🎧 Spotify Music Preference Analysis
🧠 Project Overview
This project analyzes Spotify music data to predict song preferences using machine learning models. The analysis is based on a dataset of 195 songs (100 liked, 95 disliked) with various audio features extracted from Spotify's API.
📂 Dataset Description
📥 Data Collection Process
Liked Songs (100 tracks):
- 🎵 Primarily French Rap
- 🎸 Some American Rap, Rock, and Electronic music
- ✅ Represents personal music preferences
Disliked Songs (95 tracks):
- 🤘 25 Metal songs (Cannibal Corpse)
- 🎤 20 Rap songs (PNL - personally disliked)
- 🎼 25 Classical songs
- 🕺 25 Disco songs
- 🚫 Pop songs excluded due to neutral preference
📊 Dataset Structure
The dataset contains 195 rows and 14 columns with the following features:
🎼 Audio Features Explanation
🔑 Key Predictive Features
🎵 Danceability (0.0-1.0)
- Measures how suitable a track is for dancing
- Based on tempo, rhythm stability, beat strength, and regularity
- Higher values indicate more danceable tracks
⚡ Energy (0.0-1.0)
- Represents intensity and activity level
- Energetic tracks feel fast, loud, and noisy
- Death metal = high energy, Bach prelude = low energy
😊 Valence (0.0-1.0)
- Musical positiveness measure
- High valence = happy, cheerful, euphoric
- Low valence = sad, depressed, angry
🎸 Acousticness (0.0-1.0)
- Confidence measure of acoustic content
- 1.0 = high confidence the track is acoustic
- Important for distinguishing electronic vs. acoustic music
🧩 Secondary Features
🗣️ Speechiness (0.0-1.0)
- Detects spoken words in tracks
- Above 0.66 = mostly speech (talk shows, audiobooks)
- 0.33-0.66 = mixed music and speech (rap music)
- Below 0.33 = mostly music
🎼 Instrumentalness (0.0-1.0)
- Predicts absence of vocals
- Values above 0.5 likely represent instrumental tracks
- “Ooh” and “aah” sounds treated as instrumental
📢 Liveness (0.0-1.0)
- Detects audience presence in recordings
- Above 0.8 = strong likelihood of live performance
- Useful for distinguishing studio vs. live recordings
🔊 Loudness (dB)
- Overall track loudness in decibels
- Typically ranges from -60 to 0 dB
- Useful for comparing relative loudness between tracks
⚙️ Technical Features
🥁 Tempo (BPM)
- Estimated beats per minute
- Indicates the speed/pace of the music
- Directly derived from average beat duration
🎹 Key (0-11)
- Musical key using Pitch Class notation
- 0 = C, 1 = C♯/D♭, 2 = D, etc.
- Important for harmonic analysis
🔑 Mode (0/1)
- Musical modality
- 1 = Major scale, 0 = Minor scale
- Affects emotional perception of music
📏 Time Signature
- Number of beats per bar/measure
- Common values: 3 (waltz), 4 (most popular music)
⏱️ Duration (ms)
- Track length in milliseconds
- Useful for identifying song structure preferences
📊 Data Visualizations
1️⃣ Box Plot Analysis
- Compares each audio feature between liked (1) and disliked (0) songs
- Helps identify:
- 🎯 Feature distributions for each preference category
- ⚠️ Outliers in the data
- 🚦 Clear separations between liked and disliked songs
- 🧪 Discriminative features
2️⃣ Correlation Heatmap
Reveals:
- 🔁 Relationships between audio features
- 🧬 Multicollinearity issues (highly correlated features)
- 🔍 Feature independence for model selection
- 🛠️ Feature engineering opportunities
Key insights:
- Energy and loudness → strong positive correlation
- Acousticness and energy → negative correlation
- Valence → emotional interpretation link
3️⃣ Pair Plot Analysis
- Shows relationships between: danceability, energy, valence, tempo, acousticness
- Detects:
- 🎯 Clustering patterns for liked vs. disliked
- 🤝 Feature interactions
- 📈 Distribution shapes
4️⃣ Model Performance Comparison
Models Evaluated:
- ✅ Logistic Regression
- 🌲 Random Forest
- 🧠 Neural Network
- 📍 K-Nearest Neighbors (KNN)
- 🛠️ Artificial Neural Network (ANN)
- 📐 Support Vector Machine (SVM)
- 📊 Naive Bayes
- 🚀 XGBoost
🧠 Key Insights
✅ Most Important Features:
- 😊 Valence – correlates with emotion
- ⚡ Energy – identifies genre intensity
- 💃 Danceability – relates to rhythm preference
- 🎸 Acousticness – distinguishes acoustic/electronic
- 🗣️ Speechiness – useful for rap detection
❌ Least Important Features:
- ⏱️ Duration – length is less relevant
- 🎹 Key – minimal influence on likability
- 📏 Time Signature – usually standard
- 📢 Liveness – studio vs. live not crucial
🧪 Model Performance Insights
- 🌲 Ensemble methods (Random Forest, XGBoost) perform best
- 🧠 Neural Networks capture complex patterns
- 🧮 Simple models like Logistic Regression provide baselines
- 📊 Performance differences highlight feature complexity
⚙️ Technical Implementation
🧼 Data Processing
- 🎯 Feature Extraction – Spotify API (audio features)
- 🧹 Data Cleaning – handle nulls/outliers
- 📐 Feature Scaling – normalize ranges
- 🔀 Train-Test Split – measure generalization
📏 Model Evaluation
- 🎯 Accuracy – main performance metric
- 🔁 Cross-Validation – stability check
- 📊 Feature Importance – understand model decisions
- 🧩 Confusion Matrix – analyze misclassifications
💡 Applications
- 🎧 Personal Music Recommendation Systems
- 🎶 Understanding Music Preference Patterns
- 📝 Automated Playlist Generation
- 📈 Music Streaming Service Improvements
- 📊 Market Research for Music Industry
🚀 Future Enhancements
- 📈 Larger Dataset – more genre diversity
- 🧩 Additional Features – genre, artist, release year
- 🤖 Advanced Models – deep learning (e.g. RNN, CNN)
- 🌐 Real-Time Prediction – stream-integrated recommendations
- 🧍 Collaborative Filtering – user-based recommendations
📎 Dataset
📁 File: [data.csv]
Content: Contains 195 rows and 14 audio features used for analysis and modeling.
🧠 This analysis demonstrates how machine learning can understand and predict personal music preferences using audio features extracted from Spotify's comprehensive music database.