Bootstrapping a Collaborative Filtering Model

The biggest challenge with using collaborative filtering models in practice is the bootstrapping problem. The most extreme version of this problem is when you have no users, and therefore no history to learn from. What products do you recommend to your very first user?

But even if you are a well-established company with a long history of user transactions, you still have the question: what do you do when a new user signs up? And indeed, what do you do when you add a new product to your portfolio? There is no magic solution to this problem, and really the solutions that we suggest are just variations of use your common sense. You could assign new users the mean of all of the embedding vectors of your other users, but this has the problem that that particular combination of latent factors may be not at all common (for instance, the average for the science-fiction factor may be high, and the average for the action factor may be low, but it is not that common to find people who like science-fiction without action). Better would probably be to pick some particular user to represent average taste.

Better still is to use a tabular model based on user meta data to construct your initial embedding vector. When a user signs up, think about what questions you could ask them that could help you to understand their tastes. Then you can create a model where the dependent variable is a user’s embedding vector, and the independent variables are the results of the questions that you ask them, along with their signup metadata. We will see in the next section how to create these kinds of tabular models. (You may have noticed that when you sign up for services such as Pandora and Netflix, they tend to ask you a few questions about what genres of movie or music you like; this is how they come up with your initial collaborative filtering recommendations.)

One thing to be careful of is that a small number of extremely enthusiastic users may end up effectively setting the recommendations for your whole user base. This is a very common problem, for instance, in movie recommendation systems. People that watch anime tend to watch a whole lot of it, and don’t watch very much else, and spend a lot of time putting their ratings on websites. As a result, anime tends to be heavily overrepresented in a lot of best ever movies lists. In this particular case, it can be fairly obvious that you have a problem of representation bias, but if the bias is occurring in the latent factors then it may not be obvious at all.

Such a problem can change the entire makeup of your user base, and the behavior of your system. This is particularly true because of positive feedback loops. If a small number of your users tend to set the direction of your recommendation system, then they are naturally going to end up attracting more people like them to your system. And that will, of course, amplify the original representation bias. This type of bias has a natural tendency to be amplified exponentially. You may have seen examples of company executives expressing surprise at how their online platforms rapidly deteriorated in such a way that they expressed values at odds with the values of the founders. In the presence of these kinds of feedback loops, it is easy to see how such a divergence can happen both quickly and in a way that is hidden until it is too late.

In a self-reinforcing system like this, we should probably expect these kinds of feedback loops to be the norm, not the exception. Therefore, you should assume that you will see them, plan for that, and identify up front how you will deal with these issues. Try to think about all of the ways in which feedback loops may be represented in your system, and how you might be able to identify them in your data. In the end, this is coming back to our original advice about how to avoid disaster when rolling out any kind of machine learning system. It’s all about ensuring that there are humans in the loop; that there is careful monitoring, and a gradual and thoughtful rollout.

Our dot product model works quite well, and it is the basis of many successful real-world recommendation systems. This approach to collaborative filtering is known as probabilistic matrix factorization (PMF). Another approach, which generally works similarly well given the same data, is deep learning.