[Paper Review] Cache-Augmented Inbatch Importance Resampling for Training Recommender Retriever

Jin Chen, Defu Lian+4 more|arXiv (Cornell University)|May 30, 2022

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes ChiIR, a cache-augmented in-batch importance resampling method for training recommender retrievers that improves softmax estimation accuracy by adaptively resampling negatives based on query-dependent similarity scores and maintaining a dynamic cache of frequently sampled, high-similarity items. The approach reduces training bias and accelerates convergence, achieving state-of-the-art performance on five real-world datasets with up to 4.87% relative improvement in NDCG@10.

ABSTRACT

Recommender retrievers aim to rapidly retrieve a fraction of items from the entire item corpus when a user query requests, with the representative two-tower model trained with the log softmax loss. For efficiently training recommender retrievers on modern hardwares, inbatch sampling, where the items in the mini-batch are shared as negatives to estimate the softmax function, has attained growing interest. However, existing inbatch sampling based strategies just correct the sampling bias of inbatch items with item frequency, being unable to distinguish the user queries within the mini-batch and still incurring significant bias from the softmax. In this paper, we propose a Cache-Augmented Inbatch Importance Resampling (XIR) for training recommender retrievers, which not only offers different negatives to user queries with inbatch items, but also adaptively achieves a more accurate estimation of the softmax distribution. Specifically, XIR resamples items for the given mini-batch training pairs based on certain probabilities, where a cache with more frequently sampled items is adopted to augment the candidate item set, with the purpose of reusing the historical informative samples. XIR enables to sample query-dependent negatives based on inbatch items and to capture dynamic changes of model training, which leads to a better approximation of the softmax and further contributes to better convergence. Finally, we conduct experiments to validate the superior performance of the proposed XIR compared with competitive approaches.

Motivation & Objective

• To address the significant bias in in-batch sampling for recommender retrievers, which treats all queries uniformly and fails to approximate the true softmax distribution.
• To improve model convergence by enabling query-dependent negative sampling within each mini-batch, rather than sharing the same set of negatives across queries.
• To reduce approximation error from the softmax loss by incorporating a cache of historically informative, high-similarity items as augmented negatives.
• To achieve better training efficiency and effectiveness without relying on uniformly sampled or static popularity-based negatives.

Proposed method

• ChiIR performs importance resampling of in-batch items based on query-specific similarity scores and item popularity, creating query-dependent negative sets.
• It introduces a dynamic cache that stores frequently sampled items—typically high-similarity, hard negatives—over training epochs to augment the candidate negative set.
• The sampling probability is derived from a modified softmax weight that combines similarity scores and popularity-based debiasing to reduce sampling bias.
• The cache is updated in real time using a fixed-size buffer, prioritizing items with higher sampling frequency, which are more informative for model training.
• The method integrates seamlessly with two-tower models and is compatible with deep feature towers, supporting complex item representations.
• The approach enables more accurate estimation of the full softmax distribution by combining in-batch adaptivity with long-term memory of informative samples.

Experimental results

Research questions

• RQ1 Can in-batch sampling be improved to provide query-dependent negative samples that better approximate the true softmax distribution?
• RQ2 How does incorporating a dynamic cache of frequently sampled items affect the convergence and performance of recommender retrievers?
• RQ3 To what extent can importance resampling with a cache reduce the bias in gradient estimation compared to uniform or popularity-based sampling?
• RQ4 Does the proposed method maintain or improve performance when using fewer sampled items per batch, enabling more efficient training?

Key findings

• ChiIR achieves a 4.87% relative improvement in NDCG@10 over SSL-Pop on the Tmall dataset when using deep feature towers, demonstrating strong effectiveness with complex representations.
• Resampling as few as 1/64 of the batch items can match or exceed the performance of full-batch sampling, indicating high training efficiency.
• The item distribution in the cache significantly departs from the long-tailed popularity distribution, favoring more informative, high-similarity items and reducing bias toward rare items.
• The cache size has a direct impact on performance: increasing cache size from 0.5x to 4x the batch size improves recommendation quality, with optimal gains observed at larger sizes.
• Gradient computation time decreases proportionally with fewer resampled items, confirming the method’s potential for efficient training with reduced computational cost.
• The method outperforms both BIR and SSL-Pop across five public datasets, validating its superiority in both accuracy and convergence speed.