AutoAssign+: Automatic Shared Embedding Assignment in Streaming Recommendation

With the rapid growth of personalized online applications, recommender systems have been widely implemented by various online businesses, including E-commerce websites, news platforms, online advertising, and so on [1, 2]. Among them, streaming recommendation [3, 4] is one of the common forms of recommender systems, where streaming data are constantly flowing into the recommendation models for training, thus better modeling the user's current preferences. In addition, streaming recommendations are particularly important for time-sensitive items, such as news, as they allow for rapid identification and distribution of relevant content to interested users, which is critical for commercial information retrieval systems. Due to the ability to effectively capture the highly nonlinear relationship between user and item end-to-end, neural network-based models are rapidly becoming the mainstream of recommender systems. As shown in Figure 1, existing deep recommendation models typically follow the "Embedding & Feature Interaction" paradigm [5]. The embedding layer serves as the encoder to represent sparse features in dense latent space, while the feature interaction layers serve to capture interactive signals among these features. In a streaming recommender system, new items and users are continually added to the data corpus, creating a highly dynamic streaming environment that presents several challenges, which can be summarized as: Cold-start: The streaming recommender system is confronted with a constant influx of new users, many of whom can be classified as visitor-type users and possess extremely limited behavior information. Furthermore, the system is constantly updated with new items, yet there has not been enough interaction with these items to generate an adequate level of training data. The consequence of employing insufficiently trained new user/item embeddings is a significant decline in the performance of the recommendation model.