Robust Representation Learning for Unified Online Top-K Recommendation