AAI_2025_Capstone_Chronicles_Combined

The confusion matrices indicate only marginal differences in predictions when time vectors are added to Model 1 classifier architecture. This observation aligns with the distribution plots shown in Figures 3 and 4, where the only discernible distribution shift between March 2011 and February 2012 was temporal in nature, and fully captured within that interval. In production, the ML system will process and classify new, unlabeled images. In the absence of ground ‑ truth labels, it is essential to understand and correctly interpret the meaning of model outputs and associated prediction statistics. To this end, we evaluated the reliability of the pipeline’s probabilistic outputs using the Top ‑ Versus ‑ All method with Histogram Binning calibration. Calibration was performed using the labeled validation set employed during model tuning.

293