Transfer learning based maximum entropy clustering

The classical maximum entropy clustering (MEC) algorithm can only work on a single dataset, which might result in poor effectiveness in the condition that the capacity of the dataset is insufficient. To resolve this problem, using the strategy of transfer learning, this paper proposed the novel transfer learning based maximum entropy clustering (TL_MEC) algorithm. TL_MEC employs the historical cluster centers and membership of the past data as the references to guide the clustering on the current data, which promotes its performance distinctly from three aspects: clustering effectiveness, anti-noise, as well as privacy protection. Thus TL_MEC can work well on those small dataset if enough historical data are available. The experimental studies verified and demonstrated the contributions of this study.