Target - Attention Network for Click - Through Rate Prediction

In Click-Through Rate (CTR) prediction scenario, deep neural network based learning algorithms provide leading-edge performance. Most of the algorithms embed target user and target item into vectors, then learning the click probability through deep neural networks. However, for sparse users, (i.e. users with less behaviour), or sparse items (i.e. new products, which are viewed by few users), they learn the embeddings not well enough due to the lack of specific samples, and thus affects the CTR prediction performance. To address the issue, we propose Target-Attention NetworK (TANK), a recommendation framework based on target attention technique. The framework includes information expansion for the user, information expansion for the item, and information expansion for the interaction between them. In user information extension part, considering that the target user's historical behaviour can be sparse, we use the historical users to extend target (sparse) user's information. In the item information expansion part, we use the historical items to expand target (sparse) item's information. In the user-item interaction expansion part, we consider both the historical user-item interactions of target user and historical interactions of target item. Our method's efficiency is ultimately proven using the public Amazon dataset.