NEAR$^2$: A Nested Embedding Approach to Efficient Product Retrieval and Ranking

Abstract: E-commerce information retrieval (IR) systems struggle to simultaneously achieve high accuracy in interpreting complex user queries and maintain efficient processing of vast product catalogs. The dual challenge lies in precisely matching user intent with relevant products while managing the computational demands of real-time search across massive inventories. In this paper, we propose a Nested Embedding Approach to product Retrieval and Ranking, called NEAR$^2$, which can achieve up to $12$ times efficiency in embedding size at inference time while introducing no extra cost in training and improving performance in accuracy for various encoder-based Transformer models. We validate our approach using different loss functions for the retrieval and ranking task, including multiple negative ranking loss and online contrastive loss, on four different test sets with various IR challenges such as short and implicit queries. Our approach achieves an improved performance over a smaller embedding dimension, compared to any existing models.

• The paper introduces NEAR2's nested embedding framework (MRL) to balance precision and efficiency in product retrieval.
• It leverages a ranking loss strategy and Multiple Negative Ranking Loss to enhance performance on ambiguous queries and vast inventories.
• Experimental results demonstrate up to 12x efficiency gains and over 100x memory reduction while maintaining strong precision, recall, and ranking metrics.

A Technical Review of NEAR2: A Nested Embedding Approach to Efficient Product Retrieval and Ranking

The paper "NEAR2: A Nested Embedding Approach to Efficient Product Retrieval and Ranking" presents an advanced method for optimizing e-commerce information retrieval (IR) systems, addressing the dual challenge of accuracy in query interpretation and computational efficiency. Real-time product retrieval within vast inventories remains a pressing problem in platforms like Amazon and eBay, which struggle with processing latency induced by deep neural networks and Transformer models. The solution proposed by Qian et al., known as NEAR2, aims to balance these trade-offs effectively while targeting essential practical problems such as short, ambiguous, or alphanumeric queries that many current models find challenging.

Methodology: Nested Embedding Framework

At the core of the proposed approach is the concept of Matryoshka Representation Learning (MRL), which facilitates a nested development of embeddings across various dimensional scopes within encoder-based Transformer networks like BERT and eBERT. NEAR2 employs a ranking loss mechanism to train these nested embeddings, accommodating constraints of both high-dimensional precision and low-dimensional efficiency. Multiple Negative Ranking Loss (MNRL) plays a critical role in discerning and optimizing the model's response to relevant and irrelevant product alignments. The NEAR2 model further incorporates a User-intent Centrality Optimization (UCO) Model that improves user-centric retrieval tasks, especially those complicated by ambiguous queries.

Experimental Results: Efficacy of NEAR2

NEAR2 has demonstrated significant improvements across several datasets, notably achieving an efficiency increase of up to 12 times in embedding size and a reduction in memory usage by over 100 times without any additional pre-training burden. The evaluations used four eBay-specific test datasets—Common Queries, CQ Balanced, CQ Common String, and CQ Alphanumeric—each embodying unique IR challenges. The models using NEAR2 substantially enhanced standard metrics like precision, recall, NDCG, and mean reciprocal rank (MRR). Even with deliberately constrained embedding dimensions, the performance maintained a robust consistency, suggesting the model's capacity to deliver optimal retrieval efficacy with minimal resource investments.

Implications and Future Directions

From a theoretical perspective, NEAR2 contributes to the discourse on scalable semantic retrieval, engaging with the intricacies of memory-efficient embeddings tailored for real-time product retrieval systems. It proposes a formation adaptable to various IR tasks, which has pertinent implications for large-scale e-commerce applications handling diverse and potentially opaque user inputs. Practically, it provides tangible benefits in secondary goals like user satisfaction, where reduced latency and more accurate query interpretation contribute directly to enhanced user experiences.

Future research directions, as delineated by the paper's authors, suggest deploying their NEAR2 model through real-world A/B testing and further evaluating its integration in generalized embedding models like NV-embed-v2. Moreover, there is potential to refine NEAR2 models using richer datasets, which could offer high precision across broader product categories and languages, thereby aligning the model more closely with the diverse, global landscape of e-commerce.

In conclusion, NEAR2 represents a significant step forward in addressing both accuracy and efficiency in product retrieval systems. It captures the advantages of modern AI in e-commerce, offering a template for balancing computational demands with the nuanced challenge of accurately matching user queries with product intents, all while minimizing system load and maximizing precision. The approach invites ongoing refinement and integration into broader retrieval frameworks, promising sustained improvement in digital marketplace interfaces.