Top-N Recommendation in Social Networks

To counter the information overloading problem on the Internet, recommender systems have been widely employed by online service websites to suggest items, e.g., movies and books, to users who might like them. In typical practical implementations, a list of N items are rendered to an active user which he may find the most interesting. This is known as the top-N recommendation task. There are a variety of recommendation algorithms including content-based recommendations and collaborative recommendations. In content-based recommendations, each item is characterized by a set of attributes, based on which item similarity is estimated, and the active user will be recommended items similar to the ones he liked in the past. However, the content-based recommendation system suffers from limited content-analysis, e.g., it is difficult to explicitly describe multimedia data using features. Instead, in collaborative recommendations, also known as collaborative filtering, the active user will be recommended items favorably rated by other users with similar tastes to the active user. The user similarity can be estimated based on user profiles including detailed personal information, but due to privacy concerns they are very difficult to obtain. Hence, the collaborative recommender systems evaluate user similarity based on users' historic rating data. This causes the cold-start problem for new users or users who do not provide enough ratings, i.e., the recommender systems cannot find similar users for them.

Recently, with the thriving of online social networks (OSN), the social collaborative recommendation has attracted significant attention. In social networks, people are more likely to connect to other people sharing similar interests, and they are influenced more by people they connect to, further fostering similarity to each other. Hence, by exploiting the social structure of social networks, social recommender systems can make satisfactory recommendations even for cold-start users when provided with their social connections. Moreover, social recommendation can also be conveniently incorporated with traditional collaborative filtering algorithms, e.g., matrix factorization and neighborhood methods, improving their recommendation performance, especially for cold-start users.

Meanwhile, people are increasingly concerned with their online privacy, adding new challenges to the recommendation problem. Although rating data do not directly tell personal details, it is still possible to infer user demographics, such as age and gender, from their ratings, and even uncover user identities and reveal sensitive personal information with access to other databases. Therefore, users are not willing to make their personal data accessible to the general public. Hence, when designing collaborative recommendation systems, we need to take into account user privacy.

We propose a social recommendation algorithm that exploits the social network to generate recommendations using only the implicit user data, e.g., whether a user has consumed an item or not, avoiding exposing explicit user rating data.