Ranking Candles
These plots show the change in ranking quality with increasing data set dimensionality for different distance functions.
| All-Relevant | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
| 10-Relevant | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
| Cyc-Relevant | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
| Half-Relevant | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
| All-Dependent | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
| 10-Dependent | Manhattan | Euclidean | L0.6 | L0.8 | Arccosine |
Notes:
- A ranking quality of 0.5 means "as good as random ranking". A value of 1.0 is a perfect ranking, a value of 0.0 is a perfect reversed ranking.
- The split central 20%, middle 60%, outer 20% is to visualize degradation with decreasing centrality (measured by distribution density at data generation)
- Lp norms with p < 1 ("fractional Lp norms") are no longer metric: they do not satisfy the triangle equality.
- Manhattan = L1, Euclidean = L2.
