1. Definition

Relative Accuracy is a comparative metric: it expresses a model’s accuracy relative to a baseline (reference model, naive predictor, or previous system).

$\text{Relative Accuracy} = \frac{\text{Accuracy(model)}}{\text{Accuracy(baseline)}}$

  • If = 1 → the model is as good as the baseline.
  • If > 1 → the model is better than the baseline.
  • If < 1 → the model is worse than the baseline.

2. Why it’s Useful

  • Absolute accuracy (e.g., 85%) tells us how often the model is correct.
  • Relative accuracy tells us how much better (or worse) it is compared to a baseline.
  • Useful in papers and reports where you need to quantify improvements over existing methods.

3. Example 1: Classification

  • Baseline (majority class predictor): 70% accuracy
  • Model: 84% accuracy

$\text{Relative Accuracy} = \frac{0.84}{0.70} = 1.20$

Interpretation: The model is 20% more accurate relative to baseline.

4. Example 2: Recommender Systems

  • Baseline recommender: Precision@10 = 0.25
  • New recommender: Precision@10 = 0.30

$\text{Relative Accuracy} = \frac{0.30}{0.25} = 1.20$

The new recommender improves performance by 20% relative to the baseline.

5. Alternative Expression

Sometimes people report Relative Improvement (a percentage gain) instead of a ratio:

$\text{Relative Improvement} = \frac{\text{Accuracy(model)} – \text{Accuracy(baseline)}}{\text{Accuracy(baseline)}} \times 100\%$

Using Example 1:

$\frac{0.84 – 0.70}{0.70} \times 100\% = 20\%$

So:

  • Relative Accuracy (ratio) = 1.20
  • Relative Improvement (percentage) = +20%

6. Key Points

  • Absolute Accuracy = direct proportion of correct predictions.
  • Relative Accuracy = performance compared to baseline (expressed as ratio).
  • Use case: Highlights improvement, especially when baselines vary.

Summary:
Relative Accuracy compares a model’s accuracy to a baseline accuracy. It is expressed as a ratio (≥0), showing whether the model is equal to, better, or worse than the baseline. Often reported alongside relative improvement (%), it’s widely used in ML papers and recommender evaluations to quantify gains.