Distributed Approximation Algorithms via LP-duality and Randomization