In our BIGDATA grant, we pushed the state of the art in modern recommendation systems.  Recommendation systems take in data about how people consume items (movies, products, or---in this grant---scientific articles) and build a predictor of each user’s future actions i.e what movie will they choose to see next or which article will be most likely to read.

 Classically, recommendation systems were built on matrix factorization methods.  Matrix factorization are machine learning methods that learn “representations” of each user and item which, loosely, can be interpreted as preferences of the users and attributes of the items.  A fitted matrix factorization predicts that a user will like an item when there is an affinity between their representations, when the user’s preferences match the items attributes. However, although powerful, matrix factorization is also limited, and overcoming its inherent limitations was one of our major goals over the life of this project.  To this end, we developed several new recommendation algorithms that perform significantly better than the classical approach.

In one line of work, we developed methods that capture both the content of the items and the various way in which users interact with them.  This was initially motivated by a desire to solve the "cold start problem," i.e., when a new item enters the collection but where there is no information about who has liked it.  Our methods do solve this problem, but they do more as well.  They help form interpretable representations of users---defining their preferences in terms of dimensions that can be visualized and explained---as well as representations of items, a function with interdisciplinary impact. As one point of its broader impact, this work has been used, for example, by sociologists studying the historical preferences of library-users in the 1800s.

In another line of work, we modeled the discovery process of users.  Matrix factorization tacitly biases its estimation of user preferences because it does not account for how people discover items to consume.  As an example, consider watching a movie that your discovery process does not easily find.  Intuitively, a recommendation system should "upweight" your vote on this movie because it was rare, relative to the movies that you typically watch.  We turned this intuition into a mathematical framework for recommendation, significantly improving on classical methods.

We described just two results from our project.  We innovated recommendation systems in several other ways as well.  For example, we built a method to capture how user preferences change over time, a method to model user preferences in the context of complex baskets of goods, a method to involve social networks in building recommendation systems, and a method to scale recommendations systems up to massive sparse datasets.

In its broader impacts, our work has significantly pushed the state of the art in recommendation systems and models of user behavior data.  These models have become core to many technological problems across science, industry, and government.  In its intellectual merit, this work provided new methods in modern machine learning and data science. We and others continue to build on and research the mathematical and algorithmic properties of these methods.

Last Modified: 03/22/2018
Modified by: David Blei