Looking for the best ecommerce research papers published in 2024? You’re in the right place! Explore our popular annual round up of the top ten must-reads.
In 2024, the buzz around Generative AI (GenAI) reached a fever pitch, moving from imaginative brainstorming to tangible, productized solutions with clear business use cases. This shift wasn’t just theoretical – it’s reflected in an explosion of innovation, as seen in the staggering growth of patents filed worldwide.
From 733 patent families in 2014 to over 14,000 by 2023, GenAI has achieved an 800% growth. Industry leaders are at the forefront, showcasing applications that range from enhancing customer interactions to secure and customizable AI platforms.
But AI is far more than just GenAI. In ecommerce, applied AI research is tackling complex challenges like personalization, semantic search, and user experience – areas where businesses simply can’t afford to fall behind. As highlighted in a recent HBR report sponsored by Coveo and SAP – 90% of businesses cite personalized customer experiences as essential for growth, and 70% view AI as critical to their success.
At Coveo, we believe that GenAI is a powerful tool, but it’s only one piece of the AI puzzle. Maslow’s Hammer reminds us that not every challenge in ecommerce needs a GenAI solution.
That’s why we’ve compiled this list of the best 10 ecommerce research papers, reflecting the broader scope of AI innovation and how these insights align with building smarter, more effective ecommerce solutions.
Let’s dive in.
1. Recent Advances in Refinement Recommendations
This paper explores improving refinement discoverability on mobile devices by suggesting contextually relevant options in-line. These recommendations are powered by both query-based and session-based classification techniques, ensuring users can refine their searches even with the limited screen space available on mobile devices.
We agree that faceted search and refinement are essential for intuitive and efficient search experiences. In fact, this very recent piece of research aligns closely with Coveo’s own work in this area, such as our 2020 paper, “Blending Search and Discovery: Tag-Based Query Refinement with Contextual Reinforcement Learning”, which tackled tag-based query refinement as a scalable, mobile-friendly alternative to standard faceted search.
2. Long or Short or Both? Behavioral Features in Product Search Ranking
How do different time windows for behavioral data shape the relevance of product search rankings? This paper answers that question with a hybrid model that combines short-term signals (immediate trends and recent activity) with long-term insights (enduring customer habits). Validated through real-world A/B testing at Walmart, the approach strikes a balance between immediate discovery experiences and routine shopping journeys, delivering smarter, more personalized search results.
The study highlights the pivotal role of behavioral data in modern ranking systems and shows how nuanced integration of signals can significantly enhance search relevance and engagement. At Coveo, we’ve observed similar benefits: combining different types of behavioral features is key to capturing evolving shopper intent while maintaining relevance.
This paper also takes us back to 2019, when we first shared our own approach to in-session re-ranking – integrating behavioral signals in real time. If you’re curious, take a nostalgic look at how we enabled real-time personalization with less than 100 lines of code: Clothes in Space: Real-Time Personalization.
Behavioral insights continue to shape the future of ecommerce search, and this paper is a testament to the power of blending immediate and historical data for exceptional shopper experiences.
3. Towards Scalability and Extensibility of Query Reformulation Modeling in Ecommerce Search
Query reformulation (QR) is a powerful technique for overcoming challenges like sparse or noisy data in long-tail queries – searches for less popular or emerging shopping intents. A recent paper from Amazon’s labs presents a scalable approach where these less common queries leverage behavioral patterns from more popular ones with similar meanings. Tested in multilingual and low-traffic environments, the solution demonstrated measurable improvements in search performance.
In ecommerce, search ranking and matching often depend on historical customer actions like clicks and purchases, which provide valuable feedback loops. However, long-tail queries lack sufficient data to benefit from these loops, making it harder to improve their rankings effectively. QR addresses this by reinterpreting or adjusting queries to align them with semantically similar, high-volume queries, enriching their relevance and results.
