Abstract: The affinity of a user to a type of items (e.g., stories from the same publisher, and movies of the same genre) is an important signal reflecting the user’s interests. Accurately estimating of the user type affinity has various applications in ranking and recommendation systems. For frequent users, simply dividing the number of interactions with content (e.g., clicks) by the number of impressions (e.g., the number of times the content is presented to each user) would be a good estimate. However, such estimates are erroneous for users who have sparse interaction history, (e.g., new users). To alleviate the problem, feature-based approaches aim to learn functions predicting the affinity score using only none-click features, such as user demographics, locations, and interests. Likewise, such approaches do not take full advantage of the interaction history of frequent users.
Motivated by the limitations of the two approaches, we propose a Gamma-Poisson model that aims at utilizing the interaction history of frequent users, as well as leveraging a feature-based model for infrequent users. Our intuition is that we should rely more on the interaction history when estimating affinity for frequent users, and weigh more on feature-based model for infrequent users. We present experimental results on large-scale real-world data in a publisher content clicks prediction task to demonstrate the effectiveness of the proposed method in estimating user type affinity scores