In group accommodation booking, delegating coordination to messaging apps and informal voting often leads to opaque preference trade-offs and social influence, which results in decisions that reflect dominance or conformity rather than genuine consensus. Conversational elicitation coupled with visual preference alignment can enable groups to surface, compare, and negotiate constraints transparently by separating private preference expression from shared compromise exploration, as instantiated in this demo.

Group decisions in recommender settings are rarely only about finding the best item. They require a process for expressing constraints, making trade-offs legible, and converging on a mutually acceptable choice under limited time and attention. In practice, many platforms still treat group accommodation search as an individual workflow, leaving coordination to external channels and informal negotiation.

This mismatch matters because group dynamics can distort what is expressed and what is selected. When preferences are elicited in front of others, social pressure can suppress minority constraints or push the group toward “least resistance” options. When trade-offs remain implicit, disagreements become harder to resolve because the group lacks a shared representation of what is conflicting and what is compatible. For group recommenders to be useful as decision support, they need interaction mechanisms that operationalize transparency and controllability, not only better aggregation strategies.

In a demo, in cooperation with Ibrahim Al-Hazwani, Oliver Robin Aschwanden, Oana Inel, and Jürgen Bernard, and published in the Proceedings of ACM RecSys ’25, we present PRISM: an interactive group recommender system for collaborative accommodation booking.

The demo addresses a concrete gap: typical group booking workflows externalize alignment and negotiation, while the recommendation interface itself remains a largely opaque list of options. PRISM repositions the recommender as an interactive artifact that helps groups articulate individual constraints and then build consensus through explicit, inspectable compromise structures.

Demo overview

PRISM is designed as a research instrument for studying and supporting preference alignment in groups. It enables a workflow where individual preferences can be expressed without immediate social influence and then translated into a shared, visual space where compromise can be inspected and discussed. This shifts “agreement” from a conversational outcome to something the system can help make observable: where preferences overlap, where they conflict, and which options sit on plausible negotiation frontiers.

For end users, the capability is practical decision support: the system helps them explore a space of options while keeping individual constraints visible and comparable. For researchers, PRISM provides a structured setting to examine how interaction design affects understanding, consensus formation, and perceived pressure to conform. For practitioners, it illustrates a design pattern for group recommenders in which explainability is not an after-the-fact justification, but a shared medium for collaborative reasoning.

Demo capabilities

PRISM turns preference expression into a spatial representation and then uses a bivariate visual encoding to make compromise regions explicit. The key conceptual elements to notice are the separation between individual and joint views, and the way the shared map acts as a negotiation surface rather than a static explanation.

Separating private elicitation from shared negotiation

A core design idea is a two-phase interaction model that decouples preference elicitation from group deliberation. The problem it targets is social influence during preference articulation: if users must state constraints in front of others, they may self-censor or defer. PRISM introduces “individual-first” interaction as an abstraction, so each participant can express needs privately before the system merges them into a shared representation.

This is important because it supports individual agency as a prerequisite for meaningful consensus. The group discussion is then grounded in a representation that reflects both users’ inputs, rather than being anchored by whoever frames the decision first.

Conversational elicitation grounded in spatial feedback

Another key mechanism is conversational preference elicitation paired with immediate, location-aware feedback. The problem here is the gap between natural language constraints (e.g., “quiet,” “near museums,” “close to nightlife”) and how those constraints shape the option space. PRISM uses conversation as the interaction primitive for expressing preferences while simultaneously projecting those preferences onto a spatial map that reflects how strongly different areas match the expressed needs.

This choice matters because it gives users a controllable mental model: they can iteratively refine constraints and see how the “shape” of the search space changes, which supports sensemaking rather than blind acceptance of ranked lists.

Bivariate compromise mapping as a shared decision surface

In the collaborative phase, PRISM introduces a bivariate preference map that encodes how well areas align with each user and where compromises emerge. The problem it addresses is that group trade-offs are hard to reason about when preferences are only presented as separate lists or averaged scores. The bivariate encoding makes the structure of disagreement explicit and, crucially, highlights “balanced” regions that satisfy both parties to a similar degree.

This matters for consensus-building because it reframes the discussion from “whose preference wins” to “where are the defensible compromise zones,” enabling negotiation to focus on a smaller, interpretable subset of the space.

Linked explanations that connect options to trade-offs

A final design idea is to connect the shared map with option-level explanations that decompose why a candidate is plausible for each user. The problem is that visual overviews can indicate where compromise might exist, but groups also need to understand why a specific listing is acceptable or problematic. PRISM therefore treats explanations as interaction primitives that connect spatial context to concrete attributes (e.g., amenities and proximity constraints) and make the reasoning inspectable in the course of discussion.

This matters because it supports diagnosability: users can identify which constraints are satisfied, which are violated, and which are negotiable, and can justify decisions in terms of visible evidence rather than persuasion alone.

A session typically starts with each participant expressing preferences individually until the system has a stable representation of their constraints. The group then transitions to a shared exploration view that surfaces overlap and compromise regions, using the same spatial frame to discuss candidate listings and the reasons they align (or not) with each person’s expressed needs.

Demonstration scenario and evidence of usefulness

The demo illustrates the workflow through a concrete travel-planning scenario in which two travelers have conflicting priorities, one oriented toward lively neighborhoods and the other toward quiet areas near cultural landmarks. The system demonstrates how private elicitation can capture these differences without immediate negotiation, and how the shared bivariate map can reveal compromise neighborhoods that would be difficult to identify through list-based browsing alone.

Evidence of usefulness is reported through a controlled comparison study with pairs of users performing a collaborative accommodation search task. The results indicate that the visual, two-phase approach makes the influence of preferences on group recommendations more understandable (roughly +1.8 points on a 5-point scale) and makes consensus-building easier (roughly +2 points). The study also suggests reduced conformity pressure, consistent with the intent of separating individual elicitation from group discussion, while overall usability remains comparable to a familiar baseline interface.

Conclusions

PRISM matters as an enabling artifact because it treats group recommendation as an interaction problem: eliciting authentic constraints, externalizing trade-offs, and providing a shared medium for compromise. By combining conversational elicitation with a compromise-oriented visual representation, the demo highlights how explainability can be operationalized as collaborative reasoning support rather than post-hoc justification.

Natural extensions follow directly from the capabilities the demo foregrounds. One direction is scaling from pairs to larger groups while preserving legibility of overlaps and conflicts, potentially requiring new aggregation-aware visual encodings and facilitation controls. Another direction is making the conversational elicitation more adaptive to different communication styles and levels of decisional certainty, so that the system can better distinguish hard constraints from negotiable preferences. Finally, PRISM’s interaction pattern suggests broader applicability beyond accommodation; any domain where options are geographically or conceptually “mappable,” and where groups need to negotiate trade-offs under social pressure, could benefit from a similar separation between private preference articulation and shared compromise exploration.