M.S. Applied Data Science - Capstone Chronicles 2025

37

severe recall is misclassified as Class II or III, could result in dangerous delays in public awareness and regulatory response. However, precision remains equally important. A false positive, where a lower-risk recall is misclassified as Class I, may result in misallocated resources, unnecessary public concern, and a reduced focus on genuinely critical recalls. For instance, if the model flags a minor packaging issue as Class I, it could overshadow a genuine life-threatening case. Thus, a balance between high recall and high precision is crucial. A true positive reflects a correctly predicted high-severity event, enabling timely interventions, while a true negative reflects correct identification of non-critical recalls as lower severity. A similar argument applies to Class II recalls, which involve moderate health risks. Failing to identify them correctly (false negatives) may lead to regulatory delays or continued consumer exposure to harmful products. In contrast, for Class III recalls, which involve minimal risk and are unlikely to cause harm, recall is less critical. Missing a Class III recall (false negative) typically does not compromise public safety. In this class, precision becomes more important, as it helps avoid incorrectly labeling these minor recalls as more severe. A false positive in this scenario might unnecessarily escalate a minor labeling issue to Class II or I, diverting attention from more important cases. Therefore, in lower-severity classes, the emphasis shifts to avoiding overreaction, whereas in Class I and II, both recall and precision are essential to protect public health without triggering false alarms. Compared to previous studies that have focused on the lengthy processing times associated with

recall classification after a recall event is identified, our approach offers a more efficient and accurate solution (Darbyet al., 2023). The models developed achieve precision and recall scores of 0.91, indicating a well-balanced performance. This means the model is not favoring precision over recall and is equally effective at correctly identifying cases while minimizing misclassifications across all three classes. 6.1 Limitations One limitation encountered was slow computation during model training. However, once the models are saved, they can be loaded quickly, enabling fast performance in our Streamlit dashboard. Future improvements could involve better computer hardware and more efficient modeling techniques. Another limitation is the inability to scrape thousands of FDA web links to determine whether a recall was initiated by the manufacturer or triggered by an inspection. The FDA website restricts high-volume requests, which limits automated data collection. Without the complete timeframe of when a recall was detected and how long a product has been in circulation in the market is difficult to accurately measure. This study highlights the importance of evaluating classification models using a combination of performance metrics. Precision and recall offer complementary perspectives that are especially crucial in high-stakes contexts such as FDA recall classification. Their joint consideration ensures that machine learning tools not only detect severe risks effectively but also maintain the trustworthiness and efficiency of the regulatory response.

41