The minimum mean squared error (MMSE) receiver is a linear filter which can achieve optimal near-far resistance in direct-sequence code-division multiple-access communications. However, one of the main problems of this receiver is the required number of filter taps, which is typically large. This is especially true in systems with a large processing gain in which case the receiver's computation burden becomes very high. As a result, methods for reducing the complexity of the MMSE receiver have been of great interest in recent years. We propose an efficient partitioned MMSE receiver based on a classification algorithm. It is shown that the computational complexity (in terms of the filter taps) of the proposed receiver can be reduced significantly while good performance is maintained. Based on the special structure of our proposed receiver, we also propose a release-merge adaptive partition algorithm which can update the partition and the receiver's coefficients simultaneously. In particular, it is demonstrated that the proposed receiver can perform much better than previously proposed reduced-rank MMSE receivers, such as the partial despreading MMSE receiver and the cyclically shifted filter bank receiver, with even a smaller number of taps.