Hai is powered by innovative recommendation engines with state-of-the-art deep learning techniques to learn and understand what makes the taste of each individual unique.
3 main approaches are defining the core technology behind Hai:
Deep Collaborative Filtering
We use deep collaborative filtering to understand people's preferences from cross-domain data sets. Unlike Matrix Factorization, our algorithms do not reduce an individual's preferences to a simplistic linear model but focus instead on all the complex patterns driving their tastes.
Semantic Graph Embedding
With graph embeddings we make sense of metadata, labels, tags, genres, actors and any ontological relation. Those approaches allow to find hidden correlations between seemingly unrelated or unpopular items.
Deep Content Extraction
We extract relevant information from text or image data using natural language processing and convolutional neural networks. Using this technology we can recommend an item that no one has rated yet by using its movie poster, synopsis or review.
Our algorithm uniquely learns the tastes and preferences of each individual while tackling the cold-start problem, that is, a recommendation of a new or rare item based on its content. Our backend uses a custom infrastructure built from scratch on Kubernetes which enables re-training a user's profile in real time every time they rate an item.