Cellular network performance depends heavily on the configuration of its network parameters. Current practice of parameter configuration relies largely on expert experience, which is often suboptimal, time-consuming, and error-prone. Therefore, it is desirable to automate this process to improve the accuracy and efficiency via learning-based approaches. However, such approaches need to address several challenges in real operational networks: the lack of diverse historical data, a limited amount of experiment budget set by network operators, and highly complex and unknown network performance functions. To address those challenges, we propose a collaborative learning approach to leverage data from different cells to boost the learning efficiency and to improve network performance. Specifically, we formulate the problem as a transferable contextual bandit problem, and prove that by transfer learning, one could significantly reduce the regret bound. Based on the theoretical result, we further develop a practical algorithm that decomposes a cell's policy into a common homogeneous policy learned using all cells' data and a cell-specific policy that captures each individual cell's heterogeneous behavior. We evaluate our proposed algorithm via a simulator constructed using real network data and demonstrates faster convergence compared to baselines. More importantly, a live field test is also conducted on a real metropolitan cellular network consisting 1700+ cells to optimize five parameters for two weeks. Our proposed algorithm shows a significant performance improvement of 20%.