Machine learning methods have been regarded as practical tools for the inverse design of nanophotonic devices. However, for the devices with complex expected targets, such as the spectrum with multiple peaks and valleys, there are still many sufferings remaining for these data-driven approaches, such as overfitting. To resolve it, we firstly propose a hybrid inverse design scheme combining supervised and unsupervised learning. Compared with the previous inverse design schemes based on artificial neural networks (ANNs), clustering algorithms and an encoder model are introduced for data preprocessing. A typical metamaterial composed of multiple metal strips that can produce tunable dual plasmon-induced transparency phenomena is designed to verify the performance of our proposed hybrid scheme. Compared with the ANNs directly trained by the entire dataset, the loss functions (mean squared error) of the ANNs in our hybrid scheme can be effectively reduced by more than 51% for both training and test datasets under the same training conditions. Our hybrid scheme paves an efficient improvement for the inverse design tasks with complex targets.