Integrating Human Knowledge into Data Mining through interactive Decision Trees

Abstract:

Machine Learning (ML) methods are efficient tools to identify complex relations in large datasets. Decision Trees (DT) represent a widely used ML method in the geo-environmental sciences for two main reasons. Geo-environmental processes span in a wide range of spatio-temporal scales and process controls vary with scale and geographical location. So, we use DT because they can effectively partition the entire domain in smaller subgroups and can identify a hierarchy of controls for each subgroup. Moreover, DT are often simple and in principle easy to interpret. However, their credibility and usefulness depend on the quality and quantity of the data available. Geo-environmental data pose challenges in this regard because they are often sparse, noisy, uncertain, and cannot capture the full complexity of the processes investigated or the full space of possible observations. Moreover, relying only on statistical metrics for the building process of DT will lead to statistically optimal trees but not necessarily to physically consistent trees. Embedding expert knowledge in DT could compensate for these shortcomings. But DT are designed to work automatically leaving little or no space for experts to be involved. To bridge this gap, we propose interactive DT (iDT) that put humans in the loop and integrate experts’ domain knowledge with the power of the algorithms to interactively learn patterns from large datasets. We design methods and visualization techniques that allow users to interact with the algorithm and demonstrate them using several geo-environmental case studies (incl. landslide predictions and water balance estimations). We believe that iDT will help experts incorporate their knowledge in the DT models achieving outcomes that are more robust, more transferable and with higher interpretability.