Abstract: Exploration and analysis of high-dimensional datasets require extracting valuable insights and generating proper charts from data. To lower the barriers of creating effective visualizations, a new insight-centric visualization recommendation framework is presented that recommends both what to visualize(data query recommendations) and how to visualize it(visual encoding recommendations). For the former,instead of using single indicators, multiple analysis methods and learning-based algorithms are utilized to automatically rank data queries based on the insights and the most informative ones are provided to the users. For the later, rule-based and learning-based methods are combined to generate visual mappings from data queries to charts and further highlight the key information in the generated charts to enhance their readability. It is demonstrated that proposed method achieves higher accuracy than existing methods for the chart type prediction task with real-world charts dataset. A user study is also conducted to show that proposed method can help users quickly explore new data tables and find important insights.