Section: New Results
Interactive Data Exploration at Scale
Building upon our prior work in active learning-based interactive database exploration system, we improved this system in terms of efficiency and effectiveness. First, we formally defined the class of user interest queries to which our proposed Dual Space Model (DSM) can bring significant improvement in accuracy. Second, we generalized the DSM to arbitrary queries by forcing our system to fall back to the traditional active learning-based techniques if the requested query properties are not satisfied. Third, we launched a user study to collect real-world datasets and user interest patterns for comparison experiments. The evaluation results showed that our new system outperformed the start-of-the-art active learning techniques and data exploration systems. Fourth, to show the robustness of our system, we added some label noise into the experiments. It turned out that our system maintained a good performance and significantly outperformed traditional active learning-based system. These results have appeared in the prestigious PVLDB journal [10]. In addition, we have been working on integrating DSM with version space algorithms and designing more advanced methods to deal with label noise. In the near feature, a new software based on our proposed techniques will be put into use for interactive database exploration.