Average Jaccard (AJ) Metric

• Average Jaccard is defined as the arithmetic mean of per-sample Jaccard indices, offering a normalized similarity measure between 0 and 1.
• It is widely applied in medical image segmentation, clustering, and network science, capturing both overlap and scale differences in diverse data types.
• Utilizing metric-sensitive surrogates like soft-Jaccard loss, the AJ metric enables robust optimization and resilience against outliers and class imbalances.

The Average Jaccard (AJ) metric generalizes the classical Jaccard index, providing a robust, normalized, and interpretable measure of similarity between finite sets, vectors, or graph neighborhoods. It is prominent in evaluating predictive performance in structured prediction tasks such as medical image segmentation, clustering, and network science, and has recently been studied in the context of robust statistics and random graphs. AJ appears in multiple equivalent guises: as the arithmetic mean of per-sample Jaccard indices, as the similarity-mean maximizing overall intersection-over-union, and as an average over vertex pairs in graph-theoretic models.

1. Formal Definitions and Variants

For sets $A, B \subset \Omega$ (for example, ground truth and predicted foreground voxels), the Jaccard index is

$$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$

Given $N$ pairs $(A_i,B_i)$, the Average Jaccard is the arithmetic mean: $\mathrm{AJ} = \frac{1}{N} \sum_{i=1}^N J(A_i,B_i).$ This definition generalizes to nonnegative vectors or multisets $$\{\mathbf{x}^{(i)}\}_{i=1}^N \subset \mathbb{R}_{\ge 0}^M$$ by using

$$\mathcal{J}(\mathbf{y}, \mathbf{x}^{(i)}) = \frac{\sum_j \min\{y_j, x^{(i)}_j\}}{\sum_j \max\{y_j, x^{(i)}_j\}},$$

and averaging over the collection.

For the Erdős–Rényi random graph $G(n,p)$, the AJ index is the mean Jaccard of all unordered pairs of vertex neighborhoods: $$\mathrm{AJ}(G) = \frac{1}{\binom{n}{2}} \sum_{i<j} J_{i,j}, \quad J_{i,j} = \frac{|N(i) \cap N(j)|}{|N(i) \cup N(j)|}.$$ The AJ metric inherits the $[0,1]$ range and is sensitive to both overlap and scale, providing a normalized measure across a variety of domains (Bertels et al., 2019, Travieso et al., 2023, Feng et al., 2023).

2. Theoretical Properties and Optimization Surrogates

The Jaccard and Dice scores are closely related: one can be written in terms of the other as $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$0 and $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$1. They approximate each other within a multiplicative factor of 2 and have a maximal absolute difference of $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$2. Metric-sensitive surrogates such as the soft-Jaccard loss,

$$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$3

have differentiable structure suitable for training via gradient-based methods. The gradient

$$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$4

is global in nature—each pixel's gradient depends on all others—reflecting the non-local overlap structure of the Jaccard index.

No finite weighting of cross-entropy can provide a uniform multiplicative bound or a relative $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$5-approximation with respect to Jaccard (or Dice), as the ratio between Dice/Jaccard and weighted Hamming similarity is unbounded for vanishingly small foregrounds (Bertels et al., 2019). Thus, directly optimizing a metric-sensitive surrogate is both theoretically justified and practically robust.

3. Robustness, Analytical Characteristics, and Computation

In vector (and scalar) data, the "Jaccard similarity mean" (also referred to as the AJ mean) is the sample element maximizing the average Jaccard similarity to all sample elements. Formally,

$$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$6

for the collection $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$7. In the one-dimensional case, the maximizer $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$8 always belongs to the sample. The Jaccard-mean is bounded in $$J(A,B) = \frac{|A\cap B|}{|A\cup B|},\quad 0 \leq J(A,B) \leq 1.$$9, preserves scale (homogeneity), and shifting the dataset asymptotically brings the mean toward the median. For a mixture contaminated by a fraction $N$0 of outliers, the displacement of the Jaccard mean is typically much less than that of the arithmetic mean,

$N$1

demonstrating enhanced robustness (Travieso et al., 2023). The computational complexity for the maximizer in the scalar case is $N$2, exploiting sorting and cumulative sums.

4. Statistical Behavior in Random Graphs

In the Erdős–Rényi model $N$3, the AJ index of a graph's vertex neighborhoods has an explicit distribution: for each pair $N$4,

$N$5

is constructed from overlapping Bernoulli random variables. The mean and variance satisfy

$N$6

when $N$7. The mean of all pairwise Jaccards ($N$8) converges, under mild regularity assumptions ($N$9, $(A_i,B_i)$0), to a normal distribution: $(A_i,B_i)$1 (Feng et al., 2023). The asymptotic regime is governed by central limit theorems for the counts of edges and length-2 paths—AJ thus behaves as a typical (weakly dependent) U-statistic in large random structures.

5. Empirical Performance and Applications

In medical image segmentation, AJ is the standard for evaluating predicted object masks against ground truth, computed by averaging per-sample Jaccard indices. In large-scale benchmarks—e.g., BRATS-2018, ISLES 2017/2018, MO17, PO18—metric-sensitive surrogates (soft-Jaccard, soft-Dice, Lovász-softmax) deliver nearly identical AJ performance, significantly outperforming cross-entropy losses. Weighted cross-entropy is unreliable, yielding inconsistent results across tasks and object sizes. Critically, metric-sensitive surrogates maintain stable AJ even in low-prevalence deciles; cross-entropy degrades markedly on small objects, highlighting the necessity of direct metric alignment for reliable segmentation (Bertels et al., 2019).

In robust statistics, the AJ mean provides a bounded, dimensionless centrality measure that is less susceptible than the arithmetic mean to outlier contamination. In empirical studies, such as analysis of solar cycles, the Jaccard mean accentuates subgroup structures that classical means may obscure, particularly under skewed or multimodal distributions (Travieso et al., 2023).

6. Practical Recommendations and Guidelines

• When optimizing for Jaccard-based evaluation (or vicariously for Dice), always select a metric-sensitive surrogate loss (soft-Jaccard, soft-Dice, Lovász-softmax) rather than (weighted) cross-entropy; no hyper-parameter search over cross-entropy weights can bridge the theoretical or empirical gap.
• Choice among metric-aligned surrogates may be governed by computational considerations, as their effects on AJ are essentially equivalent up to a minor constant factor.
• Compute AJ by averaging per-sample indices to prevent structural bias; avoid using a pooled global confusion matrix.
• For low-prevalence objects or highly imbalanced data, the invariance of AJ to object scale ensures reliability of evaluation.
• The AJ mean offers an alternative to outlier-sensitive averaging, remaining within the data’s convex hull and providing interpretable measures of centrality and variability (Bertels et al., 2019, Travieso et al., 2023).

7. Summary Table: Average Jaccard in Key Applications

Domain	AJ Definition	Analytical/Empirical Role
Medical Segmentation	Mean per-sample Jaccard index	Evaluation metric and risk surrogate; stable under class imbalance (Bertels et al., 2019)
Random Graphs	Mean of pairwise neighborhood Jaccards over vertices	Limiting normality under $(A_i,B_i)$2; explicit moments (Feng et al., 2023)
Robust Statistics	Maximizer of average Jaccard similarity to dataset	Outlier robustness; enhanced modal selectivity (Travieso et al., 2023)