Helpfulness Prediction of Online Drug Reviews

Online drug reviews are increasingly available for drug users and medical professionals to collect first-hand patient experiences on drugs. The challenge is, however, to provide them with quick and accurate access to the helpful reviews mingled with an extensive amount of less helpful ones. One way to facilitate information access to helpful reviews is helpfulness prediction. By analyzing the features extracted from reviews, the helpfulness of a review can be objectively evaluated. However, the user sentiments and opinions contained in drug reviews, compared with the ones in other product reviews, are exceedingly complicated to be analyzed due to the differences in users' physical and mental conditions. Hence, in this paper, we proposed a mechanism that combines medical-domain features (medical word sentiments, proportion of medical words in reviews) into a helpfulness prediction model. Besides, we noticed that the number of useful counts received by a review cannot sufficiently validate its helpfulness since useful counts are significantly influenced by the length of duration it was exposed to readers and total read times. Hence, to address this selection bias issue, we proposed four methods and examined their performances. Our findings show 1) the added medical features significantly elevated the accuracy of helpfulness prediction results. 2) readability, proportion of medical words and proportion of nouns are the three features showing the highest importance in helpfulness prediction. 3) the method to select sample based on useful counts achieves the best performance possibly because the sample data generated by this method contains more unbiased feature characteristics of review helpfulness.