The classifier

In Voyado Product Discovery, the product discovery engine first needs to understand both the product and the search intent before it can match a customer with the right product. To do this, the engine analyzes and enriches your product data, organizes it into a structured knowledge model, and uses that understanding to interpret searches and recommendations. This article explains the core components that transform product data into product intelligence and power search and discovery across Product Discovery.

From product data to search results

Before the product discovery engine can return a search result or recommendation, it needs to understand both the products in your catalog and what the customer is looking for.

To achieve this, the product discovery engine processes your product data through several connected systems. Product information is analyzed and enriched, structured into a knowledge model, and then matched against customer searches. Finally, the most relevant products are ranked and returned to the shopper.

At the heart of this process is the Product Discovery ontology, a retail-specific knowledge model that helps the product discovery engine understand products, attributes, and concepts in the same way shoppers do. Rather than relying only on keywords, the ontology enables the engine to understand product types, relationships, and customer intent.

If you'd like to learn more about the ontology itself, see Understanding ontology in Product Discovery.

The diagram below shows how product data moves through Product Discovery, from initial analysis to the final search results shown to the customer.

Pipeline diagram showing how product data flows through the classifier, knowledge graph, and query analyzer before being ranked and returned as search results

The process starts when product data is sent to Product Discovery. The classifier analyzes and enriches the data before placing products into the knowledge graph, where they are connected to product types, attributes, and concepts. When a customer performs a search, the query analyzer interprets the search intent and matches it against the structured product data. Finally, Product Ranking determines the order in which relevant products are shown.

The classifier

The classifier automatically analyzes and enriches your product data when it is sent to Product Discovery. It interprets the information in your product feed and maps each product to a structured product model within the knowledge graph.

By extracting attributes from structured fields, product descriptions, and sometimes images, the classifier ensures that the product discovery engine understands what each product is and how it should behave in search. The quality of the classification depends heavily on the quality and consistency of your product data.

Learn more about how the classifier works

The knowledge graph

The knowledge graph structures products, attributes, and concepts into a connected model that allows the product discovery engine to understand meaning rather than only text. It organizes product types into hierarchies and connects related concepts so that search queries can be interpreted more accurately.

Because these concepts are linked, customers can search at different levels of specificity and still receive relevant results. The knowledge graph also supports multilingual search by connecting equivalent concepts across languages.

Learn more about how the knowledge graph works

The query analyzer

The query analyzer interprets your customers' search queries in real time. Instead of matching text directly, it breaks each query into meaningful concepts such as product type, brand, color, and contextual attributes.

These concepts are then matched against the classified product catalog. If no exact match exists, the product discovery engine can automatically relax less important parts of the query to return the most relevant results instead of showing an empty search page.

Learn more about the query analyzer

Product ranking

Product ranking determines the order in which products appear in search results and product listings. Once relevant products have been identified, the product discovery engine evaluates how likely each product is to convert and ranks them accordingly. Ranking considers factors such as customer interactions, recent trends, stock availability, price changes, and product lifecycle signals. This ensures that customers see products that both match their search and perform well commercially. Read more here.

The classifier processes every product in your catalog and works out what it actually is. It reads your product data, analyzes images, and maps each item to the right place in the Knowledge Graph, giving the product discovery engine the structured information it needs to match products accurately against what your customers search for.

How it works

Reading your product feed

The first thing the Classifier does is read the structured fields in your product feed. If you've sent a brand name in the brand field, the engine trusts that value and uses it directly. Attributes you provide explicitly are treated as reliable and given priority over anything the Classifier might infer from other sources.

Analyzing product images

If you provide images, the Classifier identifies which image best represents the product's colors, then runs a pixel-level analysis to determine the primary and secondary colors. This works well for clean, front-facing images against a neutral background. Products with complex lighting, many shades, or patterned backgrounds can produce less accurate results. If the automatic color detection doesn't look right for a product, you can review and manually override it in the Product Discovery application.

Extracting attributes from text

Not every attribute needs to be in your feed for the Classifier to find it. Using advanced language models, it reads your product titles and descriptions to infer things like material, style, or fit, even when those aren't sent as separate fields. A detailed, well-written product description gives the Classifier more to work with and generally produces better results.

Assigning concept precision

Once it has gathered everything it can, the Classifier maps the product to concepts in the Knowledge Graph and assigns a quality score to each match. A product correctly identified as a "tank top" gets a stronger classification than one only identified as "clothing." That precision directly affects how well the product surfaces in relevant searches, as more precise matches rank more strongly.

Because the Classifier is data-driven, the quality of what comes out is closely tied to the quality of what goes in. Incorrect or inconsistent product data is the most common cause of misclassification.

How to improve classification accuracy

Send clean, consistent data

Send brand values in a dedicated brand field, and make sure they're accurate and consistent across your catalog. If brand names are embedded in descriptions or vary in spelling, the Classifier is more likely to miss or misread them.

Write clear, focused product titles and descriptions. Descriptions that reference other products, such as "pairs perfectly with our linen shorts", can confuse the Classifier, which may pick up the wrong product type from those references. Describe the product itself, not what it goes with.

Use good product images

Use high-quality, front-facing images with a neutral background. The cleaner the image, the more reliable the color analysis. Flat product shots with even lighting give the best results.

Investigate the feed first if something looks wrong

If products are appearing in unexpected search results, reviewing your product feed is the right starting point. Inconsistent data, such as mismatched brand names, descriptions that mix product types, or low-quality images, is usually the cause. If you believe a product type or classification model itself needs adjustment, contact your Customer Manager. Classification models are continuously improved, and specific feedback from customers helps drive that process.

Was this article helpful?

/