Training datasets often contain subtle but irrelevant patterns that bias image classifiers and limit their generalization. Prior research has focused on detecting biases in misclassified data using manually defined dimensions, such as age or gender. However, since these approaches rely on manually predefined dimensions, they are labor-intensive and bound to be limited in terms of their coverage. Moreover, existing methods overlook biases present in correctly classified data. To address these issues, we propose an unsupervised framework that leverages commonsense knowledge graphs and open-source foundation models to automatically derive semantic dimensions and their values, identifying biases that influence correct and incorrect classifications of data. Using these dimensions and values, we construct scene graphs and identify distinctive graph patterns for correctly and incorrectly classified data, uncovering how combinations of semantic dimensions impact classifier decisions. Evaluations confirm that our scene graphs are of high quality, as they include information from manual annotations while being denser and more informative than manually constructed graphs. Moreover, our framework demonstrates high predictive accuracy, effectively identifying patterns responsible for correct and incorrect classifications, with F1 scores ranging from 0.72 to 0.96. It also proves effective for error analysis and proactive bias detection in datasets.