← Back to Research

🏆 Best Pictorial — ACM CHItaly 2025

Interactive Explanatory Recommender System

Blooming
Beats

Venue ACM CHItaly 2025
Track Pictorial
Domain Music Recommendation

Blooming Beats transforms a decade of personal Spotify listening data into visual narratives grounded in Data Humanism principles.

Music is among the most personal data a person generates. A listening history is not a log of events — it is a record of moods, memories, relationships, and moments in time. Yet streaming platforms treat this record as a raw signal for optimization.

Full Title Blooming Beats: An Interactive Explainable Recommender System for Exploring Personal Music Narratives through Data Humanism Principles
Authors Ibrahim Al-Hazwani, Oliver Aschwanden, Daniel Lutziger, Carlos Kirchdorfer, Luca Huber, Nimra Ahmed, Anton Fedosov, Jürgen Bernard
Dataset 202,988 song plays — one user's 10-year Spotify archive
Recommender Systems Content-based filtering
Award 🏆 Best Pictorial — ACM CHItaly 2025
"What if a music recommender system treated listening histories not as input for a model, but as material for a story?"
Research question driving Blooming Beats

Two Unresolved Challenges

01

The Analytical Reduction of Music Data

A song played during a period of grief, an album associated with a particular summer, a playlist assembled for a journey — all carry meaning that play counts and timestamps cannot encode. Existing platforms treat listening data as a behavioral signal and nothing more. Spotify's annual "Wrapped" offers statistical aggregations, but provides no mechanism for users to interrogate, annotate, or make personal meaning visible.

02

The Explainability Gap in Music Recommendation

Music recommendation research has focused primarily on model transparency — revealing feature weights and genre-based rationales — but rarely grounds explanations in the user's personal history. Recommender systems position users as decision-makers evaluating algorithmic output, not as people whose lives and stories the system might reflect. Users want to understand the connection, not the algorithm.

Data Humanism × Explanatory AI

Blooming Beats is grounded in the conviction that the principles Giorgia Lupi developed for making personal data visible are the same principles that should drive how recommender systems explain themselves. Each Data Humanism principle directly addresses a structural failure in how music recommendation currently works.

Data Humanism

Small Data

This principle arguments that individual, intimate datasets — a person's specific history — can reveal what population-scale statistics erase. The personal record is irreducible and meaningful on its own terms.

XAI Dimension

Explanation Aim & Personalization

One user's 10-year archive shifts the explanation aim from "why the algorithm chose this" to "how your own listening history connects to this song." Personalization is not a feature layer, it is the entire dataset.

Data Humanism

Subjective Data

The principle that moods, relationships, and memories encoded in personal data are valid and not noise to normalize. Subjectivity carries meaning that objective analysis loses.

XAI Dimension

Personalization

Personal memory and emotional association become first-class explanation inputs. The annotation system makes the user's subjective layer the primary explanatory layer, not supplementary decoration.

Data Humanism

Design-Driven Data

The commitment to establishing a visual metaphor before building the data pipeline, letting form drive methodology rather than emerging as a styling decision after the fact.

XAI Dimension

Cognitive Load

The visual vocabulary makes the recommendation immediately perceivable without reading text. The feature values are read as visual rhythms rather than numbers where the form reduces the cognitive cost of understanding the explanation.

Data Humanism

Data to Depict Complexity

The commitment to rendering multiple data dimensions navigable progressively rather than collapsing them into a single summary view that erases the relationships between variables.

XAI Dimension

Output Format & Cognitive Load

The explanation operates at multiple output levels — overlaid flower comparison, three impact scores, behavioral line encoding — navigable on demand. Complexity is distributed across layers, not hidden.

Data Humanism

Spend Time with Data

The practice of slow, reflective engagement with personal archives, resisting the impulse toward rapid statistical consumption and inviting users to dwell within the data's temporal depth.

XAI Dimension

Cognitive Load & Personalization

Three temporal views (day, week, month) distribute attention across a 10-year archive rather than overwhelming with everything at once. Exploration surfaces pattern while selection turns those patterns into recommendations.

Data Humanism

Serendipitous Data

The recognition that unexpected patterns are meaningful signals worth surfacing, not anomalies to suppress or errors to correct.

XAI Dimension

Output Format

Unexpected connections across the archive are rendered as navigable recommendation pathways, extending the output format beyond nearest-neighbor retrieval to temporal discovery across a decade of listening.

Seven-Stage Design Process

01
Visual Exploration
Sketching and exploring visual metaphors before touching data
02
Data Sketching
Rapid hand-drawn sketches mapping audio features to visual forms
03
Contextual Enrichment
Layering personal and global events onto the timeline
04
Refinement & Testing
Iterative feedback cycles with real users and the data owner
05
Visual Dictionary
Formalizing the flower encoding system and its grammar
06
Development
Building the system in React, Python, and D3.js
07
User Study
Think-aloud evaluation with participants exploring an unfamiliar user's archive

Songs as Flowers

Each song is represented as a flower-like graph with six petals — one for each selected audio feature. The number of colored dots forming each petal encodes the feature value: a danceability score of 0.780 becomes 78 dots. This choice enables natural clustering that emerges from similar dot densities, matching the rhythmic nature of music itself.

The visual metaphor emerged during the data sketching stage, drawing from Giorgia Lupi's "Bruises — The Data We Don't See" and the OECD Better Life Index.

Danceability
Energy
Valence
Tempo
Acousticness
Instrumentalness
Continuous Wave
Uninterrupted listening — the user let the song play through completely, signaling engagement or flow state.
Interrupted Curve
Song played more than 30 seconds then skipped — exploratory behavior, testing but not committing.
Dashed Line
Skipped within 30 seconds — active curation or disinterest, a deliberate editorial choice.

What the Think-Aloud Study Found

Four participants from diverse academic backgrounds explored the listening archive of an unfamiliar user, identified a personal pattern, generated recommendations, and engaged with the explanation view The task was designed to test whether the narrative and explanation design remain legible to someone who did not produce the data.

Music-memory associations are strong but undocumented

All four participants formed connections between songs and memories — concerts, milestones, seasons, sensations. Yet three out of four did not actively track them, describing the effort as overwhelming. The desire to engage with listening history as personal narrative exists, although the cost of doing so manually is prohibitive.

Contextualization preferences are varied

Participants demonstrated three distinct approaches: high contextualization (always wanting others to know why a song matters), moderate (selective sharing), and none (sharing without commentary). The three-level contextual architecture accommodates all three without imposing a single mode of engagement.

Users want the connection, not the algorithm

Participants did not want to understand the model. They wanted to understand why this song, why now, in relation to what they had already listened to. The flower comparison view — overlaying the aggregated profile of selected songs against the recommendation — delivers exactly that: a visual argument, not a technical disclosure.

Blooming Beats: An Interactive Explainable Recommender System for Exploring Personal Music Narratives through Data Humanism Principles

Ibrahim Al-Hazwani, Oliver Aschwanden, Daniel Lutziger, Carlos Kirchdorfer, Luca Huber, Nimra Ahmed, Anton Fedosov, Jürgen Bernard

ACM CHItaly 2025 🏆 Best Pictorial

Blooming Beats contributes a novel personal data visualization approach that transforms music listening histories into personal narratives. The flower-based visual dictionary, the three-level contextual enrichment architecture, and the temporal multi-view design collectively demonstrate how Data Humanism principles can be operationalized in an interactive music recommender system.

By treating explanation not as a transparency mechanism but as a narrative layer of the user experience, Blooming Beats marks a shift in how listening histories are conceptualized: as narrative material rather than behavioral traces or optimization targets.