This letter considers optimizing user association in a heterogeneous network via utility maximization, which is a combinatorial optimization problem due to integer constraints. Different from existing solutions based on convex optimization, we alternatively propose a cross-entropy (CE)-based algorithm inspired by a sampling approach developed in machine learning. Adopting a probabilistic model, we first reformulate the original problem as a CE minimization problem which aims to learn the probability distribution of variables in the optimal association. An efficient solution by stochastic sampling is introduced to solve the learning problem. The integer constraint is directly handled by the proposed algorithm, which is robust to network deployment and algorithm parameter choices. Simulations verify that the proposed CE approach achieves near-optimal performance quite efficiently.