Online transferable representation with heterogeneous sources

Learning from streaming data has gained a lot of attention and interest in the past decades. These improvements have shown promising results when the models are trained and test on a single streaming source. However, the trained model often fail to produce the reliable results due to the difficulty of data shift and knowledge transfer with heterogeneous streaming domains. In this paper, we propose an architecture that is based on autoencoders. Specifically, we use online feature learning based on denoising autoencoder to learn more robust representations from streaming data. In order to tackle with data shift between source and target streaming data, we develop an ensemble weighted strategy, which can effectively handle the concept drifts of streaming data. Moreover, we develop the transfer mechanism, which is capable of transferring label information across heterogeneous domains. Finally, we combine online learning, data shift adaption and knowledge transfer with heterogeneous domains into a single process, which makes our proposed architecture powerful in learning and predicting for multistream classification problem. Experiments on heterogeneous datasets validate that the proposed algorithm can quickly and accurately classify instances on a stream together with a small number of labeled examples. Compared with a few related methods, our algorithm achieves some state-of-the-art results.