Dataset Description
Summary of Dataset Bias analysis with quantitative Validation
This section provides an overview of the dataset composition, the statistical evidence for representational bias, and the rationale for the weighted calculation method used to correct it. The dataset comprises 7,362 images distributed across 20 categories. Statistical analysis showed measurable inequality in category sizes and object representation, which would compromise analytical validity if uncorrected.
Key Bias Findings
Category size imbalance: Chi-square goodness of fit, χ² = 58.1, p ≈ 8×10⁻⁶, with a small effect size (Cohen’s w = 0.089). Category sizes range from 301 to 415 images (ratio 1.38:1).
Object representation skew: “People” objects were overrepresented in technology-related categories (up to 2.82× expected) and severely underrepresented in nature-related categories (as low as 0.16× expected).
Impact on analysis: Larger categories exert disproportionate influence on trends, correlations, and model training, while smaller categories produce less reliable estimates due to lower statistical power.
Why Weighting Was Needed
Two distinct structural biases were identified:
-
-
- Category size bias — larger categories would dominate any aggregate analysis simply due to having more images.
- Object density bias — images with many detected objects would contribute disproportionately more than sparsely tagged images, regardless of relevance
-
Weighted Calculation Method
Category size adjustment
$$ w_{\mathrm{img}}(i) = \frac{1}{N_{c_i}} $$
Where Nci is the number of images in the category of image i. This ensures every category contributes equally overall, regardless of its size.
Object density adjustment
$$ w_{\mathrm{obj}}(i) = \frac{w_{\mathrm{img}}(i)}{O_i} $$
where Oi is the number of detected objects in image i. This ensures that each detected object within an image contributes equally, preventing densely tagged images from dominating.
Weighted value per group
$$ x_{ig} = n_{ig}\, w_{\mathrm{obj}}(i) $$
where nig is the count of objects in image i belonging to group g.
This double weighting directly implements the bias-correction strategy validated in the statistical analysis, controlling for both unequal category sizes and inconsistent object densities.
Conclusion
The weighted calculation method is a mathematically direct and empirically validated response to the structural imbalances identified in the dataset. It produces fairer, bias-minimised comparisons across categories, supporting robust statistical inference and valid interpretation.