One Person, One Model, One World: Learning Continual User Representation without Forgetting
share
Learning generic user representations which can then be applied to other user-related tasks (e.g., profile prediction and recommendation) has recently attracted much attention. Existing approaches often derive an individual set of model parameters for each task by training their own data. However, the representation of a user usually has some potential commonalities. As such, these separately trained representations could be suboptimal in performance as well as inefficient in terms of parameter sharing. In this paper, we delve on the research to continually learn user representations task by task, whereby new tasks are learned while using parameters from old ones. A new problem arises since when new tasks are trained, previously learned parameters are very likely to be modified, and thus, an artificial neural network (ANN)-based model may lose its capacity to serve for well-trained previous tasks forever, termed as catastrophic forgetting. To address this issue, we present Conure which is the first continual, or lifelong, user representation learner – i.e., learning new tasks over time without forgetting old ones. Specifically, we propose iteratively removing unimportant weights by pruning on a well-optimized backbone representation model, enlightened by fact that neural network models are highly over-parameterized. Then, we are able to learn a coming task by sharing previous parameters and training new ones only on the empty space after pruning. We conduct extensive experiments on two real-world datasets across nine tasks and demonstrate that Conure performs largely better than common models without purposely preserving such old "knowledge", and is competitive or sometimes better than models which are trained either individually for each task or simultaneously by preparing all task data together.
READ FULL TEXT