I want to build a recommendation system for accounts purchasing some items. The ratio of purchase events to view events is very low (less than 1% items that are viewed get bought). Right now I am using a binary class classification model where I heavily downsample instances from negative class to bring the distribution of labels in training dataset to 50-50. Is there a way to use something like one-class classification model or anomaly detection model that can model positive instances only (as modeling negative instances is challenging since reason for not purchasing an item could be anything and might not be represented in the training dataset as well as compared to those for purchase events).
Looks like one-class classification is typically built for scenarios where we have large number of normal/positive instances and very small number of negative instances (or anomalies). Can I use it for this scenario with the assumption that purchase event is a negative/anomaly instance or as positive instance?