User Multi-Interest Modeling for Behavioral Cognition

Representation modeling based on user behavior sequences is an important direction in user cognition. In this study, we propose a novel framework called Multi-Interest User Representation Model. Specifically, the model consists of two sub-models. The first sub-module is used to encode user behaviors in any period into a super-high dimensional sparse vector. Then, we design a self-supervised network to map vectors in the first module to low-dimensional dense user representations by contrastive learning. With the help of a novel attention module which can learn multi-interests of user, the second sub-module achieves almost lossless dimensionality reduction. Experiments on several benchmark datasets show that our approach works well and outperforms state-of-the-art unsupervised representation methods in different downstream tasks.