At Coveo, we see this approach as highly aligned with our vision. Scalability and adaptability are essential for tackling long-tail challenges in search, especially in complex, multilingual environments. We’ve similarly focused on developing query understanding systems that handle sparse data while maintaining precision and relevance across all query types.
For more on how we tackle these challenges, check out our thoughts on building the best ecommerce search solution: How to Pick the Best E-Commerce Search Solution.
4. Large Language Model-Based Long-Tail Query Rewriting in Taobao Search
Taobao brings a fresh perspective to the challenge of managing long-tail queries with a fascinating application of large language models (LLMs). This paper stands out as the first to fine-tune LLMs specifically for industrial query rewriting—a significant milestone in search optimization.
The authors introduce BEQUE, a framework designed to rewrite long-tail queries and bridge semantic gaps in search. By employing supervised fine-tuning and contrastive learning, Taobao achieved impressive results, including marked improvements in gross merchandise volume (GMV) and user engagement.
This approach showcases the transformative potential of LLMs in tackling complex ecommerce challenges, setting a new benchmark for managing long-tail queries and enhancing search relevance.
5. Advertising as Information for Ranking E-Commerce Search Listings
While the previous paper focused on long-tail queries, let’s shift to the fascinating AI research around long-tail products – an equally challenging domain.
In ecommerce, a core issue lies in predicting the most relevant product listings for users. Platforms rely heavily on data like historical clicks, reviews, and purchases to inform ranking algorithms. However, this task is daunting given the sheer scale of ecommerce platforms, which typically host millions of products for millions of consumers with diverse preferences. The challenge becomes even more acute for new products, where platforms face a cold-start problem due to a lack of historical data.
An intriguing academic paper explores how advertising can help solve this cold-start problem. By using ad-derived signals as information for ranking algorithms, the authors demonstrated at JD.com that this approach improved organic rankings for new products. High-quality listings gained greater visibility, benefiting both customers and the platform.
At Coveo, while we find this approach compelling, cold-start challenges are not new to us. Beyond advertising, we’ve tackled these problems with scalable, content-based inference techniques that ensure new products are both discoverable and accurately ranked from the moment they’re introduced. For a deeper dive into our approach, check out our paper: Solving the Cold-Start Problem.
6. How Recommendation Affects Customer Search: A Field Experiment
As customer journeys grow increasingly complex, understanding the role and importance of different touchpoints is crucial for creating seamless experiences.
This study investigates the relationship between recommendations and search behavior, drawing on experiments with over half a million customers. The findings reveal an intriguing dynamic:
- Less relevant recommendations often drive users to engage in more search activity, particularly with long-tail queries.
- Highly relevant recommendations, on the other hand, complement search by enabling smoother, more efficient customer journeys.
These insights highlight the importance of understanding how recommendations and search interact to create cohesive user experiences that meet customer needs effectively.
At Coveo, our analytics capabilities empower businesses to track these interactions and optimize both search and recommendations for maximum impact. By leveraging data-driven insights, businesses can ensure that these touchpoints work in harmony to drive conversions and satisfaction.
Learn more about Coveo’s approach to analytics and tracking here: Conversion Rate Analytics.
7. Search Intention Network for Personalized Query Auto-Completion in E-Commerce
Looking to enhance your customer experience and boost conversion rates? Query suggestions could be the solution. They refine user queries, improve search functionality, and guide customers toward relevant results effortlessly.
In a standout paper, researchers from Taobao introduce SIN (Search Intention Network), a neural framework designed to tackle two key challenges in query auto-completion:
- Ambiguous user intent during typing.
- Shifting search behavior that deviates from historical patterns.
By leveraging transformer-based models, SIN achieves impressive results, setting a new standard for improving query auto-completion.
At Coveo, we recognize the transformative potential of auto-completion and query suggestions as a cornerstone of modern search experiences. Our investment in this area spans years, including our 2020 paper, “An Image is Worth a Thousand Features: Scalable Product Representations for In-Session Type-Ahead Personalization”. This research addressed similar challenges by incorporating dense product vectors, including image-based features, to deliver real-time, personalized suggestions.
