Elevating e-commerce through the strategic use of AI and machine learning

Tackling product matching accuracy with machine learning

The retail and tech experts behind the deal hunting platform had a clear mission: to help shoppers find the best prices by providing accurate and unbiased price comparisons across multiple online retailers and marketplaces. Early in the company’s journey, machine learning models were developed to match products across different sites, enabling users to easily compare prices and secure the best deals.

As the platform gained popularity and a growing user base, there was a need to further improve the accuracy and performance of these product-matching models. Product matching plays a significant role in e-commerce as it ensures that shoppers are comparing the same products across different retailers. However, it is often challenged by inconsistent data, lack of standardised identifiers, and variations in product descriptions across retailers. Differences in naming, incomplete or incorrect information, and product variations (e.g. size or colour) further complicate the process. Additionally, dynamic catalogues, duplicate listings, and cross-border differences add complexity. These challenges scale with the number of retailers. Additionally, the accuracy of the current models was found to be significantly lower in real-world implementations compared to controlled environments.

To address these issues, the retail scaleup sought to explore innovative AI and machine learning techniques that would not only further improve product matching but also make the models more scalable and easier to deploy. DiUS was brought in to provide specialist machine learning expertise to enhance product matching accuracy without compromising the platform’s speed or the quality of the user experience.

Enhanced machine-learning powered product discovery experience

DiUS and the retail scaleup partnered on evolving the overall product matching process, whether using the company’s website, mobile app, or browser extension to identify the best deal for an increasing range of products. Key to this approach was combining the company’s deep e-commerce expertise with DiUS’ machine learning expertise and experience to enable creative problem-solving.

To address the complexity of product matching in retail caused by the variance and sheer size of the data involved, the team explored various methods, from text-only to more advanced approaches that combine text, images, and metadata. This allowed the team to improve product matching by tackling the inconsistencies head on. The solution ultimately involved a custom multimodal machine learning model that delivered the desired ability to match products across retailers with greater precision.

In parallel, a custom machine learning model to improve product categorisation using the retail platform’s product taxonomy. A key challenge was accurately labelling products, which was achieved by training the model with high-quality data and fine tuning it with multimodal features. Semantic search was also implemented to enhance the context and subtlety understanding of search queries, making the search process more intuitive.

In addition to improving the product matching models, DiUS helped to streamline and automate parts of the machine learning model deployment pipeline. This included applying best-practice MLOps approaches and techniques, enabling faster model training and inference. This flexible deployment pipeline supported both provisioned and serverless deployments. To further enhance efficiency, data pipeline and storage were migrated to more scalable solutions, reducing operational costs and increasing overall system performance.

Shoppers get what they’re looking for faster

By using the latest machine learning advances, the retail scaleup’s is able to expand across more products, retailers and markets with confidence, even when there are gaps in product data. The improvements to the product matching and search accuracy deliver precise and reliable price comparisons, bolstering user retention and engagement. In addition, the machine learning models performed as well in the field as in test situations. The team has a leaner, more efficient operation that is able to swiftly adapt and update its product to stay ahead of market trends and competitor movements.

Together, the partnership laid the groundwork for the company’s continuous innovation and growth in the fast-evolving online retail space. DiUS is now regarded as the retail scaleup’s specialist machine learning partner and its go-to innovation partner for bringing new approaches and technology to life.

Recently, DiUS supported the scaleup by developing a proof of concept (PoC) to explore the impact on the platform’s product discovery experience of shifting from traditional searches to conversational interactions. By leveraging large language models (LLMs) and retrieval-augmented generation (RAG), this PoC demonstrated the potential for users to find the best deals through natural, conversational queries without needing to specify exact product names. Although this innovation has not yet been implemented into the live platform, it showcased how future enhancements could make the platform more user-friendly and laid the groundwork for possible future developments that could increase user engagement and conversion rates.