<![CDATA[YouTube Shorts has rapidly emerged as a dominant short-form video platform, yet small creator channels often experience an unusual viral phenomenon best described as flash viral—a sudden surge of views that peaks within 24 to 48 hours and then collapses almost as quickly. Detecting and explaining this pattern is challenging because traditional statistical detectors miss multivariate signatures, while classical Association Rule Mining (ARM) such as Apriori loses information through mandatory discretization. This study proposes a hybrid framework that combines a semi-supervised Deep Learning Autoencoder with Nature-Inspired Numerical Association Rule Mining (NiaARM) using Differential Evolution and Particle Swarm Optimization. The framework is validated on six temporal snapshots of the Indonesian Boburu YouTube Shorts channel, comprising 63 unique videos (42 active) collected between February 22 and March 19, 2026. Experimental results show that the Autoencoder achieves an F1-score of 0.667 with 100% recall, matching the best classical baseline (Z-Score) while providing learnable representational capacity for future scaling. NiaARM-PSO discovered 3,115 high-quality numerical association rules with a maximum lift of 63.00, compared to only 43 rules and a maximum lift of 2.52 obtained by Apriori, an improvement of approximately 25 times. Traffic source decomposition further revealed that 99.9% of viral views originated from external platforms rather than YouTube's recommendation system, indicating that flash viral on micro-channels is externally driven. This research contributes a methodological framework that simultaneously detects and explains flash viral phenomena in short-form video analytics]]>
