Titolo della tesi: Unveiling the Effects of Recommendation Agents on Online Behaviour: An Inquiry Into the Users’ Decision-Making Process, Implicit Social Networks and Algorithms Over-Specialization
In the last few decades, the term “algorithm” has become central to the social sciences, albeit its roots are more consolidated and evolved. The origins of the term are indeed dated back to al-Khwārizmī – an ancient Persian mathematician – and to some Euclidean scripts (Striphas, 2015) to refer to a set of mathematical procedures and rules that iteratively transform a group of input in a predefined output (Gillespie, 2014). The profusion of the term and the subsequent application coincide with the development of the Internet and new information technologies (Arnoldi, 2015). Nowadays, this simultaneous processing of users’ information – sent often passively during online activities – is leveraged by companies to direct users to contents/items related to their interests. Importantly, the results of this activity are the videos presented to users on their homepage, the featured posts on social networks, the pop-up advertisements on the visited websites, the appearance order of birthdays of our contacts lists on Facebook and, in general, the majority of things that surrounds our digital existence (Airoldi, 2015).
Specifically, the recommendation agents (RAs), that are central to this dissertation, refer to a particular category of algorithms, which has been implemented on websites with the aim of recommending contents relevant to users and their interests. MovieLens was the precursor of such agents, a platform which automatically processed the opinions of the users (expressed in the form of ratings) to offer contents of interest to the target user (Konstan and Riedl, 2012). Such algorithms change users’ digital experience into a tailor-made path built according to their interests. The growth of e-commerce has led to the exponential implementation of RAs, which, today, are fully integrated in the websites and it is not uncommon to come across the well-known prepositions "you may also like” " or " users who bought this product also bought". The results underlying these prepositions are the result of association rules and mathematical calculations that act as a filter in contexts with enormous assortments (e.g. Amazon.com)(Lash, 2007; Beer, 2009). The main aim of RAs is the reproduction of word-of-mouth that typically occurs between individuals (Ansari, 2000) and, at the same time, they replace the role of "cultural intermediaries" which had the function of disseminating information on new media and / or cultural products (Morris, 2015). However, RAs are not non-evaluative (as might be expected from their mathematical nature), on the contrary they might create a "normalized" culture in which elements beyond the user's interest are excluded (Mackenzie, 2015). Such systems are beneficial for both service providers and users (Pu et al., 2011). They reduce search costs and facilitate the selection of items in online shopping (Hu et al., 2009) and improve the decision-making and decision quality (Pathak et al., 2010). As a tool for e-commerce, RAs improve revenues, as an effective means of selling more products (Pu et al., 2011).
Although computer science and information technology literature on RAs is extensive, it is still an under-researched topic in the marketing perspective. In the manifold literature on recommender agents, only few relevant contributions have been outlined by marketing scholars with the aim to understand the phenomenon from a consumer and a firm’s perspective. Although some topics have been clarified and explained in detail, to date there are still many questions about the effectiveness of RAs.
With the aim to contribute to the extant literature related to RAs, the present thesis collects 3 articles - in 3 chapters - and reflects the evolution of 3 years of investigation on the topic. The findings of Chapter I laid down the foundations for Chapter II and, in turn, the theoretical implications of the Chapter II for the Chapter III.
In Chapter I, I carried out a systematic literature review on the topic, in order to get an organized representation of the phenomenon assuming a 22-year timeframe research period from 2000 to 2022 based on 128 articles. The contributions were then classified according to two theoretical perspectives used by marketing researchers to analyse consumers in RAs- mediated environments, (1) cognitive psychology and (2) social psychology. Then, the potential similarities among the articles were assessed through a co-citation analysis and multidimensional scaling. I found 26 theoretical frameworks which are recurrently adopted by marketing scholars to conduct research on this topic and refer to three sub fields. The findings contribute to the extant literature by providing an updated understanding of the research on recommender agents.
According to the literature gaps found in the Chapter I, no contributions have been outlined to investigate the implicit social networks enabled by recommendation algorithms, the connection among users inside the network (i.e., neighbours), their role in wide spreading marketing messages and whether dominant users exists in these implicit structures that aim at favouring customization processes.
To this end, in Chapter II, I (1) present a discussion about the role of RAs in the stages of the decision journey and through (2) an analysis of a real-world RAs-enabled network of 37,427 Amazon’s users and 1300 products (3) I assess how such agents enable implicit networks of influence inhabited by neighbourhoods of users and (4) the role of consumers in such networks. Therefore, the results emphasize the social nature of RAs-enabled networks and identify most influential users in wide spreading recommendations, according to a set of centrality and community-driven measures. Lastly, some relevant managerial implications are highlighted.
Drawing on such premises, I wondered if implicit influence social networks enabled by RAs really benefit users when associate them to similar ones or not. While prior research has primarily focused on the improvement of accuracy measures as a way to increase the match between users’ preferences and recommended items (Song et al., 2019; Dzyabura et al., 2019; Isufi et al., 2021; Hamedani and Kaedi, 2019; Panniello et al., 2014; Zhou et al., 2010; Ansari et al., 2000; Haübl et al., 2000; Knijnenburg et al., 2012; Lombardi et al., 2017; Tsekouras et al., 2020; Aggarwal. 2016), the effects of overspecialization on users’ outcomes and their antecedents are currently under-researched. In my idea, higher degrees of RAs accuracy (i.e., the attempt, for some RAs, to match users with similar interests and trigger them with the same recommendations) reduce the information overloading but increase the overspecialization and confines users within their preferences and negatively affect the outcomes of the choice. To respond to this question, in a sequence of four studies reported in Chapter III, 1) I manipulated the RAs specialization level (i.e., overspecialised vs. specialised vs. generalised (Study 1) and degree of novelty of a Recommendation set (RS; novel-based RS vs. accurate RS (Study 4), assessed the perceived reciprocity and intimacy of the RA (Study 2) and the effect on user’s expertise (Study 3), but keeping the underlying algorithms unvaried. Study 1 implies three conditions to assess how the increasing levels of RAs learning affects choice outcomes. The results, highlight that higher levels of specialization are associated to lower choice outcomes. Studies 2 and 3 reveal the antecedents of the avoidance of overspecialization. In Study 2, I assess how the RAs learning affects the perceived reciprocity and intimacy of users – as mediators - and in turn the choice outcomes. The results show that users feel a lack of reciprocity and intimacy when RAs increase the knowledge about them. Study 3 investigates how the effects of RAs specialization are detrimental for users due to a reduced chance to form new preferences. The results of this study indicate that RAs are associated to higher choice outcomes when favour the breadth of knowledge rather than the depth. Finally, Study 4 involves an online experiment in which I manipulate two degrees of novelty (high vs. low) and measure their effects on perceived novelty, as a mediator, and choice outcomes. Results show that algorithmic novelty (i.e., the ability of the algorithm to provide items far from users’ preferences ) is a viable solution to the overspecialization problem and related to higher choice outcomes. The findings contribute to the extant literature (i) by providing an updated understanding of the research on recommender agents and offers insights about the extant research gaps; (ii) emphasizing the nature of RAs-enabled networks, identify most influential users in wide spreading recommendations, according to a set of centrality and community-driven measures, and some relevant managerial implications are highlighted; (iii) measuring the effects of algorithmic overspecialization on users choice outcomes, discover the value of unlearning as a beneficial process to improve product recommendations and shed light on the main antecedents of such issue and discuss the algorithmic novelty as the viable solution.