Augmenting AI-powered Product Matching with Human Expertise to Achieve Unparalleled Accuracy
2nd May, 2024
By Namrata
In today’s expansive omnichannel commerce landscape, pricing intelligence has become indispensable for retailers seeking to stay competitive and refine their pricing strategies. The sheer magnitude of eCommerce, spanning thousands of websites, billions of SKUs, and various form factors, adds layers of complexity. Consequently, ensuring the accuracy and reliability of competitive insights presents a formidable challenge for retailers aiming to leverage pricing data effectively.
At the core of any robust pricing intelligence system lies product matching. This process enables retailers to recognize identical or similar products across competitors. Once these matches are identified, tracking prices is a relatively more straightforward task, facilitating ongoing analysis and informed decision-making.
Accurate matching is crucial for meaningful price comparisons and tailoring product assortments. The challenge is matching products is often complicated, especially for non-local brands, niche categories, or items lacking consistent global identifiers. It becomes even trickier when trying to match very similar but not identical products. A comprehensive approach that compares and analyzes multiple attributes like product titles, descriptions, images and more is essential.
Artificial intelligence algorithms are commonly used to automate product matching, leveraging machine learning techniques to analyze patterns in images and text data. While AI can adapt and improve over time, the question remains: Can it fully address the complexities of product matching on its own?
The reality is that many retailers still struggle with incomplete, inaccurate, or outdated product data, despite these AI-powered product matching solutions. This can lead to suboptimal pricing decisions, missed opportunities, and reduced competitiveness.
Challenges in an ‘AI-only’ Approach to Product Matching
While AI plays a vital role in automated product matching solutions, there are complexities that AI alone cannot fully address:
Subjectivity in Matching Criteria
Some product categories have subjective or hard-to-quantify criteria for determining similarity. AI learns from historical data, so it may struggle with nuanced aspects like:
Aesthetics, style, and design: In the Fashion and Jewellery vertical, for example, products are matched according to attributes like style, aesthetics, design – all of which have some subjectivity involved.
Quantity/packaging variations: In the grocery sector, variations in product packaging and quantities can introduce complexities that require subjective decision-making. For example, apples may be sold in different packaging like a 0.5 kg bag or a pack of 4 individual apples. Determining if these different packaging options should be considered equivalent often involves making a qualitative judgment call, rather than a clear-cut objective decision.
Matching product sets: For categories like home furnishings, the focus is often on matching coordinated sets rather than individual items. For example, in the bedroom category, matching may involve grouping together an entire set of complementary furniture like a bed frame, dresser, and wardrobe based on their cohesive design and style. This goes beyond simply making one-to-one product associations, requiring more nuanced judgments about aesthetic coordination.
Contextual Factor
Products can have regional preferences, cultural differences, or evolving trends that impact how they are matched. AI may miss important context like Local/regional product names or distinct brand names across countries.
For instance, in the image we see Sprite (in the US) is branded Xubei in China. Continuous human curation is needed to help AI adapt to this context.
High Accuracy & Coverage Expectations
Retailers rely on AI powered and automated pricing adjustments based on product matching for insight. To ensure that pricing recommendations and updates are accurate, accurate product matching is crucial. For this, simply identifying similar top results is not enough – the process must comprehensively capture all relevant matches. While AI excels at finding the top groupings with around 80% accuracy, even small matching errors can have significant consequences.
As AI matching improves, customer expectations may rise even higher. If AI achieves 90% accuracy, for instance, SLAs may demand over 95%. Reaching such a high level of accuracy is very challenging for AI alone, especially when faced with incomplete data, contextual nuances, evolving trends, and subjective matching criteria across products and categories.
The solution is to combine the power of AI with human expertise. This is the key to achieving true data veracity – the accuracy, freshness, and comprehensive coverage required for precise and reliable product matching.
