Bayesian Deep Learning with Small or Unbalanced Datasets with Applications to Marketing

Bayesian Deep Learning with Small or Unbalanced Datasets with Applications to Marketing

Abstract:

Deep learning techniques have become popular in marketing applications to analyse datasets containing a large number of variables, some of which under the form of unstructured data such as product pictures, user reviews, or network graphs.

Commonly used deep learning methods require a large number of observations for each segment of the population studied. However, in many applications, new data points (e.g., retail location, SKU, loyalty program member) are not easily acquired and datasets might be unbalanced with some categories being underrepresented.

Models trained on small or unbalanced datasets tend to show poor out-of-sample performance (overfitting), inaccurate predictions in the tails of the data distribution (e.g., top 5% of consumers), but also wrongly assess the uncertainty associated with inferences (overconfidence). A robust inference method should both be accurate and correctly assess the level of confidence in a prediction.

In this research, we propose to adopt a Bayesian approach reducing overfitting, increasing accuracy for categories with sparse data, and producing more reliable estimations of uncertainty. We compare our approach to traditional regularized optimization methods where uncertainty is estimated through cross-validation. We apply our method to various datasets, including AirBnB market data and product pictures in the beverage industry. We show that cross-validation does not solve the issues of drastically underestimated uncertainty and biased predictions.

The Seminar will be held in Seminar Room RuW 1.201 as well as broadcasted via Zoom with the following link:

https://uni-frankfurt.zoom.us/j/62958556058?pwd=b01qWXlrbVMzc1JFVEhjSWNvdGZPQT09

Meeting-ID: 629 5855 6058
Kenncode: 101686

More information on Remi Davietcan be found here.