Deep Learning for Recommendations

Complimentary recommendations are an important form of recommendations in retail and e-commerce. When a customer purchases a product, such as a patio table, they usually also need to purchase complimentary products, such as patio chairs, or an umbrella. We developed a hybrid content-based complimentary recommendation system that combines textual and visual features to produce recommendations that are complimentary in both style and function. I describe how this model is powered by a combination of Siamese Neural Network and a simple visual similar model. Finally, I will describe how we scale this model to produce recommendations for 4 million different products.

Key Takeaways: • Effective use of DL models for recommendations • Efficacy of combining multiple models to produce more robust recommendations • How to scale a DL recommendation model on a large item set.

Simon is a senior data scientist at The Home Depot where he works on personalization and ranking for the core recommendations team. In his time at Home Depot, he has built a number of different recommender systems including a personalized deals recommender, and a system for recommending how to guides for customers working on home improvement projects. Simon has a PhD in Machine Learning and Natural Language Processing from DePaul University, over 6 years’ experience working in data science, and 15 years’ experience in software development.