DinTucker: Scaling up Gaussian process models on multidimensional arrays with billions of elements