[New Feature] Introducing In-Category Product Data Mapping & Analysis - Improve your understanding of products relevance in a category

We have rolled out a new endpoint in our PIM automation platform Pumice.ai focused on providing customers with a clearer understanding of their product data within a specific category. This endpoint provides visual and text data about a specific category that shows the similarities and differences between products that exist in that given category. This is incredibly important when it comes to deciding to restructure product catalogs, where to add new products, and where categories might not be necessary based on other existing categories. Let’s take a look at the pipeline and explain how this works.

‍

‍

The pipeline is simple, users first provide a set of products from their PIM or product database that are all categorized to the same leaf of their taxonomy. This can be product data as simple as just a title, or as much data as possible. We recommend at least a title and description to extract the most information from the model.

‍

That product data is passed over to the Pumice.ai API via a JSON request containing rows of product data records. Our first ai model generates important metadata information for each product record as well as a few metrics that will help us understand this data. These models can be fine-tuned by customers as well to produce higher quality metrics. Users are given a file back with a few key items for analytics.

‍

The first item that you get back is the map you see above. This maps the product’s similarity to each other in a 3d space. This allows you to understand how similar your products in a single category are to each other on a semantic level. You should see a pretty high level of products clustered tightly together. This should make sense, as all the products are under the same category and therefore should be semantically similar to an extent. What you really want to look for are products that extend far away from the cluster. These are products that you want to review if they belong in the category. At the end of the day, the high level idea of categories is just a way to group products that are similar. If a few products are very different from the others, maybe you need a new category for it that better represents them. Running each category in your taxonomy tree through gives you a good idea of where your product grouping stands relative to your taxonomy.

‍

Another way customers are using this is to evaluate new products and where to store them in their taxonomy. This is different from just categorizing them based on existing relationships between categories and products. When onboarding a ton of new products its worth it to understand if this grouping is close enough semantically to the existing products in a category vs creating a new category.

‍

A great example of this is if you run a store that sells common sport athletic shoes (basketball,soccer, baseball, tennis). All the shoes in your current category are related to each other to some extent (at least they should be if your categories make sense). You onboard a catalog from a wholesaler that is mostly athletic shoes but has a bit of a wider range and even has some more general shoes (rock climbing shoes, hunting boots, fishing boots, etc). While these would fall under your current categories of shoes, they’d probably be very different from your current category outline. Customers that are looking for hunting boots, or boots in general aren’t the same as customers looking for baseball cleats, soccer cleats etc. You need to evaluate how similar the new product catalog is to your existing clusters of products under the same category to decide if you need new ones.

‍

‍

The next metric that comes from this endpoint is an anomaly detection graph. This shows the clusters of similar products like we saw before, but lets us know which ones are considered “anomalies” by our ai algorithm. It also provides scores (you can see above) that calculate the distance of points to others. The higher the score, the less data points are near that data point. This provides a bit of out of the box analytics for which data points you should focus on as potential products to recategorize. The example above is a great example of how to use the tool. If you passed a group of products from the same category into the endpoint, you would see two distinct clusters of similar products. You might want to dive deeper into them and split them into different categories or subcategories. For example you might see that you have two large groups of products in your shoes category, and one is all products related to cleats, and one is athletic shoes.

‍

The final metric that is generated from this endpoint is a keyword breakdown. You’re able to see the most common keywords used for products in this category. You might be surprised at the language used for products, and you need to make sure that it aligns with your SEO and keyword mapping strategy for these product pages. This output is generally used as a safety check for your desired language.

‍

How to access

This endpoint is available to customers on the Fine-tuned plan. The fine-tuned plan gives you access to all Pumice.ai endpoints and your own custom fine-tuned models for systems such as product categorization and metadata generation. With our Fine-tuned plan, we guarantee at least 90% accuracy of your model during development.

‍

Once you’re onboarded we hand you an API key with full access to this endpoint. Most customers only run it when adding new products to their catalog, but you could also run it when making changes to your taxonomy.

‍

Need Ai for your product data and PIM?

Pumice.ai focuses on automating product data tasks with ai. Our ai automates work for PIM specialists such as product categorization, metadata generation, product deduping and more. Interested in giving it a try? Reach out to us for a demo and access to our base models. Once the base models look good we’ll fine-tune your model specific to your product data and taxonomy. We guarantee 90% accuracy during fine-tuning or you don’t pay a dime.

‍