Toward Social-Semantic Recommender Systems

Introduction

Social networks are becoming more important but difficult to be searched, especially when they involve huge numbers of users as Facebook and LinkedIn. In this context, every day huge amount of information is shared between users in social networks, and these users show a wide range of variation in tastes, interests and activities. Clearly, investigating the preferences of users and their different relationships in social networks is a relevant issue in many domains such as sociology and psychology, searching for experts (Zhang, Ackerman, & Adamic, 2007), searching the web (Page, Brin, Motwani, & Winograd, 1999) and in recommender systems (McDonald & Ackerman, 2000).

In this article we are interested in recommender systems applied on social networks, based on the fact that social networks may have significant effects on user decisions and purchases (e.g. 53% of people using Twitter recommend products via their tweets; and 42% of marketers find Facebook an important environment for them). Moreover, we argue that integrating social network analysis measures like user centralities and methods related to user profiles with recommender system algorithms can lead to better performance for the computation of the recommendation list.

However, in most cases, searching social networks for recommendation is costly and complex, using the classical recommendation methods.

To cope with this problem, we propose a novel approach for analyzing social networks of connected users.

More precisely, in our approach to recommender systems, which we call social-semantic recommender systems, we consider the following measures:

• 1.

  Social measures based on user’s centrality and the power of social connections between users.

• 2.

  Semantic measure using user-item semantic relevance measure to determine the correlation between user taste and the features of the recommended items.

These measures are used to avoid searching the whole collaboration network by:

• •

  Starting the graph exploration through the vertices having the “best” centrality.

• •

  Stopping the graph exploration when the current user taste is not relevant to the features of the item to be recommended.

In this setting, we propose two kinds of recommendation algorithms, namely social-semantic depth-first search algorithm (SSDFS) and social-semantic breadth-first search algorithm (SSBFS). These two algorithms partially explore the collaboration network respectively in a depth-first manner or in a breadth-first manner. We then report our experiments on two real data sets, one from Amazon.com and one from MovieLens, and we compare our algorithms with the existing recommender systems. Our approach outperforms existing systems in terms of computation time, while giving recommendation lists of similar quality. We notice in this respect that improving existing approaches on accuracy or on computation time are two distinct key issues of equal importance. However, since it is likely that the size of social networks will increase significantly in the next future, improving computation time while preserving accuracy should be seen as an important achievement.

Regarding contributions, we notice that this article extends our previous work (Sulieman, Malek, Kadima, & Laurent, 2012; Sulieman, Malek, Kadima, & Laurent, 2013) in the following two respects:

• •

  We consider several centrality measures, not only degree centrality.

• •

  We provide additional experiments using two distinct real data sets.