Algorithm-Dependent Bounds for Representation Learning of Multi-Source Domain Adaptation