# Lessons Learned Addressing Dataset Bias in Model-Based Candidate Generation at Twitter

arXiv: https://arxiv.org/abs/2105.09293 · HTML: https://mlsanity.com/paper/2105.09293/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 13 May 2021.

## Built on by (most cited first, 4 of 4)

- [ItemSage: Learning Product Embeddings for Shopping Recommendations at Pinterest](https://mlsanity.com/paper/2205.11728/) (2022)
- [A Transformer-Based Substitute Recommendation Model Incorporating Weakly Supervised Customer Behavior Data](https://mlsanity.com/paper/2211.02533/) (2022)
- [Finding Interest Needle in Popularity Haystack: Improving Retrieval by Modeling Item Exposure](https://mlsanity.com/paper/2503.23630/) (2025)
- [A Multi-Source Ensemble Approach to Candidate Generation for Alternative Vacation Rental Property Recommendations](https://mlsanity.com/paper/2609.05748/) (2026)

## Abstract (arXiv)

Traditionally, heuristic methods are used to generate candidates for large scale recommender systems. Model-based candidate generation promises multiple potential advantages, primarily that we can explicitly optimize the same objective as the downstream ranking model. However, large scale model-based candidate generation approaches suffer from dataset bias problems caused by the infeasibility of obtaining representative data on very irrelevant candidates. Popular techniques to correct dataset bias, such as inverse propensity scoring, do not work well in the context of candidate generation. We first explore the dynamics of the dataset bias problem and then demonstrate how to use random sampling techniques to mitigate it. Finally, in a novel application of fine-tuning, we show performance gains when applying our candidate generation system to Twitter's home timeline.