A decision-analytic framework for interpretable recommendation systems with multiple input data sources: a case study for a European e-tailer

Abstract

Recommendation systems help companies construct online personalization strategies for customers who are often overwhelmed by the abundance of product choices available. To extend existing operations research literature on recommendation systems, this article proposes a decision analytic framework for interpretable recommendation systems with multiple input data sources for e-commerce settings. The impact of multiple data sources on recommendation performance is investigated and two hybridization data fusion strategies, i.e., a posteriori weighting and input data source combination using factorization machines are benchmarked. Furthermore, a new importance score mechanism is introduced to provide insight into the input data sources’ and underlying variables’ impact on recommendation performance. The framework is empirically validated on 164,338 customers and 51,367 products across eight real-life data sets with four input data sources (product, customer, raw behavioral, and aggregated behavioral data) obtained from a large European e-commerce company. With this new decision analytic framework, e-commerce companies are able to open their recommendation system’s black box, to identify the most predictive input data sources and the best hybridization strategy for their business context.