PIREC: personalized in-session recommendation engine for real-time retrieval of short-term user preferences

Session based recommendation systems are highly effective in e-commerce recommendations since they take into account a buyer’s most recent interactions, thereby providing timely and personalized recommendations. In this work, we outline a personalized recommendation retrieval system that leverages the user’s short-term interaction history. Our proposed system consists of a deployed machine learning model that: 1. generates user embeddings leveraging their site-wide short term interaction history in real time, 2. generates multimodal item embeddings using textual and image data, 3. employs contrastive learning to map users and items into a shared embedding space. We further illustrate the usage of approximate nearest neighbour algorithm for efficient real time retrieval of items, discuss strategies for deploying and maintainin