Section: Research Program
The following perspectives are at the convergence of the four aforementioned research axes and can be seen as ideal towards our goals:
Automating data science workflow discovery. The current methods for knowledge extraction and construction of decision support systems require a lot of human effort. Our three research axes aim at alleviating this effort, by devising methods that are more generic and by improving the interaction between the user and the system. An ideal solution would be that the user could forget completely about the existence of pattern mining or decision support methods. Instead the user would only loosely specify her problem, while the system constructs various data science / decision support workflows, possibly further refined via interactions.
We consider that this is a second order AI task, where AI techniques such as planning are used to explore the workflow search space, the workflow itself being composed of data mining and/or decision support components. This is a strategic evolution for data science endeavors, were the demand far exceeds the available human skilled manpower.
Logic argumentation based on epistemic interest. Having increasingly automated approaches will require better and better ways to handle the interactions with the user. Our second long term goal is to explore the use of logic argumentation, i.e., the formalisation of human strategies for reasoning and arguing, in the interaction between users and data analysis tools. Alongside visualization and interactive data mining tools, logic argumentation can be a way for users to query both the results and the way they are obtained. Such querying can also help the expert to reformulate her query in an interactive analysis setting.
This research direction aims at exploiting principles of interactive data analysis in the context of epistemic interestingness measures. Logic argumentation can be a natural tool for interactions between the user and the system: display of possibly exhaustive list of arguments, relationships between arguments (e.g., reinforcement, compatibility or conflict), possible solutions for argument conflicts, etc.
The first step is to define a formal argumentation framework for explaining data mining results. This implies to continue theoretical work on the foundations of argumentation in order to identify the most adapted framework (either existing or a new one to be defined). Logic argumentation may be implemented and deeply explored in ASP, allowing us to build on our expertise in this logic language.
Collaborative feedback and knowledge management. We are convinced that improving the data science process, and possibly automating it, will rely on high-quality feedback from communities on the web. Consider for example what has been achieved by collaborative platforms such as StackOverflow: it has become the reference site for any programming question.
Data science is a more complex problem than programming, as in order to get help from the community, the user has to share her data and workflow, or at least some parts of them. This raises obvious privacy issues that may prevent this idea to succeed. As our research on automating the production of data science workflows should enable more people to have access to data science results, we are interested in the design of collaborative platforms to exchange expert advices over data, workflows and analysis results. This aims at exploiting human feedback to improve the automation of data science system via machine learning methods.