Adaptive multi-view clustering via cross trace lasso

We propose a novel multi-view clustering method by learning auto-regression problems under structural constraints and treating the regression coefficients as new feature representations for the cluster partition. In particular, we take the data intrinsic correlation structure into account. Correlated data under one view tend to be also related under another view and are likely to fall into the same group. Therefore we pair the data matrix from one view and the regression coefficient from a different view together to meet a trace Lasso constraint, which adaptively adjusts the sparsity of regression coefficients in order to promote consistent data correlations across views. Then a joint low-rank constraint is further imposed to encourage similar regression coefficients for the same samples under distinct views. Finally, we develop an effective algorithm to optimize the objective function. And experimental results demonstrate that our method is useful and fairly competitive compared with other state-of-the-art multi-view clustering methods.