I. Introduction
For different users, a personalized advertising system feeds different ads based on the estimated relevance between the ad and the user’s interest. Normally, the relevance between a user and an ad is measured by the similarity between their embeddings, which are learned jointly from the users’ historical behaviors on the ads. Nevertheless, for new users, there are no historical user-ad behaviors for learning effective user embedding. This issue of modeling new users is normally defined as the cold start problem. To solve the cold start problem, we usually exploit the user’s demographic attributes, such as age, region, and gender. The attribute embedding has been effectively learned based on the ordinary users’ rich experience accumulated in the past and can readily generalize well to the new users. Since the attribute embedding does not rely on the historical user behaviors, they are useful for tackling the cold start problem.