When AI recommends what to buy

The talk first introduced recommendation systems in general. They are built on connecting users with products and user interactions in the form of the so-called user-item or interaction matrix. Such interactions include, for example, rating of items by users, products that the user has only viewed, or the time spent viewing each item. Subsequently, the recommendation system is able to offer to the user items that, based on the data collected, they might be interested in. Similarity between users and products can be calculated using several methods, such as the nearest neighbour method, matrix factorisation or a neural network.

Multi-armed bandit as a solution to a classic dilemma?

When filling in the user-item matrix, we may encounter several problems. The most well-known is the cold-start problem, where we have very few interactions with a new user or product. In practice, it looks like the system is trying to recommend a product to the user, but it does not know what to recommend, since it does not know the user’s preferences. Therefore, it will, for example, offer them popular products. However, the user may not be satisfied with these and so does not interact with the so-called long-tail products, which again delays the completion of the relevant parts of the matrix. An example of a solution to this problem for new products is to examine the similarity of their attributes to those products that the user has already seen or bought, so that we can place the new product in certain categories and start to make at least a partially informed recommendation. Here we encounter the classic dilemma between exploration and exploitation based on the reinforcement learning theory.

A common solution to such problems is the so-called multi-armed bandit approach. For streaming platforms, the recommending system gets rewarded for the number of times the user views the videos. When a recommendation is requested, it offers the user an item and determines whether the user has watched it; if so, the RS is rewarded. Gradually, the system refines its strategy, which is based on the fixed assumption that there are popular items that do not change with time for most users. Of course, this may not be true, which is why our researchers like Rodrigo Alves are working on better approaches to bandit recommendations that seek a balance between so-called loyal users and curious ones who do not respond to “all time favorite” items. Of course, there are a number of other approaches besides bandit algorithms. You can see the latest ones, for example, at the RecSyS conference.

It is important to test research on real data

As Recombee is a startup that runs a joint research lab with the Faculty of Information Technology, there was also a discussion about the applicability of published research in practice. Participants agreed that it is important for researchers to compare new models and approaches on well-grasped metrics and have the opportunity to test them on real data. After all, we don’t do research to outperform other models by some tiny fraction on some obscure metric, but we strive to ensure that the results of our research are used as much as possible.