Mobile recommendations based on interest prediction from consumer's installed apps–insights from a large-scale field study

Abstract

Recommender systems are essential in mobile commerce to benefit both companies and individuals by offering highly personalized products and services. One key pre-requirement of applying such systems is to gain decent knowledge about each individual consumer through user profiling. However, most existing profiling approaches on mobile suffer problems such as non-real-time, intrusive, cold-start, and non-scalable, which prevents them from being adopted in reality. To tackle the problems, this work developed real-time machine-learning models to predict user profiles of smartphone users from openly accessible data, i.e. app installation logs. Results from a study with 904 participants showed that the models are able to predict interests on average 48.81% better than a random guess in terms of precision and 13.80% better in terms of recall. Since the effectiveness of such predictive models is unknown in practice, the predictive models were evaluated in a large-scale field experiment with 73,244 participants. Results showed that by leveraging our models, personalized mobile recommendations can be enabled and the corresponding click-through-rate can be improved by up to 228.30%. Supplementary information, study data, and software can be found at https://www.autoidlabs.ch/mobile-analytics.