Query suggestions are more than just a helpful tool—they are a critical part of creating relevant, intuitive, and personalized search experiences that customers love.
8. ChatGPT Goes Shopping: LLMs Can Predict Relevance in E-Commerce Search
If you’ve been scrolling through this list waiting for a mention of ChatGPT, here it is!
This paper evaluates the potential of large language models (LLMs) to predict product relevance in e-commerce search. The findings are compelling: LLM-generated relevance judgments align with human annotations 82% of the time. The study also highlights how LLMs can streamline the evaluation process by creating annotation guidelines, paving the way for greater efficiency in relevance validation.
LLMs are not just transforming shopper experiences – they’re revolutionizing backend operations like relevance prediction and optimization workflows. At Coveo, we’ve been exploring a range of ecommerce use cases for LLMs, from improving search accuracy to enhancing operational efficiency.
Want to learn more? Check out our blog on Generative AI in E-Commerce or dive deeper into how we’re solving real-world challenges in search optimization workflows with AI. For more insights, download our ebook on Generative AI risks.
LLMs are shaping the future of ecommerce, and we’re excited to be at the forefront of this transformation.
9. An Interpretable Ensemble of Graph and Language Models for Improving Search Relevance
Interpretability in complex AI models is often a challenge, yet it’s crucial for e-commerce applications. Transparent AI systems not only foster trust but also drive adoption, particularly in creative domains where diverse stakeholders bring competing priorities.
For instance:
- Buyers, merchandisers, and financial planners take a data-driven approach focused on maximizing sales and minimizing inventory waste.
- Designers rely on intuition, drawing insights from market trends, social media, and runway shows.
To align these goals, explainable tools capable of interventional analysis are essential for bridging intuition with data-driven decision-making.
This paper from Amazon Labs introduces PP-GLAM, an ensemble of graph and language models designed to improve search relevance while maintaining explainability. The use of additive explanation metrics ensures that the model’s predictions are not only accurate but also understandable to both technical and business users, making it more deployable in real-world settings.
At Coveo, we share this commitment to transparency. Our blog, Black Box Begone: Introduction to Interpretable Product Discovery, highlights how explainability in AI fosters adoption and trust across all stakeholders. Ensuring models are interpretable enables organizations to harness advanced AI without sacrificing usability or confidence.
10. Semantic Retrieval at Walmart
Hybrid search, a combination of keyword and vector search algorithms, is a powerful method for improving relevance – but it comes with challenges. While these systems bring together the precision of keywords with the contextual depth of vectors, they don’t always “speak with one voice,” requiring careful tuning to avoid inconsistencies.
This paper from Walmart explores their hybrid search system, which combines traditional inverted indexing with neural retrieval methods. A standout feature of their approach is its focus on tail queries, often the most challenging to optimize. By employing techniques like embedding reduction and efficient negative sampling, Walmart improved both relevance and efficiency – a significant achievement in balancing scalability with precision.
At Coveo, we greatly appreciate this approach and share Walmart’s vision for hybrid search. However, we believe it’s just one piece of the equation. By integrating neural and semantic search with advanced techniques like Learning-to-Rank (LTR), Coveo goes beyond hybrid search to deliver even greater relevance. Hybrid search may excel in retrieving results, but LTR models refine rankings by dynamically orchestrating multiple layers of signals:
- User behavior (e.g., clicks, dwell time).
- Contextual data (e.g., location, time).
- Business objectives (e.g., profitability, inventory levels).
By intelligently weighing these signals, LTR balances relevance, diversity, and strategic goals, enabling personalized experiences that drive measurable business outcomes. Learn more about our use of LTR: Business-Aware Product Ranking.
Dig Deeper
Hybrid search is just the beginning – the future lies in signal orchestration to consistently achieve superior results.