A Recommender for Match-making: Item-based CF, PageRank, Evaluation techniques & Deep-Learning
Online match-making has a lot of challenges where Machine-Learning can help. When we look at a profile what is it that makes us swipe right or left? Is there something about a profile that attracts us and if so what can a person’s historical interactions say about their preferences.
I believe the contents would resonate with the audience quite well and help them appreciate the challenges of doing machine-learning to help in finding love in the digital era.
I worked at Coffeemeetsbagel.com, when I had a chance to build a recommendation engine which had all the interesting ingredients: Item-based collaborative filtering, PageRank and also Deep-Learning. Building it was a very fun and exciting activity since it seemed like unraveling the laws of attraction between people. This variant has the added complexity that the recommendation needs to work both ways. Choosing the right input data was key. Evaluating the recommender using the usual MAE was not sufficient. I needed to also consider Precision@N and a few other metrics. PageRank certainly helped in this regard. The Cold-start problem still persisted and Deep-Learning came to the rescue. Finally scalability was a real concern. Fortunately we had chosen the right tools.
Did we make an impact? We will know in a few months. The Recommender is on A/B test in a city near you.
My name is Prabhakar Srinivasan. I am currently working for Apple as a Data Scientist. I would like to submit this topic to FifthElephant on Recommendation Engines. I worked on this topic when I was employed at Coffeemeetsbagel.
Over the last decade, I gained experience working on Recommendation Engines,
and Deep-Learning and Supervised and Unsupervied Machine-learning techniques. I was able to successfully develop ML products for companies like Cisco, Coffeemeetsbagel and Apple.
I would like to share my experiences and knowledge in this space with the audience.