The microscopic cascade prediction task explores how information diffuses on social media. Existing methods typically model the connections among users as static relationship, limiting their ability to reflect the evolving and event-driven nature of user interactions. Moreover, subsequent target users are usually recognized as negative samples, leading to a misclassification situation. In response, we propose EMAO, an Expectation-Maximization and Adaptive Objective optimization algorithm for microscopic cascade prediction. By iteratively updating edge weights using the EM algorithm, it captures evolving user relationships, effectively addressing the oversimplification of static modeling. Furthermore, we propose an adaptive objective that incorporates both hard and soft labels. Hard labels guide the optimization of the current prediction target, whereas soft labels provide informative priors for subsequent targets by assigning reasonable expected probabilities. Experimental results across four datasets demonstrate that EMAO outperforms the state-of-the-art models with the average improvements of 3.04% and 2.20% in Hits@ \(\kappa \) and MAP@ \(\kappa \) metrics, respectively, validating its effectiveness.