Results

We validate the results of our community detection technique in four ways. First, we show images of the embedding of a sample of 60,000 user profiles using each of the four modalities independently, the result of merging all four modalities into a single similarity measure, and the result of our multimodal embedding. Second, we compare the performance of our technique against a wide variety of current community detection techniques, using a number of different quality metrics. Third, we compute the silhouette coefficients of the clusters in our multimodal embedding and compare them to the silhouette coefficients of each of the single modality embeddings. Fourth, we compute the topics associated with several of the clusters from the multimodal embedding using LSA, LDA, and NMF, and show that they are internally and semantically consistent. From all four views of the communities detected, our technique performs best. Separating the attribute-attribute similarity into three different modalities is also clearly important to achieve overall performance.