# An Incremental Update Framework for Online Recommenders with Data-Driven Prior

**Authors:** Chen Yang, Jin Chen, Qian Yu, Xiangdong Wu, Kui Ma, Zihao Zhao, Zhiwei Fang, Wenlong Chen, Chaosheng Fan, Jie He, Changping Peng, Zhangang Lin, Jingping Shao  
**arXiv:** [2312.15903](https://arxiv.org/abs/2312.15903)  
**Date:** 2023-12-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.15903v1/

## Key points

- The incremental update approach is popular for learning large-scale models in online recommenders
- Incremental updates can lead to overfitting on recent data and neglect of long-term information
- The Data-Driven Prior (DDP) framework consists of Feature Prior (FP) and Model Prior (MP) components
- DDP aims to improve recommendation performance by enhancing training stability and providing a robust prior for updates
- Existing studies have used distillation logit and meta-learners to address challenges in online recommenders, but lack analysis of unique characteristics like extreme data sparsity and feature diversity
- The DDP framework incorporates both feature prior and model prior components to address challenges in online recommenders
- Feature Prior estimates click-through rates at the feature level for stable learning, especially beneficial for long-tail items
- Model Prior approximates posterior probabilities on complete data while minimizing distances from prior models

## Summary

The field of online recommenders has seen a surge in interest and revenue generation for businesses in recent years. With a large number of users and items to consider, the incremental update approach has become a popular paradigm for learning large-scale models in industrial settings. This method involves feeding only newly arrived data within a sliding window into the model, allowing for quick responses to changes in data distribution. However, this strategy can lead to overfitting on recent data and result in the neglect of long-term information crucial for recommendation performance. To address these challenges, a new framework called Data-Driven Prior (DDP) is proposed for online recommenders. DDP consists of two components: Feature Prior (FP) and Model Prior (MP). The FP focuses on click estimation for specific feature values to enhance training stability, while the MP incorporates previous model outputs into current updates following Bayes rules to provide a theoretically sound prior for robust updates. By integrating both FP and MP into a unified framework that is model-agnostic and adaptable to various interaction models, the proposed approach aims to improve recommendation performance. The increasing complexity of models designed to capture intricate interactions between features has posed challenges in deploying and updating these models quickly in industrial online recommendation systems. While incremental updates using newly arrived data have been effective in reducing time overheads and adapting to dynamic data distributions, they are susceptible to overfitting when significant changes occur in data distribution. Existing studies have applied continual learning techniques such as distillation logit and meta-learners to mitigate these issues by utilizing model-based priors. However, these approaches lack an analysis of the unique characteristics of online recommendations, including extreme data sparsity and feature diversity. Limited user clicks make it challenging to accurately estimate user preferences during incremental updates, leading to less attention being paid to long-tail items. To address these challenges, the proposed DDP framework incorporates both feature prior and model prior components. The Feature Prior explicitly estimates average click-through rates at the feature level, providing more stable learning for model updates especially beneficial for long-tail items. The Model Prior approximates posterior probabilities on complete data by maximizing likelihood functions on incremental data while minimizing distances from prior models. Overall, the refined summary highlights the importance of addressing evolving data distributions in online recommenders through innovative frameworks like DDP that leverage both feature-specific information and past model outputs for enhanced recommendation performance across diverse datasets.