Human-in-the-Loop Approach for Elevated Product Matching
Human intelligence and quality testing can elevate the AI powered product matching process by addressing key challenges:
- Matching Validation: AI algorithms may identify product matches with 80-90% accuracy initially. Having humans validate these AI-suggested matches allows for correcting errors and pushing the accuracy close to 100%. As humans flag issues, provide context, and re-label incorrect predictions, it allows the AI model to learn and enhance its reliability for complex, high-stakes decisions.
- Applying Contextual Judgment: For subjective matching criteria like aesthetics, design, and categorizing product sets, human discernment is needed. Humans can make nuanced judgments beyond just quantitative rules, ensuring meaningful apples-to-apples product comparisons. Their contextual understanding augments AI’s capabilities.
- Continuous Learning Via Feedback Loop: Product experts possess rich category knowledge across markets. Integrating this human insight through an iterative feedback loop helps AI models quickly learn and adapt to changing trends, preferences, and context. As humans explain their match assessments, the AI continuously enhances its precision over time.
By combining AI’s automation and scale with human validation, judgment, and knowledge curation, pricing intelligence solutions can achieve the accuracy and coverage demanded for actionable competitive pricing insights.
DataWeave’s Data Veracity Framework: A Scalable Workflow Combining AI and Human Expertise
Given the vast number of products, retailers, and brands that exist today, any product matching solution must be highly scalable. At DataWeave, we bring you such a scalable workflow to address these complexities by integrating human expertise with AI-driven automation. The image below outlines our approach for combining AI with human intelligence in a seamless, scalable workflow for accurate product matching:
Retailers and brands can benefit in several ways with this workflow, as listed below.
Several Rounds of Data Verification Due to Hierarchical Validation Teams
The workflow employs a hierarchical validation team of Leads and Executives to efficiently integrate human expertise without creating bottlenecks. Verification Leads play a pivotal role in managing the distribution of product matches identified by DataWeave’s AI model to the Verification Executives.
The Executives then meticulously validate these AI-suggested matches, adding any missing product associations and removing inaccurate matches. After validation, the matched product groups are sent back to the Leads, who perform random sampling checks to ensure quality.
Throughout this entire workflow, feedback and suggestions are continuously gathered from both the Executives and Leads. This curated input is then incorporated back into DataWeave’s AI model, allowing it to learn and improve its matching accuracy on an ongoing basis.
This hierarchical structure ensures that human validation seamlessly scales alongside the AI’s matching capabilities. Leveraging the respective strengths of AI automation and human expertise in an iterative feedback loop prevents operational bottlenecks while steadily elevating overall accuracy.
Confidence-based Distribution of Matched Articles for Validation
The AI model assigns confidence scores, differentiating high-confidence (>95%) and low-confidence matches. For high-confidence groups, executives simply remove incorrect matches – a quicker process. Low-confidence matches require more human effort in adding/removing matches.
As the AI model improves over time with feedback, the share of high-confidence matches increases, making validation more efficient and swift.
Automated, Standardized Process with Iterative Feedback Loop
The entire workflow is standardized and automated, with verification metrics seamlessly tracked. At each step, feedback captured from both leads and executives flows back into the AI, enhancing its matching accuracy and coverage iteratively.
DataWeave’s closed-loop system of AI automation with hierarchical human validation allows product matching to achieve comprehensive accuracy at a vast scale.
Unleash the Power Accurate and Comprehensive Product Matching
In summary, combining AI and human expertise in product matching is crucial for retailers navigating the complexities of omnichannel retail. While AI algorithms excel in automation, they often struggle with subjective criteria and contextual nuances. DataWeave’s approach integrates AI-driven automation with human validation, delivering the industry’s most accurate product matching capabilities, enabling actionable competitive pricing insights.
To learn more, reach out to us today!
Thank you for Subscribing - Team DataWeave
Book a Demo
Login
For accounts configured with Google ID, use Google login on top. For accounts using SSO Services, use the button marked "Single Sign-on".