Convex Latent Effect Logit Model via Sparse and Low-rank Decomposition