Learning to Rank Hotels

Deep Learning has revolutionized the fields of computer vision and natural language processing. However, deep learning methods have struggled to perform well on recommender system / information retrieval problems. One of the main differences between these fields is that recommender system / information retrieval data are much more sparse. Furthermore, as shown in Dacrema et al 2019, many modern, published, deep learning models fail to outperform simpler methods (such as linear models or nearest neighbor models).

In this project, my team and I implemented MultVAE and word2vec (modified for hotels) on a proprietary information retrieval dataset from RocketMiles. The RocketMiles dataset was even more sparse than other traditional recommender systems/ information retrieval datasets (by about 100x more sparse). Our github repository is available here, and our final paper here.