In our previous posts, we walked through the sampled softmax and in-batch negatives as loss functions for training a retrieval model. In this post, we will explore the best of both worlds in the so-called mixed negative sampling. This method will prove the to be the most data efficient and yield the best performance of the previous two methods. The post contains a new implementation of all three methods which will allow all three to be trained from (approximately) the same data loader and compared.

Consider the retrieval problem in a recommendation system. One way to model the problem is to create a big classification model predicting the next item a user will click or watch. This basic approach could be extended to large catalogs via a sampled softmax loss, as discussed in our last post. Naturally, the resulting dataset will only contain positive examples (i.e., what was clicked). In this post, we explore a more optimized version of this approach for more complex models via in-batch negatives. We derive the LogQ Correction to improve the estimate of the gradient and present code in PyTorch to implement the method, building on the previous blog post.

Sometimes I’m asked how to start learning about recommender systems, and there’s a few go-to papers I always mention; however, without a proper map, it could be alittle difficult for the uninitiated. So, to try to make a gentle introduction, I will walk through an implementation of the candidate generation model based on Deep Neural Networks for YouTube, which I will sometimes refer to as the “Deep Nets” paper. This paper (and its talk) is jam-packed with all sorts of interesting practical recommendation system knowledge and is the perfect starting place for people hoping to understand large-scale recommendation systems. I will implement the key components of the candidate generation model and train the model on the MovieLens dataset in PyTorch, a typical benchmark dataset in the RecSys space. There will also be a few modern flourishes and comments on new developments as applicable. I’ll conclude with outlining a natural extension to this approach to predict multiple outcomes.

In this post, we explore a situation typical of a website, where a user is allowed to view the page multiple times and may click on a button of interest on any visit. We would like to perform an A/B test on the non-user level metric, the click-through rate (CTR), defined via total clicks divided by total page views. In order to do the final analysis, we would need to estimate the variance for the variable via the delta method. However, what does that mean for a priori statistical power calculations?

In the original CUPED paper, the authors mention that it is straightforward to generalize the method to multiple covariates. However, without understanding exactly the mathematical technique to find the CUPED estimate, it may be confusing to attempt the multiple covariates case. In this post, we will explain the thought process behind the CUPED estimate and demonstrate the analytic formula for a multiple covariate extension to CUPED. We then discuss the estimate for non-user level metrics, and we will need to use the delta method for the variance. In this case, the book keeping when using the delta method would be tedious unless you use a simplified calculation which we empirically demonstrate in the second section of the post.