Group recommendation inspector

This page is a case study for the grouprec toolkit: a group recommendation pipeline built end-to-end from the library's public API. Everything shown is real — no mock score matrices — computed offline once and then baked into this single self-contained file (no server, no API calls at view time).

1 · The real framework calls

Every ranking you see is produced by a real grouprec call — the aggregators through GroupRecommender, the deep model through GroupIM.group_scores (both steered by the per-member weights):

data   = gr.datasets.load("ml-latest")
groups = gr.groups.synthetic(data, kind="divergent", size=3, n=3)
gints  = gr.groups.derive_group_interactions(data, groups)   # per-group signal (majority, overridable)
ease   = EASE(reg=200.0).fit(data)

# results-aggregation (utilitarian / fairness), steered by member weights:
rec = GroupRecommender(ease, EPFuzzDAAggregator(member_weights=w)); rec.dataset_ = data
rec.recommend(members, k=5, candidates=cands)

# profile-aggregation deep model, steered by member weights:
gim = GroupIM(groups, gints).fit(data)
gim.group_scores(members, cands, member_weights=w)

This is the core part, not the whole script. The full generator (grouprec.inspector.build, ~400 lines) wraps those calls with: parsing MovieLens movies.csv for titles/genres; candidate sampling; a Top-K sparse autoencoder for the latent concepts (adapted from an external repo — not part of grouprec); and baking a 125-point member-weight grid into JSON. So the framework does the recommendation and the group-interaction derivation; the script does data prep, the SAE explanation, and packaging.

2 · How the data is shipped (offline → one file)

1. Fit the recommenders; for each group call the framework over a 125-point member-weight grid (GroupRecommender.recommend orderings for the aggregators, real GroupIM.group_scores forward passes for the deep model), plus the SAE concepts and the displayed subsets.
2. Assemble one Python dict and serialise it with json.dumps.
3. Substitute that JSON into a placeholder (const DATA = /*…*/;) in an HTML template → one self-contained file (~1 MB; the JSON is a single long line, which is what makes the file large — the generation logic itself is ~400 lines).
4. In your browser, the page reads DATA and renders/re-ranks entirely client-side; a slider just looks up the nearest precomputed grid point.

3 · Synthetic groups

Movielens latest small was chosen due to high familiarity of the items, however, no groups were available in the dataset => synthetic groups were generated.

Membership comes from gr.groups.synthetic in three regimes — similar, divergent, outlier (three each), badged with the measured mean pairwise rating correlation. A group's interactions are derived, not simulated, by the framework call gr.groups.derive_group_interactions: by default an item is a group's consensus item if rated ≥4 by ≥2 of the 3 members — a deterministic function of the real ratings — and the rule is overridable with a predicate (e.g. unanimity, or "any member").

4 · Algorithms & faithful interactivity

The slider is each member's importance weight (normalised to 100%). Outputs are precomputed on a {0,¼,½,¾,1}³ = 125-point grid and the slider snaps to the nearest point, so every ranking is a genuine output:

• Weighted average — GroupRecommender + grouprec WeightedAverageAggregator(member_weights=w).
• EP-FuzzDA — GroupRecommender + grouprec EPFuzzDAAggregator(member_weights=w) (fairness; the weight is each member's target share).
• GroupIM — the framework's GroupIM.group_scores(members, items, member_weights=w), a real forward pass that reweights the model's attention pooling (α'm ∝ wm·αm), reducing to the native model at equal weights. Steering is in the model's pooling — not via the SAE.

5 · Latent concepts (how the SAE was adopted)

We reuse a Top-K sparse autoencoder (standardised input, unit-norm decoder, ReLU encoder, hard top-K=6, MSE + small L1) and fit it on GroupIM's item embeddings — specifically the rows of its encoder.user_predictor weight, one vector per movie. Each latent feature is named by the dominant genre of its top-activating movies, turning opaque dimensions into readable concepts. A member is tagged with a concept because the films they watched activate it; the films listed are the concept's global exemplars, so they may differ from the member's own history. The SAE is only an explanation of members.

Why the encoder head, and why "lift". Fitting the SAE on group_predictor instead gave every member the same concepts: that head is trained on only as many target distributions as there are groups, so its item space collapses onto popularity (one direction explained ~66% of the variance, correlating ≈−0.68 with item popularity) and the SAE simply decomposed popularity. The encoder.user_predictor head is pretrained over all users, so it carries real taste structure. Independently, ranking a member's concepts by raw mean activation returns whatever is globally strongest, so we rank by lift — the member's mean activation divided by the global mean, i.e. what is distinctive about that member.

6 · Every displayed subset is deterministic

Subset	Rule (ties: ascending index, argsort kind="stable")
Member history (3)	top-3 by rating; ties by item id
Candidate pool (50)	consensus positive + 49 negatives sampled uniformly (per-group seed) from items no member rated
Recs received (3)	member's top-3 candidates by EASE score
Top-5 recs	top-5 by group score / selection order
Latent concepts (3)	SAE features with the highest lift over the member's items (member mean activation ÷ global mean, i.e. most distinctive)
Concept exemplars (3)	items with highest activation for the feature

Reproduce with grouprec-build-inspector. Full write-up: docs/INSPECTOR.md in the repository.