# Learning Large-scale Universal User Representation with Sparse Mixture of Experts

Preprint · Caigao Jiang, Siqiao Xue, James Zhang, Lingyue Liu, Zhibo Zhu, Hongyan Hao

- Archive id: arxiv:2207.04648
- Published: 2022-07-11
- Venue: International Conference on Machine Learning (ICML), First Workshop of Pre-training, 2022
- arXiv: https://arxiv.org/abs/2207.04648
- Source URL: https://arxiv.org/abs/2207.04648
- PDF: https://arxiv.org/pdf/2207.04648
- Provenance: arxiv

## Abstract

Learning user sequence behaviour embedding is very sophisticated and challenging due to the complicated feature interactions over time and high dimensions of user features. Recent emerging foundation models, e.g., BERT and its variants, encourage a large body of researchers to investigate in this field. However, unlike natural language processing (NLP) tasks, the parameters of user behaviour model come mostly from user embedding layer, which makes most existing works fail in training a universal user embedding of large scale. Furthermore, user representations are learned from multiple downstream tasks, and the past research work do not address the seesaw phenomenon. In this paper, we propose SUPERMOE, a generic framework to obtain high quality user representation from multiple tasks. Specifically, the user behaviour sequences are encoded by MoE transformer, and we can thus increase the model capacity to billions of parameters, or even to trillions of parameters. In order to deal with seesaw phenomenon when learning across multiple tasks, we design a new loss function with task indicators. We perform extensive offline experiments on public datasets and online experiments on private real-world business scenarios. Our approach achieves the best performance over state-of-the-art models, and the results demonstrate the effectiveness of our framework.

_Abstract reproduced under CC0-1.0 (arXiv descriptive metadata)._

---
Space Kitty Archive · https://space-kitty.com · API: https://space-kitty.com/api/v1 · MCP: https://space-kitty.com/mcp · Premium: $0.20/request via x402 (Base Sepolia (testnet)). Metadata: arXiv (CC0), OpenAlex (CC0), Europe PMC (CC BY/CC0 works), ClinicalTrials.gov.