Users of social networking services can connect with each other by forming communities for online interaction. Yet as the number of communities hosted by such websites grows over time, users have even greater need for effective commu- nity recommendations in order to meet more users. In this paper, we investigate two algorithms from very different do- mains and evaluate their effectiveness for personalized com- munity recommendation. First is association rule mining (ARM), which discovers associations between sets of com- munities that are shared across many users. Second is latent Dirichlet allocation (LDA), which models user-community co-occurrences using latent aspects. In comparing LDA with ARM, we are interested in discovering whether modeling low-rank latent structure is more effective for recommen- dations than directly mining rules from the observed data. We experiment on an Orkut data set consisting of 492, 104 users and 118, 002 communities. Our empirical comparisons using the top-k recommendations metric show that LDA performs consistently better than ARM for the community recommendation task when recommending a list of 4 or more communities. However, for recommendation lists of up to 3 communities, ARM is still a bit better. We analyze exam- ples of the latent information learned by LDA to explain this finding. To efficiently handle the large-scale data set, we parallelize LDA on distributed computers [1] and demon- strate our parallel implementation’s scalability with varying numbers of machines.
Social bookmarking systems and their emergent information structures, known as folksonomies, are increasingly important data sources for Semantic Web applications. A key question for harvesting semantics from these systems is how to extend and adapt traditional notions of similarity to folksonomies, and which measures are best suited for applications such as navigation support, semantic search, and ontology learning. Here we build an evaluation framework to compare various general folksonomy-based similarity measures derived from established information-theoretic, statistical, and practical measures. Our framework deals generally and symmetrically with users, tags, and resources. For evaluation purposes we focus on similarity among tags and resources, considering different ways to aggregate annotations across users. After comparing how tag similarity measures predict user-created tag relations, we provide an external grounding by user-validated semantic proxies based on WordNet and the Open Directory. We also investigate the issue of scalability. We find that mutual information with distributional micro-aggregation across users yields the highest accuracy, but is not scalable; per-user projection with collaborative aggregation provides the best scalable approach via incremental computations. The results are consistent across resource and tag similarity.
We present a novel method for key term extraction from text documents. In our method, document is modeled as a graph of semantic relationships between terms of that document. We exploit the following remarkable feature of the graph: the terms related to the main topics of the document tend to bunch up into densely interconnected subgraphs or commu- nities, while non-important terms fall into weakly intercon- nected communities, or even become isolated vertices. We apply graph community detection techniques to partition the graph into thematically cohesive groups of terms. We introduce a criterion function to select groups that contain key terms discarding groups with unimportant terms. To weight terms and determine semantic relatedness between them we exploit information extracted from Wikipedia. Using such an approach gives us the following two ad- vantages. First, it allows effectively processing multi-theme documents. Second, it is good at filtering out noise infor- mation in the document, such as, for example, navigational bars or headers in web pages. Evaluations of the method show that it outperforms exist- ing methods producing key terms with higher precision and recall. Additional experiments on web pages prove that our method appears to be substantially more effective on noisy and multi-theme documents than existing methods.