The main focus of recommender systems research has been recommending fitting items to the users. However, in some domains, not only which item but also the quantity the target user should consume could be part of the recommendation. In this work, we tackle the under-researched problem of recommending the duration of stay in the domain of destination recommendation. Using two data sets, one based on hotel bookings and the other on mobility derived from geotagged Tweets, we perform extensive feature engineering with unsupervised learning to discover types of users and graph embeddings of the cities. In our experiments, we compare the performance of supervised learning algorithms with varying features to statistical baselines for predicting the duration of stay at a destination. The results underline the task’s difficulty: we obtain the best results for the hotel bookings data set using personalized mobility embeddings with CatBoost. At the same time, the simple strategy of recommending the mode duration of all users is competitive in the noisy Twitter data set.
About computer science, clean code and open source.
Check out my book Java by Comparison: Become a Java Craftsman in 70 Examples!