FANS: Fast Non-autoregressive Sequential Generation for Item List Continuation

Jan. 25^th, 2023: Got accepted by TheWebConf ‘23!

Qijiong LIU, Jieming ZHU, Jiahao WU, Tiandeng WU, Zhenhua DONG, and Xiao-Ming WU^#
[Code] [Paper]

Abstract

User-curated item lists, such as video-based playlists on Youtube and book-based lists on Goodreads, have become prevalent for sharing content on online platforms. Item list continuation is proposed to model the overall trend of lists and predict subsequent lists. Recently, Transformer-based models have shown promise in comprehending contextual information and capturing item relationships in a list. However, it is difficult to deploy them in real-time industrial scenarios, mainly due to the time-consuming nature of the autoregressive generation mechanism used in them. In this paper, we propose a novel fast non-autoregressive sequence generation model, namely FANS, to accelerate inference efficiency and improve inference quality for item list continuation. First, we use a non-autoregressive generation mechanism to decode next $K$ items simultaneously instead of one-by-one as in existing models. Then, we design a two-stage classifier to replace the vanilla classifier used in current Transformer-based models to further reduce the decoding time. Moreover, to improve inference quality of non-autoregressive generation, we employ a curriculum learning strategy to optimize training. Extensive experiments on four real-world item list continuation datasets including Zhihu, Spotify, AotM, and Goodreads demonstrate that our FANS model can significantly improve inference efficiency (up to 8.7x) while achieving competitive or better inference quality compared with state-of-the-art autoregressive models. We also validate the efficiency of FANS in an industrial system. The source code and data are provided for reproducing the reported results.

Citation

@inproceedings{liu2023fans, 
    title = {FANS: Fast Non-Autoregressive Sequence Generation for Item List Continuation},
    author = {Liu, Qijiong and Zhu, Jieming and Wu, Jiahao and Wu, Tiandeng and Dong, Zhenhua and Wu, Xiao-Ming},
    booktitle = {Proceedings of the ACM Web Conference 2023},
    month = {may},
    year = {2023},
    address = {Austin, Texas, USA}
}