Design: the rift, the families, and the bridge

The one-sentence pitch

grouprec puts deep-learning SOTA group models and classic, fairness-oriented preference aggregators behind one API, so you can compare them head-to-head on the same data, splits, and metrics.

The rift

Group recommendation research has split into two communities that rarely meet (Peska et al., Bridging the Rift, UMAP 2025):

• Results aggregation — run a single-user recommender, score each member, then combine those scores/lists into one group list. This line is where the fairness work lives (least-misery, GFAR, EP-FuzzDA, RLProp/LTP, …): given per-member utilities, how do we pick a list that's good and fair to everyone? Evaluated on synthetic groups over MovieLens-style data with coupled/decoupled protocols.
• Profile aggregation — learn a group representation directly (attention, mutual information, hypergraphs, consensus graphs) and predict for the group as a unit. This line is where the deep SOTA lives (AGREE, GroupIM, ConsRec, …). Evaluated on datasets with real/inferred groups (CAMRa2011, Mafengwo, Yelp, Douban, Weeplaces) with group-level leave-one-out HR/nDCG.

They drifted apart into disjoint datasets, baselines, metrics, and protocols — so a "new" method in one community is rarely compared against the other. That's the rift.

Why we keep the two families distinct (but unified)

They genuinely consume different things and answer different questions, so we don't pretend they're identical — we make them interoperable:

	Results aggregation	Profile aggregation
paradigm	"results"	"profile"
input	base-RS per-member scores	group/member interactions
strength	fairness, interpretability, zero training	accuracy/SOTA
API	GroupRecommender(base, aggregator)	grouprec.models.*

Both expose the same contract — fit(dataset) + recommend(members, k, …) — so the evaluator and benchmark drive them identically.

Two evaluation regimes

• Member-level (synthetic groups; ML/KGRec/Last.fm): ground truth is each member's held-out feedback. Supports coupled (vs feedback) and decoupled (vs the base RS's predicted r̂, isolating the aggregator). This is the aggregator/fairness arena.
• Group-level (real/inferred groups; CAMRa/Mafengwo/Yelp/Douban/Weeplaces): ground truth is the group's held-out choice (sampled 1-vs-N HR/nDCG). Deep models can only be trained/evaluated here.

Deep models cannot run on synthetic groups (no group-level signal to learn from), so the two regimes are real, not cosmetic.

The bridge

In the group-level, coupled regime, both families emit a group top-k list scored against the same group choice — so a results-aggregator (GFAR) and a deep model (AGREE) land in the same leaderboard, comparable for the first time. Conversely, in the member-level regime the same recommender scored under coupled vs decoupled can flip rank, and the relevance–fairness trade-off becomes a Pareto front (gr.bench.viz.plot_pareto). Making both of these one-liners is the point of the library.

See Concepts for the API surface and Evaluation for the protocol/metric details.