Integration of Explainability in Recommender Systems to Enhance Enterprise Value Strategies

Companies are often looking for strategies to achieve business goals in the most efcient way. In their journey, they rely on systems and algorithms to support their decisions. One family of algorithms, that helps companies in choosing which actions to take, is Recommender Systems (RSs). RSs are a family of algorithms that generate suggestions of the item-user type. They are widely used in social networks, e-commerce, news and advertising, and online streaming applications. However, one area that has not yet been widely addressed in the literature is related to the business efects of recommendations. The business efect can be interpreted as the generation of value for the company, which can take many shapes and involve many users. For example, the business value may consist not only of a direct measure of company metrics, such as revenue or proft, but also of changing the sales distribution, increasing Clicks-Through Rates for an advertisement, or even keeping a user’s interest in an item high. This topic in the literature is called Value-Aware Recommender Systems (VARSs). VARSs are a particular class of RSs that aim to maximize one or more business indicators to achieve a well-established business objective. Through VARSs, a company can adopt strategies to increase efciency and answer business questions by driving the market dynamically. However, end-users must understand a recommendation to get the best tradeof between demand and response. In the frst part of this thesis, we explore the state-of-the-art of VARSs, collecting and classifying all VARSs available in literature in a frst-of-its-topic systematic review. Additionally, we propose a more specifc taxonomic categorization for the diferent models provided, highlighting the model outcome and the economic perspective. Furthermore, in order to improve the comprehension of the suggestions provided by these algorithms, we focus on Explainable Artifcial Intelligence (XAI) that aims to promote transparency in RSs and thus incentivize user adoption. XAI focuses on diferent perspectives, such as industrial, social, and end-user, giving explanations based on the context. In the second part of the thesis, we suggest a novel, explainable, value-aware recommender system that aims to balance XAI and VARS perspectives. Applying XAI to VARSs is still exploratory and has several potential evolutions and academic-industrial interests. Scientifc research has shown that there are many advantages to complementing a recommendation with a convincing explanation. For example, such explanations often lead to improved user trust, which in turn improves the efectiveness of recommendations and system adoption. In particular, for this reason, many research works are studying explainable recommendation algorithms based on graphs, i.e., exploiting Knowledge Graph (KG) or Graph Neural Networks (GNNs) methods. The use of graphs is very promising since algorithms can, in principle, combine the benefts of personalization and graph reasoning, thus potentially improving the effectiveness of both recommendations and explanations. However, although graph-based algorithms have been repeatedly shown to bring improvements in terms of ranking quality, not much literature has yet studied how to properly evaluate the quality of the corresponding explanations. In the fnal part of this thesis, we discuss a problem that afects explainability features applied on KG and GNN models, examining how they are currently assessed and suggesting the direction for a future more quantitative and comparable evaluation.