Cognitive Embeddings Search

Introduction

Cognitive Embeddings Search (CES) improves recall for zero and low result searches, using a vector-based search algorithm to extrapolate and cluster the relationship between products that do not share traditional product data. 

At a high level, if we consider each item in a catalog as a “star,” we can then visualize a product catalog as a “sea of stars” formed by many, many items. From this line of thinking, a search query would be some arrow (the vector) inside this “sea of stars”. Thus, we can understand the relations between the “stars” better through clustering, and we can extrapolate which ones are the closest neighbors to a given search query, the arrow, through this representation.

When do we use Cognitive Embeddings Search?

No Results

At Constructor, we use Cognitive Embeddings Search to solve the common problem of zero and low result searches. A search for something like “healthy snacks” would generally result in zero results as it doesn’t match well with products in a traditional catalog since “healthy” is a very generic and abstract concept for a search engine. Cognitive Embeddings Search changes this through relational and contextual understanding  between the search term “healthy snacks” and the result “no sugar added freeze-dried bananas” by training our Cognitive Embeddings model on product data, such as categories, product name, and text descriptions. Therefore, we can map out algorithmically the relations from “healthy snacks” to “healthy foods” to “healthy & organic dried banana chips” to finally “no sugar added freeze-dried bananas” and narrow down our understanding of a query by mapping a specific term to neighboring concepts. So inversely, if “no sugar added freeze-dried bananas” are not available in a product catalog, we can return other healthy snacks, like “organic dried banana chips” instead.

Low Results

Cognitive Embeddings Search also augments our existing search when the engine returns less than 10 products in the result set to promote additional product discoverability. These additional products returned by Cognitive Embeddings Search will be ranked alongside previous results by our ML ranking algorithm and can be differentiated in Interact within the Customer Dashboard. 

Result Validation

You can always tell if a product is returned in the result set as a result of the Cognitive Embeddings algorithm in Interact through the Dashboard. There will be a star badge () in the top right corner of the product card if the product is returned from Cognitive Embeddings. 

Analytics

Cognitive Embeddings Search analytics provide insights into what queries return Cognitive Embeddings Search results, whereas before the query would have resulted in zero or low results. Zero result analytics help merchandisers determine where to start stocking new products, place redirects, create synonyms, or perhaps introduce UI elements to help direct shopper queries. With the introduction of Cognitive Embeddings Search, zero results are greatly reduced, necessitating new analytic insights to quantify where these low result opportunities exist. Cognitive Embeddings Search analytics have been added to the opportunities analytics page in the timeline and table view below to show which queries result in Cognitive Embeddings Search results.

The timeline view (below) allows merchandisers to toggle between Cognitive Embeddings Search Results and Zero Results (toggle on the top right of the chart) to view the number of low result queries per day.

Underneath the time series chart is a new Low Result tab replacing Zero Results (see image below). Merchandisers can toggle between Cognitive Embeddings Search Results and Zero Results to view and export tables displaying counts for top queries. 

Do zero results persist with Cognitive Embeddings Search turned on?

While Cognitive Embeddings Search greatly reduces zero results, it is still possible for queries to return zero results. Zero results with Cognitive Embeddings Search turned on, often indicate that complex filters have been layered onto the query, for example searching for a coat at a specific store location or items that match a query are out of stock. With Cognitive Embeddings Search turned on, we recommend watching both sets of top queries.