Make it personal: A social explanation system applied to group recommendations

Highlights

• We propose personalized social individual explanations for group recommenders.

• We propose both a textual and a graphical social explanation approach.

• We study the benefits of including explanations in group recommender systems.

• We study the benefits of including social components to these explanations.

• Results show a significant increase in users’ intent to follow our recommendations.

Abstract

Recommender systems help users to identify which items from a variety of choices best match their needs and preferences. In this context, explanations act as complementary information that can help users to better comprehend the system’s output and to encourage goals such as trust, confidence in decision-making or utility. In this paper we propose a Personalized Social Individual Explanation approach (PSIE). Unlike other expert systems the PSIE proposal novelly includes explanations about the system’s group recommendation and explanations about the group’s social reality with the goal of inducing a positive reaction that leads to a better perception of the received group recommendations. Among other challenges, we uncover a special need to focus on “tactful” explanations when addressing users’ personal relationships within a group and to focus on personalized reassuring explanations that encourage users to accept the presented recommendations. Besides, the resulting intelligent system significatively increases users’ intent (likelihood) to follow the recommendations, users’ satisfaction and the system’s efficiency and trustworthiness.

Introduction

Recommender systems (Jameson, Smyth, 2007, Ricci, Rokach, Shapira, 2015) are expert systems which support human decision-making. They commonly use real or inferred preferences to suggest to their users items that they might like to consume. Depending on the number of users that will employ the product, we can speak of individual recommenders (Ricci et al., 2015) or group recommenders (Jameson & Smyth, 2007). In this paper we focus on the latter, and, more specifically on how to improve user’s acceptance of these systems’ outcome.

In the literature, e.g. (Golbeck, 2006, Jamali, Ester, 2009, Massa, Avesani, 2007) it was shown that using Social Network information in addition to feedback data (e.g. ratings) can significantly improve group recommendations’ accuracy. Besides, there is an agreement about the need to adapt group recommendation processes to group composition (Cantador, Castells, 2012, Ricci, Rokach, Shapira, 2015, Salamó, McCarthy, Smyth, 2012). Recent work has focused on modelling users’ social behaviour within a group to enhance the recommendation’s outcome (Mccarthy, Mcginty, Smyth, Salamó, 2006, Quijano-Sánchez, Recio-García, Díaz-Agudo, Jiménez-Díaz, 2013, Salehi-Abari, Boutilier, 2015). However, there is a lack of explanation methods for these social group recommendation results. Although some explanation components have been included in group recommender systems (Boratto, Carta, 2011, Jameson, 2004, McCarthy, Salamó, Coyle, McGinty, Smyth, Nixon, 2006) none of them have focused on using the social reality within a group for explanation generation.

Explanations and recommender systems have frequently been considered as part of the studies developed in the area of knowledge-based systems (Lopez-Suarez & Kamel, 1994) where both of them can be used to support decision-making processes. It was found that explanations can help increase users’ acceptance of the proposed recommendations, helping them make faster decisions, convincing them to buy the proposed items or even develop trust in the system as a whole (Herlocker, Konstan, & Riedl, 2000). Besides, it has been acknowledged that for the users, many recommender systems function as black boxes, and therefore, do not provide transparency into how the recommendation process works and do not offer further information to go along with the recommendations. This situation can lead to the user being startled by a given recommendation, producing the need for an explanation (Herlocker et al., 2000). For instance, explanations are able to provide transparency by presenting the reasoning and data behind a recommendation. There are some individual explanation approaches (Christakis, Fowler, 2009, Guy, Zwerdling, Carmel, Ronen, Uziel, Yogev, Ofek-Koifman, 2009), that provide the names of particular friends who liked the proposed item to induce a better acceptance of that specific item; especially if the chosen names refer to good friends, tapping into the idea that people we like are more likely to persuade us. We can consider this type of explanations to be social, as they induce a positive reaction by recalling social bonds.

To date, our main line of research has focused on improving current state-of-the art research on group recommenders through the inclusion of social factors in the generation of recommendations that satisfy a group of users with potentially competing interests. To do so, we have reviewed different ways of combining people’s personal preferences and proposed an approach that takes into account the social reality within a group. Quijano-Sánchez et al. (2013)’s Social Recommendation Model (SRM), defines a set of recommendation methods that include the analysis and use of several social factors such as the personality of group members, the tie strength between them and users’ satisfaction with past recommendations.

Departing from this starting point, this research takes a step forward and novelly translates the previously mentioned social explanations (Christakis, Fowler, 2009, Guy, Zwerdling, Carmel, Ronen, Uziel, Yogev, Ofek-Koifman, 2009) to group recommender systems. This is done by including not only friend-related information, as the previous mentioned works propose, but also all the social information that the adopted system (SRM) is able to retrieve, that is: personal ratings, user’s personality, tie strength between users and previous satisfaction. Hence, this paper’s goal is to provide to each group member a Personalized Social Individual Explanation (PSIE) about the system’s proposed group recommendation and, by doing so, to induce positive reactions that lead to a better perception of the received group recommendation and of the system in general.

Thus, this work aims to improve the performance of Quijano-Sánchez et al. (2013)’s system through the inclusion of PSIE. To do so, two different approaches are proposed, Textual Social Explanations (TSE) (Section 4.1) and Graphical Social Explanations (GSE) (Section 4.2). Then, the effects of including simple non-social explanations in group recommender systems, the effects of including just one social component to these group explanations and the effects of including all of SRM’s social information to these explanations, that is, the complete PSIE approach, are studied. To address these questions two experiments have been designed, for the textual approach (Section 5.2) and for the graphical approach (Section 5.3). By performing these experiments we evaluate which of the two presented approaches is preferable (Section 5.3).

Consequently, this research has the following main contributions:

• Study of the benefits of including explanations in group recommender systems.

• Study of the benefits of including a social component to explanations in group recommender systems.

• Proposal of a Personalized Social Individual Explanation approach (PSIE):

  • Through a Textual Social Explanation approach (TSE).

  • Through a Graphical Social Explanation approach (GSE).

The remainder of this paper is structured as follows: In the next section we introduce some of the state-of-the-art research regarding explanations. In Section 3 we present the main theoretical concepts needed to develop this work. Section 4 presents our PSIE proposal. In Section 5 we present experiments and results. Finally, Section 6 concludes the paper.