Value-Aware Recommendation: Algorithms and Applications

In a high variety of application domains, the amount of data generated daily has grown more and more over time to the point that its use now exceeds the computational capacity of humans. For example, in the case of e-commerce or online streaming platforms with many new joining users and new items marketed every day, it is complex and time-consuming to manually process and exploit hidden information in order to promptly intercept user interests. In these contexts, machine learning algorithms capable of learning from data have been successfully adopted in the industry by all major market players for their ability to identify patterns in user interactions and generate recommendations that can trigger possible purchases or views on the platforms. These algorithms, known in the literature as recommender systems, are essentially information filtering technologies designed to process a very large number of alternatives in situations of information overload with the aim of funneling the user’s attention to a subset of potentially more interesting items. Over time, alongside the development of gradually more complex machine learning models, e.g., based on deep neural networks, these systems have become increasingly effective at predicting users’ interests. Intuitively, the underlying assumption is that a higher-performing service that can provide recommendations of greater interest to users will in turn positively impact business goals as well, e.g., in the form of higher customer retention or loyalty. However, although in some cases this assumption holds, in many others the recommendation of products or services despite being of great interest to users may bring only partial benefits to the business, e.g., certain products may be unprofitable for the company while others may encourage the purchase of complementary competing products. In reality, recommender systems can be designed to target organizational economic goals more directly by incorporating monetary considerations such as profitability and business value aspects into the underlying machine learning models. Such systems, that are denoted in the literature as value-aware recommender systems, are highly relevant because typically organizations aim to generate recommendations of interest to users only as long as they can increase business value performance indicators. However, although these value-aware systems are of great interest for business purposes, research is still highly scattered and composed of many works proposed in isolated contexts, i.e., where such systems are designed to target only certain application domains and their reuse in other contexts requires major readaptations of the underlying models. Hence, more in-depth research is required. With this thesis we aim to focus on the study of value-aware recommendation systems, investigating benefits and potential harms of using these algorithms in practical business applications.