Abstract: In the big data era, the widespread of massive multi-modal data has caused huge changes in the data characteristics, namely wide variety and low value density.Different types of data are characterized by both function independently and complement each other.Discovering the hidden value behind multi-modal data has become the key problem in big data mining tasks.Therefore, to tackle the shortcomings of the low-quality multimodal data, this paper proposes a new multimodal robust feature learning method by introducing the modal specific error matrix.The effect of noisy information on the fusion result can thus be effectively reduced.Moreover, a dual graph-regularization mechanism for data manifolds and feature manifolds is designed to describe the spatial structure of multimodal data, which can ensure the data stability during multimodal feature learning.On six real-world multi-modal data sets, the results are compared with several classical algorithms in recent years based on three evaluation indexes, namely accuracy (ACC), normalized mutual information (NMI) and purity (PUR).Experimental results show that the proposed method is superior to all other compared algorithms, especially in network data sets Webkb containing large amounts of noise information, its ACC and NMI are improved by about 10% compared with the baseline algorithms.It can be seen that the proposed algorithm can accurately learn the sharing features of multi-modal data.

Key words: Adaptive weight, Dual graph-regularization, Multimodal data, Noisy data, Robust feature learning