Demand Excellence

Item Matching Recall Enhancement by Image and Text Embedding

Eway, Amir

#DataScience

Item matching in large scale with high recall has critical business impact and presents unique technical challenges for e-commerce companies. In our efforts to enhance item matching recall, we discovered a phenomenon that, while image and text embedding each provides almost non-overlapping true matches, they both compliment elastic search, meaning that they provide true matches that elastic search does not. The rate of enhancement may vary, but this phenomenon never stops to exist.

We enhance matching recall largely by tapping into this complementary phenomenon. We describe challenges in matching recall enhancement by deep embedding, including how to leverage GPU to achieve massive parallelism in embedding computation, how to leverage FAISS as a computation engine for visual and text item cluster computation, and software architecture implications for matching pipeline design. Finally we present three experimental results showcasing the complimentary phenomenon.

• 
• 
• 
•