Viscosity is one of the most investigated properties of glass materials in the past decades. Despite the large number of studies involved with glass viscosity, the huge design space of glass leaves the conventional experimental and simulation methods deficient to explore the new glass compositions yielding advanced viscosity performance. In this regard, machine learning methods provide promising solutions for mapping the oxide composition of unknown glasses to their viscosities. Here, based on a large glass dataset (>100,000 glasses), we study multiple machine learning models to predicting viscosity as a function of the glass composition and temperature, with a special focus on explaining the data pattern as learned by various machine learning approaches. These models allow us to decipher the influence of individual oxide on viscosity and to determine the range of feasible glass compositions satisfying a specific viscosity.