MPM is an independent research prototype exploring structure-based music recommendation.
Instead of relying primarily on user behavior similarity or genre classification, MPM attempts to model how listeners experience musical structure and predicts reward compatibility between a listener and a piece of music.
The project explores how cognitive mechanisms such as prediction, attention, and reward may influence music preference.
Most modern music recommendation systems rely on:
- collaborative filtering
- behavioral similarity
- large-scale listening history
MPM explores a different direction.
The model assumes that music preference emerges from how listeners perceive and predict musical structure.
When musical patterns are predictable but still stimulating, listeners may experience reward responses.
MPM attempts to model this interaction between musical structure and listener personality.
MPM is organized into three conceptual layers.
Theory Layer Defines the cognitive assumptions behind structure-based music preference.
Personality Layer Models listener perception across four musical dimensions: melody, rhythm, timbre, and arrangement.
Engineering Layer Implements the inference pipeline used to evaluate music tracks and predict reward compatibility.
The model evaluates candidate tracks using a discovery pipeline that integrates external music catalogs (such as Spotify) with the MPM reward prediction model.
flowchart TD
subgraph THEORY["Theory Layer"]
T1[Music Perception Theory]
T2[Prediction-Based Reward]
T3[Structural Listening Model]
end
subgraph PERSONALITY["Personality Layer"]
P1[LSQ Questionnaire]
P2[Personality Scoring]
P3[Listener Personality Vector\nWm Wr Wt Wa]
end
subgraph ENGINEERING["Engineering Layer"]
E1[Spotify Candidate Generation]
E2[Signal Mapping]
E3[Melody Signal]
E4[Rhythm Signal]
E5[Timbre Signal]
E6[Arrangement Signal]
E7[MPM Inference Engine]
E8[Reward Prediction]
E9[Prediction Stability]
E10[Intro Reward Bank Check]
E11[Structural Chaos Detection]
E12[Recommendation Safety Filter]
E13[Final Recommendation]
end
T1 --> P1
T2 --> P1
T3 --> P1
P1 --> P2
P2 --> P3
P3 --> E7
E1 --> E2
E2 --> E3
E2 --> E4
E2 --> E5
E2 --> E6
E3 --> E7
E4 --> E7
E5 --> E7
E6 --> E7
E7 --> E8
E7 --> E9
E7 --> E10
E7 --> E11
E8 --> E12
E9 --> E12
E10 --> E12
E11 --> E12
E12 --> E13
MPM currently consists of several conceptual modules.
A listener is represented using four perceptual dimensions:
- Melody sensitivity
- Rhythm sensitivity
- Timbre sensitivity
- Arrangement sensitivity
These dimensions form a music personality profile.
Music is represented through structural signals:
- melody signal
- rhythm signal
- timbre signal
- arrangement signal
These signals approximate how a listener perceives different musical dimensions.
The model simulates several mechanisms related to musical reward:
- prediction stability
- reward density
- reward peak intensity
- intro latency
- structural chaos
These factors influence the predicted reward compatibility between a listener and a track.
MPM prioritizes avoiding negative listening experiences.
The system evaluates:
- intro reward viability
- timbre gating effects
- structural instability
- prediction collapse
Tracks that fail safety checks are filtered out before recommendation.
The recommendation pipeline currently follows this structure:
Spotify candidate generation ↓ signal mapping ↓ MPM inference scoring ↓ recommendation safety filter ↓ final recommendation
This allows the system to integrate external music catalogs (such as Spotify) with the MPM evaluation model.
An interactive demonstration of the model is available through a GPT-based interface.
The demo allows users to:
- generate a music personality profile
- receive structure-based recommendations
- analyze how MPM evaluates specific songs
Demo access: click here
MPM is currently a research prototype.
The project focuses on conceptual modeling rather than production deployment.
Some components are simplified due to limited data access and the absence of large-scale listening datasets.
Deng Qing
Background in cognitive psychology and behavioral economics.
MPM was developed as an independent research exploration of how music recommendation might work if listener perception and reward mechanisms were modeled directly.