Start Date
10-12-2017 12:00 AM
Description
Extracting actionable information from complex data is a key challenge for business analytics researchers (Hedgebeth, 2007). This is particularly difficult for high-dimensional datasets, to which an increasing number of businesses have access (Martens et al., 2016). In this study, we develop a customized neural network for extracting interpretable features from very high-dimensional datasets. These features can be interpreted both at an aggregated as well as a very fine-grained level. Interpreting non-linear interactions is no more difficult than interpreting a linear regression. We apply the algorithm to a dataset related to product returns in online retail which contains a total of 3,637,654 transactions and 13,533 dimensions. Comparing 75 different models, we demonstrate that, in addition to being interpretable, our algorithm yields higher predictive accuracy than extant methods. The approach is sufficiently abstract to be applicable to a wide variety of business analytics datasets.
Recommended Citation
Urbanke, Patrick; Uhlig, Alexander; and Kranz, Johann Joachim, "A Customized and Interpretable Deep Neural Network for High-Dimensional Business Data - Evidence from an E-Commerce Application" (2017). ICIS 2017 Proceedings. 13.
https://aisel.aisnet.org/icis2017/DataScience/Presentations/13
A Customized and Interpretable Deep Neural Network for High-Dimensional Business Data - Evidence from an E-Commerce Application
Extracting actionable information from complex data is a key challenge for business analytics researchers (Hedgebeth, 2007). This is particularly difficult for high-dimensional datasets, to which an increasing number of businesses have access (Martens et al., 2016). In this study, we develop a customized neural network for extracting interpretable features from very high-dimensional datasets. These features can be interpreted both at an aggregated as well as a very fine-grained level. Interpreting non-linear interactions is no more difficult than interpreting a linear regression. We apply the algorithm to a dataset related to product returns in online retail which contains a total of 3,637,654 transactions and 13,533 dimensions. Comparing 75 different models, we demonstrate that, in addition to being interpretable, our algorithm yields higher predictive accuracy than extant methods. The approach is sufficiently abstract to be applicable to a wide variety of business analytics datasets.