Detecting Signals in Noisy Data - Can Ensemble Classifiers Help Identify Adverse Drug Reaction in Tweets?

In this paper, we describe our system for detecting adverse reactions from tweets, a task organized as part of Pacific Symposium of Biocomputing Social Media Mining Shared Task. The shared task primarily involves a binary classification of the tweets, whether they contain description of adverse drug reaction or not. We propose an ensemble machine learning classifier to tackle the unbalanced distribution of the classes in the data provided for the task. A feature set containing unigrams, bigrams, and trigrams (a selected list, using mutual information), co-occurrence of drug and side effect, negation, sentiment score, etc. were used to train the classifiers in identifying the tweets with the description of adverse drug reactions. Our system obtained 0.4195 F-score and ranked first in the shared task.