In this paper, a channel estimation technique for reconfigurable intelligent surface (RIS)-aided multi-user multiple-input single-output communication systems is proposed. By deploying a small number of active elements at the RIS, the RIS can receive and process the training signals. Through the partial channel state information (CSI) obtained from the active elements, the overall training overhead to estimate the entire channel can be dramatically reduced. To minimize the estimation complexity, the proposed technique is based on the linear combination of partial CSI, which only requires linear matrix operations. By exploiting the spatial correlation among the RIS elements, proper weights for the linear combination and normalization factors are developed. Numerical results show that the proposed technique outperforms other schemes using the active elements at the RIS in terms of the normalized mean squared error when the number of active elements is small, which is necessary to maintain the low cost and power consumption of RIS.