Mitigating Pooling Bias in E-commerce Search via False Negative Estimation

高效准确的产品相关性评估对用户体验和业务成功至关重要。训练出一个熟练的相关性评估模型需要高质量的查询-产品对，通常通过负采样策略获得。不幸的是，当前的方法通过错误地采样假负例引入了池化偏差，降低了性能和业务影响。为了解决这个问题，我们提出了减轻偏差的硬负采样（BHNS），这是一种新的负采样策略，旨在识别和调整假负例，建立在我们原始的假负例估计算法之上。我们在Instacart搜索设置中的实验证实了BHNS在实际电子商务中的有效性。此外，对公共数据集的比较分析展示了它在不同领域的潜力。

The paper aims to address the issue of pooling bias in negative sampling strategies, which introduces false negatives and diminishes the performance of relevance assessment models for e-commerce search.

The paper proposes a novel negative sampling strategy called Bias-mitigating Hard Negative Sampling (BHNS) that identifies and adjusts for false negatives using the False Negative Estimation algorithm. BHNS is shown to be effective in practical e-commerce use and has domain-agnostic potential for diverse applications.

The experiments in the Instacart search setting confirm the effectiveness of BHNS for practical e-commerce use. The paper also includes comparative analyses on public datasets and showcases the domain-agnostic potential of BHNS. The authors provide open-source code for their approach. The False Negative Estimation algorithm is a key innovation that helps to mitigate pooling bias in negative sampling strategies.

Some related studies in this area include 'Negative Sampling for Large-Scale Neural Content-based Retrieval' by Gao et al. and 'Learning to Match Using Local and Distributed Representations of Text for Web Search' by Shen et al.