Scalable Bundle Recommendations: A Large-Scale Field Experiment

We develop an end-to-end, scalable machine learning framework for designing bundle recommendations in a high-dimensional choice setting. We leverage historical purchases and consideration sets determined from clickstream data to generate dense representations (embeddings) of products. We impose minimal structure on these embeddings and develop heuristics for complementarity and substitutability among products. Subsequently, we use the heuristics to create multiple bundle recommendations for each of 4,500 focal products and test their performance using a field experiment with a large retailer. We use the experimental data to optimize the recommendation design policy with offline policy learning. Our optimized policy is robust across product categories, generalizes well to the retailer’s entire assortment, and provides an expected improvement of 35% ([Formula: see text] per 100 visits) in revenue from bundle recommendations over the baseline policy. This paper was accepted by Hemant Bhargava, information systems. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2022.01864 .