Large-scale Multi-label Classification and Its Application to Unstructured Text Data

This proposed research project will develop a large-scale multi-label classification frame-work for text summarization, which aims at creating a set of tags to capture the mostessential aspects of the original text documents. A novel tagging loss function is in-troduced to measure the discrepancy between predicted and actual tag sets, which isexpressed in terms of a weighted sum of pairwise margins between two tags, weightedby their degrees of similarity. On this ground, a regularized empirical loss is constructedto incorporate certain linguistic knowledge, and identify a tagger maximizing the sepa-rations between the pairwise margins. One salient feature of the proposed method is itscapability of detecting novel tags absent from a training sample by exploring similarityamong existing tags. This is in sharp contrast to most existing summarization methodsthat may completely ignore the novel tags. The PI will investigate the theoretical proper-ties of the proposed summarization method, and establish asymptotic and finite-sampleupper bounds of its tagging error. The PI will also develop efficient computing algorithmsto facilitate large-scale optimization, integrating the strength of inexact alternating di-rection method of multipliers and parallel computing platform. The proposed methodwill be applied to summarize the Reuters dataset consisting of over 800,000 news stories.