Jaccard Index
Also known as: Jaccard Similarity, Intersection over Union (IoU)
The Jaccard index measures the similarity between predicted and true label sets by computing the ratio of intersection to union. It is widely used in multi-label classification and set-based similarity tasks where partial overlap is important.
Key highlights
- Symmetric metric treating false positives and false negatives equally.
- Directly interpretable as the fraction of union that is correctly predicted.
- Provides partial credit for partially correct predictions.
- Works naturally for multi-label and set-based problems.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Jaccard index for multi-label classification when you want to emphasize that both false positives and false negatives matter equally. It is suitable for image segmentation (intersection over union of pixel sets), document tagging, multilabel classification, and any set-based comparison task.
Strengths & limitations
- Symmetric metric treating false positives and false negatives equally.
- Directly interpretable as the fraction of union that is correctly predicted.
- Provides partial credit for partially correct predictions.
- Works naturally for multi-label and set-based problems.
- Penalizes both false positives and false negatives equally, which may not match domain costs.
- Can be overly strict when only a few labels are expected but many are possible.
- Less intuitive to practitioners unfamiliar with set theory.
- May be sensitive to label set size; larger sets can result in lower Jaccard scores.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does Jaccard differ from Hamming loss?
Hamming loss averages label-by-label errors (how many labels are wrong out of all labels). Jaccard compares predicted and true label sets (what fraction of the union is correct). Jaccard is stricter because omitting a label or predicting an extra label both hurt the score. Hamming loss can give credit for partially correct predictions at the label level.
Is Jaccard the same as IoU (Intersection over Union)?
Yes, they are equivalent. Jaccard index IS intersection over union (IoU). Both terms refer to the same metric: |intersection| / |union|. IoU is more commonly used in computer vision (image segmentation), while Jaccard is more common in statistics and multi-label learning.
What is a good Jaccard score?
It depends on the problem difficulty. A Jaccard of 0.5 is often considered moderate. On hard problems with many possible labels, Jaccard can naturally be lower. Compare against appropriate baselines (e.g., predicting the most common labels for all instances) and domain expectations.
Sources
- 1.Jaccard, P. (1901). Etude comparative de la distribution florale dans une portion des Alpes et des Jura. Bulletin de la Société Vaudoise des Sciences Naturelles, 37, 547-579.
- 2.Tsoumakas, G., & Katakis, I. (2007). Multi-label classification: An overview. International Journal of Data Warehousing and Mining, 3(3), 1-13.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Jaccard Index. ScholarGate. https://scholargate.app/model-evaluation/jaccard-index