To find a particular solution for a given problem or to answer any question we need to find a solution from the given set of solution by pair wise comparisons. A historical set of data help us to find different answers for the same question asked multiple times. In our proposed system there are two models one is "offline model" and another one is "online model". In offline learning component we establish training samples with the help of data driven observations. In online search component a pool of candidate answer is collected for a given question by finding similar types of questions. With the help of offline trained model candidate answer are sorted in preference order. A comparative demonstration of the experiments on the real-world community-based question answering is shown.