Group recommendation

On this page 7

Group recommendation is the task of suggesting items to a group of users who will consume them together, such as friends choosing a film, a family planning a trip or colleagues picking a restaurant. In recommender systems, it extends single-user personalization with an extra step: the system must combine several, often conflicting, preference profiles into one ranked list that the group as a whole will accept.

The core difficulty is that no single list maximizes every member’s satisfaction at once. A group recommender therefore has to decide whose preferences count, how much, and with what fairness guarantees, which makes the problem partly a matter of prediction and partly a matter of social choice.

How group recommendation combines individual preferences

A group recommender works by aggregating member-level information into a group-level decision, and the point at which aggregation happens defines the architecture. The two classic designs are preference aggregation, which merges the members’ profiles into a single pseudo-user before recommending, and score aggregation, which first predicts each member’s rating for every candidate item and then merges the predicted scores.

Both designs usually reuse an ordinary single-user model, for example matrix factorization or a neural ranking model, to produce individual predictions. The group logic sits on top as an aggregation function that turns a vector of member scores into one group score. Newer approaches learn the aggregation itself, so that the model discovers how influential each member is for a given item.

The table below summarizes the main architectural families and the trade-offs that come with each one.

ArchitectureWhere aggregation happensMain strengthMain weakness
Profile aggregationBefore prediction, on user profilesSimple and fast at query timeLoses information about disagreement
Score aggregationAfter prediction, on predicted ratingsTransparent and easy to tuneRequires a prediction for every member
Learned aggregationInside the model, through attention weightsCaptures member influence per itemNeeds group interaction data to train
Consensus negotiationThrough interactive rounds with membersHigh acceptance of the final choiceSlow and demanding for users
Architectures for group recommendation

Aggregation strategies from social choice theory

Most practical systems choose an aggregation strategy borrowed from social choice theory, the study of how collective decisions follow from individual preferences. Each strategy encodes a different idea of what a good group outcome is, so the choice is a design decision about fairness, not merely a technical detail.

The following table compares the strategies that appear most often in research and in deployed systems, using predicted ratings on a one-to-five scale.

StrategyGroup score ruleWhat it protectsTypical risk
AverageMean of member scoresOverall satisfactionOne member may be very unhappy
Least miseryMinimum of member scoresThe least satisfied memberBland, low-risk recommendations
Most pleasureMaximum of member scoresEnthusiasm of one memberIgnores strong dislikes
Average without miseryMean after removing items below a thresholdBalance and a veto on bad itemsThreshold must be tuned
Borda countSum of rank positions across membersRobustness to rating scale differencesIgnores preference intensity
MultiplicativeProduct of member scoresBalanced, moderate agreementSensitive to very low scores
Common aggregation strategies and their behavior

User studies on group decisions have repeatedly found that average and average without misery match human group choices reasonably well in many leisure domains, while least misery suits situations where one strongly unhappy member would spoil the shared experience. No strategy is universally best; results depend on group size, relationships and the cost of a poor choice.

Group types and the factors that shape them

The right design depends heavily on what kind of group the system serves. Researchers distinguish groups by how stable they are, whether members know each other and whether the system observes the group’s past joint decisions.

Group typeExampleData availableDesign implication
Persistent groupA household sharing a streaming accountLong history of joint interactionsLearn a dedicated group representation
Ephemeral groupStrangers on a guided city tourLittle or no joint historyRely on aggregating individual models
Established small groupThree friends choosing a restaurantIndividual histories, some joint eventsModel social ties and influence
Large audienceMusic for a gym or a shopWeak signals from many peopleUse broad, low-risk aggregation
Group types and their implications

Beyond group type, several social factors influence the outcome. Member expertise matters because a wine enthusiast may reasonably carry more weight for wine choices, social relationships matter because close friends compromise more easily, and emotional contagion means that the mood of one member can shift the satisfaction of others. Modeling these effects explicitly tends to improve acceptance, but it also requires sensitive personal data that users may not wish to share.

Group recommendation is often confused with neighboring tasks that also involve several people or several goals. The key distinction is that the group consumes one shared outcome, whereas related tasks still serve each person separately.

TaskWho receives the outputShared consumptionCentral question
Group recommendationSeveral users jointlyYesWhich item satisfies the whole group
Collaborative filteringOne userNoWhat similar users liked
Multi-objective recommendationOne userNoHow to balance business goals
Recommendation fairnessMany users or providersNoWhether outcomes are equitable across groups
Conversational recommender systemUsually one userSometimesHow dialogue refines preferences
Group recommendation compared with related tasks

Fairness in group recommendation is therefore fairness within one decision, often measured as the gap between the most and least satisfied member. This differs from population-level recommendation fairness, which compares outcomes across demographic groups or item providers.

Evaluation, limitations and common mistakes

Evaluating a group recommender is harder than evaluating a single-user one because ground truth for group satisfaction is scarce. Offline studies often build synthetic groups from individual datasets such as MovieLens and then measure accuracy per member, which says little about whether real groups would agree. Stronger evidence comes from user studies and logged joint decisions, which are expensive to collect.

  • Treating the group as one user hides disagreement and can produce choices that nobody actually wants.
  • Ignoring preference intensity lets a mild majority overrule a strong objection from one member.
  • Using only average accuracy as a metric misses fairness; reporting the minimum member satisfaction is a useful complement.
  • Assuming equal influence ignores that hosts, experts or parents often decide in practice.
  • Neglecting sequences matters for repeated decisions, where alternating which member is favored can keep satisfaction balanced over time.

Group recommendation also inherits the cold start problem in an amplified form: an ephemeral group of new users offers almost no data, so systems fall back on popularity or content-based signals. Privacy is a further constraint, since explaining a group choice can reveal one member’s preferences to the others.

The idea in brief

Picking a film for one person is easy for a streaming service because it knows that person’s taste. Picking one film for four friends on a sofa is much harder, because their tastes overlap only partly. Group recommendation is the technology that looks for the best shared choice.

It works a little like a fair host at a dinner party. The host asks what everyone likes, avoids the dish that one guest cannot stand, and serves something most people will enjoy. The system does the same with numbers, and its main job is balancing happiness across the whole group rather than pleasing just one person.

A practical scenario

A travel platform offers a feature that lets a family of four plan a weekend city break together. Each member has a separate profile: the parents favor museums and quiet restaurants, one teenager prefers sports venues, and the youngest child enjoys parks and animals. The platform predicts each member’s rating for about 200 candidate activities using its standard ranking model.

Instead of averaging the scores directly, the team applies average without misery with a threshold of 2.5 on a five-point scale. Any activity that one member is predicted to rate below 2.5 is removed, and the rest are ranked by their mean score. A science museum with an interactive zoo section and a riverside park with a climbing area rise to the top, while an art gallery that the teenager would dislike drops out.

In an internal test, the team compared this strategy with plain averaging on families who used the feature. The least satisfied member’s rating improved noticeably, while average satisfaction stayed roughly the same, which convinced the team to keep the veto threshold as the default.