspectable Group Recommendation Research. Demo track — under review at RecSys 2026.
This page is a reviewer-facing tour of the demo. The star of the demonstration is the interactive inspector: a live, in-browser view of a group recommendation being produced, where every displayed number is a genuine call into the library.
Open the inspector ↗ How it's built GitHub
What the demo shows¶
- Real, not mocked. Rankings come from
GroupRecommender(aggregators) andGroupIM.group_scores(deep model). Groups are sampled withgr.groups.syntheticand their interactions derived — not simulated — viagr.groups.derive_group_interactions. - Steerable & inspectable. Per-member influence sliders reweight the aggregation / the model's attention pooling; a Top-K SAE explanation layer surfaces each member's most distinctive latent concepts.
- Self-contained. The whole thing is computed offline and baked into a single HTML file — no server, no API calls at view time.
Reproduce it¶
pip install grouprec[torch]
grouprec-build-inspector --out group_rec_inspector.html
Full write-up of every framework call and every displayed subset: Interactive inspector — how it's built.
The library behind the demo¶
- Design (the rift) — the two-community split the toolkit bridges
- Evaluation — coupled / decoupled / sampled protocols
- Reproducibility —
gr.Experiment(config / git / env / seeds), citations - README