Tag-based Recommender System for Context-Aware Content Dissemination in Opportunistic Networks

Content dissemination in opportunistic networks is a hot research topic that attracted a lot of interest in the last few years. The key idea is to optimise the diffusion of content among nodes in opportunistic networks to ensure that users are always able to obtain the most relevant items according to their interests. The classical approach is to statically define a set of interests for each user, and make sure that they receive items matching those interests. In this paper, we propose a novel approach, based on the dynamic and automatic identification of interests. To do so, we exploit the tags that users assign to the items they create, and the tags of the items that they download. We model these actions through a folksonomy and the related tripartite graph, with different nodes for users, items, and tags. We use this graph as the basis for identifying the relevance of the items. Specifically, we use a tag-based recommender system on the graph, called PLIERS, that is able to calculate the relevance of an item for a certain user, with respect to the items that are already linked to this user. We validate our approach through a series of simulations. We emulate the presence of a variable number of agents which randomly move, create and tag items, and possibly encounter other agents. Each agent maintains a tripartite graph locally, representing its actions, and it integrates this graph with information received from other encountered nodes. The agents use PLIERS on their local graph to assess the relevance of the items they find, and they decide whether these items are relevant for them or not. We evaluate the accuracy of the results by comparing the recommendations on the local graphs with the relevance of the items (calculated through PLIERS) on a global graph obtained by merging together all the local graphs of the nodes. This graph represents the complete knowledge of all actions in the network and it allows us to obtain the best possible recommendations for a target user, that could be obtained if all the nodes had the full knowledge of the actions of other nodes. The results indicate that the recommendations on the local graph are accurate and that the local knowledge of nodes reaches the global knowledge in the network through a sufficiently high number of contacts.