Ecological interactions among biological species are important factors that regulate the dynamics of communities in ecosystems and determine their properties. Gaining a knowledge of how species in a natural community interact with each other is thus critical for predicting and controlling the community dynamics. A practical methodology to evaluate the effect of species interactions will then bring researchers and practitioners wider opportunities for understanding and managing natural communities. When we want to quantify the ecological interactions with empirical data, one of key aspects to consider is their state-dependence. When species influence one another in a state-dependent way, the interaction strengths can temporally fluctuate as each species abundance shifts on the non-equilibrium attractor. While some methods such as S-map or regularised S-map have been widely used to infer the effects of state-dependent interactions in recent ecological studies, these existing methods have some limiting aspects. Here, we propose a novel inference method based on a nonparametric Bayesian machine learning technique called Gaussian process regression (GPR). Analogously to the S-map methods, the proposed GPR-based method allows for the inference of the effects of state-dependent interactions without any prior knowledge about governing equations underlying the community dynamics. To test the validity and advantage of the proposed method, we applied it to synthetic and empirical time series data. Throughout the analyses of synthetic and empirical time series data, the results suggest that our proposed method has several advantages over the existing S-map methods. First, the proposed method may have performance advantages over the existing methods since it could achieve better inference accuracies in the analyses of synthetic time series data under different noise conditions. Second, our method could recover the response of the interaction strengths to hypothetical abundance changes at least in the nearby region of each data point. That is, we can locally simulate the state-dependent variability of interaction strengths to perform scenario explorations of the effects of ecological interactions. This suggest that our method can provide more detailed pictures of the mechanisms of state-dependent interactions than the existing methods. Third, unlike the S-map models which can only provide the point estimates of interaction strengths, the GPR model can easily evaluate the uncertainties associated with the inferences using the posterior distributions of interaction strengths. Overall, our results illustrate that the proposed method can be a practical tool to evaluate the effects of state-dependent interactions with time series data.