Recent imaging techniques enable us to observe activities of hundreds of neurons simultaneously as spike sequences. The objective of this study is to estimate the network structure based on such spike sequences. Our method is an extension of existing sparse regression technique, in which we have implemented the following three ideas: (1) Each spike time -series obeys a non -stationary Poisson process whose Poisson intensity is given by an auto -regression model. (2) Spike response functions are represented by a linear summation of smooth basis functions. (3) A group -LASSO regularization is applied to obtain a sparse regression solution. When applied to simulation datasets, our method showed a better estimation performance than that by an existing state-of-the-art method.