Incremental term representation learning for social network analysis

Abstract

Term representation methods as computable and semantic tools have been widely applied in social network analysis. This paper provides a new perspective that can incrementally factorize co-occurrence matrix to query latest semantic vectors. We divide the streaming social network data into old and updated training tasks respectively, and factorize the training objective function based on stochastic gradient methods to update vectors. We prove that the incremental objective function is convergent. Experimental results demonstrate that our incremental factorizing can save a substantial amount of time by speeding up training convergence. The smaller the updated data is, the faster the update factorizing process can be, even 30 times faster than existing methods in certain cases. To evaluate the correctness of incremental representation, social text similarity/relatedness, linguistic tasks, network event detection, social user multi-label classification and user clustering for social network analysis are employed as benchmarks in this paper.