1: \begin{abstract}

2: In many application areas, data are collected on a categorical response and high-dimensional

3: categorical predictors, with the goals being to build a parsimonious model for classification while doing inferences on the important predictors.

4: In settings such as genomics, there can be complex interactions among the predictors. By using a

5: carefully-structured Tucker factorization, we define a model that can characterize any

6: conditional probability, while facilitating variable selection and modeling of higher-order

7: interactions. Following a Bayesian approach, we propose a Markov chain Monte Carlo

8: algorithm for posterior computation accommodating uncertainty in the predictors to be

9: included. Under near sparsity assumptions, the posterior distribution for the conditional

10: probability is shown to achieve close to the parametric rate of contraction even in ultra

11: high-dimensional settings. The methods are illustrated using simulation examples

12: and biomedical applications.

13: \keywords{Classification; Convergence rate; Nonparametric Bayes;

14: %Prediction;

15: Tensor factorization; Ultra high-dimensional; Variable selection.}

16: \end{abstract}

17: