In this paper we have presented a classification framework for classifying tweets relevant to some specific target sectors. Due to the imposed length restriction on an individual tweet, tweet classification faces some additional challenges which are not present in most other short text classification problems, needless to say in classification of standard written text. Hence, bag-of-word classifiers, which have been successfully leveraged for text classification in other domains, fail to achieve a similar level of accuracy in classifying tweets. In this paper, we have proposed a collocation feature selection algorithm for tweet classification. Moreover, we have proposed a strategy, built on our selected collocation features, for identifying and removing confounding outliers from a training set. An Evaluation on two real world datasets shows that the proposed model yields a better accuracy than the unigram model, uni-bigram model and also a partially supervised topic model on two different classification tasks.