Compressing Embedding Table via Multi-dimensional Quantization Encoding for Sequential Recommender Model

Sequential recommender models have become a research hotspot in the field of current recommender systems due to its excellent ability to describe users’ dynamic preferences. Sequential recommender models based on deep learning have achieved state-of-the-art results. However, with the increasing number of users and items, the traditional item embedding table may consume a huge amount of memory so that the model may be more difficult to deploy to resource-limited devices. In this paper, we propose a multi-dimensional quantization encoding(MDQE) method to resolve this issue. MDQE mainly consists of two compression techniques. We first divide items into several groups according to the interaction frequency of items and assign different dimensions to each group to construct multi-dimensional group-wise embedding tables. Then, we use mapping matrices to transform the multi-dimensional group-wise embedding tables into quantized codebooks for further compressing. The experiments on three real-world datasets demonstrate that the proposed MDQE can achieve up to 13.86x compression ratio with negligible accuracy loss during inference.