Exploiting Shared‐Specific Patterns via Memory Networks for Cross‐City Cellular Traffic Forecasting

Forecasting the spatiotemporal distribution of cellular traffic is of great significance for proactive urban management. Given the quality issues of cellular traffic data in certain cities, various cross‐city transfer strategies have been adopted to ensure accurate forecasting. However, existing methods generally neglect geospatial effects, limiting their transfer performance. Specifically, geographical similarity can guide the reliable discovery of patterns shared across cities, while spatial heterogeneity emphasizes the complementary exploration of patterns unique to the target city. To fill these gaps, this study proposes a Shared‐specific learning framework (CrossSSL) for Cross‐city cellular traffic forecasting. In CrossSSL, a key component, namely the pattern memory network, is introduced to globally store cellular traffic patterns. The pattern memory network (1) adopts a specialized memory structure to incorporate prior geographical environment knowledge, (2) employs a geographical contrastive optimization strategy to leverage geographical similarity, and (3) is configured with a dual‐integration mechanism to accommodate spatial heterogeneity. Therefore, both city‐shared and city‐specific patterns can be fully exploited in the overall framework, facilitating the ultimate goal of cross‐city forecasting. Experiments on two real‐world cellular traffic datasets demonstrate that, CrossSSL can outperform state‐of‐the‐art baselines under different forecasting scenarios. Even under data‐scarce conditions, CrossSSL exhibits robust performance.