RFBoost: An improved multi-label boosting algorithm and its application to text categorisation

The AdaBoost.MH boosting algorithm is considered to be one of the most accurate algorithms for multi-label classification. AdaBoost.MH works by iteratively building a committee of weak hypotheses of decision stumps. In each round of AdaBoost.MH learning, all features are examined, but only one feature is used to build a new weak hypothesis. This learning mechanism may entail a high degree of computational time complexity, particularly in the case of a large-scale dataset. This paper describes a way to manage the learning complexity and improve the classification performance of AdaBoost.MH. We propose an improved version of AdaBoost.MH, called RFBoost. The weak learning in RFBoost is based on filtering a small fixed number of ranked features in each boosting round rather than using all features, as AdaBoost.MH does. We propose two methods for ranking the features: One Boosting Round and Labeled Latent Dirichlet Allocation (LLDA), a supervised topic model based on Gibbs sampling. Additionally, we investigate the use of LLDA as a feature selection method for reducing the feature space based on the maximal conditional probabilities of words across labels. Our experimental results on eight well-known benchmarks for multi-label text categorisation show that RFBoost is significantly more efficient and effective than the baseline algorithms. Moreover, the LLDA-based feature ranking yields the best performance for RFBoost.