Explaining deep learning models for behavioral data.

As a consequence of digitalization, more aspects of people's lives are being captured. Examples include visiting particular physical locations or webpages, liking Facebook pages, etc. This behavioral data holds significant predictive power. For example, what you like on Facebook can be predictive for your IQ, product interest, and even creditworthiness. Deep learning has been shown to outperform other prediction techniques in making accurate predictions using behavioral data. Combining behavioral data and deep learning unfortunately results in incomprehensible black box predictions. Three reasons why: (1) behavioral data is very high-dimensional (up to millions of features), (2) the data is sparse, so every feature is only of relevance for a few data instances, and (3) the deep learning model is complex and non-linear. Consequently, although the combination of deep learning and behavioral data is so predictive, it is very difficult to understand why the model is making certain predictions, leading to skepticism to use it in practice. The main contribution of this research proposal is to design new algorithms that explain the complex deep learning prediction models. This comprehensibility issue is a research area that has gained attention in the data mining community because of the implications it has on model deployment and transparency towards users. We will validate our findings in several applications.

Keywords:RISK MANAGEMENT, MARKETING

Disciplines:Applied economics, Economic history, Macroeconomics and monetary economics, Microeconomics, Tourism