Predicting usefulness of online reviews using stochastic gradient boosting and randomized trees

Kumar, M. and Upadhyay, S.

This paper presents our analysis of online user reviews from different business categories posted on the internet rating and review services website Yelp. We use business, reviewer, and review level data to generate predictive features for estimating the number of useful votes an online review is expected to receive. Unstructured text data are mined using natural language processing techniques and combined with structured features to train two different machine learning algorithms - Stochastic Gradient Boosted Regression Trees and Extremely Randomized Trees. The results from both of these algorithms are ensembled to generate better performing predictions. The approach described in this paper mirrors the one used by one of the authors in a Kaggle competition hosted by Yelp. Out of 352 participants, the author stood 3rd on the final leaderboard.

Cite as: Kumar, M. and Upadhyay, S. (2013). Predicting usefulness of online reviews using stochastic gradient boosting and randomized trees. In Proc. Eleventh Australasian Data Mining Conference (AusDM13) Canberra, Australia. CRPIT, 146. Christen, P., Kennedy, P., Liu, L., Ong, K.L., Stranieri, A. and Zhao, Y. Eds., ACS. 65-72

(from crpit.com) (local if available)