Authors: David Gotz, Rashnil Chaturvedi
Abstract: As organizations gather ever larger and more detailed datasets, predictive modeling is becoming a widely used technology in support of data-driven decision making. In a diverse set of disciplines, ranging from advertising to medicine, temporal event data (such as click streams and electronic health records) are increasingly being used as the basis for training these predictive models. In these cases, temporal relationships between events (e.g., one event occurring before another vs. the same events in opposite order) can be highly predictive. However, existing methods for feature construction make it difficult to incorporate this sort of information, and often require domain experts to manually specify patterns of interest. This poster introduces Interactive Temporal Feature Construction (ITFC), a visual analytics technique designed to enable more effective, data-driven temporal feature construction. The primary contributions for this work include a new interactive workflow for model refinement, a set of algorithms and visual representations designed to support that workflow, and a use case which demonstrates how ITFC can result in more accurate predictive models when applied to complex cohorts of electronic health data.