[Paper Review] A Comparison of Supervised Learning to Match Methods for Product Search

Fatemeh Sarvi, Nikos Voskarides+3 more|arXiv (Cornell University)|Jul 20, 2020

Web Data Mining and Analysis39 references4 citations

TL;DR

This paper evaluates 12 supervised learning-to-match methods for product search, comparing their effectiveness and efficiency on two datasets—CIKM 2016 (public) and a proprietary European e-commerce dataset. It finds that models designed for short text matching, such as MV-LSTM and DRMMTKS, consistently rank among the top three, while ARC-I offers the best balance of accuracy and efficiency for real-world deployment; surprisingly, a state-of-the-art BERT-based model underperforms due to domain mismatch in pre-training data.

ABSTRACT

The vocabulary gap is a core challenge in information retrieval (IR). In e-commerce applications like product search, the vocabulary gap is reported to be a bigger challenge than in more traditional application areas in IR, such as news search or web search. As recent learning to match methods have made important advances in bridging the vocabulary gap for these traditional IR areas, we investigate their potential in the context of product search. In this paper we provide insights into using recent learning to match methods for product search. We compare both effectiveness and efficiency of these methods in a product search setting and analyze their performance on two product search datasets, with 50,000 queries each. One is an open dataset made available as part of a community benchmark activity at CIKM 2016. The other is a proprietary query log obtained from a European e-commerce platform. This comparison is conducted towards a better understanding of trade-offs in choosing a preferred model for this task. We find that (1) models that have been specifically designed for short text matching, like MV-LSTM and DRMMTKS, are consistently among the top three methods in all experiments; however, taking efficiency and accuracy into account at the same time, ARC-I is the preferred model for real world use cases; and (2) the performance from a state-of-the-art BERT-based model is mediocre, which we attribute to the fact that the text BERT is pre-trained on is very different from the text we have in product search. We also provide insights into factors that can influence model behavior for different types of query, such as the length of retrieved list, and query complexity, and discuss the implications of our findings for e-commerce practitioners, with respect to choosing a well performing method.

Motivation & Objective

• To investigate the performance of recent supervised learning-to-match methods in the context of product search, where the vocabulary gap is more pronounced than in traditional IR tasks.
• To evaluate the trade-offs between effectiveness (ranking quality) and efficiency (training and inference speed) across 12 learning-to-match models.
• To understand how query characteristics—such as length and popularity—affect model performance and to guide model selection for e-commerce platforms.
• To assess the impact of pre-training domain mismatch on BERT-based models in product search, where training data differs significantly from typical web text.
• To provide actionable recommendations for e-commerce practitioners on selecting optimal learning-to-match models based on real-world deployment constraints and query types.

Proposed method

• The study evaluates 12 supervised learning-to-match models, including neural architectures like MV-LSTM, DRMMTKS, DUET, and BERT-based models, trained to predict relevance between product queries and item descriptions.
• Models are evaluated using NDCG@5 and NDCG@25 as primary metrics, with comparisons conducted on two datasets: a public CIKM 2016 benchmark dataset and a proprietary query log from a European e-commerce platform.
• The analysis includes ablation studies on query length and popularity, identifying query-dependent performance variations across models.
• Efficiency is measured via training and inference time, with a focus on real-world deployability, especially for production systems with latency constraints.
• The study uses a lexical baseline (BM25) to quantify improvements from semantic matching, with performance gains reported as percentage improvements over this baseline.
• Model behavior is analyzed on queries where semantic matching outperforms or underperforms lexical matching, identifying characteristics of such queries.

Experimental results

Research questions

• RQ1 How do different supervised learning-to-match models compare in effectiveness (NDCG@5 and NDCG@25) on product search tasks across two diverse datasets?
• RQ2 What is the trade-off between model effectiveness and computational efficiency (training and inference time) for learning-to-match methods in real-world e-commerce settings?
• RQ3 How do query-specific characteristics—such as length and popularity—affect the performance of different learning-to-match models?
• RQ4 Why does a state-of-the-art BERT-based model underperform in product search despite strong performance in other IR domains?
• RQ5 Which learning-to-match model offers the best balance of accuracy, efficiency, and robustness for deployment in real-world e-commerce platforms?

Key findings

• On the CIKM 2016 dataset, learning-to-match methods improved NDCG@5 by up to 134.46% over a BM25 lexical baseline, indicating significant potential for semantic matching in low-lexical-overlap scenarios.
• On the proprietary dataset, performance gains were more modest (≤29.93%), attributed to higher lexical overlap between queries and product titles, reducing opportunities for semantic matching to improve rankings.
• MV-LSTM and DRMMTKS consistently ranked among the top three models across both datasets, demonstrating strong robustness for short text matching in product search.
• ARC-I emerged as the preferred model for real-world deployment due to its optimal balance of effectiveness, efficiency, and consistent performance across query types and datasets.
• A BERT-based model performed poorly, which the authors attribute to the mismatch between the pre-training data (general web text) and the short, keyword-rich, non-sentence-like product queries and titles.
• A substantial fraction of queries (over 50%) were improved by semantic matching, but a non-negligible portion were hurt, suggesting that query-dependent model selection could enhance overall performance.