This talk aims to introduce our app push notifications recommender system. You will learn how the Financial Times (FT) data science team used various unsupervised machine learning techniques to boost recommendation click-through rate (CTR) by 30%. This talk is for you if you want to know more about near real time and batch deployment, how our data science team deals with uncertainty caused by unsupervised machine learning, and how hybrid models can produce superior results.
The FT is a digital subscription business. More than 80% of our users are only digital subscribers. I will introduce the FT’s digital business, describe the whole spectrum of statistical models we deal with, and focus on our most recent model – Breadth of Reading Article Recommender System (BoR RecSys).
BoR RecSys is a hybrid recommender model, that uses a whole spectrum of data science techniques
* Bayesian shrinkage * Paragraph vectorization and other NLP techniques * Hierarchical clustering * Breadth of Reading (BoR)
Representing articles in a quantitative way is a typical data science problem, however we needed to be able to implement a method which would allow for a near-real time vectorization. I will discuss how we validated and implemented the start-of-the-art paragraph vectorization method. In order to keep the model relevant and useful to the business we needed to ensure clusters remained consistent and up-to-date.
In this talk, I’ll go through our infrastructure in order to enable this. Then, I will explain how we merge article vectorization, article clustering, and user page views together, track users’ activity and map users’ preferences into article clusters. I will go through the use of Bayesian shrinkage to account for the strength of evidence in cluster page views. I will also mention the BoR concept, and discuss how we used it to surface relevant recommendations. In the last part, I will evaluate the model’s performance compared to its predecessor, and the challenges we face with embedding this model in other FT products.