Overview
PRISM is a pair-based explanatory recommender system that helps two people find shared accommodation through individual preference elicitation and joint visual deliberation.
Recommender systems for accommodation are built for individuals. Yet most travel involves groups — couples, friends, colleagues — who must negotiate preferences that are inherently subjective, partially hidden, and often in tension. PRISM addresses this coordination problem through a two-phase process: first eliciting each person's preferences individually through a conversational AI interface, then rendering both preference sets onto a shared bivariate hexagonal map that makes compromise zones visible and negotiable.
"Recommending and explaining to a pair is not recommending to two individuals. It is designing for negotiation."Research premise driving PRISM
The Problem
When two people search for shared accommodation, each brings a different set of preferences like location priorities, amenity requirements, or neighborhood atmosphere, that are rarely articulated in advance. Existing platforms offer no mechanism for joint preference elicitation. Each person searches separately, compares screenshots, and negotiates through informal channels disconnected from the recommendation interface. The coordination happens outside the system, and the system cannot support it.
Explainable recommendation research has focused almost entirely on explaining recommendations to individuals. In a collaborative context, an explanation must address both users simultaneously. It needs to account for whose preference drove a recommendation, where preferences aligned, and where trade-offs were made. Existing explainable approaches lack the visual and structural vocabulary to make these group-level dynamics legible, leaving pairs unable to understand or negotiate algorithmic suggestions together.
Research Foundation
PRISM applies Data Humanism principles to the design of pair-based explanations. Giorgia Lupi's framework, developed for making individual data meaningful, reveals new design imperatives when extended to collaborative recommendation — where subjectivity, complexity, and serendipity must be navigated by two people at once.
Subjective Data
The principle that personal preferences are valid data, not noise to be normalized away. In a pair context, subjectivity is multiplied: two people bring two distinct emotional frameworks to the same decision.
Explanation Aim & Personalization
Each user's preference profile should elicited individually before any joint view is shown. The explanation aim shifts from "why this listing" to "how this listing reflects each of your preferences" — making personalization a precondition for collaborative transparency.
Small Data
This principle arguments that intimate, specific datasets — what a particular person cares about in a particular context — carry more meaning than aggregate population signals. For PRISM, the relevant "small data" is the elicited preference footprint of two specific people planning a specific trip.
Explanation Aim & Personalization
The bivariate map visualizes precisely two preference profiles — not a generalized user model. The system's explanatory power comes from specificity: this pair, these preferences, this trip. Generalizing across users would erase the reasoning that makes the map legible.
Spend Time with Data
The practice of slow, reflective engagement rather than rapid consumption. Lupi argues that staying with data — exploring it, annotating it, returning to it — produces understanding that instant visualization cannot. In collaborative settings, this patience has a second function: it creates space for conversation.
Cognitive Load & Communication Style
The two-phase structure deliberately separates individual elicitation from joint deliberation, giving each person time to form their preferences before negotiating. The shared map is designed as a conversational artifact — something to discuss over, not a final answer to accept. Cognitive load is managed by distributing decision-making across time.
Data to Depict Complexity
The commitment to rendering multi-dimensional relationships navigable rather than collapsing them into a single metric. Accommodation decisions involve spatial, social, and contextual dimensions simultaneously — a reduction to price or rating erases the structure of the decision.
Output Format & Cognitive Load
The bivariate hexagonal map encodes two-dimensional preference density in a single shared view — each hex cell represents a cluster of listings situated along both users' preference axes simultaneously. Complexity is rendered legible through spatial encoding rather than compressed into scores, distributing cognitive load across the visual field.
Serendipitous Data
The recognition that unexpected overlaps — patterns the user did not anticipate when describing their preferences — are meaningful signals worth surfacing. In a collaborative context, serendipity can emerge not just within one person's preferences but at the intersection of two.
Output Format & Cognitive Load
Purple compromise zones — hexagonal cells where both users' preference densities overlap — are rendered as visually distinct regions on the shared map. These zones are not derived from explicit joint input; they emerge from the spatial intersection of two independently elicited profiles, surfacing agreements neither person anticipated.
System Design
Each participant begins independently. A conversational interface powered by Gemini-2.0-flash-exp engages the user in a structured dialogue about their accommodation priorities — location preferences, neighborhood character, amenity requirements, and travel context.
Preferences are not entered through checkboxes or sliders. They emerge through natural language, which the system parses into a structured preference profile. This profile is then projected onto a personal hexagonal map, giving the user a visual representation of their own priority space before any comparison occurs.
Once both users have completed individual elicitation, their preference profiles are projected onto a shared bivariate hexagonal map. Each axis represents one user's preference density; hex cell color encodes the concentration of listings within that region of the joint preference space.
Purple compromise zones — cells where both preference densities are high — emerge from the spatial intersection of the two individual profiles. These zones represent accommodation options neither person explicitly requested but that both independently rated highly. Listings within these zones become the primary recommendation set.
Technical Foundation
PRISM's recommendation engine operates on a heterogeneous graph built from 21,707 Airbnb listings in Copenhagen, enriched with Foursquare OS Places POI data. GraphSAGE learns node embeddings by aggregating neighborhood information across five distinct node types — capturing the relational structure that scalar features cannot encode.
Cosine similarity over GraphSAGE embeddings retrieves candidate listings for each user's preference profile. The bivariate map then filters and visualizes candidates in the joint preference space, making the graph's internal structure visible through spatial positioning rather than numerical scores.
Evaluation Results
Participants using PRISM showed a +1.83 improvement in understanding trade-offs relative to the Airbnb baseline. The bivariate map made it possible to see not just where preferences aligned but where they diverged — transforming implicit compromise into a visible, legible artifact that both users could reference.
Pairs reached consensus +2.00 faster on average compared to baseline, with PRISM achieving a System Usability Scale score of 82.50. Rather than eliminating disagreement, the shared map gave pairs a structured space to resolve it — pointing at specific zones rather than describing preferences in the abstract.
Post-study interviews revealed that visible divergence between the two individual maps was experienced as productive rather than problematic. Seeing where preferences diverged prompted reflection on priorities that participants had not previously articulated. Conflict became a design resource rather than a system failure.
Publications
PRISM: From Individual Preferences to Group Consensus through Conversational AI-Mediated and Visual Explanations
ACM RecSys 2025 Demo TrackWe present PRISM, an interactive recommender system for collaborative accommodation search. PRISM employs a two-phase methodology — individual conversational preference elicitation followed by joint visual deliberation on a shared bivariate hexagonal map — to support pairs in negotiating shared accommodation preferences. A GraphSAGE model operating over a heterogeneous graph of 21,707 Airbnb listings in Copenhagen powers the recommendation engine. A user study with N=12 participants (6 pairs) demonstrates improved trade-off understanding and faster consensus relative to the Airbnb baseline.