Time-Dependent Demographic Prediction Based on Time-Back-Propagation Method

In various applications like personalized search and recommendation, full demographic information is a precondition for many applications' well performance, but such ideal dataset rarely exists in practical scenarios. What's worse, absence of key characteristics (e.g., baby's age in maternal and infant commodity recommendation) makes these applications struggle. We design a novel solution to solve the problem of time-dependent demographic prediction. The key insight behind our approach is, we leverage a Time-back-propagation (TBP) method to take the internal time correlation of historical behaviours into consideration and collect all available data to train a classifier, which is a mapping from user's historical behaviours to the demographic information. We demonstrate the effectiveness of our method through experiments of baby's age prediction. Our algorithm performs more balanced on each age group, and can predict Baby's age (BA) accurately in 78.2% on a real-world dataset with 2,058,909 items of a major E-commerce site.