In this section, we summarize key SRS-related topics that were raised throughout the previous two sections on people and content recommendation and suggest directions for future work. Explanations. The public nature of social media data enables to provide more transparency into recommendations by showing how they were formed. In some of the enterprise examples we reviewed for both content and people recommendations, explanations were found to have a key role in increasing the instant acceptance rate of recommendations \cite{Guy_2009,Guy:2009:PRS:1639714.1639725}. Beyond that, explanations in RS have been shown to have longer-term effects of building trust relationships with the user \cite{Herlocker_2000}.
There are also a few challenges with regards to explanations. First, as we have seen, explanations do not always increase accuracy. For example, in our mixed content recommendation study, tag-based explanations did not increase recommendations’ ratings. Second, not every recommendation method can provide intuitive explanations; there is usually a trade-off between the method’s complexity and the clarity of explanations it can provide. For example, recommendations that are based on clustering techniques are usually harder to explain. Third, explanations pose challenges in terms of privacy. For example, the YouTube explanations \cite{Davidson_2010} explicitly show videos previously watched by the user, which directly exposed information that might be sensitive if watched by another person. Fourth, explanations require extra real-estate on the user interface, which might be particularly challenging on mobile devices; therefore their cost-to-value ratio should be carefully considered when designing the recommender system. Privacy. As mentioned several times in this chapter, one of the key benefits of social media data is that large portions of it are public and thus can be used for analysis without infringing user privacy, as is the case, for example, with email or file system data. It should be noted, however, that in some countries, public social media information is still considered personal information (PI), when linked to an identity of a real person. This means that analysis and inference from such data may still require explicit user consent. Indeed, aggregation of public data, even if it was previously accessible, may reveal sensitive information the user did not intend to expose. In addition, as just mentioned, explanations aimed for a specific user might reveal very sensitive data, such as browsing or viewing history, when exposed to another person who may watch the screen alongside. Finally, there is much social media data that is still access-restricted. Recommender systems should pay special attention not to infringe the privacy model of the data, to avoid the exposure of sensitive information \cite{dwyer2011privacy}. Tags. The work we reviewed indicated that tags, a mechanism introduced by social media to annotate content, such as web pages, photos, or people, can be particularly effective as a basis for recommendation. Tags’ ability to concisely summarize user perspective over large content pieces make them a highly valuable resource for producing recommendations. Aside from recommendations, tags have been shown to be useful for other purposes, such as enhancing search or generating “tag clouds” that summarize the common topics of a group of items to the user \cite{kaser2007tag}. Unfortunately, despite their value, tag usage is on the decrease in recent years, with sites such as Delicious becoming less popular and other sites giving less prominence to tags. Tag recommendation techniques \cite{Sigurbjornsson:2008:FTR:1367497.1367542,jaschke2007tag}, which are another type of SRS not discussed in this chapter, should be used to promote tag usage and close the loop: tag recommendations help generate more tags, while these tags, in turn, used to produce other recommendations. Social relationships. One of the most important contributions of social media to recommender systems is the introduction of the explicit (articulated) network. Social network sites, such as Facebook, LinkedIn, and Twitter, allow people to explicitly articulate their connections. As mentioned, there are two main types of connections, one expresses familiarity and the other expresses interest. Both of these articulated networks are very useful for content recommendation, and were shown to enhance traditional CF techniques. They also have other benefits: (1) sparing the need for explicit feedback in the form of ratings to determine the network of similarity, (2) help coping with the new-user cold start problem, in case the network can be used across social media websites, and (3) helping users judge the recommendations, since they originate from people they know or are interested in (also making explanations more effective). On the other hand, as we have seen, recommendations of people to connect with or to follow are essential for enhancing the formation of such explicit relationships. This is a classic demonstration of the mutual relationship between recommender systems and social media discussed in the introduction: on the one hand social media introduces a new type of data that enhances RS; on the other hand RS are essential for generating this type of data. Trust and reputation. The topic of trust has a tremendous importance in the RS domain. Obviously, the best recommendations come from a trusted person. But on the other hand, trust is very challenging to compute as it represents a very abstract and subjective quality between two individuals. Reputation represents a more general concept about a person’s perception by others. One way to define it is the aggregation of trust in this person across the entire set of users. Social media and the “wisdom of the crowd” enable to estimate trust and reputation in ways that have not been possible before. Online social relationships and content feedback forms (comments, ‘likes’, etc.) introduce more signals that can be used to calculate trust and reputation. That said, many of the studies still use rough estimations that are based on controversial assumptions, for example, that a friend on an SNS is a trustworthy individual. Evaluation of trust and reputation is also particularly challenging, as even in the real world people have hard time figuring out who they trust or who has a good reputation. Assuming a network of trust is given, there are growing amounts of research that explore how to use it to enhance CF. The early work of Globeck \cite{Golbeck:2005:CAT:1104446} suggested to adapt the CF formula in a way that would boost similar users whom the user trusts. More advanced approaches incorporate trust in matrix factorization techniques. Evaluation. As reviewed throughout this chapter, evaluation of SRS typically uses the common methods in the broader RS domain. These include offline evaluation, user studies (especially common for SRS), and live field studies or A/B testing. Evaluation measures include RMSE, NDCG, precision, and other commonly used metrics from the RS domain. Looking forward, since social media is characterized by the “wisdom of the crowd”, it will only be natural to see more crowdsourcing techniques used for evaluation of SRS. These have become common in many domains in the recent years, including information retrieval (e.g., \cite{alonso2009can,buhrmester2011amazon,Kittur:2008:CUS:1357054.1357127}), however they are not as common yet in RS evaluation. Evaluation that goes beyond accuracy to include serendipity (“surprise”), diversity, novelty, coverage, and other factors is also due in the SRS area. Finally, evaluation over time, which also examines the broader effect of the recommendation on the surrounding ecosystem of users, as demonstrated in \cite{Daly_2010,said2014you}, is a highly desirable direction. Rather than focusing mostly on recommendation effectiveness, their broader and longer-term influence on the environment should also be considered. This requires new tools and creative thinking to be brought into the existing set of evaluation methods. Recommending content to produce. We extensively discussed content recommendation in Section \ref{sec:content}. Our examples focused on content the user consumes: video, news, questions, social media items, etc. As explained in the introduction, one of the key characteristics of social media is that users are not just the consumers, but also the producers of content. There is a body of research that attempts to recommend users content they may want to produce. Question recommendation in CQA sites, which has already been mentioned in Section \ref{sec:content}, has a role in encouraging users to produce content in the form of answers. Other works attempted to encourage users to create more profile entries \cite{Geyer:2008:RTS:1454008.1454019}, inspire users to write blogs \cite{Dugan:2010:LLB:1753326.1753623}, and prompt them to edit articles on Wikipeida \cite{Cosley:2007:SUI:1216295.1216309}. Recommending content to generate is a particularly challenging task since the entry barrier is higher as many social media users are lurkers (only consume content). It is rooted in the area of persuasive technologies and theories such as self determination \cite{ryan2000self} and behavioral models \cite{Fogg:2009:BMP:1541948.1541999}. Clearly, recommending content to produce has a central role in the symbiosis between recommender systems and social media.