Overview
HUMMUS reimagines group music recommendation as a collective act — where each song is a flower and the garden grows with the group's shared listening journey.
Sequential music recommenders are designed for individuals. They model temporal listening patterns to predict the next song, but treat the listener as a singular entity. HUMMUS asks a different question: what happens when recommendation must serve a group? And what if the system explained its choices using the visual language of nature rather than numbers?
"Music is rarely a solitary experience — yet music recommender systems are designed as if users are always alone."— Thesis Chapter 14, HUMMUS
The Problem
Sequential recommenders expose feature weights or nearest-neighbor distances — not the emotional, contextual, or relational dimensions of music experience. Users see only the output, never the reasoning. When the recommendation fails in a social setting, no one can question it because no one can see its logic.
Group recommendation strategies — average voting, least misery, most pleasure — aggregate individual preferences but erase individual differences. There is no mechanism for the group to negotiate, to steer, or to understand why a candidate song was proposed for their collective listening.
Research Foundation
HUMMUS sits at the intersection of two fields that rarely speak to each other. Each Data Humanism principle applied here directly addresses a structural limitation in how recommender systems currently explain themselves.
Small Data
Giorgia Lupi's call to work with intimate, human-scale datasets rather than population-level aggregations — preserving individual and group specificity over statistical generalization.
Explanation Aim · Personalization
The explanation aim shifts from population-level rationale to group-specific transparency. Personalization operates at the collective level — the group's own playlist is the primary narrative unit, prioritizing human-scale stories over algorithmic completeness.
Subjective Data
The principle that personal, emotional, and relational dimensions of data are valid — not noise to be normalized away. Subjectivity is a feature, not a flaw.
Personalization
Individual musical preferences within the group are irreconcilable and both valid. HUMMUS makes these differences visible and negotiable rather than averaging them away — differences in taste become part of the interaction, not an obstacle to it.
Serendipitous Data
The recognition that unexpected patterns and surprising connections are meaningful — not anomalies to suppress but signals worth surfacing and exploring.
Output Format · Cognitive Load
Unexpected song proposals are rendered as flowers before the vote — immediately inspectable in the shared visual vocabulary. The output format frames serendipity as discoverable, and the voting mechanism keeps cognitive load low even when the candidate is surprising.
Data to Depict Complexity
The commitment to rendering data's full complexity rather than reducing it to a single metric — embracing multiplicity as a valid and informative state.
Output Format · Cognitive Load
Five audio features as five petals, connecting lines exposing similarity on hover, a line chart tracking feature trajectories across the session — the layered explanation architecture makes complexity navigable progressively, on demand, without collapsing it.
Design-Driven Data
The commitment to letting aesthetic and experiential considerations inform every stage of design — the visual form is not decoration, it is the primary medium of understanding.
Personalization · Cognitive Load
The flower form ensures that the explanation is immediately perceivable without algorithmic knowledge — aesthetic design reduces cognitive load. The organic visual vocabulary makes the explanation feel personal, not technical.
Spend Time with Data
The practice of slow, reflective engagement — allowing users to observe how collective preferences evolve rather than simply consuming a static recommendation output.
Personalization · Cognitive Load
The session artifact — the completed garden preserved at session end — transforms the recommendation history into something the group can return to and reflect on. The explanation does not end when the music stops; it becomes a shared record worth revisiting.
The Flower Garden
Each song is visualized as a flower where every petal encodes one audio feature. Petal length represents the feature's value on a 0–1 scale. White connecting curves link petals sharing the highest feature similarity between consecutive songs — rendering recommendation relationships directly inspectable without algorithmic knowledge.
System Design
The HUMMUS interface is organized around two linked views — the flower garden and a temporal line chart — combined with a host–guest framework and real-time voting mechanism.
Songs are positioned chronologically from left to right. Each new flower appears at the right edge as a song is added. White connecting curves link petals with the highest similarity score between songs, making recommendation logic navigable without algorithmic expertise. The garden is a living artifact of the group's musical journey, updated in real time.
A complementary line chart runs below the garden, displaying the evolution of each audio feature across the full session. Where the flowers show the character of individual songs, the chart shows the trajectory of the group's collective taste over time. The two views are linked: hovering over a flower highlights the corresponding point on the chart, and vice versa.
A host initiates the session and generates a shareable QR code. Guests join without an account. When the user-contributed queue empties, the system transitions to algorithmic recommendation — and the voting mechanism activates. All participants simultaneously see the candidate song's flower and vote. The outcome determines which song joins the garden, transforming recommendation into a negotiated outcome.
Evaluation · N = 19 · Groups of 3–5 · Mixed Methods
Participants acted as a group of friends creating a shared playlist for a party, contributing songs, voting on algorithm-proposed candidates, and reflecting together at the end. Transparency, scrutability, and user experience were measured on 7-point Likert scales, complemented by qualitative group discussion.
Participants understood the system's general recommendation approach well (M = 4.68 / 7). But predictive understanding — the ability to anticipate what the system would suggest next — remained low (M = 2.84 / 7). This gap is not a design failure: comprehending a system's logic and predicting its next output are different cognitive tasks, and the latter may be irreducibly difficult in an open-ended group context.
The voting mechanism was effective at the group level (M = 4.74 / 7), but revealed a structural tension: some participants felt their contributions were overrepresented; others underrepresented. Group-level scrutability is not the same as individual scrutability — a finding with implications for all multi-stakeholder recommenders.
Satisfaction (M = 4.93 / 7) and ease of learning (M = 5.42 / 7) were the strongest scores. Participants described joy, found the experience relaxing, and expressed that it was fun to watch the garden evolve. The flower garden became a shared aesthetic object — evidence that Data Humanism principles can produce measurable shifts in how algorithmic recommendations are experienced.
HUMMUS: A Human-Centered Approach to Sequential Recommendation Explainability through Data Humanism
MRS Workshop · ACM RecSys 2025HUMMUS is a collaborative sequential music recommender that embeds Data Humanism principles directly into its recommendation and explanation interface. Rather than numeric feature scores, it renders each recommended song as a botanical flower glyph — petal length encoding audio features, connecting curves encoding similarity between consecutive songs. A host–guest framework and real-time group voting mechanism allow participants to collectively steer the recommendation process. Evaluated with N = 19 participants across groups of 3–5, HUMMUS achieved a satisfaction score of 4.93/7 and ease of learning of 5.42/7, with qualitative responses emphasizing the system's emotional resonance and its capacity to transform recommendation into a shared social experience.
← Back to Research