This paper contributes in detecting chaotic behaviors in dynamic complex social networks using a new feature diffusion-aware model from two perspectives of abnormal links as well as abnormal nodes. The proposed approach constructs a probabilistic model of dynamic complex social networks and subsequently, applies it to detect chaotic behaviors by measuring deviations from the model. The predictive model considers the main processes of features' dynamics, evolution of nodes' features, feature diffusion, and link generation processes in dynamic complex social networks. The feature diffusion process indicates the process in which each node former features influence the future features of its neighbors. The proposed approach is validated by experiments on two real dynamic complex social network datasets of Google+ and Twitter. The approach uses some Markov Chain Monte Carlo sampling methods like Metropolis-Hastings algorithm and Slice sampling strategy to extract the model parameters, given these real datasets. Experimental results indicate the improved performance characteristics of the proposed approach in comparison with baseline approaches in terms of the performance measures of accuracy, F-score, Matthews Correlation Coefficient, recall, precision, area under ROC curve, and log-likelihood.