Ranking tweets with local and global consistency using rich features

Abstract

Ranking tweets is more challenging in Microblog search because of content sparseness and lack of context. Traditional ranking methods essentially using Euclidean distance are limited to local structure. Manifold structure helps to rank with local and global consistency. However such structure is empirically built on content similarity in an unsupervised way, suffering from sparseness while being adopted in tweet ranking. In this paper, we explore rich features to alleviate content sparseness problem, where time locality feature is proposed to consider context dependency. We then propose a supervised learning model that aggregates rich features to construct manifold structure. A learning algorithm is then designed for solving the model by minimizing the loss of labeled queries. At last we conduct a series of experiments to demonstrate the performance on 109 labeled queries from TREC Microblogging. Compared with the well-known baselines and empirical manifold structure, our algorithm achieves consistent improvements on the metrics. © 2014 Springer International Publishing.