The ability to accurately predict the timing of the next purchase is critical for business decision-making yet challenging. Shopping regularity is often disrupted by negligible transaction costs of e-commerce and the ease of responding to promotions, fostering increasingly irregular behavior and thereby complicating prediction efforts. In addition, existing predictive models often overlook how buying in one product category affects future purchases in others. To address these issues, this study proposes a deep learning framework that integrates the purchase irregularity, captures category-specific purchase patterns, and learns cross-category interactions for effective next purchase time prediction at product category level. Specifically, we model purchase irregularity as a latent state that adaptively captures whether a purchase in each product category tends to follow a routine pattern or not. Then we utilize the LSTM networks to capture recurring purchase patterns based on past inter-purchase intervals. Finally, a self-attention mechanism is applied to capture interactions of shopping behaviors among distinct product categories, learning how the timing of purchases in one category can affect purchasing behavior in others. Experimental evaluations on a large-scale retail dataset demonstrate the effectiveness of our approach. The proposed model improves purchasing time prediction and enables businesses to better anticipate demand fluctuations and optimize resource allocation in online marketplaces.