DiverseTune — Recommendation Diversity Experiment

How It Works

Three Algorithms, One Question

Which algorithm best balances relevance, diversity, and fairness?

Greedy

Popularity-Based

Ranks all songs by popularity score and selects the top K=10. Simple but heavily biased toward mainstream hits.

Low Diversity High Bias Fast

Time: O(n log n) Space: O(n)

Content Filtering

Computes cosine content_filtering between liked songs and candidates. Returns top K=10 most similar tracks. Better relevance, but echo chamber risk.

Moderate Diversity Relevant O(n²)

Time: O(n·m·d) Space: O(n)

★ Best Results

Graph DPP Rerank

Content Filtering + DPP Re-ranking

Gets top N=50 via content_filtering, then applies Determinantal Point Process re-ranking to maximise diversity. Enforces ≥20% niche song fairness constraint.

High Diversity Fair Balanced

Time: O(N·K²·d) Space: O(N²)

Evaluation

Key Metrics at a Glance

Averaged results over 8 independent experiment runs with random seed songs

Greedy

Popularity-Based

ILD (Diversity) 0.0347

Popularity Concentration Index 0.0070

Avg Popularity 95.90

Niche Songs 0.00%

Content Filtering

ILD (Diversity) 0.0064

Popularity Concentration Index 0.1632

Avg Popularity 48.69

Niche Songs 26.25%

Graph DPP Rerank

Content Filtering + DPP

ILD (Diversity) 0.0187

Popularity Concentration Index 0.2156

Avg Popularity 46.64

Niche Songs 33.75%

Experiment Output

Algorithm: Greedy (Popularity)
  ILD (Diversity):     0.0347
  Popularity Concentration Index:          0.0070
  Avg Popularity:      95.90
  Niche Songs (%):     0.00%

Algorithm: Content Filtering (CF)
  ILD (Diversity):     0.0064
  Popularity Concentration Index:          0.1632
  Avg Popularity:      48.69
  Niche Songs (%):     26.25%

Algorithm: Graph DPP Rerank (DPP)
  ILD (Diversity):     0.0187
  Popularity Concentration Index:          0.2156
  Avg Popularity:      46.64
  Niche Songs (%):     33.75%

────────────────────────────────────────────────────────────
🏆 Highest Diversity (ILD):  Greedy (Popularity)  (0.0347)
🏆 Best Niche Coverage:      Graph DPP Rerank (DPP)         (33.75%)
🏆 Lowest Popularity Bias:   Graph DPP Rerank (DPP)         (46.64)
────────────────────────────────────────────────────────────

Visual Analysis

Experiment Visualizations

Interactive charts generated from 8 experiment runs across all algorithms

📊 Diversity Comparison (ILD)

Higher ILD = more diverse recommendations

⚖️ Fairness — Popularity Concentration Index

Lower Popularity Concentration = more equitable popularity distribution

🔥 Average Popularity Score

Measures mainstream bias in recommendations

📈 Diversity vs Popularity Trade-off

Dual-axis view showing how diversity relates to popularity bias

🌱 Niche Song Representation

Percentage of recommended songs with popularity < 40

Side by Side

Algorithm Comparison Table

Complete metric breakdown across all evaluation dimensions

Metric	Greedy (Popularity)	Content Filtering (Content Filtering)	Graph DPP Rerank (Content Filtering + DPP)	Best
ILD (Diversity) ↑	0.0347	0.0064	0.0187	Greedy
Popularity Concentration Index ↓	0.0070	0.1632	0.2156	Greedy
Avg Popularity	95.90	48.69	46.64	Graph DPP Rerank
Niche Songs % ↑	0.00%	26.25%	33.75%	Graph DPP Rerank
Popularity Bias	Very High	Medium	Low	Graph DPP Rerank
Genre Coverage	2 genres	3-4 genres	5+ genres	Graph DPP Rerank
Fairness Constraint	❌ None	❌ None	✅ ≥20% niche	Graph DPP Rerank

Findings

Conclusions & Key Takeaways

🔬

Key Finding

The Greedy algorithm achieves the highest raw ILD diversity (0.0347) because it selects from a global pool of popular songs spanning different genres. However, it recommends zero niche songs and has extreme popularity bias (avg 95.9).

💡

Critical Insight

Content Filtering-based filtering creates an echo chamber effect with the lowest ILD (0.0064), confirming that pure content filtering traps users in narrow taste profiles. However, it naturally discovers some niche content (26.25%).

🏆

Best Overall: Graph DPP Rerank

The Graph DPP Rerank algorithm achieves the best balance: competitive diversity (ILD 0.0187), lowest popularity bias (46.64), and highest niche representation (33.75%). The DPP re-ranking combined with the fairness constraint creates the most equitable recommendations.

📌

Practical Implication

Traditional recommendation systems (Greedy, CF) inherently reduce diversity. Graph DPP Rerank-based systems significantly improve diversity and fairness in music recommendations, promoting independent artists and reducing filter bubbles.

📐 Metrics Formulas

Intra-List Diversity (ILD)

ILD = 1|L|(|L|−1) ∑_i≠j (1 − sim(i,j))

Where sim(i,j) = cosine content_filtering between feature vectors

Popularity Concentration Index

G = ∑_i (2i − n − 1) · x_in · ∑ x_i

Applied to popularity distribution of recommended songs

DPP Kernel

L_ij = q_i · q_j · S_ij

Quality-weighted content_filtering kernel for determinantal point process