Microbiota is viewed as a network of diverse microorganisms connected by various interspecific interactions. While the stress gradient hypothesis (SGH) predicts that positive interactions is favored in stressful environments, the prediction has been less explored in complex microbiota due to the challenges of identifying interactions. Here, by applying a nonlinear time-series analysis to the metagenomic data of the soil microbiota cultured under suitable (30℃) or stressful (37℃) conditions, we show how the microbial network responds to temperature stress. The higher temperature lowers the network complexity in terms of genera diversity and connectance. At local scale, there were no correlations in degrees, interaction strengths or interacting partners of individual genera between the treatments. While the genera that persisted only under the suitable condition gave fewer positive effects, the genera that appeared only under the stressful condition received more positive effects in agreement with SGH. However, temperature changes also induce reconstruction of a community network, leading to an increased proportion of negative interactions at the whole community level. The anti-SGH pattern can be explained by the stronger competition caused by increased metabolic rate and population densities.